Figure 1. Image: detail from Abraham Ortelius, Theatrum orbis terrarum (1570). Library of Congress, Geography and Map Division. https://www.loc.gov/resource/g3200m.gct00126/?sp=106
WARNING!
We have extensively audited this manuscript using large language models, including adversarial reviews by multiple models. However, the proof is too complex for Lean formalization at the present time, and it has not yet undergone verification by a human expert.
Authors’ note
With the exception of this note, the entire document is machine-written. We have taken care to supervise this writing to ensure that the proofs meet a minimum standard of readability and that previous literature is appropriately cited.
The manuscript began as an attempt to pursue the program of Magnen, Rivasseau, and Sénéor [132] using large language models. Our hope was that understanding the problem on the torus might offer some insight into the general case. The construction that resulted from this starting point is three-dimensional and lattice-based and draws extensively on Bałaban’s renormalization group methods.
As we were verifying these results, we learned that Alpöge and Sussman announced that they appear to have a YM4 existence argument, following Bałaban’s program (https://github.com/ethanwsussman/Yang--Mills-LLM-experiment ). We are thrilled and look forward to learning more about these exciting developments. We hope that this manuscript nonetheless contributes to understanding constructive approaches to three-dimensional Yang–Mills theory. The first version of these results was posted to the first author’s GitHub on October 17, 2026 (https://github.com/pmaronow/3d-ym-torus ). Prior to posting the first version of this work, we did not inspect the ongoing work of Alpöge and Sussman. Further, to the best of our knowledge, the LLMs we used did not access their work during the mathematical development and drafting of this manuscript. This version contains additional clarifications relative to the first version; it has also been partitioned into five volumes.
We welcome suggestions concerning mathematical corrections, omitted citations, and attribution of ideas.
P. L. was partially supported by NSF grant DMS-2450004.
Part VI: Large-field control and the small-field event (continued)
Comparison of reconstructions on block regions
This section compares one renormalization group step on a block region with the same step on the whole torus. Fix a composition T M s T^{s}_{M} T M s of s ≤ J s \le J s ≤ J averaging steps and a union Ω \Omega Ω of blocks of its output lattice. In Section 18 we reconstructed the fine configuration on Ω \Omega Ω from the coarse field on Ω \Omega Ω and free logarithmic coordinates, with the coordinates outside Ω \Omega Ω held fixed. We also constructed the stationary point of the effective classical action in these coordinates and proved a gap for its free Hessian. Here we prove that the regional construction differs from the torus construction with the coarse field extended by zero outside Ω \Omega Ω only by terms which decay exponentially with the distance to the boundary of Ω \Omega Ω . We also prove that the regional construction is stable under small analytic perturbations of the incoming action. All estimates are bounds on absolute Taylor coefficients. Their constants do not depend on the shape of Ω \Omega Ω , the number of its components, the depth, the number of preceding renormalization group steps, or the values of previously frozen coordinates.
We use these estimates in the localized renormalization group analysis behind the exponential-moment bound for blocked actions (Step 1 in Section 3.4). There the configurations are decomposed according to the regions where some averaged plaquette has large curvature, and the renormalization group is run outside these regions with the link variables inside them held fixed. § 2.2 integrates the fluctuations on block regions using the reconstructions and perturbation estimates proved here. Section 21 uses the gauge transformation of Lemma 19.5 and the perturbation theorem Theorem 19.9 to construct large-field decompositions, and Section 23 uses the domain conditions of Lemmas 19.3 and 19.7 and Theorem 19.9 to choose the parameters.
The results, in order, are as follows: Lemma 19.3 treats successive minimization; Theorem 19.4 compares the regional and torus stationary points, together with the minimized actions, the Hessians, and the Haar disintegration densities; Lemma 19.5 constructs a gauge transformation which is the identity at prescribed vertices; Theorem 19.6 compares the gauge-transformed reconstructions and bounds their input columns; Lemma 19.7 collects the conditions on the radii; and Theorem 19.9 proves stability under perturbations of the incoming action.
The inputs are Theorem 18.12 (regional coordinates, the gap of the free Hessian, and the regional stationary point), Lemma 18.15 (coefficient bounds for connected networks which reach the fixed complement), Lemma 18.8 (derivatives of constrained stationary values), Lemma 18.9 (freezing of coordinates), the coefficient algebra of Definition 18.1, the target-dependent gauge transformation of Theorem 13.2, the classical coefficient bounds of Theorem 10.11, and the equality of effective classical actions on compact gauge overlaps, Theorem 11.6.
We compare constrained minimizers through the implicit equation for their response, and we control the response by an exponentially decaying inverse of the constrained Hessian. This approach follows Bałaban’s analysis of the variational problem for background fields [16, (182)–(190), pp. 307–308, and Proposition 9, p. 309], although our constraints, coordinate domains, and coefficient norms differ from his. The comparison between the regional and the torus problems and the gauge transformation fixing the prescribed endpoints are constructed here. The exponentially weighted resolvent bound, proved in Lemma 18.3, is the finite-dimensional version of the exponential-conjugation method of Combes and Thomas [61].
The mechanism. The regional problem and the torus problem with coarse field extended by zero have the same constraints away from a fixed neighborhood of the boundary of Ω \Omega Ω . If we restrict the torus stationary point to the regional coordinates, the gradient of the regional objective at this point is therefore concentrated near the boundary, up to exponentially decaying tails. This residual gradient has two sources, which are the two terms of the formula for r R r_R r R in Theorem 19.4 . First, the reconstructed fine field changes. Second, the tangent map of the constraints changes, since the regional problem omits the coarse constraints on links leaving Ω \Omega Ω . Writing p p p for the restricted torus stationary point and z R z_R z R for the regional stationary point, we have z R − p = − A R − 1 r R z_R-p=-A_R^{-1}r_R z R − p = − A R − 1 r R , where r R r_R r R is the residual gradient and A R A_R A R , the regional Hessian averaged along the segment from p p p to z R z_R z R , has a positive gap. Since this inverse decays exponentially in the distance, the two stationary points differ by an amount which decays exponentially away from the boundary. The remaining work is to make this mechanism stable under differentiation, composition, and successive minimization. This requires bounds on all Taylor coefficients, with any one external index fixed. A bound on the values of a holomorphic map does not give the spatial weights and the fixed-index sums which are needed when these functions are composed later.
Figure 19.1 shows the sets which appear below. Two of them play different roles: the prescribed vertices, where the gauge transformation must be the identity, and the complement of the incoming list, which controls the spatial decay of the perturbation in Theorem 19.9 .
Figure 19.1. The geometry in the output lattice (schematic). The solid inner rectangle is the fixed complement H = Ω c H = \Omega^c H = Ω c , and its boundary is the interface W W W . The dashed rectangle bounds the enlargement H L J H_{L_J} H L J of H H H by the reconstruction stencil. The cutoff χ \chi χ of Lemma 19.5 vanishes inside the inner dotted rectangle and equals one outside the outer dotted rectangle. The black dots are protected vertices, here the endpoints of the paths of two based transports of frozen coordinates. The small dotted rectangle inside H H H is the complement of the incoming list D i n D_{\mathrm{in}} D in of Theorem 19.9 , which lies within a fixed distance of H H H .
Setting and conventions
Geometry. Throughout this section M ≥ 3 M \ge3 M ≥ 3 is odd, J ≥ 1 J \ge1 J ≥ 1 is fixed, 1 ≤ s ≤ J 1 \le s \le J 1 ≤ s ≤ J , and B s = M s B_s=M^s B s = M s . The composition T M s T_M^s T M s of s s s averaging steps maps configurations of side M s m M^s m M s m to configurations of side m ≥ 4 m \ge4 m ≥ 4 , and the prior depth d 0 ≥ 0 d_0 \ge0 d 0 ≥ 0 is an arbitrary integer. When s = J s = J s = J the composition is one renormalization group step, with blocking factor B = M J B = M^J B = M J . Positions are measured in units of the output lattice, of side m m m . Let Ω \Omega Ω be a union of blocks of the output lattice, let H = Ω c H = \Omega^c H = Ω c be its fixed complement, let W W W be the interface between Ω \Omega Ω and H H H , and let H L J H_{L_J} H L J be the enlargement of H H H by the fixed reconstruction stencil L J L_J L J of Theorem 18.12. A coordinate is a label carrying a position, a direction, a color, and a scale index; its position is taken at the lattice where the coordinate was introduced and embedded in output units. In particular the newly frozen coordinates are counted with the density of the lattice on which they were introduced, not as coordinates of the output lattice. The density constant v J v_J v J of Definition 18.1 includes the colors, the directions, and the at most J J J scale indices. For a label i i i in Ω \Omega Ω we write δ ( i ) = dist ( i , W ) \delta(i) = \operatorname{dist}(i, W) δ ( i ) = dist ( i , W ) . Here dist \operatorname{dist} dist and t t t are the graph distance and the ℓ 1 \ell^1 ℓ 1 network length of the refined position graph fixed before Definition 18.1; coefficient bounds imported at a Euclidean rate b b b are used at rate b / 3 b/\sqrt{3} b / 3 . If Ω \Omega Ω is empty, every statement below with a fixed index in Ω \Omega Ω is void. If H H H is empty, every boundary difference considered below is zero; we never use an expression of the form 0 ⋅ e ∞ 0 \cdot e^\infty 0 ⋅ e ∞ .
Coordinates. We use the following coordinates.
The coarse field c c c is the logarithm of the output links internal to Ω \Omega Ω , and c ⊕ 0 c \oplus0 c ⊕ 0 is its extension by zero to all output links of the torus.
The free logarithmic coordinates z = ( z i ) 0 ≤ i < s z = (z_i)_{0 \le i < s} z = ( z i ) 0 ≤ i < s of Theorem 18.12 parametrize the solutions of the block averaging constraints in Ω \Omega Ω , with one block z i z_i z i for each intermediate scale i i i ; here i = 0 i = 0 i = 0 is the finest scale. On the torus we write z = ( z L , z H ) z = (z_L, z_H) z = ( z L , z H ) , where z L z_L z L are the coordinates of the regional list and z H z_H z H are the remaining torus coordinates.
Ψ R ( c , z ; ξ ~ ) \Psi_R(c, z; \widetilde{\xi}) Ψ R ( c , z ; ξ ) is the logarithm of the regional reconstruction of the fine field before gauge transformation, including its prescribed small values outside Ω \Omega Ω , and Ψ T ( c T , z T ) \Psi_T(c_T, z_T) Ψ T ( c T , z T ) is the logarithm of the torus reconstruction before gauge transformation. We call the links of the input lattice of T M s T_M^s T M s fine links.
In the successive reconstruction the free coordinates of scale i i i are a i + K i − 1 / 2 q i a_i + K_i^{-1/2}q_i a i + K i − 1/2 q i , where a i a_i a i is the conditional stationary point and K i K_i K i the constrained Hessian of that scale, and q i q_i q i is the fluctuation increment. The increments q i q_i q i replace g i w i g_iw_i g i w i , where w i w_i w i are the normalized fluctuation variables; no coupling is absorbed into K i K_i K i .
In § 2.1.7 we also use the jointly normalized coordinates u u u , defined by z = z 0 + K 0 − 1 / 2 u z = z_0 + K_0^{-1/2}u z = z 0 + K 0 − 1/2 u , where z 0 z_0 z 0 is the joint stationary point and K 0 K_0 K 0 the Hessian there.
The newly frozen coordinates ξ ~ \widetilde{\xi} ξ are the small logarithmic coordinates frozen at the present step; they enter Ψ R \Psi_R Ψ R through its prescribed values outside Ω \Omega Ω . The previously frozen coordinates b r e c b_{\mathrm{rec}} b rec are fixed real parameters, kept at the lattices on which they were introduced.
The points z = 0 z = 0 z = 0 , q = 0 q = 0 q = 0 , and u = 0 u = 0 u = 0 are different. Both q = 0 q = 0 q = 0 and u = 0 u = 0 u = 0 correspond to the joint stationary point z = z 0 z = z_0 z = z 0 (for q q q this is part of Lemma 19.3 ). The point z = 0 z = 0 z = 0 is the origin of the constraint coordinates.
Frozen coordinates and stationary points. No derivative in the previously frozen coordinates b r e c b_{\mathrm{rec}} b rec is taken or asserted in this section. In Theorems 19.4 and 19.6 the newly frozen coordinates ξ ~ \widetilde{\xi} ξ are analytic variables, which may be complexified and differentiated. In Theorem 19.9 the perturbed objects are differentiated only in the coarse field, the free coordinates, and the interpolation parameter, at fixed real ( ξ ~ , b r e c ) (\widetilde{\xi}, b_{\mathrm{rec}}) ( ξ , b rec ) . All stationary points below are the holomorphic stationary branches in the coordinate domains of Theorem 18.12. At real data the positive free Hessian makes them local minima, and the conditional stationary points are the conditional constrained minimizers of the successive minimization.
Dependence of constants. All constants and radii in this section have the dependence described in item (3) of Definition 4.14. They may depend on M M M and J J J , on the larger classical and reconstruction domains and the ratios of their radii, on the coefficient bounds on these domains, on the gaps of the free Hessians and the bounds on their weighted inverses, on the fixed spatial decay exponents, on ν J \nu_J ν J , and on any fixed bound on the size of the perturbation. In Lemma 19.7 and Theorem 19.9 they may also depend on the fixed analytic radii, on K ∗ K_\ast K ∗ , and on the ratios of these radii. They do not depend on L L L , γ \gamma γ , the lattice side, the depth d 0 d_0 d 0 , the shape or topology of Ω \Omega Ω , the number of components of Ω \Omega Ω or H H H , the values of the previously frozen coordinates, or the number of preceding renormalization group steps. They are evaluated before the later choices of the boundary-layer width w b d w_{\mathrm{bd}} w bd , the erosion width, the widths of the protected neighborhoods, the bound β r e c \beta_{\mathrm{rec}} β rec on newly frozen coordinates, and the enclosing product domains of Section 20. Each constant includes a finite maximum over 1 ≤ s ≤ J 1 \le s \le J 1 ≤ s ≤ J and over the finitely many initial segments. We do not assert uniformity as J → ∞ J \to\infty J → ∞ . Theorem 18.12 provides the gap of the free Hessian at every integer depth d 0 d_0 d 0 , and when only this classical estimate is used, no condition on an effective interaction is imposed.
The torus construction domain. Every use of Theorem 13.2 below takes place on its construction product. For a composition of s s s averaging steps, we denote the radius and coefficient constant of that theorem by r s r_s r s and C s T C_s^{\mathrm{T}} C s T . Before the later profiles and widths are chosen, we fix positive U T , Q T U_{\mathrm{T}}, Q_{\mathrm{T}} U T , Q T in the common larger product of coarse fields and fluctuation increments such that
U T + Q T < 1 2 min 1 ≤ s ≤ J r s , C T ( U T + Q T ) < 1 2 , C T = max 1 ≤ s ≤ J C s T . (1) U_{\mathrm{T}}+Q_{\mathrm{T}} < \frac{1}{2}\min_{1\le s\le J}r_s,\qquad C_{\mathrm{T}}(U_{\mathrm{T}}+Q_{\mathrm{T}})<\frac{1}{2},\qquad C_{\mathrm{T}}=\max_{1\le s\le J}C_s^{\mathrm{T}}. \tag*{(1)} U T + Q T < 2 1 1 ≤ s ≤ J min r s , C T ( U T + Q T ) < 2 1 , C T = 1 ≤ s ≤ J max C s T . (1)
Here Q T Q_{\mathrm{T}} Q T bounds max i ∥ q i ∥ ∞ \max_i\lVert q_i\rVert_\infty max i ∥ q i ∥ ∞ for the fluctuation increments, not the raw free coordinates a i + K − 1 / 2 q i a_i+K^{-1/2}q_i a i + K − 1/2 q i . Decreasing U T , Q T U_{\mathrm{T}},Q_{\mathrm{T}} U T , Q T if necessary, we may assume that the intermediate classical, inverse, and Haar-density margins of Theorem 13.2 hold. Later radii for coarse fields and fluctuation increments are chosen below U T U_{\mathrm{T}} U T and Q T Q_{\mathrm{T}} Q T , so each use of Theorem 13.2 has u + v < r s u+v<r_s u + v < r s , a positive denominator, and its intermediate margins. The rational row bound of that theorem alone does not show that a larger product is allowed.
Coefficient bounds. We use the coefficient algebra of Definition 18.1, Theorem 18.12, and Lemma 18.15 and its coefficient bounds with each external index fixed in turn. A rooted coefficient sum of a tensor fixes one external index, its root, which is either an output or an input, and sums the absolute values of the Taylor coefficients over the remaining indices, with the weight e a t e^{at} e a t , where t t t is the ℓ 1 \ell^1 ℓ 1 length of a shortest connected network spanning the positions involved. For a fixed input index the sum is the volume-normalized column sum of Definition 4.6, with the mesh of that input. A coefficient bound is a bound on all these sums, with any external index fixed. Taylor tensors at complex arguments use bilinear transposes. A majorant series is the series of absolute coefficients after the prescribed rooted sums and weights; the original coefficients need not be positive. In a scalar majorant the variable t t t is a radius after these rooted sums, never a sum of radii over all lattice sites. We say that a tensor, or a convergent series of tensors, has boundary-localized coefficient bounds if it satisfies the bounds of Lemma 18.15: with any index i i i in Ω \Omega Ω fixed, its rooted coefficient sums carry the additional weight e c δ ( i ) e^{c\delta(i)} e cδ ( i ) for a fixed c > 0 c>0 c > 0 .
A product or contraction joins the connecting networks of its factors at the contracted indices. We root the joined network at any external index and sum over the internal indices, using the corresponding rooted bounds of the factors. The boundary-localized estimate applies if one factor is boundary-localized, or if the joined network visits H L J H_{L_J} H L J . An index deep in H H H is bounded there, and the decay comes from its connection to the fixed index in Ω \Omega Ω . In particular, for decay exponents b > a > 0 b>a>0 b > a > 0 , the sum over such indices is bounded by
e a δ ( e ) ∑ h ∈ H L J e − b dist ( e , h ) ≤ e a δ L J ′ S J ( b − a ) , S J ( u ) = C V J ∑ r ≥ 0 ( r + 1 ) 3 e − u r , (2) e^{a\delta(e)}\sum_{h\in H_{L_J}}e^{-b\operatorname{dist}(e,h)}\le e^{a\delta_{L'_J}}\mathcal{S}_J(b-a),\qquad\mathcal{S}_J(u)=C_{V_J}\sum_{r\ge0}(r+1)^3e^{-ur}, \tag*{(2)} e a δ ( e ) h ∈ H L J ∑ e − b dist ( e , h ) ≤ e a δ L J ′ S J ( b − a ) , S J ( u ) = C V J r ≥ 0 ∑ ( r + 1 ) 3 e − u r , (2)
where L J ′ L'_J L J ′ also includes the fixed convention for link endpoints. This is the case of a fixed index in Ω \Omega Ω in Lemma 18.15; there is no factor of the volume of H H H .
Implicit series. The implicit equations below are solved in the absolute coefficient algebra. After the construction constant C J C_J C J is increased, the scalar majorant of the solution without boundary weight can be taken to satisfy
Y = C J t + 1 2 Y + C J ( t + Y ) 2 1 − C J ( t + Y ) . (3) Y=C_Jt+\frac{1}{2}Y+\frac{C_J(t+Y)^2}{1-C_J(t+Y)}. \tag*{(3)} Y = C J t + 2 1 Y + 1 − C J ( t + Y ) C J ( t + Y ) 2 . (3)
Fix t 0 > 0 t_0>0 t 0 > 0 such that, on 0 ≤ t ≤ t 0 0\le t\le t_0 0 ≤ t ≤ t 0 and 0 ≤ Y ≤ 4 C J t 0 0\le Y\le4C_Jt_0 0 ≤ Y ≤ 4 C J t 0 , the rational term has Y Y Y -derivative at most 1 / 4 1/4 1/4 and denominator at least 1 / 2 1/2 1/2 , and its value at ( t , 4 C J t ) (t,4C_Jt) ( t , 4 C J t ) is at most C J t C_Jt C J t . Such a t 0 t_0 t 0 exists and depends only on C J C_J C J : the derivative and denominator conditions have strict margins at ( t , Y ) = ( 0 , 0 ) (t,Y)=(0,0) ( t , Y ) = ( 0 , 0 ) , while the rational term at ( t , 4 C J t ) (t,4C_Jt) ( t , 4 C J t ) equals C J ( 1 + 4 C J ) 2 t 2 / [ 1 − C J ( 1 + 4 C J ) t ] = O ( t 2 ) C_J(1+4C_J)^2t^2/[1-C_J(1+4C_J)t]=O(t^2) C J ( 1 + 4 C J ) 2 t 2 / [ 1 − C J ( 1 + 4 C J ) t ] = O ( t 2 ) as t ↓ 0 t\downarrow0 t ↓ 0 . Iteration from zero produces series with nonnegative coefficients which stay below 4 C J t 4C_Jt 4 C J t . The iteration converges uniformly on every smaller disc, since the derivative in Y Y Y of the right side of (19.3) is at most 3 / 4 3/4 3/4 on its majorant disc. The degree-l l l coefficients of the solution are bounded by 4 C J t 0 1 − l 4C_Jt_0^{1-l} 4 C J t 0 1 − l , and dividing Taylor tensors by l ! l! l ! before this comparison gives factorial tensor bounds. The argument works in the absolute multivariate coefficient algebra because each product satisfies the rooted contraction bound just described. Uniform convergence of the power series implies holomorphy.
A derivative which inserts one difference factor, localized near H H H , has the majorant
Z ⪯ H ( t , Y ( t ) ) + A ( t ) Z , A ( t ) = 1 2 + ∂ Y C J ( t + Y ) 2 1 − C J ( t + Y ) ∣ Y = Y ( t ) , (4) Z\preceq H(t,Y(t))+A(t)Z,\qquad A(t)=\frac{1}{2}+\left.\partial_Y\frac{C_J(t+Y)^2}{1-C_J(t+Y)}\right|_{Y=Y(t)}, \tag*{(4)} Z ⪯ H ( t , Y ( t )) + A ( t ) Z , A ( t ) = 2 1 + ∂ Y 1 − C J ( t + Y ) C J ( t + Y ) 2 Y = Y ( t ) , (4)
where ⪯ \preceq ⪯ denotes coefficientwise domination and H ( t , Y ) H(t,Y) H ( t , Y ) is the series with the difference factor inserted. On a smaller fixed radius A ( t ) ≤ 3 / 4 A(t)\le3/4 A ( t ) ≤ 3/4 , and the solution is majorized coefficientwise by H ( t , Y ( t ) ) / ( 1 − A ( t ) ) H(t,Y(t))/(1-A(t)) H ( t , Y ( t )) / ( 1 − A ( t )) ; this is a convergent majorant series with one inserted difference factor (compare (18.55)). Fixing an analytic input differentiates this same series and keeps the index of that input. By Lemma 18.15, the connecting network and the route to the boundary are bounded once for the whole network, and their exponent is not divided by the Taylor order.
Matrix functions. We use matrix functions only on free coordinate spaces where a gap has been proved. The next lemma collects the facts we need.
Lemma 19.1. Let A b ≥ λ I A^{b} \geq\lambda I A b ≥ λ I be a real symmetric free matrix whose two-sided weighted Schur norm at exponent b b b is at most Q Q Q , and set
μ = b λ 8 ( Q + λ ) , C R = 2 S J ( μ / 2 ) . \mu= \frac{b\lambda}{8(Q+\lambda)}, \qquad C_{R} = 2S_{J}(\mu/2). μ = 8 ( Q + λ ) bλ , C R = 2 S J ( μ /2 ) . For every t ≥ 0 t \geq0 t ≥ 0 , the entries of ( A b + t ) − 1 (A^{b}+t)^{-1} ( A b + t ) − 1 are bounded by 2 e − μ dist ( x , y ) / ( λ + t ) 2e^{-\mu\operatorname{dist}(x,y)}/(\lambda+t) 2 e − μ dist ( x , y ) / ( λ + t ) , and its weighted Schur norm at exponent μ / 2 \mu/2 μ /2 is at most C R / ( λ + t ) C_{R}/(\lambda+t) C R / ( λ + t ) ; the constant C R C_{R} C R is the same for every subset of the label list.
If L L L is a complex coefficient perturbation with C R ∥ L ∥ μ / 2 / λ ≤ 1 / 4 C_{R}\lVert L\rVert_{\mu/2}/\lambda\leq1/4 C R ∥ L ∥ μ /2 / λ ≤ 1/4 , then the series
( A b + L + t ) − 1 = ∑ k ≥ 0 [ − ( A b + t ) − 1 L ] k ( A b + t ) − 1 (5) (A^{b}+L+t)^{-1}=\sum_{k\geq0}\left[-(A^{b}+t)^{-1}L\right]^{k}(A^{b}+t)^{-1} \tag*{(5)} ( A b + L + t ) − 1 = k ≥ 0 ∑ [ − ( A b + t ) − 1 L ] k ( A b + t ) − 1 (5) converges normally, and its coefficient Schur norm is at most 4 C R / [ 3 ( λ + t ) ] 4C_{R}/[3(\lambda+t)] 4 C R / [ 3 ( λ + t )] .
For positive definite real matrices A , B A,B A , B ,
A − 1 / 2 = 1 π ∫ 0 ∞ t − 1 / 2 ( A + t ) − 1 d t (6) A^{-1/2}=\frac{1}{\pi}\int_{0}^{\infty}t^{-1/2}(A+t)^{-1}\,dt \tag*{(6)} A − 1/2 = π 1 ∫ 0 ∞ t − 1/2 ( A + t ) − 1 d t (6) A − 1 / 2 − B − 1 / 2 = − 1 π ∫ 0 ∞ t − 1 / 2 ( A + t ) − 1 ( A − B ) ( B + t ) − 1 d t (7) A^{-1/2}-B^{-1/2}=-\frac{1}{\pi}\int_{0}^{\infty}t^{-1/2}(A+t)^{-1}(A-B)(B+t)^{-1}\,dt \tag*{(7)} A − 1/2 − B − 1/2 = − π 1 ∫ 0 ∞ t − 1/2 ( A + t ) − 1 ( A − B ) ( B + t ) − 1 d t (7) log A − log B = ∫ 0 ∞ ( A + t ) − 1 ( A − B ) ( B + t ) − 1 d t (8) \log A-\log B=\int_{0}^{\infty}(A+t)^{-1}(A-B)(B+t)^{-1}\,dt \tag*{(8)} log A − log B = ∫ 0 ∞ ( A + t ) − 1 ( A − B ) ( B + t ) − 1 d t (8) and these identities extend holomorphically to the complex perturbations in (2).
If T = ( A B C D 0 ) T=\begin{pmatrix} A & B \\ C & D_{0} \end{pmatrix} T = ( A C B D 0 ) is invertible with invertible blocks A A A and D 0 − C A − 1 B D_{0}-CA^{-1}B D 0 − C A − 1 B , then
( T − 1 ) L L − A − 1 = A − 1 B ( D 0 − C A − 1 B ) − 1 C A − 1 . (9) (T^{-1})_{LL}-A^{-1}=A^{-1}B(D_{0}-CA^{-1}B)^{-1}CA^{-1}. \tag*{(9)} ( T − 1 ) LL − A − 1 = A − 1 B ( D 0 − C A − 1 B ) − 1 C A − 1 . (9)
Proof. Parts (1) and (2) are the Combes–Thomas bounds proved in Lemma 18.3(3); see (18.5). We recall the choice of constants. For a real one-Lipschitz function f f f on the label metric, conjugation by diag ( e μ f ) \operatorname{diag}(e^{\mu f}) diag ( e μ f ) changes A b A^{b} A b by a matrix of Schur norm at most Q μ / ( b − μ ) ≤ λ / 8 Q\mu/(b-\mu)\leq\lambda/8 Q μ / ( b − μ ) ≤ λ /8 ; here we used e μ r − 1 ≤ μ r e μ r e^{\mu r}-1\leq\mu r e^{\mu r} e μ r − 1 ≤ μ r e μ r and sup r ≥ 0 r e − ( b − μ ) r ≤ ( b − μ ) − 1 \sup_{r\geq0}re^{-(b-\mu)r}\leq(b-\mu)^{-1} sup r ≥ 0 r e − ( b − μ ) r ≤ ( b − μ ) − 1 . By the geometric resolvent series, the conjugated resolvent has operator norm at most 2 / ( λ + t ) 2/(\lambda+t) 2/ ( λ + t ) for t ≥ 0 t\ge0 t ≥ 0 . Taking f ( x ) = dist ( x , y ) f(x)=\operatorname{dist}(x,y) f ( x ) = dist ( x , y ) , we obtain the entry bound, and summing at exponent μ / 2 \mu/2 μ /2 gives C R / ( λ + t ) C_R/(\lambda+t) C R / ( λ + t ) , uniformly over every subset of the label list. Under the condition in (2) each term of (5) after the first is smaller than the preceding one by the factor 1 / 4 1/4 1/4 , which gives the bound 4 C R / [ 3 ( λ + t ) ] 4C_R/[3(\lambda+t)] 4 C R / [ 3 ( λ + t )] .
For (3), we diagonalize the positive matrices and substitute t = λ u 2 t=\lambda u^2 t = λ u 2 in each eigenvalue integral, which gives (19.6. The resolvent identity ( A + t ) − 1 − ( B + t ) − 1 = − ( A + t ) − 1 ( A − B ) ( B + t ) − 1 (A+t)^{-1}-(B+t)^{-1}=-(A+t)^{-1}(A-B)(B+t)^{-1} ( A + t ) − 1 − ( B + t ) − 1 = − ( A + t ) − 1 ( A − B ) ( B + t ) − 1 fixes the order of the three factors and gives (7) after integration against π − 1 t − 1 / 2 d t \pi^{-1}t^{-1/2}\,dt π − 1 t − 1/2 d t . For (8) , we integrate ( 1 + t ) − 1 I − ( A + t ) − 1 (1+t)^{-1}I-(A+t)^{-1} ( 1 + t ) − 1 I − ( A + t ) − 1 to obtain log A \log A log A and subtract the corresponding formula for B B B . The series (5) extends these identities holomorphically to the complex perturbations in (2). Their majorants are integrable: they are bounded by C J t − 1 / 2 ( λ + t ) − 2 C_Jt^{-1/2}(\lambda+t)^{-2} C J t − 1/2 ( λ + t ) − 2 and C J ( λ + t ) − 2 C_J(\lambda+t)^{-2} C J ( λ + t ) − 2 for the two differences, whose integrals are π / ( 2 λ 3 / 2 ) \pi/(2\lambda^{3/2}) π / ( 2 λ 3/2 ) and 1 / λ 1/\lambda 1/ λ . Thus a boundary-localized middle factor A − B A-B A − B gives boundary-localized differences of these functions, rooted at every index, and differentiation under the integral is justified by the same normally convergent majorants. Restricted diagonal sums are taken after this calculation; only a complete scalar trace permits a cyclic permutation of the factors. Part (4) follows by block multiplication.
Additive constants. The coefficient seminorm with boundary decay X \mathcal{X} X of Definition 18.1 vanishes on constants, so it never estimates additive constants. The constant term of a difference of minimized actions remains in the full action. Scalar determinants, Haar disintegration densities, and traces over H H H , which have no index in Ω \Omega Ω , are kept as their exact finite values. The difference between two exterior extensions of the classical action is kept with all its constants (Theorem 19.9 ). After freezing at c c c , replacing an effective interaction R ( v , c ) R(v,c) R ( v , c ) by R ( v , c ) − R ( 0 , c ) R(v,c)-R(0,c) R ( v , c ) − R ( 0 , c ) changes the normalization constant from Λ \Lambda Λ to Λ − R ( 0 , c ) \Lambda-R(0,c) Λ − R ( 0 , c ) and preserves the density, as in Lemma 18.9.
Conditional minimization and successive substitution
The reconstruction of the fine field proceeds from the coarsest intermediate scale to the finest. At scale i i i the free coordinates are a i ( c , z > i ) + K i − 1 / 2 q i a_i(c,z_{>i})+K_i^{-1/2}q_i a i ( c , z > i ) + K i − 1/2 q i , where a i a_i a i is a conditional stationary point which depends on the coarser coordinates z > i z_{>i} z > i already reconstructed. We must check two things. First, for every fixed value of the coarser coordinates in some domain, the conditional stationary point exists. Second, the coarser coordinates produced by the earlier substitutions lie in that domain. The first does not imply the second, as the following example shows.
Example 19.2. Let F 0 ( x , y ) = ( x 2 + y 2 ) / 2 F_0(x,y)=(x^2+y^2)/2 F 0 ( x , y ) = ( x 2 + y 2 ) /2 and F 1 ( x , y ) = F 0 ( x , y ) − δ y F_1(x,y)=F_0(x,y)-\delta y F 1 ( x , y ) = F 0 ( x , y ) − δ y with δ > 0 \delta>0 δ > 0 , and eliminate x x x first, so that y y y is the coarser coordinate. For every fixed y y y the conditional stationary point of F 1 F_1 F 1 in x x x is x = 0 x=0 x = 0 , and the joint Hessian is the identity. The joint
stationary point of F 1 F_1 F 1 is ( 0 , δ ) (0,\delta) ( 0 , δ ) . If the conditional stationary points are constructed only for ∣ y ∣ < η |y|<\eta ∣ y ∣ < η with η < δ \eta<\delta η < δ , the successive substitution never reaches the joint stationary point, however small the fluctuation radius in x x x is chosen. The quantity δ V \delta_V δ V of Lemma 19.3 detects this: for the initial segment consisting of both coordinates it equals δ \delta δ , so the condition in part (3) of the lemma fails.
In the next lemma, I I I is an initial segment of the finer scale coordinates eliminated in the successive minimization, and y = z I c y=z_{I^c} y = z I c denotes the coarser coordinates held as parameters. The segment containing all free coordinates is included.
Lemma 19.3. Fix the larger construction domains of Theorem 18.12, within (19.1), and the margins of their coefficient, inverse, and image bounds and of the lower bound for the real constrained Hessian. Let F 0 ( c , z ; ξ ) F_0(c,z;\xi) F 0 ( c , z ; ξ ) be the unperturbed regional objective and z 0 ( c ; ξ ) z^0(c;\xi) z 0 ( c ; ξ ) its joint stationary point. For each initial segment I I I write y = z I c y=z_{I^c} y = z I c , and let a I ( c , y ; ξ ) a_I(c,y;\xi) a I ( c , y ; ξ ) be the conditional stationary point of F 0 F_0 F 0 in the variables z I z_I z I .
One can first choose a solution radius r I > 0 r_I>0 r I > 0 , and then a strictly smaller radius η I \eta_I η I for the deviation y − z I c 0 ( c ; ξ ) y-z^0_{I^c}(c;\xi) y − z I c 0 ( c ; ξ ) , uniformly over the coarse fields in the domain, such that
a I ( c , z I c 0 ; ξ ) = z I 0 , D y a I = − F 0 , I I − 1 F 0 , I I c , ∥ a I − z I 0 ∥ c o e f f ≤ C J ∥ y − z I c 0 ∥ c o e f f . a_I(c,z^0_{I^c};\xi)=z^0_I,\qquad D_y a_I=-F_{0,II}^{-1}F_{0,II^c},\qquad\|a_I-z^0_I\|_{\mathrm{coeff}}\le C_J\|y-z^0_{I^c}\|_{\mathrm{coeff}}. a I ( c , z I c 0 ; ξ ) = z I 0 , D y a I = − F 0 , I I − 1 F 0 , I I c , ∥ a I − z I 0 ∥ coeff ≤ C J ∥ y − z I c 0 ∥ coeff . The derivative identity is evaluated at the conditional stationary point. We write T I T_I T I for the resulting domain of y y y .
Let F λ = F 0 + λ V F_\lambda=F_0+\lambda V F λ = F 0 + λV , 0 ≤ λ ≤ 1 0\le\lambda\le1 0 ≤ λ ≤ 1 , be holomorphic and real at real arguments, with the same coefficient bounds on the larger domains. Set A I = F 0 , I I ( c , a I , y ; ξ ) A_I=F_{0,II}(c,a_I,y;\xi) A I = F 0 , I I ( c , a I , y ; ξ ) and assume that, throughout the domains of part (1), in the operator norm of the coefficientwise majorant,
∥ A I − 1 F λ , I ( c , a I , y ) ∥ ≤ r I / 4 , ∥ A I − 1 ( F λ , I I − A I ) ∥ ≤ 1 / 2. \|A_I^{-1}F_{\lambda,I}(c,a_I,y)\|\le r_I/4,\qquad\|A_I^{-1}(F_{\lambda,II}-A_I)\|\le1/2. ∥ A I − 1 F λ , I ( c , a I , y ) ∥ ≤ r I /4 , ∥ A I − 1 ( F λ , I I − A I ) ∥ ≤ 1/2. Assume also that the real Hessian is perturbed by less than half of the lower bound for the constrained Hessian, and that the complex resolvent condition of Lemma 19.1(2) holds. Then for each I I I and each y ∈ T I y\in T_I y ∈ T I there is a unique holomorphic stationary point a λ , I ( c , y ; ξ ) a_{\lambda,I}(c,y;\xi) a λ , I ( c , y ; ξ ) of F λ F_\lambda F λ in z I z_I z I within distance r I / 2 r_I/2 r I /2 of a I a_I a I . It is real at real data and a local real minimum there, and it is unique in the convex domain on which the second inequality holds.
Put V = ∂ λ F λ V=\partial_\lambda F_\lambda V = ∂ λ F λ and
δ V = max I sup y ∈ T I ∥ A I − 1 V I ( c , a I ( c , y ; ξ ) , y ) ∥ c o e f f . \delta_V=\max_I\sup_{y\in T_I}\|A_I^{-1}V_I(c,a_I(c,y;\xi),y)\|_{\mathrm{coeff}}. δ V = I max y ∈ T I sup ∥ A I − 1 V I ( c , a I ( c , y ; ξ ) , y ) ∥ coeff . where the coefficient norm includes the common product of coarse fields, uniformly at fixed real frozen coordinates, and let ε t u b e \varepsilon_{\mathrm{tube}} ε tube be the minimum over I I I of the solution and parameter margins at I I I . Then, for every I I I and every y ∈ T I y\in T_I y ∈ T I ,
∥ a λ , I ( c , y ; ξ ) − a I ( c , y ; ξ ) ∥ c o e f f ≤ 2 λ δ V . \|a_{\lambda,I}(c,y;\xi)-a_I(c,y;\xi)\|_{\mathrm{coeff}}\le2\lambda\delta_V. ∥ a λ , I ( c , y ; ξ ) − a I ( c , y ; ξ ) ∥ coeff ≤ 2 λ δ V .
Assume further that δ V < ∞ \delta_V < \infty δ V < ∞ , that the mixed coefficient bounds hold uniformly on the larger domains, and that r q > 0 r_q > 0 r q > 0 satisfies
C s e q , J ( 2 δ V + r q ) < 1 2 ε t u b e , C_{\mathrm{seq},J}(2\delta_V+r_q)<\frac{1}{2}\varepsilon_{\mathrm{tube}}, C seq , J ( 2 δ V + r q ) < 2 1 ε tube ,
where C s e q , J C_{\mathrm{seq},J} C seq , J is the constant given in the proof. For I i I_i I i consisting of scales 0 , … , i 0,\ldots,i 0 , … , i , let a λ , i a_{\lambda,i} a λ , i be the scale-i i i component of the conditional stationary point a λ , I i a_{\lambda,I_i} a λ , I i of F λ F_\lambda F λ , and let K λ , i K_{\lambda,i} K λ , i be M i M^i M i times the Schur block of F λ F_\lambda F λ at scale i i i in input units. Then, whenever max i ∥ q i ∥ ∞ ≤ r q \max_i\lVert q_i\rVert_\infty\leq r_q max i ∥ q i ∥ ∞ ≤ r q , the successive substitutions in the original free coordinates
z i = a λ , i ( c , z > i ; ξ ) + K λ , i − 1 / 2 ( c , z > i ; ξ ) q i z_i=a_{\lambda,i}(c,z_{>i};\xi)+K_{\lambda,i}^{-1/2}(c,z_{>i};\xi)q_i z i = a λ , i ( c , z > i ; ξ ) + K λ , i − 1/2 ( c , z > i ; ξ ) q i
stay in every parameter domain and every solution domain, with half of each margin unused. At q = 0 q=0 q = 0 they give the joint stationary point.
The radii are chosen in the following order: first the unperturbed product, for which δ V = 0 \delta_V=0 δ V = 0 ; then a bound on the size of the perturbation on the fixed larger unperturbed products, and a smaller perturbed product; only afterwards the smaller enclosing and fluctuation radii. Reducing r q r_q r q does not reduce the radius of the coarse field, since the chain at q = 0 q=0 q = 0 is the joint stationary point. All radii and constants have the dependence of Section 19.1, independently of the later widths, the depth, and the geometry of Ω \Omega Ω .
Proof. Part (1) is a quantitative implicit function argument about the joint stationary point, and part (2) is the same argument about a I a_I a I . The new point is part (3), which first bounds the displacement of the conditional stationary points at every fixed y y y and then propagates these bounds through the successive substitution by a finite recurrence.
Step 1: the unperturbed conditional stationary points. Fix I I I and write y = z I c 0 ( c ; ξ ) + v y=z^0_{I^c}(c;\xi)+v y = z I c 0 ( c ; ξ ) + v . This shift is an analytic change of coordinates with the coefficient bounds of Theorem 18.12. In these coordinates we set u = z I − z I 0 ( c ; ξ ) u=z_I-z^0_I(c;\xi) u = z I − z I 0 ( c ; ξ ) and
B I ( c ; ξ ) = F 0 , I I ( c , z 0 ( c ; ξ ) ; ξ ) . B_I(c;\xi)=F_{0,II}(c,z^0(c;\xi);\xi). B I ( c ; ξ ) = F 0 , I I ( c , z 0 ( c ; ξ ) ; ξ ) .
By the bounds on the inverse and on the joint third derivative, there exist constants I 0 , C 0 I_0,C_0 I 0 , C 0 on the fixed larger products such that
∥ B I − 1 F 0 , I ( c , z I 0 , z I c 0 + v ; ξ ) ∥ c o e f f ≤ I 0 C 0 ∥ v ∥ c o e f f , (10) \left\lVert B_I^{-1}F_{0,I}(c,z_I^0,z_{I^c}^0+v;\xi)\right\rVert_{\mathrm{coeff}}\leq I_0C_0\lVert v\rVert_{\mathrm{coeff}}, \tag*{(10)} B I − 1 F 0 , I ( c , z I 0 , z I c 0 + v ; ξ ) coeff ≤ I 0 C 0 ∥ v ∥ coeff , (10)
∥ B I − 1 ( F 0 , I I ( c , z I 0 + u , z I c 0 + v ; ξ ) − B I ) ∥ c o e f f ≤ I 0 C 0 ( ∥ u ∥ c o e f f + ∥ v ∥ c o e f f ) . (11) \left\lVert B_I^{-1}\left(F_{0,II}(c,z_I^0+u,z_{I^c}^0+v;\xi)-B_I\right)\right\rVert_{\mathrm{coeff}}\leq I_0C_0\left(\lVert u\rVert_{\mathrm{coeff}}+\lVert v\rVert_{\mathrm{coeff}}\right). \tag*{(11)} B I − 1 ( F 0 , I I ( c , z I 0 + u , z I c 0 + v ; ξ ) − B I ) coeff ≤ I 0 C 0 ( ∥ u ∥ coeff + ∥ v ∥ coeff ) . (11)
For the first inequality we integrate F 0 , I I c F_{0,I I^c} F 0 , I I c in v v v , starting from the joint stationarity F 0 , I ( c , z 0 ; ξ ) = 0 F_{0,I}(c,z^0;\xi)=0 F 0 , I ( c , z 0 ; ξ ) = 0 ; for the second we integrate the third derivative tensor in ( u , v ) (u,v) ( u , v ) . These are coefficient inequalities, since integration along a segment from the origin multiplies each term of positive degree by a number at most one, and the rooted bounds on the larger domains control the products.
Fix r I > 0 r_I>0 r I > 0 within a fixed fraction of the solution margin of the chart, with I 0 C 0 r I < 1 / 8 I_0C_0r_I<1/8 I 0 C 0 r I < 1/8 . Then fix 0 < η I < r I 0<\eta_I<r_I 0 < η I < r I within the parameter margin, with I 0 C 0 η I < r I / 8 I_0C_0\eta_I<r_I/8 I 0 C 0 η I < r I /8 and I 0 C 0 ( r I + η I ) < 1 / 4 I_0C_0(r_I+\eta_I)<1/4 I 0 C 0 ( r I + η I ) < 1/4 . For ∥ v ∥ ≤ η I \lVert v\rVert\le\eta_I ∥ v ∥ ≤ η I , the map
u ⟼ u − B I − 1 F 0 , I ( c , z I 0 + u , z I c 0 + v ; ξ ) u\longmapsto u-B_I^{-1}F_{0,I}(c,z_I^0+u,z_{I_c}^0+v;\xi) u ⟼ u − B I − 1 F 0 , I ( c , z I 0 + u , z I c 0 + v ; ξ )
has derivative − B I − 1 ( F 0 , I I − B I ) -B_I^{-1}(F_{0,II}-B_I) − B I − 1 ( F 0 , I I − B I ) , whose norm is less than 1 / 4 1/4 1/4 on ∥ u ∥ ≤ r I / 2 \lVert u\rVert\le r_I/2 ∥ u ∥ ≤ r I /2 by (19.11), and its value at u = 0 u=0 u = 0 has norm less than r I / 8 r_I/8 r I /8 by (19.10). It is therefore a coefficient contraction of the ball ∥ u ∥ ≤ r I / 2 \lVert u\rVert\le r_I/2 ∥ u ∥ ≤ r I /2 into itself. Its Picard iterates converge in the coefficient norm, and their power series converge locally uniformly on the parameter product, which implies that the limit a I − z I 0 a_I-z_I^0 a I − z I 0 is holomorphic. At v = 0 v=0 v = 0 the stationary point is the joint one, which proves the first identity of part (1). Differentiating the gradient equation at the conditional stationary point and using Lemma 18.8, we obtain
D y a I = − F 0 , I I − 1 F 0 , I I c . D_y a_I=-F_{0,II}^{-1}F_{0,II_c}. D y a I = − F 0 , I I − 1 F 0 , I I c .
The inverse and the mixed matrices have convergent coefficient bounds on a product larger than the chosen one. Integrating this identity from y = z I c 0 y=z_{I_c}^0 y = z I c 0 to y y y , we obtain the Lipschitz bound of part (1), with C I C_I C I a common bound for the product of these matrices. The choices are uniform in the coarse field, because the estimates were made on its fixed larger domain before v v v was introduced.
Step 2: the perturbed conditional stationary points. We write z I = a I + Y z_I=a_I+Y z I = a I + Y and consider on the solution domain the map
T λ ( Y ) = Y − A I − 1 F λ , I ( c , a I + Y , y ) . (12) T_\lambda(Y)=Y-A_I^{-1}F_{\lambda,I}(c,a_I+Y,y). \tag*{(12)} T λ ( Y ) = Y − A I − 1 F λ , I ( c , a I + Y , y ) . (12)
Since D Y T λ = − A I − 1 ( F λ , I I − A I ) D_YT_\lambda=-A_I^{-1}(F_{\lambda,II}-A_I) D Y T λ = − A I − 1 ( F λ , I I − A I ) , the two inequalities of part (2) imply
∥ T λ ( 0 ) ∥ ≤ r I / 4 , ∥ D Y T λ ∥ ≤ 1 / 2. \lVert T_\lambda(0)\rVert\le r_I/4,\qquad\lVert D_YT_\lambda\rVert\le1/2. ∥ T λ ( 0 )∥ ≤ r I /4 , ∥ D Y T λ ∥ ≤ 1/2.
Then T λ T_\lambda T λ maps the coefficient ball of radius r I / 2 r_I/2 r I /2 into itself and contracts differences by the factor 1 / 2 1/2 1/2 . Iterating from zero proves existence and convergence. The sum of the successive differences is a geometric series in the coefficient norm, so the limit is holomorphic. The differentiated series converge on a fixed smaller domain: they satisfy an equation Z = H + T Z Z=H+TZ Z = H + T Z in which the majorant of T T T is at most 1 / 2 1/2 1/2 , and the solution is ∑ j ≥ 0 T j H \sum_{j\ge0}T^jH ∑ j ≥ 0 T j H . This gives the coefficient bounds with any parameter index fixed.
If two stationary points lie in the convex domain on which the derivative bound holds, their difference is unchanged by (19.12) and has norm at most half of itself, so the two points coincide. Since F λ F_\lambda F λ is real at real arguments, every Picard iterate is real at real data, and so is the limit. The margin on the real Hessian keeps at least half of the lower bound for the constrained Hessian on a neighborhood of the stationary point. Integrating the real Hessian along segments shows a positive quadratic increase away from it, so it is a local real minimum. The complex resolvent condition is a hypothesis, which provides the inverse square roots used below.
Step 3: the displacement at fixed coarser coordinates. For part (3), note that F 0 , I ( c , a I , y ) = 0 F_{0,I}(c,a_I,y)=0 F 0 , I ( c , a I , y ) = 0 , so the map (19.12) satisfies T λ ( 0 ) = − λ A I − 1 V I ( c , a I , y ) T_\lambda(0)=-\lambda A_I^{-1}V_I(c,a_I,y) T λ ( 0 ) = − λ A I − 1 V I ( c , a I , y ) . Writing Y λ , I = a λ , I − a I Y_{\lambda,I}=a_{\lambda,I}-a_I Y λ , I = a λ , I − a I in the fixed point equation, we obtain for every y ∈ T I y\in T_I y ∈ T I
∥ Y λ , I ∥ c o e f f ≤ λ δ V + 1 2 ∥ Y λ , I ∥ c o e f f , hence ∥ Y λ , I ∥ c o e f f ≤ 2 λ δ V . (13) \lVert Y_{\lambda,I}\rVert_{\mathrm{coeff}}\leq\lambda\delta_V+\frac{1}{2}\lVert Y_{\lambda,I}\rVert_{\mathrm{coeff}},\qquad\text{hence}\qquad\lVert Y_{\lambda,I}\rVert_{\mathrm{coeff}}\leq2\lambda\delta_V. \tag*{(13)} ∥ Y λ , I ∥ coeff ≤ λ δ V + 2 1 ∥ Y λ , I ∥ coeff , hence ∥ Y λ , I ∥ coeff ≤ 2 λ δ V . (13)
This estimate is uniform at each fixed higher argument in T I T_I T I , with the common product of coarse fields. A supremum of these fixed-argument coefficient norms does not itself bound substitution of a moving coefficient series. We use it for the joint initial segment, which has no higher argument; the moving substitutions below use the joint mixed-derivative majorants instead. No derivative in a frozen coordinate is involved.
Step 4: the Lipschitz bound in the coarser coordinates. By the coefficient bounds on the larger domains, the mixed matrices F λ , I I c F_{\lambda,II^c} F λ , I I c are bounded by a common constant Q m i x Q_{\mathrm{mix}} Q mix and the reference inverse by I 0 I_0 I 0 . The second inequality of part (2) and the Neumann series imply that ∥ F λ , I I − 1 ∥ ≤ 2 I 0 \lVert F_{\lambda,II}^{-1}\rVert\leq2I_0 ∥ F λ , I I − 1 ∥ ≤ 2 I 0 . Implicit differentiation at the perturbed conditional stationary point yields
D y a λ , I = − F λ , I I − 1 F λ , I I c , ∥ D y a λ , I ∥ c o e f f ≤ 2 I 0 Q m i x . (14) D_y a_{\lambda,I}=-F_{\lambda,II}^{-1}F_{\lambda,II^c},\qquad\lVert D_y a_{\lambda,I}\rVert_{\mathrm{coeff}}\leq2I_0Q_{\mathrm{mix}}. \tag*{(14)} D y a λ , I = − F λ , I I − 1 F λ , I I c , ∥ D y a λ , I ∥ coeff ≤ 2 I 0 Q mix . (14)
The rooted versions follow from the same ordered products on a fixed smaller domain. From now on let C 0 C_0 C 0 denote a bound for these maps and their unperturbed versions, and let C S C_S C S bound the rows of the inverse square roots. These constants depend on the common bound for the mixed tensors, not on δ V \delta_V δ V or on the later free radius.
Step 5: successive substitution. Let ε t u b e \varepsilon_{\mathrm{tube}} ε tube be as in part (3). We use the successive maps in the original free coordinates, with K λ , i K_{\lambda,i} K λ , i equal to M i M^i M i times the Schur blocks in input units. Let z λ ( c ) z_\lambda(c) z λ ( c ) be the perturbed joint stationary point. For the initial segment of all free coordinates there is no higher argument, so (19.13) gives
∥ z λ − z 0 ∥ c o e f f ≤ 2 δ V . \lVert z_\lambda-z^0\rVert_{\mathrm{coeff}}\leq2\delta_V. ∥ z λ − z 0 ∥ coeff ≤ 2 δ V .
The assumed tube inequality and the unperturbed Lipschitz bound of part (1) place every projection of this joint point strictly in the corresponding parameter and solution domains. Joint stationarity and uniqueness of each conditional branch therefore give
a λ , I ( c , z λ , I c ( c ) ; ξ ~ ) = z λ , I ( c ) . a_{\lambda,I}\bigl(c,z_{\lambda,I^c}(c);\widetilde{\xi}\bigr)=z_{\lambda,I}(c). a λ , I ( c , z λ , I c ( c ) ; ξ ) = z λ , I ( c ) .
Let D i D_i D i bound the deviations of the already reconstructed coarser coordinates from z λ z_\lambda z λ , in the joint coefficient algebra of the coarse field and the remaining increments. Integrating D y a λ , I D_y a_{\lambda,I} D y a λ , I from z λ , I c z_{\lambda,I}^c z λ , I c to the moving higher argument uses its joint mixed-derivative majorant on the larger product. The small coefficient displacement of z λ z_\lambda z λ permits this recentering on a fixed smaller product; enlarge C 0 C_0 C 0 once to include the composition. The increment adds at most C S r q C_S r_q C S r q . Thus
D i − 1 ≤ ( 1 + C 0 ) D i + C S r q , D s = 0 , (15) D_{i-1} \le(1+C_0)D_i+C_Sr_q,\qquad D_s=0, \tag*{(15)} D i − 1 ≤ ( 1 + C 0 ) D i + C S r q , D s = 0 , (15)
as an inequality of majorants, rather than a substitution into the fixed-y y y supremum. After the finite conversion between the original free coordinates and these centered coordinates is included in C 0 C_0 C 0 and C S C_S C S , we may retain
C s e q , J = 2 ( 1 + C S ) ( 1 + C 0 ) J + 1 ( 1 + ∑ j = 0 J − 1 ( 1 + C 0 ) j ) . C_{\mathrm{seq},J}=2(1+C_S)(1+C_0)^{J+1}\left(1+\sum_{j=0}^{J-1}(1+C_0)^j\right). C seq , J = 2 ( 1 + C S ) ( 1 + C 0 ) J + 1 ( 1 + j = 0 ∑ J − 1 ( 1 + C 0 ) j ) .
Indeed, iterating (15) and then adding the joint displacement, every deviation A i A_i A i from z 0 z^0 z 0 satisfies
A i ≤ 2 δ V + C S r q ∑ j = 0 J − 1 ( 1 + C 0 ) j , A_i\le2\delta_V+C_Sr_q\sum_{j=0}^{J-1}(1+C_0)^j, A i ≤ 2 δ V + C S r q j = 0 ∑ J − 1 ( 1 + C 0 ) j ,
which is at most C s e q , J ( 2 δ V + r q ) C_{\mathrm{seq},J}(2\delta_V+r_q) C seq , J ( 2 δ V + r q ) . The condition of part (3) therefore leaves half of each margin unused. At each admitted higher argument the Picard iterations stay in the fixed solution balls of part (2). This proves admissibility before proceeding to the next scale; it does not assert that every point of those fixed balls has deviation bounded by A i A_i A i .
Step 6: zero increments. At q = 0 q=0 q = 0 , the stationarity at the coarsest scale and the successive stationarities at the finer scales give the joint gradient equation. By uniqueness of its branch, this chain is the joint stationary point; Step 5 shows that the chain stays in the required domains.
Step 7: the order of choices. For δ V = 0 \delta_V=0 δ V = 0 we first choose r q r_q r q by the condition of part (3). On the fixed larger unperturbed products, we then choose the perturbed product and the bound on the perturbation so that the same condition holds, and all later radii are chosen inside that product. Decreasing r q r_q r q changes only the terms of positive degree in q q q , and the image at q = 0 q=0 q = 0 is the joint stationary point for every coarse field in the domain; hence no later reduction of the coarse-field radius follows. Finite minima and maxima over the initial segments preserve the dependence of the constants.
Comparison of the regional and torus stationary points
We now compare the regional problem with the torus problem at a nonzero coarse field. Recall that F R ( c , z ; ζ ) = E d 0 ( Ψ R ( c , z ; ζ ) ) F_R(c,z;\zeta)=E_{d_0}(\Psi_R(c,z;\zeta)) F R ( c , z ; ζ ) = E d 0 ( Ψ R ( c , z ; ζ )) is the regional objective of Theorem 18.12, and that the torus objective is the effective classical action of the torus reconstruction with the coarse field extended by zero. For a target v v v of the torus we write z T ( v ) z_T(v) z T ( v ) for the torus stationary point at v v v , and P P P for the restriction from the torus free list to the regional list.
Theorem 19.4. Fix the larger domains of Theorem 18.12, an odd M ≥ 3 M \ge3 M ≥ 3 , J ≥ 1 J \ge1 J ≥ 1 , 1 ≤ s ≤ J 1 \le s \le J 1 ≤ s ≤ J , m ≥ 4 m \ge4 m ≥ 4 , a depth d 0 ≥ 0 d_0 \ge0 d 0 ≥ 0 , and a union Ω \Omega Ω of output blocks; positions are in output units. Use a smaller fixed product of ( c , z ; ζ ) (c,z;\zeta) ( c , z ; ζ ) on which the gap, inverse, and image margins of that theorem hold, with the previously frozen coordinates fixed. Write
F R = E d 0 ∘ Ψ R ( c , z ; ζ ) , F T = E d 0 ∘ Ψ T ( c ⊕ 0 , z L , z H ) , ( p , h ) = z T ( c ⊕ 0 ) , F_R = E_{d_0} \circ\Psi_R(c,z;\zeta), \qquad F_T = E_{d_0} \circ\Psi_T(c \oplus0,z_L,z_H), \qquad(p,h) = z_T(c \oplus0), F R = E d 0 ∘ Ψ R ( c , z ; ζ ) , F T = E d 0 ∘ Ψ T ( c ⊕ 0 , z L , z H ) , ( p , h ) = z T ( c ⊕ 0 ) , put
a T 0 = Ψ T ( c ⊕ 0 , p , h ) , a ^ R = Ψ R ( c , p ; ζ ) , d a = a ^ R − a T 0 , a_T^0 = \Psi_T(c \oplus0,p,h), \qquad\widehat{a}_R = \Psi_R(c,p;\zeta), \qquad d_a = \widehat{a}_R-a_T^0, a T 0 = Ψ T ( c ⊕ 0 , p , h ) , a R = Ψ R ( c , p ; ζ ) , d a = a R − a T 0 , J R = Ψ R , z ( c , p ; ζ ) , J T = Ψ T , z L ( c ⊕ 0 , p , h ) , J_R = \Psi_{R,z}(c,p;\zeta), \qquad J_T = \Psi_{T,z_L}(c \oplus0,p,h), J R = Ψ R , z ( c , p ; ζ ) , J T = Ψ T , z L ( c ⊕ 0 , p , h ) , and let r R = F R , z ( c , p ; ζ ) r_R = F_{R,z}(c,p;\zeta) r R = F R , z ( c , p ; ζ ) be the regional gradient at the restricted torus stationary point. Then the following statements hold.
(Gradient and response.) Let z R z_R z R be the regional stationary point. Then
r R = ( J R − J T ) T E d 0 ′ ( a T 0 ) + J R T ∫ 0 1 E d 0 ′ ′ ( a T 0 + u d a ) d a d u , r_R = (J_R-J_T)^{\mathsf T} E_{d_0}'(a_T^0) + J_R^{\mathsf T}\int_0^1 E_{d_0}''(a_T^0+u d_a)d_a\,du, r R = ( J R − J T ) T E d 0 ′ ( a T 0 ) + J R T ∫ 0 1 E d 0 ′′ ( a T 0 + u d a ) d a d u , A R ( z R − p ) = − r R , A R = ∫ 0 1 F R , z z ( c , p + u ( z R − p ) ; ζ ) d u . A_R(z_R-p)=-r_R, \qquad A_R=\int_0^1 F_{R,zz}(c,p+u(z_R-p);\zeta)\,du. A R ( z R − p ) = − r R , A R = ∫ 0 1 F R , z z ( c , p + u ( z R − p ) ; ζ ) d u . The first term of r R r_R r R is the chain-rule contribution of the coarse constraints omitted in the regional problem. With a R = Ψ R ( c , z R ; ζ ) a_R=\Psi_R(c,z_R;\zeta) a R = Ψ R ( c , z R ; ζ ) , for every free label or fine link i i i in Ω \Omega Ω ,
∣ ( r R ) i ∣ + ∣ ( z R − p ) i ∣ + ∣ ( a R − a T 0 ) i ∣ ≤ C J ( ∥ c ∥ ∞ + ∥ ζ ∥ ∞ ) e − c J dist ( i , W ) , |(r_R)_i|+|(z_R-p)_i|+|(a_R-a_T^0)_i| \le C_J(\lVert c\rVert_\infty+\lVert\zeta\rVert_\infty)e^{-c_J\operatorname{dist}(i,W)}, ∣ ( r R ) i ∣ + ∣ ( z R − p ) i ∣ + ∣ ( a R − a T 0 ) i ∣ ≤ C J (∥ c ∥ ∞ + ∥ ζ ∥ ∞ ) e − c J dist ( i , W ) , each summand being read on its own list. These three differences have factorial boundary-localized coefficient bounds, as in Lemma 18.15. In particular, for a fixed input x x x and a fixed output e e e in Ω \Omega Ω , their differentiated coefficient series are bounded by C J e − c J [ dist ( e , W ) + dist ( e , x ) ] C_J e^{-c_J[\operatorname{dist}(e,W)+\operatorname{dist}(e,x)]} C J e − c J [ dist ( e , W ) + dist ( e , x )] .
(Full output.) Let u ∗ ( c , ζ ) = T M s exp a R u_*(c,\zeta)=T_M^s\exp a_R u ∗ ( c , ζ ) = T M s exp a R be the full output of the s s s averaging steps, including the values of the coarse links which are not constrained in Ω \Omega Ω . It lies in the common torus target chart on this product. The bounds of part (1) on Ω \Omega Ω , including those with a fixed input, remain true if c ⊕ 0 c\oplus0 c ⊕ 0 is replaced by Log u ∗ \operatorname{Log}u_* Log u ∗ and a T 0 a_T^0 a T 0 by the torus stationary field a T ∗ a_T^* a T ∗ at the target Log u ∗ \operatorname{Log}u_* Log u ∗ . All derivatives through u ∗ u_* u ∗ , including those through z R z_R z R , are included.
(Minimized action.) Let V R ( c ; ζ ) = B s F R ( c , z R ; ζ ) V_R(c;\zeta)=B_sF_R(c,z_R;\zeta) V R ( c ; ζ ) = B s F R ( c , z R ; ζ ) . There are fixed ρ 0 ( J ) , a W , α W , A J c l > 0 \rho_0(J),a_W,\alpha_W,A_J^{\mathrm{cl}}>0 ρ 0 ( J ) , a W , α W , A J cl > 0 such that
χ Ω , a W , α W ρ 0 , c l ( V R − E d 0 + s ( c ⊕ 0 ) ) ≤ A J c l . \chi_{\Omega,a_W,\alpha_W}^{\rho_0,\mathrm{cl}}\left(V_R-E_{d_0+s}(c\oplus0)\right)\le A_J^{\mathrm{cl}}. χ Ω , a W , α W ρ 0 , cl ( V R − E d 0 + s ( c ⊕ 0 ) ) ≤ A J cl . This is the seminorm of Definition 18.1 with the displayed constant classical radius and rates; it includes all degrees l ≥ 1 l \ge1 l ≥ 1 . The constant term of the difference is kept in the full action.
(Newly frozen coordinates. ) For the newly frozen coordinates ξ \xi ξ , at their positions and with their density,
D ξ z R = − F R , z z − 1 F R , z ξ , ∑ f ∣ ( D s f z R ) e ∣ ≤ C J e − a W dist ( e , W ) , D_{\xi}z_R=-F_{R,zz}^{-1}F_{R,z\xi},\qquad\sum_f\left|(D_{sf}z_R)_e\right|\le C_J e^{-a_W\operatorname{dist}(e,W)}, D ξ z R = − F R , z z − 1 F R , z ξ , f ∑ ∣ ( D s f z R ) e ∣ ≤ C J e − a W dist ( e , W ) ,
∥ F R , z ξ ∥ a W + ∥ F R , z c ∥ a W ≤ C J . \|F_{R,z\xi}\|_{a_W}+\|F_{R,zc}\|_{a_W}\le C_J. ∥ F R , z ξ ∥ a W + ∥ F R , z c ∥ a W ≤ C J .
(Hessians. ) The constrained Hessians at the two stationary points satisfy
∣ [ F R , z z ( c , z R ; ξ ) − P F T , z z ( c ⊕ 0 , z T ) P T ] e f ∣ ≤ C J e − a W [ dist ( e , f ) + dist ( e , W ) + dist ( f , W ) ] . \left|\left[F_{R,zz}(c,z_R;\xi)-PF_{T,zz}(c\oplus0,z_T)P^{\mathsf T}\right]_{ef}\right|\le C_J e^{-a_W[\operatorname{dist}(e,f)+\operatorname{dist}(e,W)+\operatorname{dist}(f,W)]}. [ F R , z z ( c , z R ; ξ ) − P F T , z z ( c ⊕ 0 , z T ) P T ] e f ≤ C J e − a W [ dist ( e , f ) + dist ( e , W ) + dist ( f , W )] .
The order-l l l coefficient tensors of this difference, with every coarse-field and free argument in the connecting network and any root in Ω \Omega Ω fixed, are bounded by ( l + 2 ) ! C J l + 2 (l+2)!C_J^{l+2} ( l + 2 )! C J l + 2 at fixed exponents.
(Successive minima, inverse square roots, and Haar disintegration densities. ) Use a common smaller full torus product of higher coordinates y = ( y L , y H ) y=(y_L,y_H) y = ( y L , y H ) for each initial segment I I I . On this product compare the regional conditional center a I R ( c , y L ; ξ ) a_I^R(c,y_L;\xi) a I R ( c , y L ; ξ ) with the regional components of a I T ( c ⊕ 0 , y L , y H ) a_I^T(c\oplus0,y_L,y_H) a I T ( c ⊕ 0 , y L , y H ) , and compare the Hessian of the regional partial minimum with the LL block of the torus partial-minimum Hessian at that same full argument. The conditional-center differences have the boundary-localized coefficient bounds of part (1), with every parameter input fixed in turn. The Hessian differences have the bounds of part (5), with the exact normalization M i M^i M i at scale i i i . The common product can be taken as ∥ y ∥ c o e f f < η I / 2 \|y\|_{\mathrm{coeff}}<\eta_I/2 ∥ y ∥ coeff < η I /2 after the coarse and frozen radii are reduced as in Step 10 of the proof; it is not a slice at an unspecified y H y_H y H . The inverse square roots of these Hessians and the nonconstant coefficients of the logarithms of the Haar disintegration densities satisfy the corresponding boundary-localized comparisons, with any index fixed. Scalar determinants and Haar disintegration densities which are not rooted at a label in Ω \Omega Ω are kept as exact quantities.
All constants have the dependence of Section 19.1 and do not depend on the later widths.
The theorem gives no inverse on fixed boundary coordinates and no rank estimate for the boundary terms at a nonzero coarse field.
Proof. We pose both problems on the torus free list by interpolating between the torus objective and the augmented regional objective. The two objectives then differ by one inserted factor localized near H H H , and all estimates follow from the response of the interpolated stationary point to this factor and from the exponential decay of the inverse Hessian.
Step 1: one coordinate list. We write z = ( z L , z H ) z=(z_L,z_H) z = ( z L , z H ) for the torus free list and set
F ~ R ( c , z ; ξ ) = F R ( c , z L ; ξ ) + 1 2 ∣ z H ∣ 2 , F ~ t = ( 1 − t ) F T + t F ~ R , 0 ≤ t ≤ 1 , D = F ~ R − F T . (16) \widetilde{F}_R(c,z;\xi)=F_R(c,z_L;\xi)+\frac{1}{2}|z_H|^2,\qquad\widetilde{F}_t=(1-t)F_T+t\widetilde{F}_R,\qquad0\le t\le1,\qquad D=\widetilde{F}_R-F_T. \tag*{(16)} F R ( c , z ; ξ ) = F R ( c , z L ; ξ ) + 2 1 ∣ z H ∣ 2 , F t = ( 1 − t ) F T + t F R , 0 ≤ t ≤ 1 , D = F R − F T . (16)
Since its minimum is zero, the added auxiliary quadratic does not change the regional value. Theorem 18.12 gives a positive gap at the identity for the Hessian of this interpolation, for all its principal blocks, and for its Schur complements. We reduce the construction product by a fixed amount so that the coefficient variation of each of these Hessians satisfies the condition of Lemma 19.1(2), uniformly in t t t . This is possible since the variation is at most Q J ( ∥ c ∥ + ∥ z ∥ + ∥ ξ ∥ ) Q_J(\lVert c\rVert+ \lVert z\rVert+ \lVert\xi\rVert) Q J (∥ c ∥ + ∥ z ∥ + ∥ ξ ∥) , and the gap at the identity of the interpolation is at least the minimum of the regional gap, the torus gap, and 1. All straight segments of fine fields used below then lie strictly inside the larger classical chart.
We retain a split resolvent margin for the later perturbation argument. Choose a common intrinsic exponent b r e f > 0 b_{\mathrm{ref}} > 0 b ref > 0 and a constant C 0 , J C_{0,J} C 0 , J such that every flat principal and scale Schur reference K i b K_i^b K i b has resolvent norm at most C 0 , J / ( λ J + t ) C_{0,J}/(\lambda_J + t) C 0 , J / ( λ J + t ) at this exponent, including the factors M i M^i M i at the scale stages. When the larger unperturbed products are fixed, reduce them once so that their conditional matrices K 0 , i K_{0,i} K 0 , i satisfy
C 0 , J ∥ K 0 , i − K i b ∥ b r e f ≤ λ J / 8 C_{0,J}\lVert K_{0,i} - K_i^b\rVert_{b_{\mathrm{ref}}} \leq\lambda_J/8 C 0 , J ∥ K 0 , i − K i b ∥ b ref ≤ λ J /8
on those products, uniformly over the finitely many stages. We decrease the unperturbed reserve c J c_J c J if necessary so that c J ≤ b r e f c_J \leq b_{\mathrm{ref}} c J ≤ b ref . These choices use only the unperturbed constructions at fixed J J J and precede the incoming rates, perturbation threshold, and later widths.
Step 2: localization of the difference of the reconstructions. We use the comparison of the nonlinear words in Theorem 18.12. At common internal coarse values and common regional free values, we set
d ( c , z ; ξ ) = Ψ R ( c , z L ; ξ ) − Ψ T ( c ⊕ 0 , z L , z H ) . d(c,z;\xi) = \Psi_R(c,z_L;\xi) - \Psi_T(c \oplus0,z_L,z_H). d ( c , z ; ξ ) = Ψ R ( c , z L ; ξ ) − Ψ T ( c ⊕ 0 , z L , z H ) .
On fine links in Ω \Omega Ω , its defining words vanish outside the fixed interface band. The prescribed values on H H H can differ, and their difference is a bounded source on H H H with the separate coefficient bounds of the reconstruction charts. The difference is supported in H H H together with that band, hence in H L J H_{LJ} H L J ; the two extensions need not agree on H H H . Its coefficient tensors have the corresponding finite-range bounds, rooted at any index. To check the statement about fine links in Ω \Omega Ω at nonzero arguments, consider an internal eliminated-edge equation at the coarsest intermediate scale. On the two sides it has the same endpoint blocks, the same coarse value, and the same free values, so by uniqueness on the common local branch its solution is the same. We repeat this down the s s s intermediate scales. Only equations meeting a block of H H H can differ, and the union of their at most J J J stencils is contained in H L J H_{LJ} H L J . This argument keeps the values of the coarse links across the boundary of the block region as functions of the common free values; it imposes no value on an omitted coarse constraint.
Step 3: the inserted difference. The difference in (16) is
D = ∫ 0 1 E d 0 ′ ( Ψ T + v d ) [ d ] d v + 1 2 ∣ z H ∣ 2 . (17) D=\int_0^1 E'_{d_0}(\Psi_T+vd)[d]\,dv+\frac{1}{2}\lvert z_H\rvert^2. \tag*{(17)} D = ∫ 0 1 E d 0 ′ ( Ψ T + v d ) [ d ] d v + 2 1 ∣ z H ∣ 2 . (17)
Every derivative of the integral either differentiates d d d , whose fine position stays in H L J H_{LJ} H L J , or leaves such a factor undifferentiated. In the latter case the classical tensor of Theorem 10.11, which is bounded at every root, connects that position to every differentiated chart argument. A derivative in a coarse field or in a free variable in Ω \Omega Ω therefore cannot disconnect the difference factor from the network. The auxiliary quadratic has only indices in H H H , and any connected contribution from it to an external index in Ω \Omega Ω passes through an index in H H H and is bounded by (2) . This verifies the hypotheses of Lemma 18.15 for the whole inserted difference, at every order and with any external root.
Step 4: the gradient formula and the bounds on values. Torus stationarity says that J T T E d 0 ′ ( a T 0 ) = 0 J_T^T E'_{d_0}(a_T^0)=0 J T T E d 0 ′ ( a T 0 ) = 0 . Adding and subtracting J R T E d 0 ′ ( a T 0 ) J_R^T E'_{d_0}(a_T^0) J R T E d 0 ′ ( a T 0 ) in F R , z ( c , p ; ξ ) = J R T E d 0 ′ ( a ^ R ) F_{R,z}(c,p;\xi)=J_R^T E'_{d_0}(\widehat a_R) F R , z ( c , p ; ξ ) = J R T E d 0 ′ ( a R ) , and applying the fundamental theorem of calculus on the complex segment from a T 0 a_T^0 a T 0 to a ^ R \widehat a_R a R , we obtain the formula for r R r_R r R . The difference J R − J T J_R-J_T J R − J T is supported in the construction band, in its regional input indices and its fine output indices. Its product with the gradient of the action on the fine links is kept: torus stationarity makes only the projection of this gradient onto the torus free directions vanish. By the classical bounds, we have
∥ E d 0 ′ ( a T 0 ) ∥ ∞ + ∥ d a ∥ ∞ ≤ C J ( ∥ c ∥ ∞ + ∥ ξ ∥ ∞ ) ; \lVert E'_{d_0}(a_T^0)\rVert_\infty+\lVert d_a\rVert_\infty\le C_J(\lVert c\rVert_\infty+\lVert\xi\rVert_\infty); ∥ E d 0 ′ ( a T 0 ) ∥ ∞ + ∥ d a ∥ ∞ ≤ C J (∥ c ∥ ∞ + ∥ ξ ∥ ∞ ) ;
for the first summand we used the bound on the torus stationary point, and for the second the word comparison of Step 2 and the same bound for ( p , h ) (p,h) ( p , h ) . Contracting the two terms of r R r_R r R with their localized sources and using (2) , we obtain
∣ ( r R ) i ∣ ≤ C J ( ∥ c ∥ ∞ + ∥ ξ ∥ ∞ ) e − c J δ ( i ) . (18) \lvert(r_R)_i\rvert\le C_J(\lVert c\rVert_\infty+\lVert\xi\rVert_\infty)e^{-c_J\delta(i)}. \tag*{(18)} ∣( r R ) i ∣ ≤ C J (∥ c ∥ ∞ + ∥ ξ ∥ ∞ ) e − c J δ ( i ) . (18)
Subtracting the regional gradient equations at z R z_R z R and at p p p and integrating along the segment between them, we obtain A R ( z R − p ) = − r R A_R(z_R-p)=-r_R A R ( z R − p ) = − r R . The segment lies in the convex larger free box, and its averaged Hessian is the positive regional Hessian at the identity plus a coefficient perturbation satisfying the condition of Lemma 19.1 (2). By that lemma, the inverse has entries bounded by C J e − b J dist ( e , i ) C_J e^{-b_J\operatorname{dist}(e,i)} C J e − b J dist ( e , i ) . Convolving this bound with (18) and using δ ( e ) ≤ δ ( i ) + dist ( e , i ) + L J ′ \delta(e)\le\delta(i)+\operatorname{dist}(e,i)+L'_J δ ( e ) ≤ δ ( i ) + dist ( e , i ) + L J ′ , we obtain the bound on z R − p z_R-p z R − p , at a fixed smaller exponent. For the fine field we have the identity
a R − a T 0 = d a + ∫ 0 1 Ψ R , z ( c , p + v ( z R − p ) ; ξ ) ( z R − p ) d v . (19) a_R-a_T^0=d_a+\int_0^1\Psi_{R,z}(c,p+v(z_R-p);\xi)(z_R-p)\,dv. \tag*{(19)} a R − a T 0 = d a + ∫ 0 1 Ψ R , z ( c , p + v ( z R − p ) ; ξ ) ( z R − p ) d v . (19)
The first term is supported in H L J H_{LJ} H L J , and the second is the product of a finite-range chart kernel, without boundary weight, with z R − p z_R-p z R − p . This proves the bound on fine links in Ω \Omega Ω . No decay is asserted at a fine link deep in H H H .
Step 5: the coefficient bounds. We now prove the coefficient bounds directly, using the interpolation (16) . On a common smaller product of parameters, the stationary point z t z_t z t of F t F_t F t is obtained by the implicit iteration with the inverse of the free Hessian at the identity, which converges coefficientwise. The radius condition and the bound on the inverse times the gradient follow from Theorem 18.12, and the scalar comparison equation is (3) . Differentiating in t t t , we obtain
z ˙ t = − F t , z z − 1 D z ( c , z t ; ξ ) . (20) \dot{z}_t=-F_{t,zz}^{-1}D_z(c,z_t;\xi). \tag*{(20)} z ˙ t = − F t , z z − 1 D z ( c , z t ; ξ ) . (20)
This expression contains exactly one inserted difference from (17) , and its convergent majorant is (4) . A further derivative in an input may hit the inverse, the inserted difference, or its argument z t z_t z t ; in each case the joined network still contains the inserted difference and the index of the new input. We apply Lemma 18.15 and integrate from t = 0 t=0 t = 0 to t = 1 t=1 t = 1 . Since the endpoints are z T z_T z T and ( z R , 0 ) (z_R,0) ( z R , 0 ) , their difference on the regional list has factorial boundary-localized coefficient bounds, rooted at every index. Reading (19) as an identity of convergent majorant series, we obtain the corresponding bounds for the fine field, and the same argument applied to the formula for r R r_R r R gives its bounds. Keeping one input x x x fixed and summing over the remaining indices, we obtain the decay with a fixed input stated in part (1).
Step 6: the full output. The composition T M S exp Ψ R T_M^S\exp\Psi_R T M S exp Ψ R is holomorphic on the common larger chart, by the construction in Theorem 18.12. At the stationary point it satisfies
∥ Log u ∗ ∥ ∞ ≤ C J ( ∥ c ∥ ∞ + ∥ ξ ∥ ∞ ) . \lVert\operatorname{Log}u_*\rVert_\infty\le C_J(\lVert c\rVert_\infty+\lVert\xi\rVert_\infty). ∥ Log u ∗ ∥ ∞ ≤ C J (∥ c ∥ ∞ + ∥ ξ ∥ ∞ ) .
The internal coordinates of Log u ∗ \operatorname{Log}u_* Log u ∗ are c c c , and its other coordinates are the values of the coarse links which are not constrained in Ω \Omega Ω . These are analytic functions of ( c , z ; ξ ) (c,z;\xi) ( c , z ; ξ ) , whose differentiated series compose the coefficient series of the chart words with the series of the stationary point from Step 5. For instance, for any analytic direction v v v the first derivative is
D v Log u ∗ = D c T D v c + D z T D v z R + D ξ T D v ξ , T = Log ( T M S exp Ψ R ) . D_v\operatorname{Log}u_*=D_cT\,D_vc+D_zT\,D_vz_R+D_\xi T\,D_v\xi,\qquad T=\operatorname{Log}(T_M^S\exp\Psi_R). D v Log u ∗ = D c T D v c + D z T D v z R + D ξ T D v ξ , T = Log ( T M S exp Ψ R ) .
Since higher derivatives are compositions of these factors, all derivatives through the stationary point are included. We reduce the parameter radius so that the segment from Log u ∗ \operatorname{Log}u_* Log u ∗ to c ⊕ 0 c\oplus0 c ⊕ 0 lies in the common torus target chart. If A T ( v ) A_T(v) A T ( v ) denotes the torus stationary fine field at the target v v v , then
a T ∗ − a T 0 = ∫ 0 1 D A T ( c ⊕ 0 + t ( Log u ∗ − ( c ⊕ 0 ) ) ) [ Log u ∗ − ( c ⊕ 0 ) ] d t . (21) a_T^*-a_T^0=\int_0^1D A_T\bigl(c\oplus0+t(\operatorname{Log}u_*-(c\oplus0))\bigr)\bigl[\operatorname{Log}u_*-(c\oplus0)\bigr]\,dt. \tag*{(21)} a T ∗ − a T 0 = ∫ 0 1 D A T ( c ⊕ 0 + t ( Log u ∗ − ( c ⊕ 0 )) ) [ Log u ∗ − ( c ⊕ 0 ) ] d t . (21)
The difference of targets is supported on coarse links which are not internal to Ω \Omega Ω . By the same implicit equation on the torus list, the lower bound for the torus constrained Hessian, the chart bounds, and the classical tensors give coefficient bounds without boundary weight for the response of A T \mathcal{A}_T A T . Each term of (21) with a root in Ω \Omega Ω therefore passes through a coarse link outside Ω \Omega Ω , and its composed network, including the derivatives of u ∗ u_* u ∗ , visits H L J H_{L_J} H L J . Lemma 18.15 and the finite word bounds then give the same amplitude and fixed-input bounds as before. The same integral with A T \mathcal{A}_T A T replaced by the torus free stationary point z T z_T z T bounds p ∗ − p p^* - p p ∗ − p on the regional free labels, where ( p ∗ , h ∗ ) = z T ( Log u ∗ ) (p^*,h^*) = z_T(\operatorname{Log} u_*) ( p ∗ , h ∗ ) = z T ( Log u ∗ ) . Thus z R − p ∗ = ( z R − p ) − ( p ∗ − p ) z_R - p^* = (z_R - p) - (p^* - p) z R − p ∗ = ( z R − p ) − ( p ∗ − p ) satisfies the same bounds. For the new residual gradient, we subtract its value at p p p and integrate F R , z z F_{R,zz} F R , z z along the segment from p p p to p ∗ p^* p ∗ . A Hessian without boundary weight multiplies this boundary-localized displacement, so the residual satisfies the same bounds. Its formula follows again from torus stationarity, now at the fixed target Log u ∗ \operatorname{Log} u_* Log u ∗ . These arguments are identities of convergent series in all variables, so they also give the fixed-input bounds at the full output. A representative in complementary forest coordinates, if needed, can use forest words of fixed depth based at the regional vertices and the coarse roots. These words change only the construction stencil outside Ω \Omega Ω , their coefficient series are among the chart bounds, and the comparison in Ω \Omega Ω is unchanged.
Step 7: the minimized action. Let f t ( c ; ζ ) = F t ( c , z t ; ζ ) f_t(c;\zeta)=F_t(c,z_t;\zeta) f t ( c ; ζ ) = F t ( c , z t ; ζ ) . By stationarity and Lemma 18.8, we have
∂ t f t = D ( c , z t ; ζ ) . (22) \partial_t f_t = D(c,z_t;\zeta). \tag*{(22)} ∂ t f t = D ( c , z t ; ζ ) . (22)
At the endpoints, B s f 1 = V R B_s f_1 = V_R B s f 1 = V R and B s f 0 = E d 0 + s ( c ⊕ 0 ) B_s f_0 = E_{d_0+s}(c\oplus0) B s f 0 = E d 0 + s ( c ⊕ 0 ) . The second identity follows by iterating the recurrence E d + 1 ( c ) = M min z E d ( b ( c , z ) ) E_{d+1}(c)=M\min_z E_d(b(c,z)) E d + 1 ( c ) = M min z E d ( b ( c , z )) which defines the effective classical action: by Lemma 18.8 the successive minima equal the joint minimum, and the product of the s s s factors M M M is B s B_s B s . A nonconstant derivative of the right side of (22) has the network structure checked in Step 3. Therefore, for k ≥ 1 k\ge1 k ≥ 1 and a root i i i in Ω \Omega Ω ,
∑ i 2 , … , i k e c J t ( i , i 2 , … , i k ) ∣ D k [ B s ( f 1 − f 0 ) ] ( 0 ; ζ ) i , i 2 , … , i k ∣ ≤ k ! C J k + 1 e − c J δ ( i ) , (23) \sum_{i_2,\ldots,i_k} e^{c_J t(i,i_2,\ldots,i_k)}\left|D^k[B_s(f_1-f_0)](0;\zeta)_{i,i_2,\ldots,i_k}\right|
\le k! C_J^{k+1}e^{-c_J\delta(i)}, \tag*{(23)} i 2 , … , i k ∑ e c J t ( i , i 2 , … , i k ) D k [ B s ( f 1 − f 0 )] ( 0 ; ζ ) i , i 2 , … , i k ≤ k ! C J k + 1 e − c J δ ( i ) , (23)
where the factor B s ≤ M J B_s\le M^J B s ≤ M J is included in C J C_J C J . For a scalar series f ( c ) = ∑ ν f ν c ν f(c)=\sum_\nu f_\nu c^\nu f ( c ) = ∑ ν f ν c ν we have, as in (18.2),
∑ ∣ ν ∣ = k ν i k ∣ f ν ∣ e a t ( ν ) = 1 k ! ∑ i 2 , … , i k ∣ D k f ( 0 ) i , i 2 , … , i k ∣ e a t ( i , i 2 , … , i k ) . (24) \sum_{|\nu|=k}\frac{\nu_i}{k}|f_\nu|e^{a_t(\nu)}
=
\frac{1}{k!}\sum_{i_2,\ldots,i_k}|D^k f(0)_{i,i_2,\ldots,i_k}|e^{a_t(i,i_2,\ldots,i_k)}. \tag*{(24)} ∣ ν ∣ = k ∑ k ν i ∣ f ν ∣ e a t ( ν ) = k ! 1 i 2 , … , i k ∑ ∣ D k f ( 0 ) i , i 2 , … , i k ∣ e a t ( i , i 2 , … , i k ) . (24)
Indeed, the multiset obtained from ν \nu ν by removing one occurrence of i i i has ( k − 1 ) ! / [ ( ν i − 1 ) ! ∏ j ≠ i ν j ! ] (k-1)!/[(\nu_i-1)!\prod_{j\ne i}\nu_j!] ( k − 1 )! / [( ν i − 1 )! ∏ j = i ν j !] ordered listings, and D k f ( 0 ) D^k f(0) D k f ( 0 ) evaluated on any of them is ν ! f ν \nu!f_\nu ν ! f ν . Fix a W , α W < c J / 4 a_W,\alpha_W<c_J/4 a W , α W < c J /4 , and then ρ 0 > 0 \rho_0>0 ρ 0 > 0 in the smaller construction domain with C J ρ 0 < 1 / 4 C_J\rho_0<1/4 C J ρ 0 < 1/4 . Inserting (23) into (24) , we bound the regional seminorm by
∑ k ≥ 1 ρ 0 k − 2 C J k + 1 < ∞ . \sum_{k\ge1}\rho_0^{k-2}C_J^{k+1}<\infty. k ≥ 1 ∑ ρ 0 k − 2 C J k + 1 < ∞.
This proves part (3), including degree one. The value B s ( f 1 − f 0 ) ( 0 ; ζ ) B_s(f_1-f_0)(0;\zeta) B s ( f 1 − f 0 ) ( 0 ; ζ ) is never subtracted in (22) ; it remains in the full action.
Step 8: derivatives in the newly frozen coordinates. Differentiating the regional gradient equation in ξ \xi ξ and using invertibility of the Hessian, we obtain D ξ z R = − F R , z z − 1 F R , z ξ D_\xi z_R=-F_{R,zz}^{-1}F_{R,z\xi} D ξ z R = − F R , z z − 1 F R , z ξ . The mixed derivative has exactly two terms,
F R , u v = Ψ R , u T E d 0 ′ ′ ( Ψ R ) Ψ R , v + E d 0 ′ ( Ψ R ) Ψ R , u v , (25) F_{R,uv}=\Psi_{R,u}^{\mathsf{T}}E_{d_0}^{\prime\prime}(\Psi_R)\Psi_{R,v}+E_{d_0}^{\prime}(\Psi_R)\Psi_{R,uv}, \tag*{(25)} F R , uv = Ψ R , u T E d 0 ′′ ( Ψ R ) Ψ R , v + E d 0 ′ ( Ψ R ) Ψ R , uv , (25)
and the bounds of Theorem 18.12 on the larger products control both terms for ( u , v ) = ( z , ξ ) (u,v)=(z,\xi) ( u , v ) = ( z , ξ ) and ( u , v ) = ( z , c ) (u,v)=(z,c) ( u , v ) = ( z , c ) . An index of ξ \xi ξ is at a fine position outside Ω \Omega Ω or in its fixed interface stencil. Contracting it with the free inverse and summing over the labels of ξ \xi ξ , we obtain the row sum C J e − a W δ ( e ) C_J e^{-aW\delta(e)} C J e − aW δ ( e ) by (2) . This proves both bounds on the mixed matrices and the bound on D ξ z R D_\xi z_R D ξ z R , with no change of the density of the inputs.
Step 9: the Hessians. We compare the endpoints of (16) at their stationary points. Before any elimination, the total derivative is
d d t F t , z z ( c , z t ; ξ ) = D z z ( c , z t ; ξ ) + F t ( 3 ) ( c , z t ; ξ ) [ z ˙ t ] . (26) \frac{d}{dt}F_{t,zz}(c,z_t;\xi)=D_{zz}(c,z_t;\xi)+F_t^{(3)}(c,z_t;\xi)[\dot z_t]. \tag*{(26)} d t d F t , z z ( c , z t ; ξ ) = D z z ( c , z t ; ξ ) + F t ( 3 ) ( c , z t ; ξ ) [ z ˙ t ] . (26)
Both terms contain the inserted difference, by (17) and (20) , and all terms of (25) , including the one with the gradient of the action, are part of these tensors. For two matrix roots in Ω \Omega Ω , Lemma 18.15 gives the three distances in the Hessian bound of part (5), and at degree l l l in additional coarse-field and free arguments it gives the bound ( l + 2 ) ! C J l + 2 (l+2)!C_J^{l+2} ( l + 2 )! C J l + 2 after C J C_J C J is increased. For example, the corresponding coefficient series of the matrix is dominated on radius R R R by
∑ l ≥ 0 ( l + 2 ) ! l ! C J l + 2 R l = 2 C J 2 ( 1 − C J R ) 3 , C J R < 1. \sum_{l\ge0}\frac{(l+2)!}{l!}C_J^{l+2}R^l=\frac{2C_J^2}{(1-C_JR)^3},\qquad C_JR<1. l ≥ 0 ∑ l ! ( l + 2 )! C J l + 2 R l = ( 1 − C J R ) 3 2 C J 2 , C J R < 1.
The same majorant with a fixed input gives every version rooted at an input.
Step 10: successive minima. We restrict to smaller products for the initial segments before this comparison is used, following the construction of Lemma 19.3 (1) uniformly in t t t . For an initial segment I I I , we take a solution radius r I r_I r I about the joint stationary point and a smaller radius η I \eta_I η I for the deviation of the coarser coordinates. The principal inverse has bound I J I_J I J and the mixed tensors have bound C J C_J C J , so at the joint stationary point the inverse times the conditional gradient is at most I J C J η I I_JC_J\eta_I I J C J η I . We require I J C J r I < 1 / 8 I_JC_Jr_I<1/8 I J C J r I < 1/8 and I J C J η I < r I / 8 I_JC_J\eta_I<r_I/8 I J C J η I < r I /8 within the margin of the chart, uniformly in t t t . Then the conditional Newton map is a contraction on the solution domain. We reduce the radii of the coarse field and of ξ \xi ξ so that sup t ∥ z t ( c ; ξ ) ∥ c o e f f < min I η I / 8 \sup_t\|z_t(c;\xi)\|_{\mathrm{coeff}}<\min_I\eta_I/8 sup t ∥ z t ( c ; ξ ) ∥ coeff < min I η I /8 . Then a common product ∥ y ∥ c o e f f < η I / 2 \|y\|_{\mathrm{coeff}}<\eta_I/2 ∥ y ∥ coeff < η I /2 of coarser coordinates lies in the allowed domain for every t t t . On this product the conditional response is z ˙ I = − F t , I I − 1 D I \dot z_I=-F_{t,II}^{-1}D_I z ˙ I = − F t , I I − 1 D I at fixed y y y , so the argument of Step 5 applies before a partial minimum is taken. Since there are finitely many initial segments, a common choice is possible, and this is the smaller product in part (6). At such a partial minimum, Lemma 18.8 gives
f y y = F y y − F y I F I I − 1 F I y , D 3 f [ η 1 , η 2 , η 3 ] = D 3 F [ V η 1 , V η 2 , V η 3 ] . f_{yy}=F_{yy}-F_{yI}F_{II}^{-1}F_{Iy}, \qquad D^3f[\eta_1,\eta_2,\eta_3]=D^3F[V\eta_1,V\eta_2,V\eta_3]. f y y = F y y − F y I F I I − 1 F I y , D 3 f [ η 1 , η 2 , η 3 ] = D 3 F [ V η 1 , V η 2 , V η 3 ] .
Differentiating the Schur formula in t t t replaces one factor by its derivative, which contains the inserted difference; an inverse contributes − F I I − 1 F ˙ I I F I I − 1 -F_{II}^{-1}\dot{F}_{II}F_{II}^{-1} − F I I − 1 F ˙ I I F I I − 1 , and evaluation at a moving stationary point contributes its response, which also contains the inserted difference. Each resulting network keeps all external roots and a visit to H L J H_{LJ} H L J . At scale i i i the formula is multiplied by the normalization M i M^i M i of the action, and a finite composition over i < J i<J i < J changes only C J C_J C J .
Step 11: inverse square roots, and fixing versus eliminating. We apply (19.7) to the Hessians of Step 10. By the common gap and the resolvent condition, the differences of the inverse square roots have integrable boundary-localized coefficient bounds, rooted at every index. To compare ( K L L T ) − 1 / 2 (K^T_{LL})^{-1/2} ( K LL T ) − 1/2 with ( ( K T ) − 1 / 2 ) L L ((K^T)^{-1/2})_{LL} (( K T ) − 1/2 ) LL , we set C = K L L T ⊕ K H H T C=K^T_{LL}\oplus K^T_{HH} C = K LL T ⊕ K H H T and apply the resolvent identity to K T − C K^T-C K T − C before integrating the resolvents. Each off-diagonal factor connects the two sides of the boundary, and hence gives the same boundary-localized estimate rooted in Ω \Omega Ω . This also treats the difference between eliminating and fixing the torus free coordinates outside the regional list. For a free matrix T T T written in blocks with respect to the regional and the remaining torus coordinates, (19.9) expresses ( T − 1 ) L L − A − 1 (T^{-1})_{LL}-A^{-1} ( T − 1 ) LL − A − 1 through the Schur complement D 0 − C A − 1 B D_0-CA^{-1}B D 0 − C A − 1 B on the remaining coordinates, which has the gap of Theorem 18.12. The factors B B B and C C C pass through the remaining coordinates and keep the visit to H L J H_{LJ} H L J , also after T T T is replaced by T + t T+t T + t . The same identity inside a larger Schur formula for one scale treats elimination instead of fixing.
Step 12: logarithms of the Haar disintegration densities. In the fixed coordinates, the Haar disintegration density of one intermediate scale is that of Lemma 18.7:
c H a a r f ∏ e n o n f o r e s t j H ( Γ e ) ∏ c det D X e c ∗ C b , c ( Γ ) , j H ( X ) = sinc 2 ( ∣ X ∣ / 2 ) , (27) c^f_{\mathrm{Haar}}\frac{\prod_{e\ \mathrm{nonforest}}j_H(\Gamma_e)}{\prod_c\det D_{X_{e_c^*}}C_{b,c}(\Gamma)}, \qquad j_H(X)=\operatorname{sinc}^2(|X|/2), \tag*{(27)} c Haar f ∏ c det D X e c ∗ C b , c ( Γ ) ∏ e nonforest j H ( Γ e ) , j H ( X ) = sinc 2 ( ∣ X ∣/2 ) , (27)
where e c ∗ e_c^* e c ∗ is the eliminated edge of the constraint at the coarse link c c c ; the Haar factors of the solved eliminated edges are included in the numerator. At the common zero the local factors of the numerator equal 1 1 1 and the local constraint determinants equal M − 6 M^{-6} M − 6 . By the word coefficients of fixed depth and the bounds on the implicit chart, there exists a uniform smaller radius on which each factor, divided by its nonzero value at zero, has nonconstant part with absolute coefficient row sum less than 1 / 2 1/2 1/2 . The series log ( 1 + x ) = ∑ k ≥ 1 ( − 1 ) k + 1 x k / k \log(1+x)=\sum_{k\geq1}(-1)^{k+1}x^k/k log ( 1 + x ) = ∑ k ≥ 1 ( − 1 ) k + 1 x k / k therefore converges absolutely, with the attached local networks. We compose these logarithms with the conditional stationary points and the inverse square roots. The words for fine links in Ω \Omega Ω agree off the interface band, and the prescribed values on H H H remain bounded sources. Thus every difference has an output in H L J H_{LJ} H L J , and the derivatives of the other factors stay joined to it. Logarithmic matrix factors are treated by (8) . A nonconstant scalar logarithmic coefficient has a differentiated index in Ω \Omega Ω at which its trace cycle can be rooted, so the rooted sums above bound it. Pure scalar determinants, Haar disintegration densities, and traces over H H H have no such index and are kept as their exact finite values. All choices involved only the fixed construction data of § 2.1.1 , which proves the stated dependence of the constants.
A gauge transformation fixing prescribed vertices
On the torus, Theorem 13.2 brings the reconstructed field into a small neighborhood of the identity by a gauge transformation which depends only on the coarse field. On a block region this gauge transformation must not change the field at the vertices to which frozen coordinates are attached. A frozen coordinate is stored as its based transport U ( σ e ) a e U(\sigma_e)a_e U ( σ e ) a e : the transport along a path σ e \sigma_e σ e from its basepoint to a link e e e , times the frozen value a e a_e a e of that link (Lemma 18.7). Under a gauge transformation G G G the frozen value transforms as a e G = G v a e G w − 1 a_e^G=G_v a_e G_w^{-1} a e G = G v a e G w − 1 , where v , w v,w v , w are the endpoints of e e e . We therefore multiply the logarithm of the torus gauge transformation by a cutoff which vanishes near H H H . To avoid a clash with the block region Ω \Omega Ω , we write Ω d 0 , s T \Omega^{\mathsf T}_{d_0,s} Ω d 0 , s T for the gauge transformation of Theorem 13.2, applied to a composition of s s s averaging steps at depth d 0 d_0 d 0 .
Lemma 19.5. Fix M M M , J J J , and the geometry of a composition of s ≤ J s\leq J s ≤ J averaging steps as above. Let the prescribed protected set contain all vertices of H H H , every fine endpoint and every basepoint of each based transport of a newly or previously frozen coordinate, and all endpoints required in finer-lattice representations of these transports. Assume that its protected vertices outside H H H lie in B B B , the two-block neighborhood of H H H in output units, so they lie in the zero region of the cutoff below. Let 0 ≤ χ ≤ 1 0\leq\chi\leq1 0 ≤ χ ≤ 1 be fixed by the geometry, zero on H H H and its neighborhood of width two blocks, one outside its neighborhood of width three blocks, and zero at every prescribed protected vertex. Define
k ( c ) = Log Ω d 0 , s T ( c ⊕ 0 ) , G Ω ( c ) = exp ( χ k ( c ) ) . k(c)=\operatorname{Log}\Omega^{\mathsf T}_{d_0,s}(c\oplus0),\qquad G^\Omega(c)=\exp(\chi k(c)). k ( c ) = Log Ω d 0 , s T ( c ⊕ 0 ) , G Ω ( c ) = exp ( χ k ( c )) . There are u J , A J , a G > 0 u_J,A_J,a_G>0 u J , A J , a G > 0 such that the following statements hold on ∥ c ∥ ∞ < u J \lVert c\rVert_\infty<u_J ∥ c ∥ ∞ < u J .
The maps k k k and G Ω G^\Omega G Ω are holomorphic, periodic, and equivariant under constant global color rotations. At real data they take values in su ( 2 ) \operatorname{su}(2) su ( 2 ) and SU ( 2 ) \operatorname{SU}(2) SU ( 2 ) , respectively, and ∥ k ( c ) ∥ ∞ ≤ A J ∥ c ∥ ∞ \lVert k(c)\rVert_\infty\leq A_J\lVert c\rVert_\infty ∥ k ( c ) ∥ ∞ ≤ A J ∥ c ∥ ∞ . Their order-l l l absolute coefficients are bounded by l ! C J l + 1 l!C_J^{l+1} l ! C J l + 1 , with the connecting-network weight at exponent a G a_G a G and with the volume-normalized column of every fixed input. Replacing the volume normalization of a column for a fine output by counting measure costs at most the factor M 3 s ≤ M 3 J M^{3s}\leq M^{3J} M 3 s ≤ M 3 J .
At fixed geometry G Ω G^\Omega G Ω depends only on c c c , not on the free coordinates or the frozen coordinates. It is exactly I I I at c = 0 c=0 c = 0 , at the coarse roots, and at all protected vertices, also for complex c c c .
Consequently, for every reconstruction U U U in the chart before gauge transformation,
( T M s U G Ω ( c ) ) i n t e r n a l = ( T M s U ) i n t e r n a l , ( U G Ω ( c ) ) ( σ e ) a e G Ω ( c ) = U ( σ e ) a e \left(T_M^s U^{G^{\Omega}(c)}\right)_{\mathrm{internal}}=\left(T_M^s U\right)_{\mathrm{internal}},\qquad\left(U^{G^{\Omega}(c)}\right)(\sigma_e)a_e^{G^{\Omega}(c)}=U(\sigma_e)a_e ( T M s U G Ω ( c ) ) internal = ( T M s U ) internal , ( U G Ω ( c ) ) ( σ e ) a e G Ω ( c ) = U ( σ e ) a e
for every based transport of a frozen coordinate. A frozen link across the boundary of the block region with both fine endpoints protected is also unchanged.
The gauge transformation is applied after the reconstruction in the rooted spanning-forest coordinates. It does not change the free coordinates, the support restrictions, the Haar disintegration density, or the determinant of the normalization of the fluctuation variables.
The constants have the dependence of § 2.1.1 and do not depend on the later widths or on the size of the protected set. The existence of a large-field decomposition whose protected vertices satisfy these hypotheses is a geometric hypothesis, not a conclusion of the lemma.
Proof. We take the torus gauge transformation of Theorem 13.2 for each s ≤ J s \le J s ≤ J , with a common smaller domain of targets. Extending the target by zero on the coarse links outside Ω \Omega Ω deletes coefficients from its majorant. After a construction constant A A A is increased, the gauge transformation minus the identity has absolute row series at most
U ( t ) = A t 1 − A t ; U(t)=\frac{At}{1-At}; U ( t ) = 1 − A t A t ;
here and below this includes the connecting network and the versions with a fixed input. Theorem 13.2 also provides the logarithm of the gauge transformation. To check the domain condition, we fix u J ≤ U T u_J \le U_T u J ≤ U T inside the larger target domain with A u J < 1 / 4 Au_J < 1/4 A u J < 1/4 . This is the case Q T = 0 Q_T=0 Q T = 0 of (19.1), and it gives u J < r s u_J < r_s u J < r s , a positive denominator, and the fixed intermediate chart margins for every s ≤ J s \le J s ≤ J . The convergent matrix logarithm is bounded by
∑ r ≥ 1 U ( t ) r r = log ( 1 − A t ) − log ( 1 − 2 A t ) ≤ A t 1 − 2 A t ≤ 2 A t ( 0 ≤ t ≤ u J ) . (28) \sum_{r\ge1}\frac{U(t)^r}{r}=\log(1-At)-\log(1-2At)\le\frac{At}{1-2At}\le2At\qquad(0\le t\le u_J). \tag*{(28)} r ≥ 1 ∑ r U ( t ) r = log ( 1 − A t ) − log ( 1 − 2 A t ) ≤ 1 − 2 A t A t ≤ 2 A t ( 0 ≤ t ≤ u J ) . (28)
In the first inequality, we wrote the difference of logarithms as log ( 1 + A t / ( 1 − 2 A t ) ) \log(1+At/(1-2At)) log ( 1 + A t / ( 1 − 2 A t )) and used log ( 1 + x ) ≤ x \log(1+x)\le x log ( 1 + x ) ≤ x for x ≥ 0 x\ge0 x ≥ 0 . Products join coefficient networks. Fixing one input differentiates this majorant series, and on a further fixed smaller radius we obtain order-l l l tensors bounded by l ! C J l + 1 l!C_J^{l+1} l ! C J l + 1 , for an output row or for the volume-normalized column of any input. The same bounds hold for k ( c ) k(c) k ( c ) , and in particular ∥ k ( c ) ∥ ∞ ≤ A J ∥ c ∥ ∞ \lVert k(c)\rVert_\infty\le A_J\lVert c\rVert_\infty ∥ k ( c ) ∥ ∞ ≤ A J ∥ c ∥ ∞ . The fine mesh in output units is M − s M^{-s} M − s , so the volume-normalized column of a fine output carries the factor M − 3 s M^{-3s} M − 3 s . Replacing it by counting measure multiplies the bound by the reciprocal, at most M 3 J M^{3J} M 3 J ; no factor depending on the depth or the volume occurs.
At fixed geometry, multiplication by χ ( v ) ∈ [ 0 , 1 ] \chi(v)\in[0,1] χ ( v ) ∈ [ 0 , 1 ] can only decrease an absolute coefficient row. With L ( t ) = ∑ r ≥ 1 U ( t ) r / r L(t)=\sum_{r\ge1}U(t)^r/r L ( t ) = ∑ r ≥ 1 U ( t ) r / r , exponentiating χ k \chi k χ k gives the coefficientwise majorant exp L ( t ) = ( 1 − A t ) / ( 1 − 2 A t ) \exp L(t)=(1-At)/(1-2At) exp L ( t ) = ( 1 − A t ) / ( 1 − 2 A t ) . This series and its fixed-input derivative series converge on a fixed smaller radius and give the asserted factorial coefficient bounds for G Ω \mathrm{G}^{\Omega} G Ω . The numerical bound L ( t ) ≤ 2 A t L(t) \leq2At L ( t ) ≤ 2 A t in (28) bounds values only and is not used in this coefficient comparison. Periodicity and global color equivariance hold for the torus gauge transformation and its logarithm by Theorem 13.2, and multiplication by the fixed scalar cutoff and exponentiation preserve them. For real targets the logarithm of the compact gauge transformation lies in s u ( 2 ) \mathfrak{su}(2) su ( 2 ) , and its exponential lies in S U ( 2 ) \mathrm{SU}(2) SU ( 2 ) ; the complex functions are their holomorphic continuations with values in s l 2 ( C ) \mathfrak{sl}_{2}(\mathbb{C}) sl 2 ( C ) and S L ( 2 , C ) \mathrm{SL}(2,\mathbb{C}) SL ( 2 , C ) .
The torus gauge transformation is I I I at c = 0 c=0 c = 0 and at each coarse root, so its logarithm vanishes at these points. At every other protected vertex χ = 0 \chi=0 χ = 0 by hypothesis. Hence G Ω = I \mathrm{G}^{\Omega}=I G Ω = I at all these points, identically as a holomorphic function and not only at real arguments.
The transport along a path σ \sigma σ from v v v to w w w transforms as
U G ( σ ) = G v U ( σ ) G w − 1 , U^{G}(\sigma)=G_{v}U(\sigma)G_{w}^{-1}, U G ( σ ) = G v U ( σ ) G w − 1 ,
since all intermediate vertex factors cancel. A based transport U ( σ e ) a e U(\sigma_{e})a_{e} U ( σ e ) a e therefore transforms only at its basepoint and at the far fine endpoint of e e e . Both values of the gauge transformation there are I I I , which implies the second identity in part (3). If both fine endpoints of a frozen link across the boundary of the block region are protected, the gauge transformation also leaves its value unchanged.
For the coarse output, every based loop holonomy in the centered average (4.3) transforms by conjugation at its initial coarse root, and its reference path transforms at its two coarse endpoints. On the branch of the logarithm used here, the logarithm commutes with conjugation, term by term in its power series. Thus one averaging step is covariant, with the gauge transformation acting only at the coarse endpoints. Iterating s s s steps cancels the intermediate gauge factors and leaves only the gauge transformation at the two coarse roots of the output link, where it is I I I . This proves the first identity in part (3), also at complex data. No exceptional value of the Borel extension of T M T_{M} T M occurs inside these charts.
Finally, the gauge transformation acts on a field already reconstructed from its free coordinates in the rooted spanning-forest coordinates, and it changes neither these coordinates nor the integration variables. Hence the support restrictions, the Haar disintegration density, and the determinant of the normalization of the fluctuation variables remain the same functions of the original variables. The cutoff enters only as a bounded fixed multiplier; neither the size of its support nor a derivative in its geometric parameters enters any bound. This proves that the constants do not depend on the size of the protected set or on the later widths, under the stated hypothesis on the protected vertices.
The gauge-transformed reconstruction and its input columns
We now apply G Ω ( c ) G^\Omega(c) G Ω ( c ) to the regional reconstruction and compare the result with the gauge-transformed torus reconstruction of Theorem 13.2. Let a R r a w ( c , q ; ζ ) a_R^{\mathrm{raw}}(c,q;\zeta) a R raw ( c , q ; ζ ) be the logarithm of the regional reconstruction before gauge transformation, built from the successive regional conditional stationary points and constrained inverse square roots, and let a T r a w a_T^{\mathrm{raw}} a T raw be the corresponding torus reconstruction. For a fine link e = ( v , w ) e=(v,w) e = ( v , w ) and a logarithmic gauge field k k k , we write
H e χ ( a , k ) = Log ( e χ ( v ) k ( v ) e e a e − χ ( w ) k ( w ) ) . H_e^\chi(a,k)=\operatorname{Log}\left(e^{\chi(v)k(v)}e_e^a e^{-\chi(w)k(w)}\right). H e χ ( a , k ) = Log ( e χ ( v ) k ( v ) e e a e − χ ( w ) k ( w ) ) .
Then Φ R = H χ ( a R r a w , k ( c ) ) \Phi_R=H^\chi(a_R^{\mathrm{raw}},k(c)) Φ R = H χ ( a R raw , k ( c )) is the logarithm of the regional reconstruction transformed by G Ω ( c ) G^\Omega(c) G Ω ( c ) , and Φ T = H 1 ( a T r a w , k ( c ) ) \Phi_T=H^1(a_T^{\mathrm{raw}},k(c)) Φ T = H 1 ( a T raw , k ( c )) is the gauge-transformed torus reconstruction of Theorem 13.2.
Theorem 19.6. On the fixed products of Theorem 19.4 and Lemma 19.5 , fix a common positive coefficient decay rate b c o m b_{\mathrm{com}} b com below the rates of both reconstructions, their conditional matrix functions, and the gauge logarithm of rate a G a_G a G . Reduce the common boundary rate so that 8 a W < b c o m 8a_W<b_{\mathrm{com}} 8 a W < b com , and use network exponents a > 0 a>0 a > 0 with a + 2 a W < b c o m / 2 a+2a_W<b_{\mathrm{com}}/2 a + 2 a W < b com /2 , leaving a positive reserve for density sums. Enlarge L J L_J L J once to include the three-block cutoff band and the fine-link endpoint buffer. On these products build the reconstruction from the free coordinates with the successive regional conditional stationary points and constrained inverse square roots, with small newly frozen coordinates ζ ~ \tilde{\zeta} ζ ~ . There are fixed larger construction radii and positive rates such that, on every smaller product of coarse fields and fluctuation increments satisfying the domain conditions,
Φ R ( c , q ; ζ ~ ) = Φ T ( c ⊕ 0 , q ⊕ 0 ) ∣ Ω + Δ ( c , q ; ζ ~ ) , \Phi_R(c,q;\tilde{\zeta})=\left.\Phi_T(c\oplus0,q\oplus0)\right|_\Omega+\Delta(c,q;\tilde{\zeta}), Φ R ( c , q ; ζ ~ ) = Φ T ( c ⊕ 0 , q ⊕ 0 ) ∣ Ω + Δ ( c , q ; ζ ~ ) , where q ⊕ 0 q\oplus0 q ⊕ 0 uses the same scale indices on the regional list and sets the remaining torus increments to zero. The following statements also hold.
All terms have convergent majorant series in all variables. The correction Δ \Delta Δ has factorial coefficient bounds with a root in Ω \Omega Ω , a connecting network through every input, and a visit to H L J H_{L_J} H L J . In particular, let R \mathbf{R} R be a radius profile on a smaller construction product, let a > 0 a>0 a > 0 be a network exponent in the range just fixed, and let Δ e , ν \Delta_{e,\nu} Δ e , ν be the Taylor coefficients of Δ e \Delta_e Δ e in its analytic variables, namely the coarse field, the increments q q q , and the newly frozen coordinates ζ ~ \tilde{\zeta} ζ ~ . For a fixed input x x x , the differentiated series
D e , x ( R ) = ∑ ν : ν x ≥ 1 ν x ∣ Δ e , ν ∣ R ν − 1 x e a t ( { e } ∪ supp ν ) , \mathcal{D}_{e,x}(\mathbf{R})=\sum_{\nu:\nu_x\geq1}\nu_x\lvert\Delta_{e,\nu}\rvert\mathbf{R}^{\nu-1_x}e^{a t(\{e\}\cup\operatorname{supp}\nu)}, D e , x ( R ) = ν : ν x ≥ 1 ∑ ν x ∣ Δ e , ν ∣ R ν − 1 x e a t ({ e } ∪ supp ν ) , with the measure conventions of § 2.1.1 , satisfies
D e , x ( R ) ≤ C J e − a W [ dist ( e , W ) + dist ( e , x ) ] . \mathcal{D}_{e,x}(\mathbf{R})\leq C_J e^{-a_W[\operatorname{dist}(e,W)+\operatorname{dist}(e,x)]}. D e , x ( R ) ≤ C J e − a W [ dist ( e , W ) + dist ( e , x )] . Unlike C e , x C_{e,x} C e , x in part (4), this series does not contain the factor R x R_x R x .
At q = 0 q=0 q = 0 , the dependence of Δ \Delta Δ on the newly frozen coordinates satisfies, in the coefficient norm in c c c ,
∥ Δ e ( ⋅ , 0 ; ξ ) − Δ e ( ⋅ , 0 ; 0 ) ∥ c o e f f ≤ C F ∥ ξ ∥ ∞ e − a W dist ( e , W ) , Δ ( 0 , 0 ; 0 ) = 0. \left\|\Delta_e(\cdot,0;\xi)-\Delta_e(\cdot,0;0)\right\|_{\mathrm{coeff}}\le C_F\|\xi\|_\infty e^{-a_W\operatorname{dist}(e,W)},\qquad\Delta(0,0;0)=0. ∥ Δ e ( ⋅ , 0 ; ξ ) − Δ e ( ⋅ , 0 ; 0 ) ∥ coeff ≤ C F ∥ ξ ∥ ∞ e − a W dist ( e , W ) , Δ ( 0 , 0 ; 0 ) = 0.
At radius s q s_q s q for the fluctuation increments, the absolute coefficient row sum of the terms of Δ \Delta Δ of positive degree in q q q is at most C Δ , J s q C_{\Delta,J}s_q C Δ , J s q . The rows of the coarse-field terms near the boundary and all rows in the free variables are only bounded by finite constants depending on J J J ; no contraction factor M − J M^{-J} M − J is asserted for them.
Input columns. Let R y ≤ K ∗ ρ R_y\le K_*\rho R y ≤ K ∗ ρ be a construction profile, and choose a 1 , b 1 > 0 a_1,b_1>0 a 1 , b 1 > 0 with b 1 > α b d b_1>\alpha_{\mathrm{bd}} b 1 > α bd and with a 1 + b 1 a_1+b_1 a 1 + b 1 smaller than a decay exponent for which the full reconstruction has coefficient bounds without boundary weight. Define
C e , x ( R ) = ∑ ν ∣ ν x ∣ ∣ Φ R , e , ν ( ξ ) ∣ R ν e a 1 t ( { e } ∪ supp ν ) , C_{e,x}(\mathbf{R})=\sum_\nu|\nu_x|\,|\Phi_{R,e,\nu}(\xi)|R^\nu e^{a_1t(\{e\}\cup\operatorname{supp}\nu)}, C e , x ( R ) = ν ∑ ∣ ν x ∣ ∣ Φ R , e , ν ( ξ ) ∣ R ν e a 1 t ({ e } ∪ supp ν ) ,
where Φ R , e , ν ( ξ ) \Phi_{R,e,\nu}(\xi) Φ R , e , ν ( ξ ) are the Taylor coefficients of Φ R \Phi_R Φ R at the fine output e e e . Then C e , x ( R ) ≤ D J R x e − b 1 dist ( e , x ) C_{e,x}(\mathbf{R})\le D_JR_xe^{-b_1\operatorname{dist}(e,x)} C e , x ( R ) ≤ D J R x e − b 1 dist ( e , x ) , and
∑ e e − α b d d o l d ( e ) C e , x ( R ) ≤ K J ρ e − α b d d o u t ( x ) , \sum_e e^{-\alpha_{\mathrm{bd}}d_{\mathrm{old}}(e)}C_{e,x}(\mathbf{R})\le K_J\rho e^{-\alpha_{\mathrm{bd}}d_{\mathrm{out}}(x)}, e ∑ e − α bd d old ( e ) C e , x ( R ) ≤ K J ρ e − α bd d out ( x ) ,
K J ≤ C M , J D J K ∗ e α b d C J ∑ l ≥ 0 ( l + 1 ) 3 e − ( b 1 − α b d ) l . K_J\le C_{M,J}D_JK_*e^{\alpha_{\mathrm{bd}}C_J}\sum_{l\ge0}(l+1)^3e^{-(b_1-\alpha_{\mathrm{bd}})l}. K J ≤ C M , J D J K ∗ e α bd C J l ≥ 0 ∑ ( l + 1 ) 3 e − ( b 1 − α bd ) l .
Here d o l d d_{\mathrm{old}} d old and d o u t d_{\mathrm{out}} d out are the distances to the complements of the specified old and output position lists, respectively, both expressed in output units, and we assume that, for every summation label e e e and every input x x x ,
d o u t ( x ) ≤ d o l d ( e ) + dist ( e , x ) + C J . d_{\mathrm{out}}(x)\le d_{\mathrm{old}}(e)+\operatorname{dist}(e,x)+C_J. d out ( x ) ≤ d old ( e ) + dist ( e , x ) + C J .
with a buffer C J C_J C J depending only on the fixed construction geometry. The sum runs over the full list of old output labels (the fine labels), at their creation-scale density. For an input among the free coordinates the same statement holds with its own radius R x R_x R x in place of K ∗ ρ K_*\rho K ∗ ρ in the last bound. For a kernel from inputs to outputs, the column of a fixed input carries the volume of the output cells divided by the volume of the cell of that input. The same calculation applies to the profile R y = 2 ρ R_y=2\rho R y = 2 ρ within the auxiliary construction ball and to the smaller spatially varying radius profiles of Definition 18.1. These column bounds are different from the bounds on the image under the reconstruction majorant in Lemma 19.7 .
The constants have the dependence of § 2.1.1 , are uniform in the depth and in the shape of Ω \Omega Ω , and do not depend on the later widths. No equality with a perturbed incoming constrained problem is asserted.
Proof. We compare the successive regional and torus reconstructions before applying the gauge transformation with cutoff, and then estimate the differentiated coefficient series. Step 1: the domain of the successive substitutions. We take a common larger construction product as in Theorem 19.4 , with radii for coarse fields and fluctuation increments at most u T u_T u T and Q T Q_T Q T from (1) . Then the torus reconstruction, its gauge transformation, and all column series below are evaluated with u + v < r s u+v<r_s u + v < r s and denominator at least 1 / 2 1/2 1/2 , and the smaller product keeps the intermediate classical and Haar-density margins. For the finitely many conditional problems at the intermediate scales, we use the smaller parameter domains of Step 10 of the proof of Theorem 19.4 . We choose the radius of the fluctuation increments so that every successive substitution lies strictly inside these domains. Such a radius exists: by Lemma 18.8, at q = 0 q=0 q = 0 the successive conditional stationary points are the joint stationary point. Each conditional stationary point and each inverse square root has a convergent majorant series without boundary weight on a larger product. If A i A_i A i bounds the deviation of the coarser coordinates from the joint stationary point after the scales above i i i have been reconstructed, then by these series we have
A i − 1 ≤ ( 1 + C 0 ) A i + C S s q , A s = 0 , (29) A_{i-1} \le(1+C_0)A_i+C_Ss_q,\qquad A_s=0, \tag*{(29)} A i − 1 ≤ ( 1 + C 0 ) A i + C S s q , A s = 0 , (29)
on a fixed smaller free ball, where the constants are maxima of the rows of the conditional derivatives and of the inverse square roots on the larger products. The same recurrence holds in the coefficient norms. Iterating it at most J J J times, we obtain A i ≤ C S s q ∑ j = 0 J − 1 ( 1 + C 0 ) j A_i\le C_Ss_q\sum_{j=0}^{J-1}(1+C_0)^j A i ≤ C S s q ∑ j = 0 J − 1 ( 1 + C 0 ) j . Taking s q s_q s q smaller than a fixed fraction of every domain margin divided by this finite sum places the successive product inside the domains. This is the case δ V = 0 \delta_V=0 δ V = 0 of Lemma 19.3 (3), and this choice does not reduce the radius of the coarse field.
Step 2: one scale of the reconstruction. On the original list of free coordinates, the reconstruction at scale i i i is
Γ i ( c i + 1 , a i ( c i + 1 ; ζ ) + K i ( c i + 1 ; ζ ) − 1 / 2 q i ; ζ ) . (30) \Gamma_i\bigl(c_{i+1},a_i(c_{i+1};\zeta)+K_i(c_{i+1};\zeta)^{-1/2}q_i;\zeta\bigr). \tag*{(30)} Γ i ( c i + 1 , a i ( c i + 1 ; ζ ) + K i ( c i + 1 ; ζ ) − 1/2 q i ; ζ ) . (30)
Here a i a_i a i is the conditional stationary point of the partial minimum after the finer scales have been eliminated, K i K_i K i is its constrained Hessian at scale i i i , and Γ i \Gamma_i Γ i is the local inverse chart of the block averaging constraint. The action at this scale is M i M^i M i times the partial minimum of the objective in input units, so K i K_i K i includes exactly the factor M i M^i M i . The variable q i q_i q i is the fluctuation increment, and no coupling is absorbed into K i K_i K i .
Step 3: the reconstructions before gauge transformation. On each common list the comparison of the stationary points is given by Theorem 19.4 . To compare inverse square roots, we extend the regional matrix to the full list by its direct sum with a principal free block of the torus matrix on the complementary coordinates. Both summands have the gap. The internal difference is the difference of the regional and torus Hessians at nonzero coarse field, estimated for the conditional partial minimum in Theorem 19.4 (6). The removed off-diagonal blocks of the torus matrix have their two roots on opposite sides of W W W , so they are boundary-localized. The identity (7) then gives the comparison on the common list, with every coarse-field and free argument. When a list of finer coordinates is eliminated on one side and fixed on the other, the identity (9) inserts a chain through that list, as in Step 11 of the proof of Theorem 19.4 , and that chain visits H L J H_{LJ} H L J . Thus the statement concerns the regional matrix itself, not the torus problem with coordinates deleted.
We subtract the regional and the torus versions of (30) , with the torus free inputs outside the regional list set to zero. First we change the chart words at the same arguments; their difference lies in the fixed interface stencil. For each conditional object O (a center or an inverse square root), let y R y_R y R and y T y_T y T be the higher arguments in the two chains, and put y m i x = ( y R , L , y T , H ) y_{\mathrm{mix}}=(y_{R,L},y_{T,H}) y mix = ( y R , L , y T , H ) . The augmented regional objective at the endpoint of (16) is independent of y H y_H y H on its regional components. Thus, on those components,
O R ( y R , L ) − O T ( y T ) = [ O R − O T ] ( y m i x ) + [ O T ( y m i x ) − O T ( y T ) ] . O_R(y_{R,L})-O_T(y_T)=[O_R-O_T](y_{\mathrm{mix}})+[O_T(y_{\mathrm{mix}})-O_T(y_T)]. O R ( y R , L ) − O T ( y T ) = [ O R − O T ] ( y mix ) + [ O T ( y mix ) − O T ( y T )] .
The first difference is the matched-argument comparison in part (6); the second is the integral of the torus derivative along the segment from y T y_T y T to y m i x y_{\mathrm{mix}} y mix . Step 1 restricts both chains to a smaller common product, so this mixed argument and its segment lie in the full product fixed in Step 10. This first changes the conditional object at the same argument and then changes the coarser argument. The difference of an analytic function at two arguments is its derivative integrated along the common segment, that is, a tensor without boundary weight applied to a boundary-localized difference of arguments. At the coarsest scale these are all the terms. Each descent by one scale adds the boundary-localized difference from the scale above as one more difference of arguments. By induction over at most J J J scales, the difference a R r a w − a T r a w a_R^{\mathrm{raw}}-a_T^{\mathrm{raw}} a R raw − a T raw is a convergent boundary-localized series. The values of the torus reconstruction on the complementary outputs, which can be nonzero at zero complementary free input, are joined by networks without boundary weight from their output in H H H to their inputs in Ω \Omega Ω . Such networks cross W W W , and they provide the visit to H L J H_{LJ} H L J at the next substitution. No sum of sources over the volume of H H H without decay occurs.
Step 4: the gauge transformation. We apply the same torus gauge transformation k ( c ) k(c) k ( c ) to both fields and cut it off only on the regional side. After the construction radius is chosen, all arguments of H e χ H^\chi_e H e χ lie in one fixed larger identity chart, and the difference is
Δ e = H e χ ( a R r a w , k ) − H e χ ( a T r a w , k ) + H e χ ( a T r a w , k ) − H e 1 ( a T r a w , k ) . (31) \begin{aligned}
\Delta_e={}&H^\chi_e(a_R^{\mathrm{raw}},k)-H^\chi_e(a_T^{\mathrm{raw}},k)\\
&+H^\chi_e(a_T^{\mathrm{raw}},k)-H^1_e(a_T^{\mathrm{raw}},k). \tag*{(31)}
\end{aligned} Δ e = H e χ ( a R raw , k ) − H e χ ( a T raw , k ) + H e χ ( a T raw , k ) − H e 1 ( a T raw , k ) . (31)
The first line contains a R r a w − a T r a w a_R^{\mathrm{raw}}-a_T^{\mathrm{raw}} a R raw − a T raw as a factor, after the derivative in a a a is integrated. The second line vanishes when both endpoint values of the cutoff equal 1, so its output root lies in a fixed interface band of width three blocks, including the buffer for the endpoints of fine links. Its other factors have the torus and gauge coefficient bounds of Theorem 13.2 and Lemma 19.5 . These facts imply the boundary-localized estimate for both lines of (31) . A coarse-field derivative which hits a gauge factor appends the network of that factor to the existing network and does not remove the visit to H L J H_{L_J} H L J . The torus term without the cutoff is Φ T \Phi_{\mathrm{T}} Φ T . This proves the decomposition and its factorial coefficient bounds.
Step 5: the cutoff at first order. The contribution of the cutoff can be read off at first order. Let k 1 = D k ( 0 ) c k_1 = Dk(0)c k 1 = D k ( 0 ) c . The cutoff part of the linearized correction on the link ( v , w ) (v,w) ( v , w ) is the difference of ( 1 − χ ) k 1 (1-\chi)k_1 ( 1 − χ ) k 1 between its endpoints, which equals
( 1 − χ ( v ) ) ( k 1 ( w ) − k 1 ( v ) ) + ( χ ( v ) − χ ( w ) ) k 1 ( w ) . (1-\chi(v))(k_1(w)-k_1(v))+(\chi(v)-\chi(w))k_1(w). ( 1 − χ ( v )) ( k 1 ( w ) − k 1 ( v )) + ( χ ( v ) − χ ( w )) k 1 ( w ) .
Both terms are contained in (31) . In particular the second term, which contains the gradient of the cutoff, is not removed by treating the gauge transformation as constant in the target.
Step 6: the local logarithms. The local logarithms in this argument have uniform complex coefficient margins. Indeed, in a submultiplicative matrix norm the exponential series are dominated by e t e^t e t , and the logarithm of I + X I+X I + X is dominated by ∑ l ≥ 1 t l / l \sum_{l\geq1}t^l/l ∑ l ≥ 1 t l / l on ∥ X ∥ ≤ t < 1 \lVert X\rVert\leq t<1 ∥ X ∥ ≤ t < 1 . We take the rows of the reconstruction before gauge transformation and of the gauge transformation small enough that the product lies in such a disc with a fixed margin. Composition of majorants and the attached networks then bound every derivative of H e χ H^\chi_e H e χ , with constants independent of χ ∈ [ 0 , 1 ] \chi\in[0,1] χ ∈ [ 0 , 1 ] .
Step 7: fixed inputs and the newly frozen coordinates. Fix an input x x x in the boundary-localized series of the reconstruction and keep its position while summing over the other indices. The connected network contains e e e , x x x , and a visit to H L J H_{L_J} H L J , so by the larger decay exponent and Lemma 18.15, we have D e , x ( R ) ≤ C J e − a W [ δ ( e ) + dist ( e , x ) ] \mathcal{D}_{e,x}(\mathcal{R})\leq C_J e^{-a_W[\delta(e)+\operatorname{dist}(e,x)]} D e , x ( R ) ≤ C J e − a W [ δ ( e ) + dist ( e , x )] . At q = 0 q=0 q = 0 the dependence on ξ \xi ξ satisfies the coefficient identity
Δ e ( c , 0 ; ξ ) − Δ e ( c , 0 ; 0 ) = ∫ 0 1 ∑ f ξ f D ξ f Δ e ( c , 0 ; t ξ ) d t . (32) \Delta_e(c,0;\xi)-\Delta_e(c,0;0)=\int_0^1\sum_f \xi_f D_{\xi_f}\Delta_e(c,0;t\xi)\,\mathrm{d}t. \tag*{(32)} Δ e ( c , 0 ; ξ ) − Δ e ( c , 0 ; 0 ) = ∫ 0 1 f ∑ ξ f D ξ f Δ e ( c , 0 ; t ξ ) d t . (32)
The derivatives in ξ \xi ξ come from its analytic input slots. The coefficient bound on their sum follows from the fixed-input bounds in parts (1) and (6) of Theorem 19.4 , with the analytic input chosen among the labels of ξ \xi ξ , and from Lemma 18.15(3). Summing these fixed-input envelopes by (2) preserves the boundary decay, including in the coefficient norm in c c c . The finite chart compositions preserve these envelopes, and the gauge transformation does not depend on ξ \xi ξ . Integrating their coefficient series in c c c in (32) , we obtain the bound C F ∥ ξ ∥ ∞ e − a W δ ( e ) C_F\lVert\xi\rVert_\infty e^{-a_W\delta(e)} C F ∥ ξ ∥ ∞ e − a W δ ( e ) of part (2). At c = q = ξ = 0 c=q=\xi=0 c = q = ξ = 0 the unperturbed stationary points vanish and both gauge transformations are the identity, so Δ ( 0 , 0 ; 0 ) = 0 \Delta(0,0;0)=0 Δ ( 0 , 0 ; 0 ) = 0 .
Step 8: terms of positive degree in q q q . If a convergent majorant series in all variables is bounded on a larger fixed free radius r Q r_Q r Q , then, for s q ≤ r Q / 2 s_q \le r_Q/2 s q ≤ r Q /2 , its terms of positive total degree in q q q are at most s q s_q s q times a fixed bound for its derivative majorant on the larger product. Applying this to (19.31), we obtain the bound C Δ , J s q C_{\Delta,J}s_q C Δ , J s q of part (3). The rows near the boundary and the rows in the free variables are bounded by the corresponding convergent majorants without boundary weight, with no factor M − J M^{-J} M − J .
Step 9: the input columns. Fix the input x x x in the differentiated coefficient series. Removing one power of its radius R x R_x R x and using the column bound with any input fixed, we see that the connecting network contains a path from e e e to x x x . Assigning b 1 b_1 b 1 of the larger decay exponent to this path gives
C e , x ( R ) ≤ D J R x e − b 1 dist ( e , x ) (33) C_{e,x}(\mathbf{R}) \le D_J R_x e^{-b_1 \operatorname{dist}(e,x)} \tag*{(33)} C e , x ( R ) ≤ D J R x e − b 1 dist ( e , x ) (33)
on every smaller construction profile in the statement; since the majorant has nonnegative coefficients, we may restrict to that profile before summing. A free input is included, since it is an input of (19.30) and of every later composition. By hypothesis the distances satisfy
d o u t ( x ) ≤ d o l d ( e ) + dist ( e , x ) + C J , d_{\mathrm{out}}(x) \le d_{\mathrm{old}}(e) + \operatorname{dist}(e,x) + C_J, d out ( x ) ≤ d old ( e ) + dist ( e , x ) + C J ,
where the additive constant is a fixed buffer. For an incoming list D i n D_{\mathrm{in}} D in whose complement lies in H H H up to the fixed stencil, with d o l d ( e ) = d D i n ( e ) / B s d_{\mathrm{old}}(e) = d_{D_{\mathrm{in}}}(e)/B_s d old ( e ) = d D in ( e ) / B s , this condition follows from Lemma 19.8 after the incoming distance and the buffer are divided by B s B_s B s ; we do not identify the incoming and the current fixed complements. Since B s ≥ 1 B_s \ge1 B s ≥ 1 , the condition also holds when d o l d d_{\mathrm{old}} d old is measured in input units. Hence
∑ e e − α b d d o l d ( e ) C e , x ( R ) ≤ D J R x e α b d C J e − α b d d o u t ( x ) ∑ e e − ( b 1 − α b d ) dist ( e , x ) ≤ C M , J D J R x e α b d C J e − α b d d o u t ( x ) ∑ l ≥ 0 ( l + 1 ) 3 e − ( b 1 − α b d ) l . \sum_e e^{-\alpha_{bd}d_{\mathrm{old}}(e)} C_{e,x}(\mathbf{R}) \le D_J R_x e^{\alpha_{bd}C_J} e^{-\alpha_{bd}d_{\mathrm{out}}(x)} \sum_e e^{-(b_1-\alpha_{bd})\operatorname{dist}(e,x)}
\le C_{M,J}D_J R_x e^{\alpha_{bd}C_J}e^{-\alpha_{bd}d_{\mathrm{out}}(x)} \sum_{l\ge0}(l+1)^3e^{-(b_1-\alpha_{bd})l}. e ∑ e − α b d d old ( e ) C e , x ( R ) ≤ D J R x e α b d C J e − α b d d out ( x ) e ∑ e − ( b 1 − α b d ) dist ( e , x ) ≤ C M , J D J R x e α b d C J e − α b d d out ( x ) l ≥ 0 ∑ ( l + 1 ) 3 e − ( b 1 − α b d ) l .
Using R x ≤ K ∗ ρ R_x \le K_*\rho R x ≤ K ∗ ρ , we obtain the bound on K J K_J K J ; keeping R x R_x R x gives the version for a free input. The sum includes every old output label at its density; restricting it to a sublist only decreases it. For a kernel from inputs to outputs it is multiplied by the volume of the fine output cells divided by the volume of the cell of the fixed input, as in Definition 4.6. The colors and the finitely many scale indices enter C M , J C_{M,J} C M , J , not a factor of the lattice volume. Setting R y = 2 ρ R_y = 2\rho R y = 2 ρ is allowed within the auxiliary ball since K ∗ ≥ 2 K_* \ge2 K ∗ ≥ 2 , and the smaller spatially varying radius profile is another restriction of the same nonnegative coefficient series. This completes the proof.
Construction radii and bounds on the reconstruction majorant
The next lemma collects the conditions on the radii under which the reconstruction maps a product domain strictly inside the analytic domain used at the next step, in the following sense. For a map with a convergent Taylor series, the absolute coefficient row sum of the Taylor series, evaluated at the radii of the product domain, bounds the image of the domain under the majorant of the map. For the reconstruction we call this bound the image under the reconstruction majorant. For c ≥ 0 c \ge0 c ≥ 0 we write S c S_c S c for this row sum with the weight e c ℓ e^{c\ell} e c ℓ , where ℓ \ell ℓ is the length of the connecting network of a monomial including its output; thus S 0 S_0 S 0 is the unweighted row sum. A bound S c < ρ / 2 S_c < \rho/2 S c < ρ /2 is stronger than a bound on the values of the map, and it is the hypothesis needed to substitute the reconstruction into a function whose coefficients are controlled at radius ρ \rho ρ . The lemma is applied twice, with the quantum radius ρ \rho ρ and with the classical radius ρ c l \rho_{\mathrm{cl}} ρ cl .
Lemma 19.7. Fix the constants of the unperturbed constructions before the later choices of the boundary-layer width, the erosion width, the widths of the protected neighborhoods, and the bound β r e c \beta_{\mathrm{rec}} β rec on newly frozen coordinates.
(Construction radius.) Let v f r e e , J > 0 v_{\mathrm{free},J} > 0 v free , J > 0 be a free construction radius, distinct from the density constant v J v_J v J , with
Q J v f r e e , J < λ J / 64 , I J Q J v f r e e , J < r z , J / 64. Q_J v_{\mathrm{free},J} < \lambda_J/64,\qquad I_J Q_J v_{\mathrm{free},J} < r_{z,J}/64. Q J v free , J < λ J /64 , I J Q J v free , J < r z , J /64. There is a radius R r a w > 0 R_{\mathrm{raw}} > 0 R raw > 0 , inside both the chart of the block averaging constraint and the chart of Theorem 11.6, which may be chosen below
min { min ( r Γ , J , r E ) 64 G J , λ J 64 Q J , r z , J 64 I J Q J , v f r e e , J 64 C l i f t , J , 16 min ( ρ J , β r e c , J ) } . \min\left\{\frac{\min(r_{\Gamma,J},r_E)}{64G_J},\frac{\lambda_J}{64Q_J},\frac{r_{z,J}}{64I_JQ_J},\frac{v_{\mathrm{free},J}}{64C_{\mathrm{lift},J}},16\min(\rho_J,\beta_{\mathrm{rec},J})\right\}. min { 64 G J min ( r Γ , J , r E ) , 64 Q J λ J , 64 I J Q J r z , J , 64 C lift , J v free , J , 16 min ( ρ J , β rec , J ) } . Here ρ J \rho_J ρ J and β r e c , J \beta_{\mathrm{rec},J} β rec , J are the coarse and newly frozen parameter radii of Theorem 18.12; thus the lifted arguments of size less than R r a w / 8 R_{\mathrm{raw}}/8 R raw /8 stay inside its parameter product. The quantities G J G_J G J and r Γ , J r_{\Gamma,J} r Γ , J are the bound and the radius of the finite constraint words, r E r_E r E is the larger radius of the torus action, and C l i f t , J C_{\mathrm{lift},J} C lift , J bounds the block-constant lift and the coarse-field lift, including all embeddings for the initial segments. If
C l i f t , J [ K ∗ max ( ρ , ρ c l ) + r + β r e c + q r a w ] < R r a w / 8 , C_{\mathrm{lift},J}\left[K_* \max(\rho,\rho_{\mathrm{cl}}) + r + \beta_{\mathrm{rec}} + q_{\mathrm{raw}}\right] < R_{\mathrm{raw}}/8, C lift , J [ K ∗ max ( ρ , ρ cl ) + r + β rec + q raw ] < R raw /8 , where r r r is the radius of the small-link domain B r B_r B r and q r a w q_{\mathrm{raw}} q raw is the radius allowed for the fluctuation increments before gauge transformation, then all images of the reconstruction before gauge transformation lie in these fixed larger charts, and their real free Hessians have gap at least λ J / 2 \lambda_J/2 λ J /2 , with the complex inverse margins. No factor M − J M^{-J} M − J is required of the lift.
(Zero fluctuations.) For each ρ ∙ ∈ { ρ , ρ c l } \rho_\bullet\in\{\rho,\rho_{\mathrm{cl}}\} ρ ∙ ∈ { ρ , ρ cl } , use its own spatially varying radius profile ϕ ∙ \phi_\bullet ϕ ∙ , equal to the boundary radius ρ ∂ , ∙ \rho_{\partial,\bullet} ρ ∂ , ∙ at labels within distance w b d w_{\mathrm{bd}} w bd of H H H and to ρ ∙ \rho_\bullet ρ ∙ elsewhere, the outer profile K ∗ ϕ ∙ K_*\phi_\bullet K ∗ ϕ ∙ , and the outer boundary radius R ∂ , ∙ = K ∗ ρ ∂ , ∙ R_{\partial,\bullet} = K_*\rho_{\partial,\bullet} R ∂ , ∙ = K ∗ ρ ∂ , ∙ . For a full renormalization group step, s = J s = J s = J and B = M J B = M^J B = M J , assume the torus contraction inequality (the first inequality below) and impose the four inequalities
B − 1 Z ∙ ( K ∗ ρ ∙ ) ≤ ρ ∙ / 16 , C Δ , ∙ C a , ∙ R ∂ , ∙ ≤ ρ ∙ / 16 , B^{-1}Z_\bullet(K_*\rho_\bullet) \le\rho_\bullet/16,\qquad C_{\Delta,\bullet}C_{a,\bullet}R_{\partial,\bullet} \le\rho_\bullet/16, B − 1 Z ∙ ( K ∗ ρ ∙ ) ≤ ρ ∙ /16 , C Δ , ∙ C a , ∙ R ∂ , ∙ ≤ ρ ∙ /16 , C Δ , ∙ C a , ∙ e − a w b d / 2 K ∗ ρ ∙ ≤ ρ ∙ / 16 , C F , ∙ β r e c ≤ ρ ∙ / 16. C_{\Delta,\bullet}C_{a,\bullet}e^{-a w_{\mathrm{bd}}/2}K_*\rho_\bullet\le\rho_\bullet/16,\qquad C_{F,\bullet}\beta_{\mathrm{rec}} \le\rho_\bullet/16. C Δ , ∙ C a , ∙ e − a w bd /2 K ∗ ρ ∙ ≤ ρ ∙ /16 , C F , ∙ β rec ≤ ρ ∙ /16.
where Z ∙ Z_{\bullet} Z ∙ is the classical row majorant of Theorem 13.2, C Δ , ∙ C_{\Delta,\bullet} C Δ , ∙ and C F , ∙ C_{F,\bullet} C F , ∙ are the constants of Theorem 19.6 , and C a , ∙ C_{a,\bullet} C a , ∙ is a summation constant fixed in the proof. These constants are evaluated on the larger auxiliary construction ball, before the radii and widths are chosen. The sum of the four left sides bounds the row sum S 0 S_0 S 0 at q = 0 q=0 q = 0 by ρ ∙ / 4 \rho_{\bullet}/4 ρ ∙ /4 .
(Fluctuation increments.) For a radius s q s_q s q of the fluctuation increments, the increment of S 0 S_0 S 0 is bounded by
W ∙ ( s q ) = C T , ∙ s q 1 − C T , ∙ ( K ∗ ρ ∙ + s q ) + C Δ , ∙ s q . W_{\bullet}(s_q)=\frac{C_{T,\bullet}s_q}{1-C_{T,\bullet}(K_*\rho_{\bullet}+s_q)}+C_{\Delta,\bullet}s_q. W ∙ ( s q ) = 1 − C T , ∙ ( K ∗ ρ ∙ + s q ) C T , ∙ s q + C Δ , ∙ s q .
Require the denominator to be at least 1 / 2 1/2 1/2 and W ∙ ( s q ) ≤ ρ ∙ / 8 W_{\bullet}(s_q)\leq\rho_{\bullet}/8 W ∙ ( s q ) ≤ ρ ∙ /8 for both radii. The same rule applies to the product in the jointly normalized coordinates u u u , with its finite-composition free radius and its own denominator, which includes β r e c \beta_{\mathrm{rec}} β rec if it occurs in that majorant. The products for the successive substitution must lie in this joint product.
(Spatial weight.) The resulting unweighted absolute row sum satisfies S 0 ≤ 3 ρ ∙ / 8 S_0\leq3\rho_{\bullet}/8 S 0 ≤ 3 ρ ∙ /8 . If S c ≤ C J ρ ∙ S_c\leq C_J\rho_{\bullet} S c ≤ C J ρ ∙ for a fixed c > 0 c>0 c > 0 , then there is a fixed 0 < ϑ < 1 0<\vartheta<1 0 < ϑ < 1 , common to both radii and to the finitely many initial segments, such that
S ϑ c ≤ S 0 1 − ϑ S c ϑ ≤ ρ ∙ ( 3 / 8 ) 1 − ϑ C J ϑ < ρ ∙ / 2. S_{\vartheta c}\leq S_0^{1-\vartheta}S_c^{\vartheta}\leq\rho_{\bullet}(3/8)^{1-\vartheta}C_J^{\vartheta}<\rho_{\bullet}/2. S ϑ c ≤ S 0 1 − ϑ S c ϑ ≤ ρ ∙ ( 3/8 ) 1 − ϑ C J ϑ < ρ ∙ /2.
Thus, for both radii, the image under the reconstruction majorant, with connecting-network weight at one positive exponent, is less than half of the radius.
All constants and thresholds have the dependence of § 2.1.1 and do not depend on the later widths. For a composition of s < J s<J s < J averaging steps, the conclusion of part (4) requires the torus contraction inequality with B s = M s B_s=M^s B s = M s in place of B B B , which must be checked separately for that composition; the finiteness of the constants at fixed J J J does not give it.
Proof. We choose the construction radius, bound the unweighted row sum at zero fluctuation and its increment, and then add the spatial weight by interpolation.
Step 1: the construction radius. We choose the finite word and lift constants on the larger products of Theorem 18.12 and Lemma 19.3 . We increase C l i f t , J C_{\mathrm{lift},J} C lift , J so that it bounds the block-constant lift of the target, the coarse-field lift, the extensions by zero and by the small newly frozen coordinates, and each of the finitely many embeddings for initial segments. Since a local lift uses only its block stencil of fixed depth, its row and input columns have one finite coefficient bound, independent of the number of components and of the later interface widths. Let v f r e e , J v_{\mathrm{free},J} v free , J satisfy the two strict inequalities in part (1), and let R r a w R_{\mathrm{raw}} R raw lie below the listed bounds and below the common radii of the local Haar-density chart and of the compact-overlap chart. The lift inequality then places all images strictly inside the larger inverse and torus action charts, their free arguments inside the radius v f r e e , J v_{\mathrm{free},J} v free , J , and their stationary points inside the solution radius r z , J r_{z,J} r z , J . Indeed, the variation of the free Hessian is bounded by the sum of its contributions from the free radius and from the lifted arguments, which is at most
Q J v f r e e , J + Q J R r a w < λ J / 32. Q_Jv_{\mathrm{free},J}+Q_JR_{\mathrm{raw}}<\lambda_J/32. Q J v free , J + Q J R raw < λ J /32.
Enlarging the fixed lift constant before the radii are chosen accounts for any finite number of such types of arguments. The displacement of the stationary point is bounded by the corresponding inverse times gradient, which is less than I J Q J ( v f r e e , J + R r a w ) < r z , J / 32 I_JQ_J(v_{\mathrm{free},J}+R_{\mathrm{raw}})<r_{z,J}/32 I J Q J ( v free , J + R raw ) < r z , J /32 . The real gap λ J \lambda_J λ J at the identity therefore implies a real gap at least λ J / 2 \lambda_J/2 λ J /2 on all these products, with room for the complex margins. The term E d 0 ′ Ψ z z E_{d_0}'\Psi_{zz} E d 0 ′ Ψ z z , which contains the gradient of the action, is included in the bound on the variation of the Hessian in Theorem 18.12. The conditional stationary points are those of Lemma 19.3 .
Step 2: the row sum at zero fluctuation. We carry out the following calculation for each ρ ∙ \rho_\bullet ρ ∙ , with its own profile and constants. The construction radii satisfy K ∗ ρ ∙ ≤ U T K_*\rho_\bullet\le U_T K ∗ ρ ∙ ≤ U T and s q ≤ Q T s_q\le Q_T s q ≤ Q T , so K ∗ ρ ∙ + s q < r s K_*\rho_\bullet+s_q<r_s K ∗ ρ ∙ + s q < r s and all intermediate margins hold by {(19.1)}; these domain checks precede each rational majorant. At q = 0 q=0 q = 0 we use Theorem 19.6 . By Theorem 13.2, the torus term has absolute row at most B − 1 Z ∙ ( K ∗ ρ ∙ ) B^{-1}Z_\bullet(K_*\rho_\bullet) B − 1 Z ∙ ( K ∗ ρ ∙ ) . We split the regional correction into its value at ξ = 0 \xi=0 ξ = 0 and its increment in ξ \xi ξ . By Theorem 19.6 (2), the latter is bounded by C F , ∙ ρ r e c C_{F,\bullet}\rho_{\mathrm{rec}} C F , ∙ ρ rec in the coefficient norm in the target. The former vanishes at zero target, and integrating its differentiated series with a fixed target input bounds it by
C Δ , ∙ ∑ x R x e − a W [ δ ( e ) + dist ( e , x ) ] , R x = ( K ∗ ϕ ∙ ) ( x ) , C_{\Delta,\bullet}\sum_x R_x e^{-a_W[\delta(e)+\operatorname{dist}(e,x)]},
\qquad R_x=(K_*\phi_\bullet)(x), C Δ , ∙ x ∑ R x e − a W [ δ ( e ) + dist ( e , x )] , R x = ( K ∗ ϕ ∙ ) ( x ) ,
where C Δ , ∙ C_{\Delta,\bullet} C Δ , ∙ is increased to include the fixed normalization of the derivative. Terms with δ ( x ) < w b d \delta(x)<w_{\mathrm{bd}} δ ( x ) < w bd have R x = R ∂ , ∙ R_x=R_{\partial,\bullet} R x = R ∂ , ∙ , and their sum is at most C a , ∙ R ∂ , ∙ C_{a,\bullet}R_{\partial,\bullet} C a , ∙ R ∂ , ∙ , where C a , ∙ C_{a,\bullet} C a , ∙ bounds the sum of the exponentials. For the other terms, by the conventions for endpoints, we have
δ ( e ) + dist ( e , x ) ≥ δ ( x ) − L J ′ ≥ w b d − L J ′ . \delta(e)+\operatorname{dist}(e,x)\ge\delta(x)-L'_J\ge w_{\mathrm{bd}}-L'_J. δ ( e ) + dist ( e , x ) ≥ δ ( x ) − L J ′ ≥ w bd − L J ′ .
We use half of the exponent for this inequality and sum the other half using the density bound. After e a W L J ′ / 2 e^{a_WL'_J/2} e a W L J ′ /2 is included in the fixed constant C a , ∙ C_{a,\bullet} C a , ∙ , these terms are at most C a , ∙ e − a W w b d / 2 K ∗ ρ ∙ C_{a,\bullet}e^{-a_Ww_{\mathrm{bd}}/2}K_*\rho_\bullet C a , ∙ e − a W w bd /2 K ∗ ρ ∙ . This bounds the row at q = 0 q=0 q = 0 by the sum of the four terms of part (2). Since each of them is at most ρ ∙ / 16 \rho_\bullet/16 ρ ∙ /16 , the row is at most ρ ∙ / 4 \rho_\bullet/4 ρ ∙ /4 . All constants in this calculation were evaluated on the larger unperturbed auxiliary ball, before w b d w_{\mathrm{bd}} w bd or the smaller boundary radius was chosen.
Step 3: terms of positive degree in the increments. For positive degree in the free coordinates, the torus row is bounded by
C T , ∙ s q 1 − C T , ∙ ( K ∗ ρ ∙ + s q ) . \frac{C_{T,\bullet}s_q}{1-C_{T,\bullet}(K_*\rho_\bullet+s_q)}. 1 − C T , ∙ ( K ∗ ρ ∙ + s q ) C T , ∙ s q .
which is the majorant of Theorem 13.2. The row of the correction of positive degree in q q q is at most C Δ , ∙ s q C_{\Delta,\bullet}s_q C Δ , ∙ s q , by Theorem 19.6 (3). Adding the two, we obtain W ∙ ( s q ) W_{\bullet}(s_q) W ∙ ( s q ) , and under the two conditions of part (3) the increment is at most ρ ∙ / 8 \rho_{\bullet}/8 ρ ∙ /8 . Combining this with Step 2, we obtain the absolute coefficient row bound
S 0 ≤ 3 ρ ∙ / 8. (34) S_0 \le3\rho_{\bullet}/8. \tag*{(34)} S 0 ≤ 3 ρ ∙ /8. (34)
The denominator cannot be discarded: its geometric series contains all higher degrees in the free coordinates and in the target.
Step 4: the jointly normalized coordinates. For the joint coordinate u u u , we substitute z = z 0 + K 0 − 1 / 2 u z=z_0+K_0^{-1/2}u z = z 0 + K 0 − 1/2 u into the maps in the original coordinates before taking the majorant. The successive increments are recovered from z z z by
q i = K i ( c , z > i ; ξ ) 1 / 2 ( z i − a i ( c , z > i ; ξ ) ) , q_i=K_i(c,z_{>i};\xi)^{1/2}\left(z_i-a_i(c,z_{>i};\xi)\right), q i = K i ( c , z > i ; ξ ) 1/2 ( z i − a i ( c , z > i ; ξ ) ) ,
which inverts the unperturbed insertion at scale i i i . The factor K i 1 / 2 = K i K i − 1 / 2 K_i^{1/2}=K_iK_i^{-1/2} K i 1/2 = K i K i − 1/2 has the same larger coefficient bounds. At u = 0 u=0 u = 0 all these increments vanish, since the chain at zero fluctuation is the joint stationary point. By finite composition of majorants, there exists a joint radius for which the resulting majorant of the increments is at most Q T Q_T Q T . We take the product of u u u inside this radius and inside the solution domains of the initial segments. Its target argument is at most U T U_T U T , so (1) also verifies the torus domain and the denominator for this composition. The finite composition has a rational majorant of the same kind for its increment, with its own constants and argument; in particular, a contribution of the newly frozen coordinates to that argument remains in its denominator. Requiring this majorant to be at most ρ ∙ / 8 \rho_{\bullet}/8 ρ ∙ /8 gives (19.34) for the joint product as well. The products for the successive substitution are restricted to the preimage of this product and to the domains of Lemma 19.3 . This restricts a coefficient series which is already bounded.
Step 5: the spatial weight. For one output row, we write its terms as b ν ≥ 0 b_\nu\ge0 b ν ≥ 0 and let ℓ ν \ell_\nu ℓ ν be the length of the attached network, including the output root. For finite partial sums, by Hölder’s inequality with exponents 1 / ( 1 − ϑ ) 1/(1-\vartheta) 1/ ( 1 − ϑ ) and 1 / ϑ 1/\vartheta 1/ ϑ , we have
∑ ν b ν e ϑ c ℓ ν = ∑ ν b ν 1 − ϑ ( b ν e c ℓ ν ) ϑ ≤ ( ∑ ν b ν ) 1 − ϑ ( ∑ ν b ν e c ℓ ν ) ϑ , \sum_\nu b_\nu e^{\vartheta c\ell_\nu}
=\sum_\nu b_\nu^{1-\vartheta}\left(b_\nu e^{c\ell_\nu}\right)^\vartheta
\le\left(\sum_\nu b_\nu\right)^{1-\vartheta}\left(\sum_\nu b_\nu e^{c\ell_\nu}\right)^\vartheta, ν ∑ b ν e ϑ c ℓ ν = ν ∑ b ν 1 − ϑ ( b ν e c ℓ ν ) ϑ ≤ ( ν ∑ b ν ) 1 − ϑ ( ν ∑ b ν e c ℓ ν ) ϑ ,
and passing to increasing partial sums gives the inequality for the convergent series. Using S c ≤ C J ρ ∙ S_c\le C_J\rho_{\bullet} S c ≤ C J ρ ∙ and (19.34), we obtain the bound in part (4). The function ϑ ↦ ( 3 / 8 ) 1 − ϑ C J ϑ \vartheta\mapsto(3/8)^{1-\vartheta}C_J^\vartheta ϑ ↦ ( 3/8 ) 1 − ϑ C J ϑ is continuous at ϑ = 0 \vartheta=0 ϑ = 0 , with value 3 / 8 < 1 / 2 3/8<1/2 3/8 < 1/2 . We fix ϑ ∈ ( 0 , 1 ) \vartheta\in(0,1) ϑ ∈ ( 0 , 1 ) at which its value is below 1 / 2 1/2 1/2 for both radii and for the finitely many initial segments. This proves both weighted bounds on the image under the reconstruction majorant at the same positive exponent. Because the weighted row bound is a bound of the unperturbed construction, this choice of exponent uses no later perturbative estimate.
Step 6: shorter compositions. For s < J s < J s < J , Step 1 and all estimates for finite compositions still apply, with B s = M s B_s = M^s B s = M s . The first inequality of part (2), however, becomes B s − 2 Z ∙ ( K ∗ ρ ∙ ) ≤ ρ ∙ / 16 B_s^{-2}Z_{\bullet}(K_{\ast}\rho_{\bullet}) \leq\rho_{\bullet}/16 B s − 2 Z ∙ ( K ∗ ρ ∙ ) ≤ ρ ∙ /16 , which follows only from a separate contraction inequality for that composition. This is the restriction for s < J s < J s < J in the statement.
Perturbations of the incoming action
The incoming action of a renormalization group step on a block region is the effective classical action plus a perturbation, defined on the list of variable labels of the previous step. This list and the current block region play different roles; in particular, the protected vertices of the gauge transformation alone do not determine the boundary of the incoming list. We first compare the distance to the boundary of the incoming list with the distance to the current boundary.
Lemma 19.8. Fix 1 ≤ s ≤ J 1 \leq s \leq J 1 ≤ s ≤ J , B s = M s B_s = M^s B s = M s , a list D i n D_{\mathrm{in}} D in of incoming positions, and a block region Ω \Omega Ω . Use the periodic graph metric, rescaled linearly on positions of coordinates introduced at finer lattices, and freeze all components of a label at one position together. Write π s x = x / B s \pi_s x = x/B_s π s x = x / B s . Suppose that every position x ∉ D i n x \notin D_{\mathrm{in}} x ∈ / D in lies in H H H up to the fixed chart stencil:
dist ( π s x , Ω c ) ≤ L b u f , J . \operatorname{dist}(\pi_s x,\Omega^c) \leq L_{\mathrm{buf},J}. dist ( π s x , Ω c ) ≤ L buf , J . Here L b u f , J L_{\mathrm{buf},J} L buf , J and the fixed buffer c e n d , J c_{\mathrm{end},J} c end , J for the convention on link endpoints depend only on the construction geometry, and are fixed before the later widths. Let d D i n ( i ) d_{D_{\mathrm{in}}}(i) d D in ( i ) be the distance of i i i to the complement of D i n D_{\mathrm{in}} D in , in input units, and d c u r r e n t ( e ) d_{\mathrm{current}}(e) d current ( e ) the distance of e e e to H H H , in output units. Then for every i ∈ D i n i \in D_{\mathrm{in}} i ∈ D in and every position e e e of a label in Ω \Omega Ω ,
d D i n ( i ) ≥ B s d c u r r e n t ( e ) − B s dist ( e , π s i ) − C b u f , J , C b u f , J = max 1 ≤ s ≤ J B s ( L b u f , J + c e n d , J ) . (35) d_{D_{\mathrm{in}}}(i) \geq B_s d_{\mathrm{current}}(e) - B_s \operatorname{dist}(e,\pi_s i) - C_{\mathrm{buf},J}, \qquad C_{\mathrm{buf},J} = \max_{1 \leq s \leq J} B_s(L_{\mathrm{buf},J}+c_{\mathrm{end},J}). \tag*{(35)} d D in ( i ) ≥ B s d current ( e ) − B s dist ( e , π s i ) − C buf , J , C buf , J = 1 ≤ s ≤ J max B s ( L buf , J + c end , J ) . (35) The same conclusion holds with dist ( e , W ) \operatorname{dist}(e,W) dist ( e , W ) in place of d c u r r e n t ( e ) d_{\mathrm{current}}(e) d current ( e ) after the fixed endpoint buffer is enlarged. It also holds if H H H is enlarged by additional blocks.
Proof. Suppose that the complement of D i n D_{\mathrm{in}} D in is nonempty, and choose a position x ∉ D i n x \notin D_{\mathrm{in}} x ∈ / D in reached by a shortest path from i i i to this complement; its length is d D i n ( i ) d_{D_{\mathrm{in}}}(i) d D in ( i ) . The scaled path has length d D i n ( i ) / B s d_{D_{\mathrm{in}}}(i)/B_s d D in ( i ) / B s , including the position of x x x at its own lattice; scale indices and colors give no extra distance. By the triangle inequality and the hypothesis, we have
d c u r r e n t ( e ) ≤ dist ( e , π s x ) + L b u f , J + c e n d , J ≤ dist ( e , π s i ) + d D i n ( i ) / B s + L b u f , J + c e n d , J . d_{\mathrm{current}}(e) \leq\operatorname{dist}(e,\pi_s x) + L_{\mathrm{buf},J} + c_{\mathrm{end},J}
\leq\operatorname{dist}(e,\pi_s i) + d_{D_{\mathrm{in}}}(i)/B_s + L_{\mathrm{buf},J} + c_{\mathrm{end},J}. d current ( e ) ≤ dist ( e , π s x ) + L buf , J + c end , J ≤ dist ( e , π s i ) + d D in ( i ) / B s + L buf , J + c end , J . Multiplying by B s B_s B s , we obtain (19.35). The endpoint buffer covers the difference between the distance to the positions in H H H and the distance to the interface W W W . Enlarging H H H can
only decrease its distance from a fixed position in Ω \Omega Ω , so additional blocks of H H H preserve both the hypothesis and the bound. If D i n D_{\mathrm{in}} D in has no complement, the finiteness of the boundary seminorm forces the nonconstant boundary part of the incoming perturbation to vanish, by the convention of Definition 18.1, and there is no incoming boundary dependence to transport. If D i n D_{\mathrm{in}} D in is empty there is no incoming root.
The constant incoming radius in the next theorem is the radius produced by freezing, in the following form of Lemma 18.9. We start with a classical output profile ϕ \phi ϕ of interior radius r c r_c r c on a set D D D of variable labels, keep I ⊂ { i ∈ D : d D ( i ) ≥ s e r } I \subset\{i \in D : d_D(i) \ge s_{\mathrm{er}}\} I ⊂ { i ∈ D : d D ( i ) ≥ s er } with s e r ≥ max ( 2 , w b d ) s_{\mathrm{er}} \ge\max(2,w_{\mathrm{bd}}) s er ≥ max ( 2 , w bd ) , and freeze F = D ∖ I F = D \setminus I F = D ∖ I , which contains the zone of reduced radius. For an analytic function f f f and frozen values c ♯ c^\sharp c ♯ , we write T c ♯ f ( v ) = f ( v ⊕ c ♯ ) − f ( 0 ⊕ c ♯ ) T_{c^\sharp}f(v) = f(v \oplus c^\sharp) - f(0 \oplus c^\sharp) T c ♯ f ( v ) = f ( v ⊕ c ♯ ) − f ( 0 ⊕ c ♯ ) . For ∣ c j ♯ ∣ ≤ η ϕ ( j ) |c^\sharp_j| \le\eta\phi(j) ∣ c j ♯ ∣ ≤ η ϕ ( j ) with η ≤ 1 / 4 \eta\le1/4 η ≤ 1/4 , Lemma 18.9 implies
ϕ ∣ I ≡ r c , X I r c ( T c ♯ f ) ≤ 2 e − α c l ( s e r − 1 ) X D ϕ ( f ) . (36) \phi|_I \equiv r_c,\qquad\mathcal{X}^{r_c}_I(T_{c^\sharp}f) \le2e^{-\alpha_{\mathrm{cl}}(s_{\mathrm{er}}-1)}\mathcal{X}^{\phi}_D(f). \tag*{(36)} ϕ ∣ I ≡ r c , X I r c ( T c ♯ f ) ≤ 2 e − α cl ( s er − 1 ) X D ϕ ( f ) . (36)
The seminorms here have the classical normalization and the stated network and boundary rates. The identity ϕ ∣ I ≡ r c \phi|_I \equiv r_c ϕ ∣ I ≡ r c holds because every remaining position lies outside the zone of reduced radius, and it is needed for the seminorm with constant profile on the left. The frozen values become fixed real parameters, not new variables of radius r c r_c r c . Applying (19.36) to the boundary correction produced at the previous step, with D i n = I D_{\mathrm{in}} = I D in = I , gives the incoming seminorm at the next step. A new classical boundary-completion difference of the torus action must in addition satisfy its own bound in the seminorm with constant profile, by the separate statement on such differences in Lemma 18.9; erosion alone does not bound such a difference at an additional component of H H H . The theorem below assumes the bound on the incoming seminorm. That the regions and frozen values of a large-field decomposition satisfy its hypotheses is verified in the construction of that decomposition in Section 21.
Theorem 19.9. Fix the domains, the protected geometry, the radius R r a w R_{\mathrm{raw}} R raw , the conditions on the image under the reconstruction majorant for both radii, and the products for the initial segments of Lemmas 19.3, 19.5, and 19.7, with s = J s = J s = J ; compositions of s < J s < J s < J averaging steps are treated at the end of the statement. Fix K ∗ K_* K ∗ , classical input rates α c l , α c l > 0 \alpha_{\mathrm{cl}}, \alpha_{\mathrm{cl}} > 0 α cl , α cl > 0 , and r c = ρ c l r_c = \rho_{\mathrm{cl}} r c = ρ cl . Let b = ( ξ , b r e c ) b = (\xi,b_{\mathrm{rec}}) b = ( ξ , b rec ) , where the newly frozen coordinates ξ \xi ξ are small and the previously frozen coordinates b r e c b_{\mathrm{rec}} b rec are arbitrary in their allowed real set.
The incoming action. Let E i n ( a ; b ) = E d 0 ( ι 0 a ) + P ( a ; b ) E_{\mathrm{in}}(a;b) = E_{d_0}(\iota_0a) + P(a;b) E in ( a ; b ) = E d 0 ( ι 0 a ) + P ( a ; b ) be holomorphic and real at real arguments in its variable input a a a , with b r e c b_{\mathrm{rec}} b rec fixed, where ι 0 \iota_0 ι 0 extends a a a by zero. Fix its incoming list D i n D_{\mathrm{in}} D in of variable labels, the positions of these labels at the lattices where they were introduced, and their cell volumes. The incoming seminorm is the seminorm X D i n , α c l , α c l r c \mathcal{X}^{r_c}_{D_{\mathrm{in}},\alpha_{\mathrm{cl}},\alpha_{\mathrm{cl}}} X D in , α cl , α cl r c of Definition 18.1, with the constant profile ϕ i n ≡ r c \phi_{\mathrm{in}} \equiv r_c ϕ in ≡ r c on every label of D i n D_{\mathrm{in}} D in at which Φ 0 \Phi_0 Φ 0 below is substituted; the labels of ξ \xi ξ are fixed parameters of this seminorm. The spatially varying radius profiles are used only for the output. Assume that the complement of D i n D_{\mathrm{in}} D in lies in H H H up to the fixed stencil, as in Lemma 19.8
(H H H may contain additional blocks), and that, in input units, for every position e e e of a label in Ω \Omega Ω and every incoming position i i i ,
d D i n ( i ) ≥ B s d c u r r e n t ( e ) − B s dist ( e , i / B s ) − C b u f , J , B s = M s , d_{D_{\mathrm{in}}}(i) \ge B_s d_{\mathrm{current}}(e) - B_s \operatorname{dist}(e,i/B_s) - C_{\mathrm{buf},J}, \qquad B_s = M^s, d D in ( i ) ≥ B s d current ( e ) − B s dist ( e , i / B s ) − C buf , J , B s = M s ,
where the positions on the right are in output units and the fixed buffer includes the chart stencils. If D i n D_{\mathrm{in}} D in has no complement, the nonconstant boundary part of the incoming perturbation is zero. All images of Φ 0 \Phi_0 Φ 0 and all connecting segments lie strictly inside the ball of the incoming seminorm. A difference of two perturbations uses the same incoming list, weights, and profile.
The exterior extension. If the unperturbed regional objective uses the extension ι ζ \iota_\zeta ι ζ , which inserts the newly frozen coordinates into their slots while all previously frozen torus slots remain zero, define the classical boundary-completion difference and the adjusted perturbation
S ζ c l ( a ) = E d 0 ( ι ζ a ) − E d 0 ( ι 0 a ) , P ‾ = P − S ζ c l , S^{\mathrm{cl}}_\zeta(a) = E_{d_0}(\iota_\zeta a) - E_{d_0}(\iota_0a), \qquad\overline{P}=P-S^{\mathrm{cl}}_\zeta, S ζ cl ( a ) = E d 0 ( ι ζ a ) − E d 0 ( ι 0 a ) , P = P − S ζ cl ,
and assume V D i n , α c l , α c l r c ( P ‾ ) ≤ 2 σ ∗ \mathcal{V}^{\mathrm{rc}}_{D_{\mathrm{in}},\alpha_{\mathrm{cl}},\alpha_{\mathrm{cl}}}(\overline{P}) \le2\sigma_* V D in , α cl , α cl rc ( P ) ≤ 2 σ ∗ . If the reference objective already uses the extension by zero, take S ζ c l = 0 S^{\mathrm{cl}}_\zeta=0 S ζ cl = 0 . All constants of S ζ c l S^{\mathrm{cl}}_\zeta S ζ cl remain in the full classical action. The corresponding difference for the effective interaction, if the reference extension is changed there, is kept with the exact sign of the constant term in Lemma 18.9.
The objective. Let z 0 z_0 z 0 be the unperturbed joint stationary point, K 0 = F 0 , z z ( c , z 0 ; ζ ) K_0=F_{0,zz}(c,z_0;\zeta) K 0 = F 0 , z z ( c , z 0 ; ζ ) , and S 0 = K 0 − 1 / 2 S_0=K_0^{-1/2} S 0 = K 0 − 1/2 , and put
F 0 ( c , z ; ζ ) = E d 0 ( ι ζ Ψ R ( c , z ; ζ ) ) , F_0(c,z;\zeta)=E_{d_0}(\iota_\zeta\Psi_R(c,z;\zeta)), F 0 ( c , z ; ζ ) = E d 0 ( ι ζ Ψ R ( c , z ; ζ )) ,
A 0 ( c , u ; ζ ) = Ψ R ( c , z 0 + S 0 u ; ζ ) , A_0(c,u;\zeta)=\Psi_R(c,z_0+S_0u;\zeta), A 0 ( c , u ; ζ ) = Ψ R ( c , z 0 + S 0 u ; ζ ) ,
Φ 0 ( c , u ; ζ ) = Log [ ( exp A 0 ( c , u ; ζ ) ) G Ω ( c ) ] , \Phi_0(c,u;\zeta)=\operatorname{Log}\left[(\exp A_0(c,u;\zeta))^{G^\Omega(c)}\right], Φ 0 ( c , u ; ζ ) = Log [ ( exp A 0 ( c , u ; ζ ) ) G Ω ( c ) ] ,
F λ ( c , u ; b ) = F 0 ( c , z 0 + S 0 u ; ζ ) + λ P ‾ ( Φ 0 ( c , u ; ζ ) ; b ) , 0 ≤ λ ≤ 1 , \mathcal{F}_\lambda(c,u;b)=F_0(c,z_0+S_0u;\zeta)+\lambda\overline{P}(\Phi_0(c,u;\zeta);b), \qquad0\le\lambda\le1, F λ ( c , u ; b ) = F 0 ( c , z 0 + S 0 u ; ζ ) + λ P ( Φ 0 ( c , u ; ζ ) ; b ) , 0 ≤ λ ≤ 1 ,
and Q λ ( c , z ; b ) = F λ ( c , S 0 − 1 ( z − z 0 ) ; b ) \mathcal{Q}_\lambda(c,z;b)=\mathcal{F}_\lambda(c,S_0^{-1}(z-z_0);b) Q λ ( c , z ; b ) = F λ ( c , S 0 − 1 ( z − z 0 ) ; b ) . The perturbation is evaluated on the gauge-transformed reconstruction Φ 0 \Phi_0 Φ 0 , and the unperturbed Hessian of F λ \mathcal{F}_\lambda F λ in u u u at u = 0 u=0 u = 0 is I I I . Initial segments, the conditional stationary points a λ , I a_{\lambda,I} a λ , I , and the Hessians K λ , i K_{\lambda,i} K λ , i refer to Q λ \mathcal{Q}_\lambda Q λ , in the original free coordinates.
Then the following statements hold.
(Incoming derivatives.) If the inputs at which P ‾ \overline{P} P is evaluated have absolute coefficient row sums at most κ ρ c l \kappa\rho_{\mathrm{cl}} κ ρ cl , with 0 < κ < 1 0<\kappa<1 0 < κ < 1 , then for every j ≥ 1 j\ge1 j ≥ 1 ,
∥ D j P ‾ ∥ r o o t ≤ j ! ( 1 − κ ) j + 1 ( 2 σ ∗ ) ρ c l 2 − j , \left\|D^j\overline{P}\right\|_{\mathrm{root}}
\le\frac{j!}{(1-\kappa)^{j+1}}(2\sigma_*)\rho_{\mathrm{cl}}^{2-j}, D j P root ≤ ( 1 − κ ) j + 1 j ! ( 2 σ ∗ ) ρ cl 2 − j ,
with the incoming network and boundary weights, and with any differentiated input as root.
(Stationary points.) There exist σ 0 , J , Q 0 , J > 0 \sigma_{0,J},Q_{0,J}>0 σ 0 , J , Q 0 , J > 0 and a fixed smaller domain of u u u such that the following holds for every 0 ≤ σ ∗ ≤ σ 0 , J 0\le\sigma_*\le\sigma_{0,J} 0 ≤ σ ∗ ≤ σ 0 , J , every 0 ≤ λ ≤ 1 0\le\lambda\le1 0 ≤ λ ≤ 1 , and every product of increments
with max i ∥ q i ∥ ∞ ≤ Q 0 , J \max_i \lVert q_i\rVert_\infty\le Q_{0,J} max i ∥ q i ∥ ∞ ≤ Q 0 , J , whose substitutions stay in that domain. Every initial segment of the original free coordinates has a unique holomorphic stationary point of Q λ Q_\lambda Q λ in the domains of Lemma 19.3, and the inequalities of Lemma 19.3(2) on the preconditioned gradient and Hessian hold, not only a spectral bound. The stationary points are real at real data, the real free Hessians satisfy
1 2 λ J I ≤ K λ , i ≤ Λ J I , ∥ u λ ∥ c o e f f ≤ C J σ ∗ ρ c l , \frac{1}{2}\lambda_J I \le K_{\lambda,i} \le\Lambda_J I,\qquad\lVert u_\lambda\rVert_{\mathrm{coeff}} \le C_J\sigma_*\rho_{\mathrm{cl}}, 2 1 λ J I ≤ K λ , i ≤ Λ J I , ∥ u λ ∥ coeff ≤ C J σ ∗ ρ cl ,
with the exact factors M i M^i M i included in K λ , i K_{\lambda,i} K λ , i , and for every initial segment I I I and every y ∈ T I y \in T_I y ∈ T I ,
∥ a λ , I ( c , y ; b ) − a I ( c , y ; ζ ) ∥ c o e f f ≤ C J σ ∗ ρ c l . \lVert a_{\lambda,I}(c,y;b)-a_I(c,y;\zeta)\rVert_{\mathrm{coeff}} \le C_J\sigma_*\rho_{\mathrm{cl}}. ∥ a λ , I ( c , y ; b ) − a I ( c , y ; ζ ) ∥ coeff ≤ C J σ ∗ ρ cl .
The quantity δ V \delta_V δ V of Lemma 19.3 satisfies δ V ≤ I 0 C P σ ∗ ρ c l \delta_V \le I_0C_P\sigma_*\rho_{\mathrm{cl}} δ V ≤ I 0 C P σ ∗ ρ cl , with constants I 0 I_0 I 0 , C P C_P C P depending only on the fixed data, and σ 0 , J \sigma_{0,J} σ 0 , J , Q 0 , J Q_{0,J} Q 0 , J are chosen so that the condition of Lemma 19.3(3) holds for all initial segments. The weighted complex resolvents are bounded by C J / ( λ J + t ) C_J/(\lambda_J+t) C J / ( λ J + t ) for t ≥ 0 t \ge0 t ≥ 0 at the working rate b w b_w b w fixed in part (4).
(Differences.) The differences from λ = 0 \lambda=0 λ = 0 of the stationary points, the Hessians, their inverses and inverse square roots, and the successive reconstructions have convergent boundary-localized coefficient bounds in all variables, including every coarse-field input in Ω \Omega Ω and every fluctuation increment; their networks carry the incoming boundary dependence. These differences are O J ( σ ∗ ) O_J(\sigma_*) O J ( σ ∗ ) , and the displacements of stationary points are O J ( σ ∗ ρ c l ) O_J(\sigma_*\rho_{\mathrm{cl}}) O J ( σ ∗ ρ cl ) on the fixed product. For example, at fixed coarser coordinates y y y ,
∂ λ z I , λ = − Q λ , I I − 1 V I , \partial_\lambda z_{I,\lambda}=-Q_{\lambda,II}^{-1}V_I, ∂ λ z I , λ = − Q λ , I I − 1 V I ,
where V I V_I V I is the I I I -component of the gradient of z ↦ P ‾ ( Φ 0 ( c , S 0 − 1 ( z − z 0 ) ; ξ ) ; b ) z\mapsto\overline{P}(\Phi_0(c,S_0^{-1}(z-z_0);\xi);b) z ↦ P ( Φ 0 ( c , S 0 − 1 ( z − z 0 ) ; ξ ) ; b ) . Along the successive chain this derivative contains the additional term − Q λ , I I − 1 Q λ , I I c ∂ λ z I c -Q_{\lambda,II}^{-1}Q_{\lambda,II^c}\partial_\lambda z_{I^c} − Q λ , I I − 1 Q λ , I I c ∂ λ z I c and the full derivative ( ∂ λ K λ , i − 1 / 2 ) q i (\partial_\lambda K_{\lambda,i}^{-1/2})q_i ( ∂ λ K λ , i − 1/2 ) q i . The function V = P ‾ ∘ Φ 0 V=\overline{P}\circ\Phi_0 V = P ∘ Φ 0 satisfies ∥ V u ∥ ≤ C J σ ∗ ρ c l \lVert V_u\rVert\le C_J\sigma_*\rho_{\mathrm{cl}} ∥ V u ∥ ≤ C J σ ∗ ρ cl and ∥ V u u ∥ ≤ C J σ ∗ \lVert V_{uu}\rVert\le C_J\sigma_* ∥ V uu ∥ ≤ C J σ ∗ ; the term P ‾ Φ 0 , u u \overline{P}\Phi_{0,uu} P Φ 0 , uu is kept in the second derivative.
(Output rates.) The estimates of part (3) hold at any fixed positive output rates a o u t , α o u t a_{\mathrm{out}},\alpha_{\mathrm{out}} a out , α out with
a o u t + α o u t < c J / C g e o m , a o u t + α o u t < B a c l / C g e o m , α o u t < B α c l / C g e o m , a_{\mathrm{out}}+\alpha_{\mathrm{out}}<c_J/C_{\mathrm{geom}},\qquad a_{\mathrm{out}}+\alpha_{\mathrm{out}}<B a_{\mathrm{cl}}/C_{\mathrm{geom}},\qquad\alpha_{\mathrm{out}}<B\alpha_{\mathrm{cl}}/C_{\mathrm{geom}}, a out + α out < c J / C geom , a out + α out < B a cl / C geom , α out < B α cl / C geom ,
Before the output rates or the perturbation threshold are chosen, reduce c J c_J c J so that c J ≤ b r e f c_J\le b_{\mathrm{ref}} c J ≤ b ref and fix
b w = min ( b r e f , c J , B a c l ) 2 C d i m , C d i m ≥ 1 , C g e o m ≥ 8 C d i m . b_w=\frac{\min(b_{\mathrm{ref}},c_J,Ba_{\mathrm{cl}})}{2C_{\mathrm{dim}}},\qquad C_{\mathrm{dim}}\ge1,\qquad C_{\mathrm{geom}}\ge8C_{\mathrm{dim}}. b w = 2 C dim min ( b ref , c J , B a cl ) , C dim ≥ 1 , C geom ≥ 8 C dim .
Here b r e f b_{\mathrm{ref}} b ref is the unperturbed resolvent exponent fixed above. Then every allowed pair of output rates satisfies a o u t + α o u t < b w / 4 a_{\mathrm{out}}+\alpha_{\mathrm{out}}<b_w/4 a out + α out < b w /4 , whereas C d i m b w ≤ min ( c J , B a c l ) / 2 C_{\mathrm{dim}}b_w\le\min(c_J,Ba_{\mathrm{cl}})/2 C dim b w ≤ min ( c J , B a cl ) /2 . The perturbation threshold is fixed at this one working rate and need not depend on the subsequently selected output rates. The constant C g e o m C_{\mathrm{geom}} C geom is fixed before J J J , the rates, and the later widths: it depends only on the dimension, the block geometry, and the allocation of exponents, not on J J J or on the construction radii. Fixed stencil buffers depending on J J J affect the amplitudes C J C_J C J , not the coefficients of the distances. In particular, the perturbation and every external argument stay in one connected coefficient network.
(Two perturbations.) For two perturbations P ‾ 0 , P ‾ 1 \overline{P}_0,\overline{P}_1 P 0 , P 1 as above, the bounds of parts (2) and (3) on the differences hold with σ ∗ \sigma_* σ ∗ replaced by X D i n , α c l , α c l r c ( P ‾ 1 − P ‾ 0 ) \mathcal{X}^{r_c}_{D_{\mathrm{in},\alpha_{\mathrm{cl}},\alpha_{\mathrm{cl}}}}(\overline{P}_1-\overline{P}_0) X D in , α cl , α cl r c ( P 1 − P 0 ) . The corresponding minimized full actions V 0 o u t , V 1 o u t V^{\mathrm{out}}_0,V^{\mathrm{out}}_1 V 0 out , V 1 out satisfy a Lipschitz bound in the output seminorm with constant K J c l K^{\mathrm{cl}}_J K J cl , namely (19.51).
(The full action.) The outgoing full action is M S F 1 ( c , u 1 ; b ) M^{S}\mathcal{F}_1(c,u_1;b) M S F 1 ( c , u 1 ; b ) , and on the connected chart
F 1 ( c , u ; b ) = E i n ( Φ 0 ( c , u ; ξ ) ; b ) . \mathcal{F}_1(c,u;b)=E_{\mathrm{in}}(\Phi_0(c,u;\xi);b). F 1 ( c , u ; b ) = E in ( Φ 0 ( c , u ; ξ ) ; b ) .
This identity uses the equality of classical actions on compact gauge overlaps only for the torus term, together with the cancellation of the exact classical boundary-completion difference. It does not assert that P ‾ ( Φ 0 ) \overline{P}(\Phi_0) P ( Φ 0 ) equals P ‾ ( A 0 ) \overline{P}(A_0) P ( A 0 ) , and it uses no gauge identity for an incoming branch.
(Frozen coordinates.) All perturbed estimates differentiate the coarse-field and free variables at fixed real b = ( ξ , b r e c ) b=(\xi,b_{\mathrm{rec}}) b = ( ξ , b rec ) ; no derivative of a perturbed quantity in ξ \xi ξ or b r e c b_{\mathrm{rec}} b rec is asserted, and the unperturbed estimates in ξ \xi ξ of Theorem 19.4 are unchanged. If the incoming coefficients and the charts are Borel (respectively continuous) functions of the real parameter b b b , then the stationary points, the matrix functions, and the minimized values are Borel (respectively continuous) in b b b .
All constants have the dependence of Section 19.1; they do not depend on the later widths, the number of preceding steps, the shape of Ω \Omega Ω , the depth, or the smaller enclosing and fluctuation radii chosen later. For a composition of s < J s<J s < J averaging steps, the same statements hold, with B s B_s B s in place of B B B , whenever the separate conditions of Lemma 19.7 for that composition hold; no repeated reduction of radii is used. The theorem is a local analytic statement and does not compare probability densities.
Proof. We extract the derivatives of the incoming perturbation from the incoming seminorm and transport its boundary dependence to the output. We then construct the stationary points with Lemma 19.3, first jointly and then for the initial segments. The remaining steps treat the differences, two perturbations, the full action, and the dependence on the frozen coordinates.
Step 1: conventions. All functions in this proof are considered at a fixed real b = ( ξ , b r e c ) b=(\xi,b_{\mathrm{rec}}) b = ( ξ , b rec ) in the allowed set, including the newly frozen coordinates ξ \xi ξ . Every estimate is uniform over this set, and the supremum over it is taken after the sums over degrees. Only the coarse field in Ω \Omega Ω , the free coordinates, and the interpolation parameter are differentiated, and the labels of ζ \zeta ζ are not roots of the incoming seminorm.
Step 2: derivatives of the incoming perturbation. We write r c = ρ c l r_c=\rho_{\mathrm{cl}} r c = ρ cl and P ‾ ( a ; b ) = ∑ ν p ν ( b ) a ν \overline{P}(a;b)=\sum_{\nu}p_{\nu}(b)a^{\nu} P ( a ; b ) = ∑ ν p ν ( b ) a ν , where every multi-index runs over D i n D_{\mathrm{in}} D in , with the positions of its labels at the lattices where they were introduced. Since φ i n ( i ) = r c \varphi_{\mathrm{in}}(i)=r_c φ in ( i ) = r c for every such label, we define, for each degree k k k ,
A k ( b ) = r c − 2 sup i ∈ D i n e α c l d D i n ( i ) ∑ ∣ ν ∣ = k ν i k ∣ p ν ( b ) ∣ r c k e α c l t i n ( ν ) . A_k(b)=r_c^{-2}\sup_{i\in D_{\mathrm{in}}}e^{\alpha_{\mathrm{cl}}d_{D_{\mathrm{in}}}(i)}\sum_{|\nu|=k}\frac{\nu_i}{k}|p_{\nu}(b)|r_c^k e^{\alpha_{\mathrm{cl}}t_{\mathrm{in}}(\nu)}. A k ( b ) = r c − 2 i ∈ D in sup e α cl d D in ( i ) ∣ ν ∣ = k ∑ k ν i ∣ p ν ( b ) ∣ r c k e α cl t in ( ν ) .
By hypothesis, sup b real ∑ k ≥ 1 A k ( b ) ≤ 2 σ ∗ \sup_b\operatorname{real}\sum_{k\geq1}A_k(b)\leq2\sigma_* sup b real ∑ k ≥ 1 A k ( b ) ≤ 2 σ ∗ . This uses the full input radius also at the incoming boundary; a reduced radius near the boundary appears only in the output product. By (19.36), the incoming action produced by freezing has this profile when the hypotheses of Lemma 18.9 on freezing the zone of reduced radius and on the frozen values hold. We consider inputs whose absolute coefficient row sums are at most κ r c \kappa r_c κ r c , and fix the first differentiated input label i i i . In degree k ≥ j k\geq j k ≥ j , summing over the other j − 1 j-1 j − 1 ordered differentiated labels gives the factor
ν i k ( k ) j , ( k ) j = k ( k − 1 ) ⋯ ( k − j + 1 ) , \frac{\nu_i}{k}(k)_j,\qquad(k)_j=k(k-1)\cdots(k-j+1), k ν i ( k ) j , ( k ) j = k ( k − 1 ) ⋯ ( k − j + 1 ) ,
and the remaining input powers are bounded by κ k − j r c k − j \kappa^{k-j}r_c^{k-j} κ k − j r c k − j . Keeping the original incoming network keeps all differentiated roots and all contracted input labels in it. Therefore the rooted derivative of order j j j is at most
r c 2 − j ∑ k ≥ j A k ( k ) j κ k − j ≤ ( 2 σ ∗ ) r c 2 − j j ! ( 1 − κ ) j + 1 . r_c^{2-j}\sum_{k\geq j}A_k(k)_j\kappa^{k-j}\leq(2\sigma_*)r_c^{2-j}\frac{j!}{(1-\kappa)^{j+1}}. r c 2 − j k ≥ j ∑ A k ( k ) j κ k − j ≤ ( 2 σ ∗ ) r c 2 − j ( 1 − κ ) j + 1 j ! .
where we used ∑ k ≥ j ( k ) j κ k − j = j ! / ( 1 − κ ) j + 1 \sum_{k\geq j}(k)_j\kappa^{k-j}=j!/(1-\kappa)^{j+1} ∑ k ≥ j ( k ) j κ k − j = j ! / ( 1 − κ ) j + 1 , which follows by differentiating the geometric series j j j times. Any differentiated root can be chosen first, so this proves part (1), with the incoming network and the boundary weight of its root. Degree one is included. The constant coefficient p 0 p_0 p 0 does not appear in these derivatives and is not restricted.
Step 3: the normalized objective and the perturbing gradient. Let A 0 A_0 A 0 , Φ 0 \Phi_0 Φ 0 , and F λ F_\lambda F λ be as in the statement, on the joint product of Lemma 19.7. The image of this product under the reconstruction majorant, with weight at one fixed positive exponent of the connecting network, is less than r c / 2 r_c/2 r c /2 , so part (1) applies with a fixed κ < 1 \kappa<1 κ < 1 , for example κ = 1 / 2 \kappa=1/2 κ = 1/2 . At u = 0 u=0 u = 0 we have
D u F 0 ( c , z 0 + S 0 u ; ζ ) ∣ u = 0 = 0 , D u 2 F 0 ( c , z 0 + S 0 u ; ζ ) ∣ u = 0 = S 0 T K 0 S 0 = I . (37) D_uF_0(c,z_0+S_0u;\zeta)\big|_{u=0}=0,\qquad D_u^2F_0(c,z_0+S_0u;\zeta)\big|_{u=0}=S_0^{\mathsf T}K_0S_0=I. \tag*{(37)} D u F 0 ( c , z 0 + S 0 u ; ζ ) u = 0 = 0 , D u 2 F 0 ( c , z 0 + S 0 u ; ζ ) u = 0 = S 0 T K 0 S 0 = I . (37)
The first identity is the joint stationarity, and the second uses the symmetric inverse square root on the branch connected to the identity. The coefficients of this function, and those of z 0 z_0 z 0 , Ψ R \Psi_R Ψ R , and G Ω G^\Omega G Ω , are bounded on a larger product by the unperturbed results above.
Let V = P ‾ ∘ Φ 0 V=\overline{P}\circ\Phi_0 V = P ∘ Φ 0 . By the chain rule, we have
V u = ∑ i P ‾ i ( Φ 0 ) ( Φ 0 ) i , u . (38) V_u=\sum_i \overline{P}_i(\Phi_0)(\Phi_0)_{i,u}. \tag*{(38)} V u = i ∑ P i ( Φ 0 ) ( Φ 0 ) i , u . (38)
V u v = ∑ i , j P ‾ i j ( Φ 0 ) ( Φ 0 ) i , u ( Φ 0 ) j , v + ∑ i P ‾ i ( Φ 0 ) ( Φ 0 ) i , u v . (39) V_{uv}=\sum_{i,j}\overline{P}_{ij}(\Phi_0)(\Phi_0)_{i,u}(\Phi_0)_{j,v}+\sum_i\overline{P}_i(\Phi_0)(\Phi_0)_{i,uv}. \tag*{(39)} V uv = i , j ∑ P ij ( Φ 0 ) ( Φ 0 ) i , u ( Φ 0 ) j , v + i ∑ P i ( Φ 0 ) ( Φ 0 ) i , uv . (39)
By part (1) with j = 1 , 2 j=1,2 j = 1 , 2 , together with the fixed rows and volume-normalized columns of the chart, we obtain
∥ V u ∥ ≤ C P σ ∗ r c , ∥ V u u ∥ ≤ C P σ ∗ . (40) \lVert V_u\rVert\le C_P\sigma_*r_c,\qquad\lVert V_{uu}\rVert\le C_P\sigma_*. \tag*{(40)} ∥ V u ∥ ≤ C P σ ∗ r c , ∥ V uu ∥ ≤ C P σ ∗ . (40)
The term P ‾ ′ Φ 0 , u u \overline{P}'\Phi_{0,uu} P ′ Φ 0 , uu in (39) contributes C J σ ∗ r c C_J\sigma_*r_c C J σ ∗ r c , which is included in C P σ ∗ C_P\sigma_* C P σ ∗ since r c r_c r c is fixed. We note that the perturbation need not have zero gradient at u = 0 u=0 u = 0 .
These bounds hold at all orders, as coefficient bounds with spatial weights. Let b J ( t ) = C J t / ( 1 − C J t ) b_J(t)=C_Jt/(1-C_Jt) b J ( t ) = C J t / ( 1 − C J t ) majorize the increment series of the chart in a fixed displacement direction, on a fixed larger domain. After division by the factorials of the derivatives, part (1) and the chain rule over partitions are bounded by the coefficients of
2 σ ∗ r c 2 { 1 1 − κ − b J ( t ) / r c − 1 1 − κ } . (41) 2\sigma_*r_c^2\left\{\frac{1}{1-\kappa-b_J(t)/r_c}-\frac{1}{1-\kappa}\right\}. \tag*{(41)} 2 σ ∗ r c 2 { 1 − κ − b J ( t ) / r c 1 − 1 − κ 1 } . (41)
Fix t ∗ > 0 t_*>0 t ∗ > 0 , a fixed multiple of min ( r c , 1 ) \min(r_c,1) min ( r c , 1 ) , with b J ( t ∗ ) ≤ ( 1 − κ ) r c / 2 b_J(t_*)\le(1-\kappa)r_c/2 b J ( t ∗ ) ≤ ( 1 − κ ) r c /2 . By nonnegativity of its coefficients, the coefficient of degree j j j of (41) is at most C J σ ∗ r c 2 t ∗ − j C_J\sigma_*r_c^2t_*^{-j} C J σ ∗ r c 2 t ∗ − j , and restoring j ! j! j ! gives the factorial bounds j ! C J j + 1 σ ∗ r c 2 − j j!C_J^{j+1}\sigma_*r_c^{2-j} j ! C J j + 1 σ ∗ r c 2 − j , with constants depending on the fixed radius. Fixing a coarse-field or free input uses the column version of the same composition, with the cell volume of that input. Thus all derivatives used here satisfy coefficient bounds on the product.
Step 4: transport of the incoming boundary dependence. A rooted input monomial carries its incoming network and a path from its chosen input root to the complement of D i n D_{\mathrm{in}} D in . The incoming list and its complement are those used in A k ( b ) A_k(b) A k ( b ) . By hypothesis the complement of D i n D_{\mathrm{in}} D in lies in H H H up to the fixed stencil, so Lemma 19.8 implies (35) with its fixed buffer, and hence the comparison assumed in the statement. We scale these paths by B − 1 B^{-1} B − 1 to output units and join every network of the chart at its fine input label. The union is connected and contains every external coarse-field and free argument. If the input root lies in Ω \Omega Ω , its path to the complement of D i n D_{\mathrm{in}} D in reaches H H H after at most one additional leg of fixed length. If it lies in a newly removed region, the chart branch joining it to a root in Ω \Omega Ω already reaches the boundary. Additional components of H H H can only shorten the required distance to the boundary.
To make this precise, let ℓ 0 \ell_0 ℓ 0 be the length of the incoming network in input units, d 0 i n d_0^{\mathrm{in}} d 0 in the distance of the chosen root to the complement of D i n D_{\mathrm{in}} D in , and ℓ β \ell_\beta ℓ β the lengths of the branches of the chart and of the free inverse in output units. Let A A A be the set of external positions and e e e any root in Ω \Omega Ω . There exists a constant C d i m C_{\mathrm{dim}} C dim , depending only on the dimension and on the comparison of elementary paths, and fixed before J J J , such that
a o u t t ( A ) + α o u t δ ( e ) ≤ C d i m ( a o u t + α o u t ) ( ℓ 0 / B + ∑ β ℓ β ) + C d i m α o u t d 0 i n / B + ( a o u t + α o u t ) L J ′ ( k + 1 ) . (42) a_{\mathrm{out}}^t(A)+\alpha_{\mathrm{out}}\delta(e)\le C_{\mathrm{dim}}(a_{\mathrm{out}}+\alpha_{\mathrm{out}})\left(\ell_0/B+\sum_\beta\ell_\beta\right)+C_{\mathrm{dim}}\alpha_{\mathrm{out}}d_0^{\mathrm{in}}/B+(a_{\mathrm{out}}+\alpha_{\mathrm{out}})L'_J(k+1). \tag*{(42)} a out t ( A ) + α out δ ( e ) ≤ C dim ( a out + α out ) ℓ 0 / B + β ∑ ℓ β + C dim α out d 0 in / B + ( a out + α out ) L J ′ ( k + 1 ) . (42)
Indeed, the scaled incoming network and the chart branches alone join all external positions, and adding the scaled path to the complement of D i n D_{\mathrm{in}} D in also reaches the boundary. The first network bounds t ( A ) t(A) t ( A ) , and the second bounds the distance of e e e to the boundary; this gives the separate coefficients of ℓ 0 \ell_0 ℓ 0 and d 0 i n d_0^{\mathrm{in}} d 0 in in (42) . Here k k k is the incoming degree, equivalently the number of attached input factors, including factors evaluated at a nonzero frozen value. The last term accounts for their fixed displacements by the stencil. To see why C d i m C_{\mathrm{dim}} C dim depends only on the dimension, we take shortest paths in the graph metric and their connected union. Using a tree on the external vertices increases the length by at most a factor two, by traversing each edge of a Steiner tree twice. Replacing segments by the fixed paths in coordinate order increases it only by the fixed comparison constant of the three coordinate directions. Scaling is exact, and neither operation depends on the number of scales.
We sum the stencil offsets using the strict image bound, not by shrinking the external Taylor variables. Reduce the intrinsic rate c J c_J c J once, before output rates are chosen, so that e c J L J ′ ≤ 4 / 3 e^{c_JL'_J}\le4/3 e c J L J ′ ≤ 4/3 . Put β = ( a o u t + α o u t ) L J ′ \beta=(a_{\mathrm{out}}+\alpha_{\mathrm{out}})L'_J β = ( a out + α out ) L J ′ and κ ′ = e β κ \kappa'=e^\beta\kappa κ ′ = e β κ . The reconstruction image has κ ≤ 1 / 2 \kappa\le1/2 κ ≤ 1/2 , so the allowed output rates give κ ′ ≤ 2 / 3 \kappa'\le2/3 κ ′ ≤ 2/3 . Multiplying the degree-k k k derivative term of Step 2 by e β ( k + 1 ) e^{\beta(k+1)} e β ( k + 1 ) gives
e β ( k + 1 ) κ k − j = e β ( j + 1 ) ( κ ′ ) k − j . e^{\beta(k+1)}\kappa^{k-j}=e^{\beta(j+1)}(\kappa')^{k-j}. e β ( k + 1 ) κ k − j = e β ( j + 1 ) ( κ ′ ) k − j .
Thus its degree sum is bounded by
( 2 σ ∗ ) r c 2 − j e β ( j + 1 ) j ! ( 1 − κ ′ ) j + 1 . (2\sigma_*)^r c^{2-j}\frac{e^{\beta(j+1)}j!}{(1-\kappa')^{j+1}}. ( 2 σ ∗ ) r c 2 − j ( 1 − κ ′ ) j + 1 e β ( j + 1 ) j ! .
In (41) the same calculation replaces κ \kappa κ by κ ′ \kappa' κ ′ and b J ( t ) b_J(t) b J ( t ) by e β b J ( t ) e^{\beta}b_J(t) e β b J ( t ) , with one additional fixed factor e β e^\beta e β . A smaller fixed t ∗ t_* t ∗ leaves a positive denominator. These changes alter only the fixed-J J J amplitudes. They cover degree-zero contributions from frozen coordinates as well, without assuming geometric decay of the incoming sequence A k A_k A k . The intrinsic-rate reduction precedes all later profiles and widths and does not change C g e o m C_{\mathrm{geom}} C geom .
We take C g e o m ≥ 8 C d i m C_{\mathrm{geom}}\ge8C_{\mathrm{dim}} C geom ≥ 8 C dim , increased according to the fixed allocation of exponents among the external network, the route to the boundary, and the internal sums, and we keep part of the remaining exponent in each chart branch for its density sum S J ( ε ) S_J(\varepsilon) S J ( ε ) . Under the three strict inequalities of part (4), the incoming factors e c 1 ℓ 0 e^{c_1\ell_0} e c 1 ℓ 0 and e α c ℓ d 0 i e^{\alpha c\ell d_0^i} e α c ℓ d 0 i , together with the exponent c J c_J c J of the unperturbed branches, dominate the output network and boundary weights in (42) , and a positive part of the exponent is left for the summations. The exponent in the bound on the image under the reconstruction majorant from Lemma 19.7 is included in the choice of c J c_J c J before these rates are specified. The allocation is applied to the whole network, so the exponent is not divided by the Taylor order or by the number of scales. This proves the boundary-localized versions of (40) and (41) at every allowed output rate.
Step 5: the joint stationary point. We now verify the domain conditions, working first in u u u near zero. By the unperturbed bound on the third derivative and (37) , we have
∥ D u 2 F 0 ( c , z 0 + S 0 u ; ξ ) − I ∥ ≤ C 1 ∥ u ∥ c o e f f . \left\lVert D_u^2F_0(c,z_0+S_0u;\xi)-I\right\rVert\le C_1\lVert u\rVert_{\mathrm{coeff}}. D u 2 F 0 ( c , z 0 + S 0 u ; ξ ) − I ≤ C 1 ∥ u ∥ coeff .
Fix u 0 > 0 u_0>0 u 0 > 0 inside the joint product with C 1 u 0 < 1 / 8 C_1u_0<1/8 C 1 u 0 < 1/8 , and fix σ 0 , J > 0 \sigma_{0,J}>0 σ 0 , J > 0 with C P σ 0 , J < 1 / 8 C_P\sigma_{0,J}<1/8 C P σ 0 , J < 1/8 and C P σ 0 , J r c < u 0 / 8 C_P\sigma_{0,J}r_c<u_0/8 C P σ 0 , J r c < u 0 /8 . Then u ↦ u − D u F λ u\mapsto u-D_uF_\lambda u ↦ u − D u F λ is a coefficient contraction with factor at most 1 / 4 1/4 1/4 on the domain of radius u 0 u_0 u 0 , and it maps the ball of radius u 0 / 2 u_0/2 u 0 /2 into itself. Picard iteration gives a unique holomorphic joint stationary point, and from the fixed point inequality we obtain
∥ u λ ∥ c o e f f ≤ 2 C P σ ∗ r c . (43) \lVert u_\lambda\rVert_{\mathrm{coeff}}\le2C_P\sigma_*r_c. \tag*{(43)} ∥ u λ ∥ coeff ≤ 2 C P σ ∗ r c . (43)
The stationary point is real at real data, since F λ F_\lambda F λ is real at real arguments and the stationary point is unique.
Step 6: initial segments in the original coordinates. To treat initial segments, we return to the original free coordinates by the invertible affine map u = S 0 − 1 ( z − z 0 ) u=S_0^{-1}(z-z_0) u = S 0 − 1 ( z − z 0 ) at fixed c c c and ξ \xi ξ . The pullback is
Q λ ( c , z ; b ) = F 0 ( c , z ; ξ ) + λ P ‾ ( Log [ ( exp Ψ R ( c , z ; ξ ) ) G Ω ( c ) ] ; b ) . (44) Q_\lambda(c,z;b)=F_0(c,z;\xi)+\lambda\overline{P}\left(\operatorname{Log}\left[(\exp\Psi_R(c,z;\xi))G^\Omega(c)\right];b\right). \tag*{(44)} Q λ ( c , z ; b ) = F 0 ( c , z ; ξ ) + λ P ( Log [ ( exp Ψ R ( c , z ; ξ )) G Ω ( c ) ] ; b ) . (44)
All subscripts referring to initial segments refer to this expression, with the original lists of scales. Because the affine map and its inverse have fixed coefficient bounds, (40) holds in the original coordinates with larger fixed constants. We take the solution domain of every initial segment from Lemma 19.3 small enough that its image under this affine map stays in the domain of radius u 0 u_0 u 0 .
At an unperturbed conditional stationary point a I ( c , y ; ξ ) a_I(c,y;\xi) a I ( c , y ; ξ ) , the preconditioned perturbed gradient is
A I − 1 Q λ , I ( c , a I , y ; b ) = λ A I − 1 V I ( c , a I , y ; b ) , A_I^{-1}Q_{\lambda,I}(c,a_I,y;b)=\lambda A_I^{-1}V_I(c,a_I,y;b), A I − 1 Q λ , I ( c , a I , y ; b ) = λ A I − 1 V I ( c , a I , y ; b ) ,
whose coefficient norm is at most I 0 C P σ ∗ r c I_0C_P\sigma_*r_c I 0 C P σ ∗ r c . This holds uniformly for every y ∈ T I y\in\mathcal{T}_I y ∈ T I , not only at the joint stationary point, since all these conditional reference points lie in the image of the same product under the reconstruction majorant. If δ V \delta_V δ V is defined for the pullback of V V V as in Lemma 19.3 , the same calculation yields
δ V ≤ I 0 C P σ ∗ r c . (45) \delta_V \le I_0 C_P \sigma_* r_c. \tag*{(45)} δ V ≤ I 0 C P σ ∗ r c . (45)
Across a conditional displacement of size r I r_I r I , the derivative bound is at most
∥ A I − 1 ( Q λ , I − A I ) ∥ ≤ I 0 ( C 1 r I + C P σ ∗ ) . (46) \left\| A_I^{-1}\left(\mathcal{Q}_{\lambda,I}-A_I\right)\right\| \le I_0\left(C_1r_I+C_P\sigma_*\right). \tag*{(46)} A I − 1 ( Q λ , I − A I ) ≤ I 0 ( C 1 r I + C P σ ∗ ) . (46)
We first take r I r_I r I with I 0 C 1 r I < 1 / 8 I_0C_1r_I<1/8 I 0 C 1 r I < 1/8 , and then reduce σ 0 , J \sigma_{0,J} σ 0 , J so that, for every initial segment,
I 0 C P σ 0 , J < 1 / 8 , I 0 C P σ 0 , J r c < r I / 8. I_0C_P\sigma_{0,J}<1/8,\qquad I_0C_P\sigma_{0,J}r_c<r_I/8. I 0 C P σ 0 , J < 1/8 , I 0 C P σ 0 , J r c < r I /8.
These inequalities are stronger than the two inequalities required in Lemma 19.3 (2). We also require that the error of the real Hessian is less than λ J / 8 \lambda_J/8 λ J /8 . The coefficient resolvent condition is imposed at the allocated working rates of Step 4, using the unperturbed resolvents as references, as detailed in Step 8 below. The larger construction products, their radii, and the unperturbed half of the resolvent margin are fixed first; after J J J and these spatial rates are fixed, all perturbative constants in this condition are fixed before σ 0 , J \sigma_{0,J} σ 0 , J is decreased and before the later widths and enclosing radii are chosen. Then Lemma 19.3 applies on each domain T I \mathcal{T}_I T I of coarser coordinates, and its estimate (19.13) implies that, for every such y y y ,
∥ a λ , I ( c , y ; b ) − a I ( c , y ; ξ ) ∥ c o e f f ≤ 2 λ I 0 C P σ ∗ r c ≤ C J σ ∗ r c . (47) \left\|a_{\lambda,I}(c,y;b)-a_I(c,y;\xi)\right\|_{\mathrm{coeff}}\le2\lambda I_0C_P\sigma_*r_c\le C_J\sigma_*r_c. \tag*{(47)} ∥ a λ , I ( c , y ; b ) − a I ( c , y ; ξ ) ∥ coeff ≤ 2 λ I 0 C P σ ∗ r c ≤ C J σ ∗ r c . (47)
This is the conditional bound of part (2), with one common C J ≥ 2 I 0 C P C_J\ge2I_0C_P C J ≥ 2 I 0 C P , and it holds before any moving coarser coordinates are substituted. The mixed coefficient bounds in the original coordinates follow from (19.41) and the affine inverse; together with the free inverse they verify (19.14) on the same products.
Step 7: successive substitution. The recurrence (19.15), with the conditional and matrix-function bounds just obtained, gives a constant C s e q , J C_{\mathrm{seq},J} C seq , J for a fixed rough bound on the size of the perturbation. Fix Q 0 , J > 0 Q_{0,J}>0 Q 0 , J > 0 and, if necessary, reduce σ 0 , J \sigma_{0,J} σ 0 , J further, so that
C s e q , J ( 2 I 0 C P σ 0 , J r c + Q 0 , J ) < 1 2 ε t u b e , Q 0 , J ≤ r c , C_{\mathrm{seq},J}\left(2I_0C_P\sigma_{0,J}r_c+Q_{0,J}\right)<\frac{1}{2}\varepsilon_{\mathrm{tube}},\qquad Q_{0,J}\le r_c, C seq , J ( 2 I 0 C P σ 0 , J r c + Q 0 , J ) < 2 1 ε tube , Q 0 , J ≤ r c ,
where the conversion between the original and the normalized coordinates is included in the minimal margin. The bound (19.45) verifies the condition of Lemma 19.3 (3) for every y y y . The recurrence bounds the substituted higher arguments and the constructed conditional centers. At each such argument, the Picard iterations stay in the fixed solution balls of Lemma 19.3 (2); these balls were chosen inside the larger reconstruction product. Thus the hypothesis on the image under the reconstruction majorant holds throughout the construction. This proves the statements on stationary points in part (2), for all initial segments simultaneously, on the product of increments.
Step 8: the Hessians. The unperturbed real Hessian in the original coordinates is at least 7 λ J / 8 7\lambda_J/8 7 λ J /8 on these products, since its variation from the Hessian at the identity is at most λ J / 8 \lambda_J/8 λ J /8 by Theorem 18.12, and the perturbation changes it by less than λ J / 8 \lambda_J/8 λ J /8 . Its principal blocks and Schur complements have the same lower bound, since for a symmetric matrix H ≥ λ I H \ge\lambda I H ≥ λ I
v T ( H y y − H y I H I I − 1 H I y ) v = min x ( x , v ) T H ( x , v ) ≥ λ ∣ v ∣ 2 , v^{\mathsf{T}}\left(H_{yy}-H_{yI}H_{II}^{-1}H_{Iy}\right)v=\min_x (x,v)^{\mathsf{T}}H(x,v)\ge\lambda\lvert v\rvert^2, v T ( H y y − H y I H I I − 1 H I y ) v = x min ( x , v ) T H ( x , v ) ≥ λ ∣ v ∣ 2 ,
as in (18.6). After the normalization by M i ≥ 1 M^i \ge1 M i ≥ 1 , we obtain K λ , i ≥ λ J / 2 K_{\lambda,i} \ge\lambda_J/2 K λ , i ≥ λ J /2 . The upper bounds on the tensors and on the inverses give a common upper bound Λ J \Lambda_J Λ J on the same product.
The complex resolvent condition is used at the working exponent b w b_w b w defined in part (4), with strict reserve for the output network and boundary weights. Its defining inequalities put it below b r e f b_{\mathrm{ref}} b ref and inside the transported incoming-rate allocation of Step 4, uniformly over every allowed output pair. For each flat reference K i b K_i^b K i b , including the factor M i M^i M i when present, the unperturbed bounds give
∥ ( K i b + t ) − 1 ∥ b w ≤ C 0 , J λ J + t ( t ≥ 0 ) . \left\|\left(K_i^b+t\right)^{-1}\right\|_{b_w}\le\frac{C_{0,J}}{\lambda_J+t}\qquad(t\ge0). ( K i b + t ) − 1 b w ≤ λ J + t C 0 , J ( t ≥ 0 ) .
Put L i = K λ , i − K i b L_i=K_{\lambda,i}-K_i^b L i = K λ , i − K i b . At each fixed coarser argument on the conditional products, the fixed-reference Picard bounds of Steps 5 and 6 give the coefficient displacement of the conditional center. For this bulk matrix estimate the offset calculation of Step 4 is used with β w = b w L J ′ \beta_w=b_wL'_J β w = b w L J ′ : since b w ≤ c J b_w\le c_J b w ≤ c J , one still has e β w κ ≤ 2 / 3 e^{\beta_w\kappa}\le2/3 e β w κ ≤ 2/3 , and C d i m b w ≤ min ( c J , B a c l ) / 2 C_{\mathrm{dim}}b_w\le\min(c_J,Ba_{\mathrm{cl}})/2 C dim b w ≤ min ( c J , B a cl ) /2 pays the transported network. Together with the unperturbed mixed tensor bounds, this bounds the perturbative matrix difference at b w b_w b w by C P , b w , J σ ∗ C_{P,b_w,J}\sigma_* C P , b w , J σ ∗ . For a Schur block, telescope its defining product and use the principal inverses already constructed. This estimate is obtained before successive substitution. Since b w < b r e f b_w<b_{\mathrm{ref}} b w < b ref , monotonicity preserves the unperturbed half-margin on the fixed products. Fix the perturbative constants at the working rates and then reduce σ 0 , J \sigma_{0,J} σ 0 , J so that
C 0 , J C P , b w , J σ 0 , J ≤ λ J / 8. C_{0,J}C_{P,b_w,J}\sigma_{0,J}\le\lambda_J/8. C 0 , J C P , b w , J σ 0 , J ≤ λ J /8.
The two halves give C 0 , J ∥ L i ∥ b w ≤ λ J / 4 C_{0,J}\lVert L_i\rVert_{b_w}\le\lambda_J/4 C 0 , J ∥ L i ∥ b w ≤ λ J /4 at every stage. With R i b ( t ) = ( K i b + t ) − 1 R_i^b(t)=(K_i^b+t)^{-1} R i b ( t ) = ( K i b + t ) − 1 , the ordered Neumann expansion
( K λ , i + t ) − 1 = ∑ n ≥ 0 [ − R i b ( t ) L i ] n R i b ( t ) (K_{\lambda,i}+t)^{-1}=\sum_{n\ge0}\left[-R_i^b(t)L_i\right]^nR_i^b(t) ( K λ , i + t ) − 1 = n ≥ 0 ∑ [ − R i b ( t ) L i ] n R i b ( t )
converges in the coefficient norm at this same working exponent and gives 4 C 0 , J / [ 3 ( λ J + t ) ] 4C_{0,J}/[3(\lambda_J+t)] 4 C 0 , J / [ 3 ( λ J + t )] . The bounds with an analytic input fixed follow by inserting the differentiated tensor in this normally convergent series on a fixed smaller product. Thus the inverse uses the intrinsic decay of the unperturbed resolvent and the allocated working decay of the incoming perturbation.
Step 9: the differences carry the perturbation. It remains to prove that the differences carry the incoming boundary dependence, including through the moving coarser arguments. At fixed y y y , differentiating the gradient equation of an initial segment, we obtain
∂ λ z I , λ = − Q λ , I I − 1 V I . (48) \partial_\lambda z_{I,\lambda}=-Q_{\lambda,II}^{-1}V_I. \tag*{(48)} ∂ λ z I , λ = − Q λ , I I − 1 V I . (48)
The free inverse just constructed is a coefficient series at the allocated working rates without boundary weight, and V I V_I V I is the boundary-localized series of Steps 3 and 4. These are joint coefficient bounds in the coarse field and the shifted higher parameters: the jet majorant of Step 3, its transported version of Step 4, and the differentiated Picard series give this joint statement on a fixed smaller product. We use these bounds for moving higher arguments; the fixed-y y y supremum defining δ V \delta_V δ V alone is not a substitution estimate. The majorant of the product is (19.4), and it is O J ( σ ∗ r c ) O_J(\sigma_*r_c) O J ( σ ∗ r c ) ; integrating in λ \lambda λ proves the bound on the difference of stationary points. Let Z λ Z_\lambda Z λ be the stationary point at which a Hessian is evaluated, and let H H H denote here this Hessian of Q λ Q_\lambda Q λ . Its total derivative is
H ˙ = V z z + Q λ ( 3 ) [ Z ˙ λ ] , \dot{H}=V_{zz}+Q_\lambda^{(3)}[\dot{Z}_\lambda], H ˙ = V z z + Q λ ( 3 ) [ Z ˙ λ ] ,
which is O J ( σ ∗ ) O_J(\sigma_*) O J ( σ ∗ ) and boundary-localized. For a set L L L of finer eliminated coordinates, the Schur block of scale i i i is S = H i i − H i L R H L i S=H_{ii}-H_{iL}RH_{Li} S = H ii − H i L R H L i with R = H L L − 1 R=H_{LL}^{-1} R = H LL − 1 , and its derivative is
S ˙ = H ˙ i i − H ˙ i L R H L i − H i L R H ˙ L i + H i L R H ˙ L L R H L i . (49) \dot{S}=\dot{H}_{ii}-\dot{H}_{iL}RH_{Li}-H_{iL}R\dot{H}_{Li}+H_{iL}R\dot{H}_{LL}RH_{Li}. \tag*{(49)} S ˙ = H ˙ ii − H ˙ i L R H L i − H i L R H ˙ L i + H i L R H ˙ LL R H L i . (49)
Every term has exactly one factor which carries the perturbation, including the term coming from the derivative of the inverse. Multiplication by M i M^i M i gives the derivative of K λ , i K_{\lambda,i} K λ , i . The derivative of the inverse is − K − 1 K ˙ K − 1 -K^{-1}\dot{K}K^{-1} − K − 1 K ˙ K − 1 , and (19.7), or its differentiated integral, gives the derivative of the inverse square root, with integrable majorant C J t − 1 / 2 ( λ J + t ) − 2 ∥ K ˙ ∥ C_Jt^{-1/2}(\lambda_J+t)^{-2}\lVert\dot{K}\rVert C J t − 1/2 ( λ J + t ) − 2 ∥ K ˙ ∥ . These identities prove the stated bounds on all differences of matrix functions. Along the successive chain, the derivative in (19.48) becomes
z ˙ I = − Q λ , I I − 1 ( V I + Q λ , I c z ˙ c ) . \dot{z}_I=-Q_{\lambda,II}^{-1}(V_I+Q_{\lambda,I c}\dot{z}_c). z ˙ I = − Q λ , I I − 1 ( V I + Q λ , I c z ˙ c ) .
An inserted increment has total derivative
d d λ [ a λ , i ( c , z > i ) + K λ , i ( c , z > i ) − 1 / 2 q i ] = ∂ λ a λ , i + D y a λ , i z ˙ > i + d d λ K λ , i − 1 / 2 q i , \frac{d}{d\lambda}\left[a_{\lambda,i}(c,z_{>i})+K_{\lambda,i}(c,z_{>i})^{-1/2}q_i\right]
=\partial_\lambda a_{\lambda,i}+D_ya_{\lambda,i}\dot{z}_{>i}+\frac{d}{d\lambda}K_{\lambda,i}^{-1/2}q_i, d λ d [ a λ , i ( c , z > i ) + K λ , i ( c , z > i ) − 1/2 q i ] = ∂ λ a λ , i + D y a λ , i z ˙ > i + d λ d K λ , i − 1/2 q i ,
where the last derivative includes both the explicit change of the objective and the change of the coarser argument. We start at the coarsest conditional stationary point and descend. Each direct derivative carries the perturbation, and each new contribution from the coarser arguments contains a factor which is already known to carry it. By (19.49) and (19.7) this remains true for every inverse square root used at the next scale. Since ∥ q i ∥ ∞ ≤ Q 0 , J ≤ r c \lVert q_i\rVert_\infty\le Q_{0,J}\le r_c ∥ q i ∥ ∞ ≤ Q 0 , J ≤ r c , the finite recurrence gives O J ( σ ∗ r c ) O_J(\sigma_*r_c) O J ( σ ∗ r c ) for the stationary points and the reconstructed displacements, and O J ( σ ∗ ) O_J(\sigma_*) O J ( σ ∗ ) for their matrix coefficients. Differentiating in a coarse-field or free input fixes the index of that input in these convergent series, and the allocation of Step 4 keeps the perturbation and every external argument in one connected network. This proves part (3) on the product, and together with Step 4 it proves part (4).
Step 10: two perturbations. For two perturbations we interpolate between them. Since the coefficient bounds are convex and the fixed bound σ 0 , J \sigma_{0,J} σ 0 , J holds for both, every interpolated objective stays in the same domains. Its derivative in the interpolation parameter is given by (48) with V = ( P ‾ 1 − P ‾ 0 ) ∘ Φ 0 V=(\overline{P}_1-\overline{P}_0)\circ\Phi_0 V = ( P 1 − P 0 ) ∘ Φ 0 . We set
Δ i n = χ D i n , α c l , χ c l φ r c ( P ‾ 1 − P ‾ 0 ) , φ i n = r c . (50) \Delta_{\mathrm{in}}=\chi^{\varphi^{r_c}}_{D_{\mathrm{in}},\alpha_{\mathrm{cl}},\chi_{\mathrm{cl}}}(\overline{P}_1-\overline{P}_0),\qquad\varphi_{\mathrm{in}}=r_c. \tag*{(50)} Δ in = χ D in , α cl , χ cl φ r c ( P 1 − P 0 ) , φ in = r c . (50)
Since both perturbations use the same D i n D_{\mathrm{in}} D in , the same positions and cell volumes, and the same incoming boundary, all the preceding difference estimates are linear in Δ i n \Delta_{\mathrm{in}} Δ in . We write Q λ \mathcal{Q}_\lambda Q λ also for the interpolated objective. The derivative of the outgoing values in the interpolation parameter is
∂ λ [ M s Q λ ( c , z λ ; b ) ] = M s ( P ‾ 1 − P ‾ 0 ) ( Φ 0 ( c , u λ ; ξ ) ; b ) . \partial_\lambda\left[M^s\mathcal{Q}_\lambda(c,z_\lambda;b)\right]=M^s(\overline{P}_1-\overline{P}_0)\left(\Phi_0(c,u_\lambda;\xi);b\right). ∂ λ [ M s Q λ ( c , z λ ; b ) ] = M s ( P 1 − P 0 ) ( Φ 0 ( c , u λ ; ξ ) ; b ) .
Its nonconstant coefficient networks contain the same incoming difference, also when a derivative hits the stationary point. To display the degree sum, denote the right side by G λ G_\lambda G λ and use a fixed outer profile ψ ≥ ζ φ c l \psi\geq\zeta\varphi_{\mathrm{cl}} ψ ≥ ζ φ cl , ζ > 1 \zeta>1 ζ > 1 , within the already chosen construction product. The distinguished reconstruction column and the transport estimate of Step 4 give
sup i ∈ Ω e α o u t d Ω ( i ) ∑ ν ≠ 0 ν i ∣ ( G λ ) ν ∣ e α o u t t ( ν ) ψ ν ≤ C J r c 2 Δ i n . \sup_{i\in\Omega}e^{\alpha_{\mathrm{out}}d_\Omega(i)}\sum_{\nu\ne0}\nu_i|(G_\lambda)_\nu|e^{\alpha_{\mathrm{out}}t(\nu)}\psi^\nu\leq C_Jr_c^2\Delta_{\mathrm{in}}. i ∈ Ω sup e α out d Ω ( i ) ν = 0 ∑ ν i ∣ ( G λ ) ν ∣ e α out t ( ν ) ψ ν ≤ C J r c 2 Δ in .
At degree n n n , restriction to φ c l \varphi_{\mathrm{cl}} φ cl supplies ζ − n \zeta^{-n} ζ − n and the root fraction in (24) supplies 1 / n 1/n 1/ n . Thus, uniformly in λ \lambda λ and the frozen values,
χ Ω , α o u t , α o u t φ c l , c l ( G λ ) ≤ C J Δ i n ∑ n ≥ 1 ζ − n n = − C J Δ i n log ( 1 − ζ − 1 ) . \chi^{\varphi_{\mathrm{cl}},\mathrm{cl}}_{\Omega,\alpha_{\mathrm{out}},\alpha_{\mathrm{out}}}(G_\lambda)\leq C_J\Delta_{\mathrm{in}}\sum_{n\geq1}\frac{\zeta^{-n}}{n}=-C_J\Delta_{\mathrm{in}}\log(1-\zeta^{-1}). χ Ω , α out , α out φ cl , cl ( G λ ) ≤ C J Δ in n ≥ 1 ∑ n ζ − n = − C J Δ in log ( 1 − ζ − 1 ) .
The fixed profile ratio is used once for this scalar sum. Integrating in λ \lambda λ gives
χ Ω , α o u t , α o u t φ c l , c l ( V 1 o u t − V 0 o u t ) ≤ K J c l Δ i n . (51) \chi^{\varphi_{\mathrm{cl}},\mathrm{cl}}_{\Omega,\alpha_{\mathrm{out}},\alpha_{\mathrm{out}}}\left(V_1^{\mathrm{out}}-V_0^{\mathrm{out}}\right)\leq K_J^{\mathrm{cl}}\Delta_{\mathrm{in}}. \tag*{(51)} χ Ω , α out , α out φ cl , cl ( V 1 out − V 0 out ) ≤ K J cl Δ in . (51)
Here the outgoing classical seminorm uses the output profile φ c l \varphi_{\mathrm{cl}} φ cl with interior radius r c r_c r c . The seminorm in (19.50) is the incoming seminorm with constant profile. The output seminorm includes degree one. A difference of incoming constants stays in the full value and has zero seminorm. This proves part (5).
Step 11: the full action. We prove the identity of part (6), with its constants. When Ψ R \Psi_R Ψ R is written on its regional list, ι ζ \iota_\zeta ι ζ inserts the newly frozen coordinates once; when it is written as a complete fine field, the same insertion is already included. In either notation, the exact identities of Lemma 19.5 at the endpoints imply the equality of complete group fields
exp ( ι ζ Φ 0 ) = ( exp ( ι ζ A 0 ) ) G Ω ( c ) \exp(\iota_\zeta\Phi_0) = \left(\exp(\iota_\zeta A_0)\right)^{G^{\Omega(c)}} exp ( ι ζ Φ 0 ) = ( exp ( ι ζ A 0 ) ) G Ω ( c )
since the basepoint and the fine endpoints of every based transport of a newly frozen coordinate are protected, and the gauge transformation is the identity on the previously frozen extension. For real data both complete fields lie in the common chart of Theorem 11.6: the field before gauge transformation by the choice of R r a w R_{\rm raw} R raw , and the gauge-transformed field by the bound on the image under the reconstruction majorant and the bound on the newly frozen coordinates. The gauge transformation is periodic and compact there. Thus Theorem 11.6 applies to the torus classical action, and only to it, and we obtain
F 1 = E d 0 ( ι ζ Φ 0 ) + P ( Φ 0 ; b ) − S ζ c l ( Φ 0 ) = E d 0 ( ι 0 Φ 0 ) + P ( Φ 0 ; b ) = E i n ( Φ 0 ; b ) . \mathcal{F}_1 = E_{d_0}(\iota_\zeta\Phi_0) + P(\Phi_0;b) - S^{\rm cl}_\zeta(\Phi_0)
= E_{d_0}(\iota_0\Phi_0) + P(\Phi_0;b) = E_{\rm in}(\Phi_0;b). F 1 = E d 0 ( ι ζ Φ 0 ) + P ( Φ 0 ; b ) − S ζ cl ( Φ 0 ) = E d 0 ( ι 0 Φ 0 ) + P ( Φ 0 ; b ) = E in ( Φ 0 ; b ) .
These equalities keep S ζ c l ( 0 ) S^{\rm cl}_\zeta(0) S ζ cl ( 0 ) as well as all nonconstant coefficients of S ζ c l S^{\rm cl}_\zeta S ζ cl ; no constant Taylor coefficient has been subtracted. Both sides are holomorphic on the connected product, and they agree on its real open part. Expanding at a real interior point and differentiating the real identity in each coordinate, we find that all Taylor coefficients of the difference vanish. Continuation through overlapping polydiscs in the connected product gives the equality everywhere on it.
If the reference extension of the effective interaction is changed, the corresponding difference is treated by Lemma 18.9. In particular, replacing a resulting interaction R ( v , c ) R(v,c) R ( v , c ) by R ( v , c ) − R ( 0 , c ) R(v,c)-R(0,c) R ( v , c ) − R ( 0 , c ) changes the normalization constant from Λ \Lambda Λ to Λ − R ( 0 , c ) \Lambda-R(0,c) Λ − R ( 0 , c ) and preserves the density; the constant of the extension is not deleted. The perturbation has been evaluated on Φ 0 \Phi_0 Φ 0 throughout.
Step 12: dependence on the frozen coordinates. Fix the geometry, the finite lists of labels, and a strictly smaller analytic product, and suppose that the incoming coefficients and the charts are Borel functions of the real parameter b b b . A finite Taylor partial sum of any chart or incoming function is then jointly Borel in its analytic arguments and b b b . Since the common coefficient bounds on the larger radius give a geometric tail on the smaller product, uniformly in b b b , the limits of the series are jointly Borel. For the scalar incoming function this argument is applied to its part of positive degree, and the unrestricted coefficient ρ 0 ( b ) \rho_0(b) ρ 0 ( b ) is kept as a separate Borel summand.
Every finite-dimensional matrix inverse used in a Picard iterate is a quotient of polynomials in its entries with nonzero determinant on the domain, hence Borel there, and each Picard iterate is a composition of such operations and of the jointly Borel functions just obtained. The uniform contraction ratio bounds the distance of the iterates to the stationary point by a geometric sequence, uniformly in b b b , which implies that the stationary point is jointly Borel. This applies first to the joint problem and then to every conditional problem, at arbitrary coarser arguments in the domains, and the finite successive substitution of these maps is Borel.
For matrix functions we use (5) and (6) . Finite Neumann sums are Borel, and their uniform geometric limit gives the resolvent. On a compact interval of integration inside ( 0 , ∞ ) (0,\infty) ( 0 , ∞ ) the Riemann sums of the integral for the inverse square root are Borel, and the uniform bounds on the gap and the resolvent make the tails at zero and at infinity small, uniformly in the parameters; hence the integral is Borel. The same argument applies to the differences and to the logarithmic factors on their fixed nonvanishing branches. Evaluating the full objective at the stationary point gives Borel minimized values, with the incoming constant restored exactly.
If the coefficients and charts are continuous in b b b , every finite partial sum, inverse, and Picard iterate in this argument is continuous. The common geometric bounds and contraction ratios make all limits uniform on the smaller products, and the tails of the integrals satisfy the same uniform bounds. The limits and the finite successive compositions are therefore continuous in b b b . This proves part (7) without any derivative in ξ \xi ξ or b r e c b_{\mathrm{rec}} b rec .
Step 13: shorter compositions. For s < J s<J s < J we replace B B B by B s B_s B s in the scaling and use the separate conditions on the radius and on the image under the reconstruction majorant for that composition; the proof for the initial segments is unchanged, with s s s in place of J J J . All constants were computed on fixed larger products, before the eventual free, enclosing, and geometric choices.
Fluctuation integration and normalization on block regions
In Sections 18 and § 2.1 we set up the averaging steps of one renormalization group step on a block region. There, the coordinates attached to a union Ω \Omega Ω of coarse blocks are analytic variables, and the coordinates on the fixed complement of Ω \Omega Ω are frozen at real values chosen earlier. Those sections constructed the conditional constrained minimizers of the classical problem on Ω \Omega Ω , proved a lower bound on the constrained Hessian independent of the shape of Ω \Omega Ω , and controlled the effect of a small perturbation of the classical objective. The goal of this section is to integrate out the normalized fluctuation variables of such a step. We prove that the resulting effective interaction differs from the corresponding torus quantity only by terms that decay exponentially with the distance to the fixed complement, and we keep track of the normalization constant Λ \Lambda Λ by a separate argument. All estimates are uniform in the shape and topology of Ω \Omega Ω , the number of its components, the preceding depth, the number of renormalization group steps already performed, and the values of the frozen coordinates.
§ 2.2.1 fixes the notation and the coefficient calculus. We then compare the conditional Hessians and Haar densities with their torus counterparts (§ 2.2.2 ) and derive the conditional fluctuation potential and its concentration bounds (§ 2.2.3 ). We obtain the effective interaction by a cluster expansion (§ 2.2.4 ) and control Λ \Lambda Λ by a separate comparison of real integrals (§ 2.2.5 ). Next, we choose the parameters and prove the finite iteration (Theorem 20.14 ). Finally, we sum the truncation errors for couplings in ( 0 , Γ ] (0,\Gamma] ( 0 , Γ ] with any finite Γ \Gamma Γ (§ 2.2.7 ).
We use the norms and the dependence of constants of Section 4, the concentration estimates of Theorem 5.4, and the triangular change of variables and the finite graph inequality of Theorems 12.6 and 12.13. From Section 18 we use the norm with boundary decay, the network estimates, and the substitution identities, and from § 2.1 the conditional domains, image bounds, perturbation theorem, and column bounds. The results are used in Section 21, which compares the integrals over the regions of the large-field decomposition with the analytic integrals constructed here, in Section 22, which proves the exponential-moment estimates for blocked actions, and in Section 23, which chooses the parameters.
All statements of this section concern analytic functions on stated domains and real integrals over stated products. We prove in Section 21 that the configurations of the large-field decomposition lie in these domains.
Heuristics. We first explain the orders of magnitude. We write the free logarithmic coordinates of one intermediate scale as z = z ∗ + g K − 1 / 2 w z = z^* + gK^{-1/2}w z = z ∗ + g K − 1/2 w , where z ∗ z^* z ∗ is the conditional constrained minimizer, K K K the constrained Hessian, g g g the coupling, and w w w the normalized fluctuation variable. Since the classical action enters the density divided by g 2 g^2 g 2 , and its expansion around the minimizer has no linear term and a quadratic term equal to ∣ w ∣ 2 / 2 \lvert w\rvert^2/2 ∣ w ∣ 2 /2 , the remaining classical terms are
g − 2 O ( ( z − z ∗ ) 3 ) = O ( g ∣ w ∣ 3 ) . g^{-2} O\left((z-z^*)^3\right) = O\left(g\lvert w\rvert^3\right). g − 2 O ( ( z − z ∗ ) 3 ) = O ( g ∣ w ∣ 3 ) .
The Haar disintegration density, the logarithmic determinants of the higher scales, and the incoming effective interaction are not divided by g 2 g^2 g 2 , and their dependence on w w w starts with a term of order g ∣ w ∣ g\lvert w\rvert g ∣ w ∣ . On fluctuation balls of radius p ≥ 1 p \ge1 p ≥ 1 , the interaction is therefore small when g [ p 3 + ( 1 + H ) p ] g[p^3 + (1+H)p] g [ p 3 + ( 1 + H ) p ] is small, where H H H bounds the incoming effective interaction; this is the smallness condition of Theorem 20.9 . The terms of first order in g g g are odd under w ↦ − w w \mapsto-w w ↦ − w . Their expectations under a symmetric reference measure vanish, and the fluctuation correction to the effective interaction is of order g 2 g^2 g 2 . The conditional measure with the frozen coordinates held fixed is interacting and need not be symmetric. For this reason, we bound the means of the fluctuation variables separately in Lemma 20.5 .
We control the normalization constant by two identities. First, a missing scalar Gaussian coordinate at an intermediate scale with coupling g i g_i g i removes one factor g i g_i g i from the density. This contributes log ( 1 / g i ) \log(1/g_i) log ( 1/ g i ) to the difference between the regional and the torus normalization. Second, fixing coordinates at values c c c uses
e Λ − R ( v , c ) = e Λ − R ( 0 , c ) e − [ R ( v , c ) − R ( 0 , c ) ] . e^{\Lambda-R(v,c)} = e^{\Lambda-R(0,c)}e^{-[R(v,c)-R(0,c)]}. e Λ − R ( v , c ) = e Λ − R ( 0 , c ) e − [ R ( v , c ) − R ( 0 , c )] .
which moves the value at v = 0 v=0 v = 0 into Λ \Lambda Λ and leaves a centered function. Since the norms of the effective interaction vanish on constants, they cannot control Λ \Lambda Λ , and we estimate Λ \Lambda Λ by comparing real integrals.
Finally, fixing the coordinates of a boundary layer affects the boundary part of the effective interaction in two ways. First, the norm of the old boundary discrepancy is bounded by 2 e − α b d ( s e r − 1 ) 2e^{-\alpha_{\mathrm{bd}}(s_{\mathrm{er}}-1)} 2 e − α bd ( s er − 1 ) times its previous norm, where s e r s_{\mathrm{er}} s er is the width of the frozen layer. Second, fixing coordinates at nonzero values creates a new mixed term from the torus interaction, of size at most a constant times E β r e c / ρ E\beta_{\mathrm{rec}}/\rho E β rec / ρ , where E E E bounds the torus interaction and β r e c \beta_{\mathrm{rec}} β rec bounds the frozen values. The first estimate needs a wide layer, the second needs small frozen values, and neither replaces the other.
Attribution. The architecture used here, namely constrained minimization in a gauge-fixed chart, separation of a background field from normalized fluctuations, and integration of the fluctuations by a cluster expansion, is Bałaban’s renormalization group for lattice gauge theories [18, 19]. We attribute the individual tools where they are used.
Order of the proofs. The dependencies of the proofs are acyclic. Parts (1)–(2) of Lemma 20.5 use the coefficient, product, and curvature bounds of Lemma 20.3 , the density sum (20.2), and Theorem 5.4; they determine the constant C 1 C_1 C 1 used in a hypothesis of Lemma 20.3 . Part (3) refers forward to the independently proved endpoint summation lemma Lemma 20.15 . The logarithmic gates are sufficient conditions for the explicit product inclusion used in parts (1)–(2), rather than inputs to that independent summation proof.
Setting and coefficient calculus
Geometry and labels. Fix an odd integer M ≥ 3 M \ge3 M ≥ 3 , an integer J ≥ 1 J \ge1 J ≥ 1 , and B = M J B = M^J B = M J . We consider a composition of s s s averaging steps, 1 ≤ s ≤ J 1 \le s \le J 1 ≤ s ≤ J , from a lattice of side M s m M^s m M s m to a lattice of side m ≥ 4 m \ge4 m ≥ 4 , together with an arbitrary union Ω \Omega Ω of blocks of the output lattice. When s = J s = J s = J , we speak of a full step; it is one renormalization group step. The intermediate scales are numbered i = 0 , … , s − 1 i = 0,\ldots,s - 1 i = 0 , … , s − 1 . A label is a scalar coordinate: it specifies a direction, a color, an intermediate scale (or the output lattice), and the lattice on which the coordinate was introduced, and it has an assigned position. A fluctuation block is the three-component vector w i , e ∈ R 3 w_{i,e} \in\mathbb{R}^3 w i , e ∈ R 3 of normalized fluctuation variables attached to one free link e e e at one intermediate scale i i i ; it is a single variable of the reference measures below, not three independent ones. Scalar labels and fluctuation blocks are counted separately.
The labels attached to Ω \Omega Ω are unfixed: their coordinates are analytic variables. These are the coarse background variable v = Log c v = \operatorname{Log} c v = Log c on the output lattice and the free logarithmic coordinates z i z_i z i at the intermediate scales. We write Ω c \Omega^c Ω c for the fixed complement of the unfixed positions, W W W for the boundary (the interface between the unfixed positions and Ω c \Omega^c Ω c ), and Ω L J c \Omega^c_{L_J} Ω L J c for the fixed enlargement of Ω c \Omega^c Ω c determined by the construction stencils, as in Lemma 18.15. The fixed complement is denoted H H H in Section 18 and § 2.1 : in their notation, Ω c = H \Omega^c = H Ω c = H and Ω L J c = H L J \Omega^c_{L_J} = H_{L_J} Ω L J c = H L J . In this section the letter H H H denotes instead the bound E + X ρ ( f ) E + \mathcal{X}^\rho(f) E + X ρ ( f ) on the incoming effective interaction introduced in Lemma 20.7 . All distances are graph distances on the refined position graph of Section 18.2, in units of the output lattice, and network lengths use its periodic ℓ 1 \ell^1 ℓ 1 metric, with internal vertices allowed; positions of labels introduced on finer lattices are embedded in these units. We write d g d_g d g and t g t_g t g for this distance and network length, and d E d_E d E , t E t_E t E for the periodic Euclidean distance and network length of Definition 4.6. Then
d E ≤ d g ≤ 3 d E , t E ≤ t g ≤ 3 t E , d_E \le d_g \le\sqrt{3}d_E,\qquad t_E \le t_g \le\sqrt{3}t_E, d E ≤ d g ≤ 3 d E , t E ≤ t g ≤ 3 t E ,
since ∣ y ∣ 2 ≤ ∣ y ∣ 1 ≤ 3 ∣ y ∣ 2 |y|_2 \le|y|_1 \le\sqrt{3}|y|_2 ∣ y ∣ 2 ≤ ∣ y ∣ 1 ≤ 3 ∣ y ∣ 2 for y ∈ R 3 y \in\mathbb{R}^3 y ∈ R 3 , applied to each periodic image and to each segment [ p , q ] [p,q] [ p , q ] of a network, with y = p − q y=p-q y = p − q . Consequently, a coefficient bound stated with Euclidean distances and network lengths, at spatial exponent c c c and boundary exponent c ′ c' c ′ , holds with the graph quantities at exponents c / 3 c/\sqrt{3} c / 3 and c ′ / 3 c'/\sqrt{3} c ′ / 3 . Conversely, a bound stated with the graph quantities implies the Euclidean bound at the same exponents. Every imported Euclidean coefficient bound is converted in this way before the rates of this section are chosen, and the fixed factor 3 \sqrt{3} 3 is included in the geometric constants. Throughout this section, dist \operatorname{dist} dist and t t t denote d g d_g d g and t g t_g t g , as in Section 18.2. The torus seminorm B ρ , a T \mathcal{B}_{\rho,a_T} B ρ , a T keeps the Euclidean convention of Definition 4.6. For an unfixed label i i i we write d Ω ( i ) d_{\Omega}(i) d Ω ( i ) for its distance to Ω c \Omega^c Ω c . We write Ω ′ \Omega' Ω ′ for the set of unfixed scalar output labels. Separately, let ∂ b l k Ω \partial_{\mathrm{blk}}\Omega ∂ blk Ω be the set of anchors of retained final blocks that share a face with a final block of the fixed complement. This is a geometric set on the output lattice, counted once per anchor, regardless of directions, colors, or the presence of an output link. In particular it includes an isolated retained block, even though that block has no internal coarse bond and contributes no label to Ω ′ \Omega' Ω ′ . It is empty when Ω \Omega Ω or its fixed complement is empty. We assume the label-density bound
# { y : R ≤ dist ( x , y ) < R + 1 } ≤ v J ( R + 2 ) 3 \#\{y:R\le\operatorname{dist}(x,y)<R+1\}\le v_J(R+2)^3 # { y : R ≤ dist ( x , y ) < R + 1 } ≤ v J ( R + 2 ) 3
where v J v_J v J accounts for directions, colors, intermediate scales, and the lattices on which labels are introduced.
Frozen coordinates. The frozen coordinates are written b = ( ξ , b r e c ) b=(\xi,b_{\mathrm{rec}}) b = ( ξ , b rec ) . Here ξ \xi ξ consists of the coordinates frozen at the end of the preceding step, written in the lattice on which they were frozen, and b r e c b_{\mathrm{rec}} b rec consists of the older frozen coordinates, each on the lattice on which it was frozen. Throughout this section the frozen coordinates are real and fixed, and only the unfixed background and free variables are differentiated or continued to complex values. The supremum over the frozen values is taken after every coefficient sum and every degree sum. We claim no derivative of a perturbed quantity in ξ \xi ξ or b r e c b_{\mathrm{rec}} b rec .
Norms. For a function f ( v ; b r e c ) = ∑ ν f ν ( b r e c ) v ν f(v;b_{\mathrm{rec}})=\sum_{\nu}f_{\nu}(b_{\mathrm{rec}})v^{\nu} f ( v ; b rec ) = ∑ ν f ν ( b rec ) v ν of the unfixed labels, a radius ρ > 0 \rho>0 ρ > 0 , a positive profile φ ≤ ρ \varphi\le\rho φ ≤ ρ , a spatial exponent a s p > 0 a_{\mathrm{sp}}>0 a sp > 0 , and a boundary exponent α b d > 0 \alpha_{\mathrm{bd}}>0 α bd > 0 , we use the coefficient norm with boundary decay of Definition 18.1, which is the graph-metric variant of (4.11):
X Ω φ ( f ) = sup b r e c r e a l ρ − 2 ∑ k ≥ 1 sup i ∈ Ω e α b d d Ω ( i ) ∑ ∣ ν ∣ = k v i k ∣ f ν ( b r e c ) ∣ e a s p t ( supp ν ) φ ν . \mathcal{X}_{\Omega}^{\varphi}(f)=\sup_{b_{\mathrm{rec}}\ \mathrm{real}}\rho^{-2}\sum_{k\ge1}\sup_{i\in\Omega}e^{\alpha_{\mathrm{bd}}d_{\Omega}(i)}\sum_{|\nu|=k}\frac{v_i}{k}|f_{\nu}(b_{\mathrm{rec}})|e^{a_{\mathrm{sp}}t(\operatorname{supp}\nu)}\varphi^{\nu}. X Ω φ ( f ) = b rec real sup ρ − 2 k ≥ 1 ∑ i ∈ Ω sup e α bd d Ω ( i ) ∣ ν ∣ = k ∑ k v i ∣ f ν ( b rec ) ∣ e a sp t ( supp ν ) φ ν .
Here t ( S ) = t g ( S ) t(S)=t_g(S) t ( S ) = t g ( S ) is the ℓ 1 \ell^1 ℓ 1 length of a shortest connected network containing the positions in S S S , internal vertices allowed, and d Ω d_{\Omega} d Ω is the graph distance, as in Definition 18.1. The supremum over the root i i i precedes the degree sum, and the supremum over the frozen values follows it. Constants and functions on an empty set of unfixed labels have norm zero, and the degree-one term is included. When Ω c \Omega^{c} Ω c is empty, a boundary difference has zero nonconstant part and norm zero. We write X ρ \mathcal{X}^{\rho} X ρ for the norm with the uniform profile ϕ ≡ ρ \phi\equiv\rho ϕ ≡ ρ , and X c l \mathcal{X}^{\mathrm{cl}} X cl for the classical version, which has its own radius ρ c l \rho_{\mathrm{cl}} ρ cl , profiles, and exponents. For a torus function r r r with D r ( 0 ) = 0 D r(0) = 0 D r ( 0 ) = 0 we use
B ρ , a ( r ) = ∑ k ≥ 2 ρ k − 2 k ! ∥ D k r ( 0 ) ∥ a , r o o t \mathcal{B}_{\rho,a}(r) = \sum_{k \ge2} \frac{\rho^{k-2}}{k!} \lVert D^{k}r(0) \rVert_{a,\mathrm{root}} B ρ , a ( r ) = k ≥ 2 ∑ k ! ρ k − 2 ∥ D k r ( 0 ) ∥ a , root
from Definition 4.6. This seminorm does not control the value r ( 0 ) r(0) r ( 0 ) , which is unrestricted. The three profiles are
ϕ ( i ) = { ρ ∂ , d Ω ( i ) < w b d , ρ , d Ω ( i ) ≥ w b d , ψ p r o f = K ∗ ϕ , ϕ ^ = K ∗ ϕ , \phi(i) =
\begin{cases}
\rho_{\partial}, & d_{\Omega}(i) < w_{\mathrm{bd}},\\
\rho, & d_{\Omega}(i) \ge w_{\mathrm{bd}},
\end{cases}
\qquad
\psi_{\mathrm{prof}} = \sqrt{K_*}\phi,
\qquad
\widehat{\phi} = K_*\phi, ϕ ( i ) = { ρ ∂ , ρ , d Ω ( i ) < w bd , d Ω ( i ) ≥ w bd , ψ prof = K ∗ ϕ , ϕ = K ∗ ϕ ,
where K ∗ ≥ 2 K_* \ge2 K ∗ ≥ 2 is a fixed analytic-radius enlargement factor, 0 < ρ ∂ ≤ ρ 0 < \rho_{\partial} \le\rho 0 < ρ ∂ ≤ ρ , and w b d w_{\mathrm{bd}} w bd is the width of the boundary layer with the smaller radius. We call ϕ ^ \widehat{\phi} ϕ the outer profile, ψ p r o f \psi_{\mathrm{prof}} ψ prof the intermediate profile, and ϕ \phi ϕ the working profile. A bound on a coefficient tensor is rooted at an external index if it holds with that index fixed and all other indices summed, where for an input index the volume-normalized column factor a o u t 3 / a i n 3 a^{3}_{\mathrm{out}}/a^{3}_{\mathrm{in}} a out 3 / a in 3 of Definition 4.6 is included. A coefficient bound with each external index fixed in turn is a bound rooted at every external index. All coefficient bounds concern absolute Taylor coefficients; we never substitute for them a bound on values of a holomorphic function or on an evaluated Hessian.
Admissible data. We work on the products of Lemmas 19.3 and 19.7: the conditional solution neighborhoods of the finitely many stages of the conditional minimization, the smaller neighborhoods of the higher parameters, and the product of sequential free coordinates contained in all of them. On these products, the images under the absolute reconstruction majorant satisfy strict bounds for the classical and quantum radii. A coarse background, a frozen value, or a product of fluctuation balls is called admissible if it lies in these products and satisfies these strict image bounds and the inequality on the conditional forces in Lemma 19.3. When fixing these analytic products, also fix h J > 0 h_J > 0 h J > 0 below the finitely many auxiliary convergence, implicit-function, resolvent, and logarithm radii used below. The outer background products and the sequential free products are restricted inside that reserve; in particular require
K ∗ max ( ρ , ρ c l ) < h J , Q < h J for every sequential free radius Q . (52) K_* \max(\rho,\rho_{\mathrm{cl}}) < h_J,\qquad Q < h_J \quad\text{for every sequential free radius } Q. \tag*{(52)} K ∗ max ( ρ , ρ cl ) < h J , Q < h J for every sequential free radius Q . (52)
This restriction is made with the larger construction products, before the preliminary perturbation threshold and the output constants, rather than inferred later from a real Hessian bound.
Dependence of constants. Unless stated otherwise, every constant in this section (C J C_J C J , c J c_J c J , λ J \lambda_J λ J , radii, and bounds) has the dependence described in item (3) of Definition 4.14. It depends only on M M M , J J J , the fixed larger analytic domains and their radius ratios, the spatial and boundary exponents, the label density ν J \nu_J ν J , and the fixed torus and perturbation bounds. It does not depend on L L L , γ \gamma γ , m m m , N N N , K K K , the preceding depth d 0 d_0 d 0 , the number of renormalization group steps already performed, the shape or topology of Ω \Omega Ω , the number of components of Ω \Omega Ω or Ω c \Omega^c Ω c , or the values of the frozen coordinates. It is also independent of the widths chosen later (the width of the boundary layer, the widths e r _{\mathrm{ser}} ser of the frozen layer, the widths of protected neighborhoods, and the size of the cells of the large-field decomposition): every constant is evaluated on the fixed larger domains before these widths are chosen. We assert no bound uniform as J → ∞ J \to\infty J → ∞ . Constants in statements about summed errors have the dependence described in item (4) of Definition 4.14.
Density sums and boundary transport. Spatial sums use the label density in units of the output lattice. For b > 0 b > 0 b > 0 we use
S J ( b ) = C ν J ∑ l ≥ 0 ( l + 1 ) 3 e − b l , sup x ∑ y e − b dist ( x , y ) ≤ S J ( b ) . (53) S_J(b) = C\nu_J \sum_{l\ge0}(l+1)^3 e^{-bl}, \qquad\sup_x \sum_y e^{-b\operatorname{dist}(x,y)} \le S_J(b). \tag*{(53)} S J ( b ) = C ν J l ≥ 0 ∑ ( l + 1 ) 3 e − b l , x sup y ∑ e − b dist ( x , y ) ≤ S J ( b ) . (53)
The constant C C C accounts for directions, colors, and the finitely many intermediate scales. This is the periodic subset estimate of Lemma 18.15; restricting any of the sets of labels only decreases the sum. Kernels between different meshes keep the volume-normalized column factors of Definition 4.6. For an unfixed root x x x and a position u u u in Ω L J c \Omega^c_{L_J} Ω L J c we have
d Ω ( x ) ≤ d Ω ( u ) + dist ( x , u ) + C J , (54) d_\Omega(x) \le d_\Omega(u) + \operatorname{dist}(x,u) + C_J, \tag*{(54)} d Ω ( x ) ≤ d Ω ( u ) + dist ( x , u ) + C J , (54)
where d Ω ( u ) = 0 d_\Omega(u)=0 d Ω ( u ) = 0 if u ∈ Ω c u\in\Omega^c u ∈ Ω c and C J C_J C J accounts for the fixed enlargement. The weights for network length and for distance to Ω c \Omega^c Ω c are both bounded by one fixed part of the available spatial exponent. The fixed stencil offset in (54) contributes a factor e b C J e^{bC_J} e b C J to the coefficient bound and no factor depending on a later width. We always use the version of Lemma 18.15 with every external index rooted. It performs all contractions with this single allocation of the exponent, independently of the number of vertices and of the Taylor order, instead of reducing an exponent at each contraction.
For a proper nonempty block region, we also use the boundary sum
∑ u u n f i x e d e − c d Ω ( u ) ≤ C J ( c ) ∣ ∂ b l k Ω ∣ , c > 0. (55) \sum_{u\ \mathrm{unfixed}} e^{-c d_\Omega(u)} \le C_J(c)|\partial_{\mathrm{blk}}\Omega|,\qquad c>0. \tag*{(55)} u unfixed ∑ e − c d Ω ( u ) ≤ C J ( c ) ∣ ∂ blk Ω∣ , c > 0. (55)
Suppose first that Ω \Omega Ω and its fixed complement are nonempty. A shortest route from an unfixed position to the fixed complement crosses an exposed face of a retained final block. The anchor of that block belongs to ∂ b l k Ω \partial_{\mathrm{blk}}\Omega ∂ blk Ω and is within a fixed distance C f a c e C_{\mathrm{face}} C face of the crossing point, in output units. The constant includes the fixed convention for link midpoints and endpoints. Thus each unfixed label u u u can be assigned to an anchor a ( u ) ∈ ∂ b l k Ω a(u) \in\partial_{\mathrm{blk}}\Omega a ( u ) ∈ ∂ blk Ω with
dist ( u , a ( u ) ) ≤ d Ω ( u ) + C f a c e . \operatorname{dist}(u,a(u)) \le d_{\Omega}(u) + C_{\mathrm{face}}. dist ( u , a ( u )) ≤ d Ω ( u ) + C face .
The density sum (53) , applied also at these anchor positions, gives
∑ u e − c d Ω ( u ) ≤ e c C f a c e ∑ a ∈ ∂ b l k Ω ∑ u e − c dist ( u , a ) ≤ e c C f a c e S J ( c ) ∣ ∂ b l k Ω ∣ . \sum_u e^{-c d_{\Omega}(u)} \le e^{c C_{\mathrm{face}}} \sum_{a\in\partial_{\mathrm{blk}}\Omega}\sum_u e^{-c\operatorname{dist}(u,a)} \le e^{c C_{\mathrm{face}}} S_J(c)|\partial_{\mathrm{blk}}\Omega|. u ∑ e − c d Ω ( u ) ≤ e c C face a ∈ ∂ blk Ω ∑ u ∑ e − c dist ( u , a ) ≤ e c C face S J ( c ) ∣ ∂ blk Ω∣.
An anchor need not itself carry a variable; the same density bound applies at every position of the refined graph, after a fixed enlargement of v J v_J v J if needed. This proves (55) , without a regularity or component assumption on the boundary. If Ω \Omega Ω is empty there is no label to sum. If its fixed complement is empty, boundary differences vanish and the bound is used only for those zero terms.
We also record the comparison with the interface labels used in the normalization count. For a fixed construction width w b a n d ≥ 1 w_{\mathrm{band}} \ge1 w band ≥ 1 , let N b a n d N_{\mathrm{band}} N band count the scalar labels of the regional step whose positions are within that width of W W W , and set it to zero when W W W is empty. Then
∣ ∂ b l k Ω ∣ ≤ N b a n d ≤ C J ∣ ∂ b l k Ω ∣ . (56) |\partial_{\mathrm{blk}}\Omega| \le N_{\mathrm{band}} \le C_J|\partial_{\mathrm{blk}}\Omega|. \tag*{(56)} ∣ ∂ blk Ω∣ ≤ N band ≤ C J ∣ ∂ blk Ω∣. (56)
For the upper bound, a position within w b a n d w_{\mathrm{band}} w band of W W W is within w b a n d + C f a c e w_{\mathrm{band}} + C_{\mathrm{face}} w band + C face of a boundary-block anchor. The label-density bound on this fixed-radius ball proves the claim. For the lower bound, choose one exposed face of each boundary block and a finest-scale M M M -block touching that face. There is a finest-scale free edge on the exposed face of this smaller block. Indeed, for the coordinate-ordered tree of Definition 18.6, edges in direction 1 are tree edges only on the line where coordinates 2 and 3 equal those of the root, and edges in direction 2 are tree edges only in the plane where coordinate 3 equals that of the root. On a face normal to direction 1, take an edge in direction 2 with coordinate 3 different from that of the root; on a face normal to direction 2 or 3, take an edge in direction 1 off the root line. Such edges exist since M ≥ 3 M \ge3 M ≥ 3 . Each lies inside a finest-scale block, so it is not an eliminated edge, which joins two such blocks. Its midpoint is within one input mesh unit, hence within 1 / B s ≤ 1 1/B_s \le1 1/ B s ≤ 1 , of a link of W W W . One fixed color component of this edge is therefore counted by N b a n d N_{\mathrm{band}} N band . The choices for distinct final blocks are distinct scalar labels. This gives the lower bound, including isolated final blocks. All constants depend only on M M M , J J J , v J v_J v J and the fixed construction width, not on any later width of an excluded or frozen neighborhood.
Two series. For 0 ≤ q < 1 0 \le q < 1 0 ≤ q < 1 and integers j ≥ 0 j \ge0 j ≥ 0 we have
∑ k ≥ 0 ( k + 1 ) ( k + 2 ) q k = 2 ( 1 − q ) 3 , ∑ k ≥ j ( k ) j q k − j = j ! ( 1 − q ) j + 1 . (57) \sum_{k\ge0}(k+1)(k+2)q^k = \frac{2}{(1-q)^3}, \qquad\sum_{k\ge j}(k)_j q^{k-j} = \frac{j!}{(1-q)^{j+1}}. \tag*{(57)} k ≥ 0 ∑ ( k + 1 ) ( k + 2 ) q k = ( 1 − q ) 3 2 , k ≥ j ∑ ( k ) j q k − j = ( 1 − q ) j + 1 j ! . (57)
where ( k ) j = k ( k − 1 ) ⋯ ( k − j + 1 ) (k)_j = k(k-1)\cdots(k-j+1) ( k ) j = k ( k − 1 ) ⋯ ( k − j + 1 ) . They follow by differentiating ∑ k ≥ 0 q k = ( 1 − q ) − 1 \sum_{k\ge0}q^k = (1-q)^{-1} ∑ k ≥ 0 q k = ( 1 − q ) − 1 twice and j j j times, respectively, on compact subintervals of [ 0 , 1 ) [0,1) [ 0 , 1 ) . In particular a tensor of order k + 2 k+2 k + 2 with factorial bound C k + 2 ( k + 2 ) ! C^{k+2}(k+2)! C k + 2 ( k + 2 )! becomes, after division by k ! k! k ! , a summable series of degree-k k k Hessian coefficients. Additional fixed external indices contribute fixed powers of k + 1 k+1 k + 1 , which are absorbed by passing to one strictly smaller auxiliary product. No smaller radius is chosen at a later renormalization group step.
Coefficient resolution. The following lemma makes precise how the connected networks of coefficients behave under the operations used below. It combines the calculus of coefficientwise majorants of Lemma 5.3 with Definition 18.1. For a tensor V V V with output index o ( V ) o(V) o ( V ) we write v i v_i v i for the cube of the mesh of the lattice of an index i i i , and ∣ V ∣ b |V|_b ∣ V ∣ b for the maximum of its output-row norm and its volume-normalized input-column norms at spatial exponent b b b , as in Definition 4.6.
Lemma 20.1. Consider a finite family of typed analytic coefficient tensors. Each tensor is supplied with a coefficient resolution whose total variation, weighted by the exponentials of the lengths of its assigned connected networks, satisfies the bounds of Definition 4.6 with each external index fixed in turn on one common larger product. Consider a sum of externally rooted connected contraction diagrams built from these tensors by substitution, contraction at shared indices (represented by equality tensors), finitely many Schur complements of positive constrained free matrices, and positive implicit or Neumann series. Products of factors with disjoint fluctuation supports and no shared index are not connected; they form separate connected components, each with its own network. Assume that every substitution F ∘ G F\circ G F ∘ G is made on a product whose absolute image majorant for G G G lies strictly inside the coefficient domain of F F F , that all diagram majorants, including the mesh conversions of part (4), converge absolutely there, and that the inverses and Schur operations satisfy the imported uniform gap and coefficient Neumann conditions on that product. Then the following hold.
The sum has an exact coefficient resolution indexed by the occurrences of its constituent tensors, its contracted indices, and its derivative partitions. In each connected diagram component, the network of a resolved occurrence is the union of the assigned networks of its constituent occurrences, joined at the shared indices. This union is connected, contains every external position of that component, and
t ( external positions of the component ) ≤ ∑ constituent occurrences ℓ ( assigned network ) . t(\text{external positions of the component}) \le\sum_{\text{constituent occurrences}} \ell(\text{assigned network}). t ( external positions of the component ) ≤ constituent occurrences ∑ ℓ ( assigned network ) . The monomial of a resolved occurrence has an exact multi-index. Its fluctuation support is the set of whole fluctuation blocks on which its multi-index is positive. Background positions need not belong to this support, and no enlargement of the support to a connected geometric set is made.
If one constituent occurrence of a connected diagram component has a network that visits Ω L j c \Omega^c_{L_j} Ω L j c , then the union visits Ω L j c \Omega^c_{L_j} Ω L j c , and the visit lies in the same connected network as every external position of that component.
Fix an external root of a connected diagram component. For a vector factor V V V attached to its parent at an input index j j j use the counting bound A V ( j ) = ( v j / v o ( V ) ) ∣ V ∣ b A_V(j) = (v_j/v_{o(V)})|V|_b A V ( j ) = ( v j / v o ( V ) ) ∣ V ∣ b , and if it is attached at its output use A V ( o ( V ) ) = ∣ V ∣ b A_V(o(V)) = |V|_b A V ( o ( V )) = ∣ V ∣ b ; for a scalar factor use its counting-root bound. Then the sum over the resolution of the absolute coefficients, weighted by the exponentials of the network lengths, is bounded by the product of these bounds, multiplied at the end by the external normalization v o / v r o o t v_o/v_{\mathrm{root}} v o / v root when the component has an output o o o . For a substitution tree the volume ratios cancel exactly and the bound is the product of the norms ∣ ⋅ ∣ b |\cdot|_b ∣ ⋅ ∣ b , in the row norm and in every volume-normalized column norm. For other diagrams the remaining ratios are kept in the scalar majorant. In particular, a counting scalar outer tensor on an old mesh, rooted at a new input l l l , carries the factor v l / v o l d v_l/v_{\mathrm{old}} v l / v old .
If a connected network of length ℓ \ell ℓ contains the external positions S S S and the unfixed index i i i , and visits Ω L J c \Omega^c_{L_J} Ω L J c , then t ( S ) ≤ ℓ t(S) \le\ell t ( S ) ≤ ℓ and d Ω ( i ) ≤ ℓ + L 0 d_{\Omega}(i) \le\ell+ L_0 d Ω ( i ) ≤ ℓ + L 0 , where L 0 L_0 L 0 is the fixed constant of the enlargement, and consequently
e a t ( S ) + α d Ω ( i ) ≤ e α L 0 e ( a + α ) ℓ . e^{a t(S)+\alpha d_{\Omega}(i)} \le e^{\alpha L_0}e^{(a+\alpha)\ell}. e a t ( S ) + α d Ω ( i ) ≤ e α L 0 e ( a + α ) ℓ .
Proof. We first work with finite Taylor polynomials and finite diagrams. For a substitution U ( o ; l 1 , … , l k ) = ∑ j 1 , … , j k T ( o ; j 1 , … , j k ) ∏ r S r ( j r ; l r ) U(o;l_1,\ldots,l_k) = \sum_{j_1,\ldots,j_k} T(o;j_1,\ldots,j_k) \prod_r S_r(j_r;l_r) U ( o ; l 1 , … , l k ) = ∑ j 1 , … , j k T ( o ; j 1 , … , j k ) ∏ r S r ( j r ; l r ) , the union of the network of T T T with the networks of the S r S_r S r , joined at the indices j r j_r j r , connects every external position, so
t ( o , l 1 , … , l k ) ≤ t ( o , j ) + ∑ r t ( j r , l r ) . t(o,l_1,\ldots,l_k) \le t(o,j) + \sum_r t(j_r,l_r). t ( o , l 1 , … , l k ) ≤ t ( o , j ) + r ∑ t ( j r , l r ) .
We take absolute values before summing. For an output row we sum the inner factors by their rows and then the outer tensor by its row. If a fixed external input belongs to the factor S r S_r S r , we sum the other inner factors by their rows, use the input column of S r S_r S r , and use the j r j_r j r -column of T T T . Along this path the volume factors multiply to ( v o / v j r ) ( v j r / v l ) = v o / v l (v_o/v_{j_r})(v_{j_r}/v_l) = v_o/v_l ( v o / v j r ) ( v j r / v l ) = v o / v l , which is exactly the normalization of the column of the composite at l l l . This proves the cancellation for substitution trees. Products with a common output are treated in the same way: the networks meet at the output, and an input-rooted estimate uses the column of the factor carrying the input and the rows of the others.
For a general connected diagram, we replace every edge outside a spanning tree by two independently summed copies of its index, which only increases the nonnegative sums. We orient the tree from the factor containing the fixed external root and, starting at the leaves, sum each factor over all its indices except the one attached to its parent. A factor attached at an input index j j j contributes at most its volume-normalized column bound times v j / v o ( V ) v_j/v_{o(V)} v j / v o ( V ) , since the column norm carries the factor v o ( V ) / v j v_{o(V)}/v_j v o ( V ) / v j ; this is the counting bound A V ( j ) A_V(j) A V ( j ) . A factor attached at its output contributes its row bound. Finally summing the root factor over all its indices except the fixed root, we obtain the product of the counting bounds, with no count of the internal indices. The external normalization is applied only at the end. In a substitution tree the ratios on the path from the root input to the output multiply to v 1 / v 0 v_1/v_0 v 1 / v 0 and cancel against this normalization; in a general diagram they are kept. Connectedness is essential: factors with disjoint supports and no shared index have no tree, and their networks remain separate. A cycle without an external index, such as an undifferentiated trace, is not bounded in this way.
For compositions we use the ordered derivative formula
D N ( F ∘ G ) ( 0 ) [ u 1 , … , u N ] = ∑ n = 0 N 1 n ! ∑ ( B 1 , … , B n ) F n [ G ∣ B 1 ∣ [ u B 1 ] , … , G ∣ B n ∣ [ u B n ] ] D^N(F \circ G)(0)[u_1,\ldots,u_N] = \sum_{n=0}^{N} \frac{1}{n!} \sum_{(B_1,\ldots,B_n)} F_n\left[G_{\lvert B_1\rvert}[u_{B_1}],\ldots,G_{\lvert B_n\rvert}[u_{B_n}]\right] D N ( F ∘ G ) ( 0 ) [ u 1 , … , u N ] = n = 0 ∑ N n ! 1 ( B 1 , … , B n ) ∑ F n [ G ∣ B 1 ∣ [ u B 1 ] , … , G ∣ B n ∣ [ u B n ] ]
where G ( 0 ) = 0 G(0)=0 G ( 0 ) = 0 , the inner sum is over ordered partitions of { 1 , … , N } \{1,\ldots,N\} { 1 , … , N } into nonempty blocks, and F n = D n F ( 0 ) F_n=D^nF(0) F n = D n F ( 0 ) , G k = D k G ( 0 ) G_k=D^kG(0) G k = D k G ( 0 ) . If G ( 0 ) ≠ 0 G(0)\ne0 G ( 0 ) = 0 , we write G = G ( 0 ) + G ~ G=G(0)+\widetilde{G} G = G ( 0 ) + G and apply this formula to G ~ \widetilde{G} G , with F n F_n F n replaced by D n F ( G ( 0 ) ) D^nF(G(0)) D n F ( G ( 0 )) ; the strict image condition justifies this translation and the infinite sums over the constant terms. Dividing by N ! N! N ! , we obtain the coefficient of s N s^N s N in F ^ ( G ^ ( s ) ) \widehat{F}(\widehat{G}(s)) F ( G ( s )) , where F ^ , G ^ \widehat{F},\widehat{G} F , G are the absolute majorants. Fixing an external index chooses its branch and then its position in that branch, as in ( F ^ ∘ G ^ ) ′ = ( F ^ ′ ∘ G ^ ) G ^ ′ (\widehat{F}\circ\widehat{G})'=(\widehat{F}'\circ\widehat{G})\widehat{G}' ( F ∘ G ) ′ = ( F ′ ∘ G ) G ′ ; no second Taylor factorial occurs. We define the resolution in part (1) by retaining a separate signed occurrence for every summand of this formula, every tuple of contracted indices, and every network refinement of the constituent tensors; grouping occurrences can only decrease the total variation. Its monomial is the product of the constituent monomials, which proves part (2). In a connected component the union of networks contains the visit of a constituent occurrence and is connected, which proves part (3). The diagram bound above, applied to every resolved occurrence, proves part (4) for finite truncations; the gap and Neumann conditions cover the Schur complements, Neumann series, and implicit series. Finally, by the image condition the nonnegative majorant sums converge on the common product. Tonelli’s theorem then justifies every grouping, contraction, substitution, and derivative, and the resulting analytic functions agree because their coefficients agree.
For part (5), the network contains every external position, so t ( S ) ≤ ℓ t(S)\le\ell t ( S ) ≤ ℓ . A route in the network from i i i to the visited position h ∈ Ω L J c h\in\Omega^c_{L_J} h ∈ Ω L J c has length at most ℓ \ell ℓ , and at most L 0 L_0 L 0 more length reaches Ω c \Omega^c Ω c . We conclude that d Ω ( i ) ≤ ℓ + L 0 d_{\Omega}(i)\le\ell+L_0 d Ω ( i ) ≤ ℓ + L 0 , and adding the two inequalities with weights a a a and α \alpha α completes the proof.
When we apply Lemma 20.1 , we sum an internal index by keeping a positive part b b b of the connecting exponent and using (53) . We convert a visit to Ω L J c \Omega^c_{L_J} Ω L J c by (54) into decay in the distance of every unfixed root to Ω c \Omega^c Ω c . In the applications below, the mixed-mesh contractions are substitutions, chains of matrices on a single lattice, and the composition with the incoming interaction, where the remaining ratio is the fixed factor Γ J ≤ M 3 \Gamma_J\le M^3 Γ J ≤ M 3 of Step 4 of the proof of Lemma 20.7 ; these factors are included in the common majorants.
Conditional minimizers, constrained Hessians, and Haar densities
We order the intermediate scales from the finest, i = 0 i=0 i = 0 , to the coarsest, i = s − 1 i=s-1 i = s − 1 . A stage of the conditional minimization is given by an initial segment I I I of the conditional integrations in this order. At that stage the free coordinates of the scales in I I I are minimized, with the coordinates y = z I c y=z_{I^c} y = z I c of the remaining scales held fixed. The fluctuation increments in the following theorem are the coupling-scaled fluctuation inputs, ζ i = q i = g i w i \zeta_i=q_i=g_iw_i ζ i = q i = g i w i in the notation of § 2.1 .
Theorem 20.2. Fix an odd integer M ≥ 3 M\geq3 M ≥ 3 , an integer J ≥ 1 J\geq1 J ≥ 1 , B = M J B=M^J B = M J , the larger analytic domains and their strict margins, the label-density bound ν J \nu_J ν J , and the bounds for the classical and quantum torus norms. Use the conditional solution neighborhoods, the smaller neighborhoods of the higher parameters, and the product of sequential free coordinates of Lemma 19.3 , and the perturbed objective of Theorem 19.9 . In joint coordinates this objective is
F ( v , u ; b r e c ) = F 0 ( v , z 0 ( v ; ξ ) + S 0 ( v ; ξ ) u ; ξ ) + P ˉ ( Φ 0 ( v , u ; ξ ) ; ξ , b r e c ) , S 0 = K 0 − 1 / 2 . (58) \mathcal{F}(v,u;b_{\mathrm{rec}})=F_0\left(v,z_0(v;\xi)+S_0(v;\xi)u;\xi\right)+\bar{P}\left(\Phi_0(v,u;\xi);\xi,b_{\mathrm{rec}}\right),\qquad S_0=K_0^{-1/2}. \tag*{(58)} F ( v , u ; b rec ) = F 0 ( v , z 0 ( v ; ξ ) + S 0 ( v ; ξ ) u ; ξ ) + P ˉ ( Φ 0 ( v , u ; ξ ) ; ξ , b rec ) , S 0 = K 0 − 1/2 . (58) Here v = Log c v=\operatorname{Log}c v = Log c is the coarse background variable, z 0 z_0 z 0 is the unperturbed joint constrained minimizer, K 0 K_0 K 0 is its constrained Hessian, u u u is the joint normalized free variable, and Φ 0 \Phi_0 Φ 0 is the unperturbed reconstruction followed by the gauge transformation G Ω ( c ) G^\Omega(c) G Ω ( c ) of Lemma 19.5 , which depends only on c c c and is the identity at all protected endpoints. The function P ˉ = P − S ξ c l \bar{P}=P-S^{\mathrm{cl}}_\xi P ˉ = P − S ξ cl is the incoming perturbation of the classical action with the classical boundary-completion difference S ξ c l S^{\mathrm{cl}}_\xi S ξ cl subtracted, including its classical constant, as in Theorem 19.9 . Assume
χ c l ( P ˉ ) ≤ 2 σ ∗ , 0 < σ ∗ ≤ σ 0 , \chi^{\mathrm{cl}}(\bar{P})\leq2\sigma_*,\qquad0<\sigma_*\leq\sigma_0, χ cl ( P ˉ ) ≤ 2 σ ∗ , 0 < σ ∗ ≤ σ 0 , where the preliminary perturbation bound σ 0 > 0 \sigma_0>0 σ 0 > 0 satisfies the margins of Lemma 19.3 for the preconditioned conditional forces, the Hessians, and the images. Here X c l \mathcal{X}^{\mathrm{cl}} X cl is the incoming classical norm on the incoming set D i n D_{\mathrm{in}} D in of unfixed labels, with the uniform profile ρ c l \rho_{\mathrm{cl}} ρ cl , as in Theorem 19.9 . Assume the inequality on the conditional forces of Lemma 19.3 , and assume that the incoming fixed complement lies on the current fixed side, in the sense that for every current unfixed position e e e and every incoming label i i i ,
d D i n ( i ) ≥ B s d c u r r e n t ( e ) − B s dist ( e , i / B s ) − C b u f , J . (59) d_{D_{\mathrm{in}}}(i)\geq B_s d_{\mathrm{current}}(e)-B_s\operatorname{dist}(e,i/B_s)-C_{\mathrm{buf},J}. \tag*{(59)} d D in ( i ) ≥ B s d current ( e ) − B s dist ( e , i / B s ) − C buf , J . (59) Here B s = M s B_s=M^s B s = M s , d D i n d_{D_{\mathrm{in}}} d D in is measured in input units, d c u r r e n t = d Ω d_{\mathrm{current}}=d_\Omega d current = d Ω , i / B s i/B_s i / B s is the position of i i i in output units, and C b u f , J C_{\mathrm{buf},J} C buf , J is a fixed buffer that includes the chart stencils. Assume the auxiliary convergence-radius restriction in the definition of admissible data, and assume finally that both images of the incoming functions under the absolute reconstruction majorant, for the classical radius and for the quantum radius, are strictly smaller than half of that radius, as in Lemma 19.7 . Derivatives are taken at fixed real ( ζ , b r e c ) (\zeta,b_{\mathrm{rec}}) ( ζ , b rec ) .
Then the following statements hold for every preceding depth d 0 ≥ 0 d_0 \ge0 d 0 ≥ 0 , every 1 ≤ s ≤ J 1 \le s \le J 1 ≤ s ≤ J , every output side m ≥ 4 m \ge4 m ≥ 4 , every union Ω \Omega Ω of output blocks, and every admissible real value of the frozen coordinates.
Every stage of the conditional minimization has a holomorphic conditional constrained minimizer, and every scale has a positive constrained Hessian K i Ω , P K_i^{\Omega,P} K i Ω , P , 0 ≤ i < s 0 \le i < s 0 ≤ i < s , which includes the exact factor M i M^i M i of its scale. Writing the unnormalized fluctuation increments as ζ i = g i w i \zeta_i = g_iw_i ζ i = g i w i , the free logarithmic coordinates at scale i i i are
z i = z i ∗ ( c i + 1 ; b r e c ) + K i ( c i + 1 ; b r e c ) − 1 / 2 ζ i z_i = z_i^*(c_{i+1};b_{\mathrm{rec}}) + K_i(c_{i+1};b_{\mathrm{rec}})^{-1/2}\zeta_i z i = z i ∗ ( c i + 1 ; b rec ) + K i ( c i + 1 ; b rec ) − 1/2 ζ i
where c i + 1 c_{i+1} c i + 1 is the reconstruction from the coarser scales, with their fluctuation increments inserted. There are Q , a 1 , α 1 , λ J > 0 Q,a_1,\alpha_1,\lambda_J > 0 Q , a 1 , α 1 , λ J > 0 and C J , L K , J < ∞ C_J,\mathcal{L}_{K,J} < \infty C J , L K , J < ∞ such that, on the outer profile ϕ ^ \hat{\phi} ϕ ^ and for ∣ ζ i , e ∣ ≤ Q |\zeta_{i,e}| \le Q ∣ ζ i , e ∣ ≤ Q , the real matrices satisfy
1 2 λ J I ≤ K i Ω , P ≤ L K , J I , ∥ ( K i Ω , P + t ) − 1 ∥ r o w , c o l , a 1 ≤ C J λ J + t ( t ≥ 0 ) . (60) \frac{1}{2}\lambda_J I \le K_i^{\Omega,P} \le\mathcal{L}_{K,J}I,\qquad\left\|\left(K_i^{\Omega,P}+t\right)^{-1}\right\|_{\mathrm{row,col},a_1} \le\frac{C_J}{\lambda_J+t}\qquad(t \ge0). \tag*{(60)} 2 1 λ J I ≤ K i Ω , P ≤ L K , J I , ( K i Ω , P + t ) − 1 row , col , a 1 ≤ λ J + t C J ( t ≥ 0 ) . (60)
The complex continuation satisfies the same resolvent bound with a fixed larger C J C_J C J .
Adjoin to the regional problem, as identical independent auxiliary variables with unit Hessian, the free labels of the matched torus problem that are absent from it, so that both problems have the same sets of free labels. Let K i T , c u t K_i^{\mathbb{T},\mathrm{cut}} K i T , cut be the restriction of the matched torus Hessian to the unfixed free labels, extended by the identity on the auxiliary labels and by zero on the mixed blocks. If
Δ i ( v , ζ ; b r e c ) = K i Ω , P − K i T , c u t = ∑ ν , κ Δ i ; ν κ ( b r e c ) v ν ζ κ , \Delta_i(v,\zeta;b_{\mathrm{rec}}) = K_i^{\Omega,P} - K_i^{\mathbb{T},\mathrm{cut}} = \sum_{\nu,\kappa}\Delta_{i;\nu\kappa}(b_{\mathrm{rec}})v^\nu\zeta^\kappa, Δ i ( v , ζ ; b rec ) = K i Ω , P − K i T , cut = ν , κ ∑ Δ i ; ν κ ( b rec ) v ν ζ κ ,
then
sup u u n f i x e d e α 1 d Ω ( u ) ∑ ν ′ , ν , κ ∣ Δ i ; u v ′ , ν κ ( b r e c ) ∣ ϕ ^ Q ∣ k ∣ e a 1 t ( u , v ′ , supp ν , supp κ ) ≤ C J . (61) \sup_{u\ \mathrm{unfixed}} e^{\alpha_1 d_{\Omega}(u)} \sum_{\nu',\nu,\kappa} \left|\Delta_{i;uv',\nu\kappa}(b_{\mathrm{rec}})\right|\hat{\phi}^{Q|k|}e^{a_1t(u,v',\operatorname{supp}\nu,\operatorname{supp}\kappa)} \le C_J. \tag*{(61)} u unfixed sup e α 1 d Ω ( u ) ν ′ , ν , κ ∑ ∣ Δ i ; u v ′ , ν κ ( b rec ) ∣ ϕ ^ Q ∣ k ∣ e a 1 t ( u , v ′ , supp ν , supp κ ) ≤ C J . (61)
Each regional or torus quantity is evaluated along its own reconstruction chain, at matched retained data and with missing torus increments set to zero; the two higher arguments need not coincide. The transposed bound holds, and so do the bounds rooted at any unfixed background or free external index, with the volume-normalized column factors of Definition 4.6. The underlying mixed tensors with k k k additional background or free derivatives have bounds C J k + 2 ( k + 2 ) ! C_J^{k+2}(k+2)! C J k + 2 ( k + 2 )! at every external index, on a fixed larger product. The auxiliary blocks have zero defect. If Ω c \Omega^c Ω c is empty, the defect has zero nonconstant part.
The same rooted comparisons with boundary decay hold for the conditional minimizers of every stage, the inverse square roots, the logarithms of the Haar disintegration densities, and the diagonal densities ( log K i ) x x / 2 (\log K_i)_{xx}/2 ( log K i ) xx /2 , after composition with the reconstruction from every value of the coarser fluctuation variables. In particular, let L i L_i L i be the negative logarithm of
the Haar disintegration density at scale i i i after its field-independent normalization has been moved into Λ \Lambda Λ , and put
L 0 , i ( c i + 1 ) = L i ( c i + 1 , z i ∗ ( c i + 1 ) ) , B i = L 0 , i + 1 2 log det K i . L_{0,i}(c_{i+1}) = L_i(c_{i+1},z_i^*(c_{i+1})), \qquad B_i = L_{0,i} + \frac{1}{2}\log\det K_i. L 0 , i ( c i + 1 ) = L i ( c i + 1 , z i ∗ ( c i + 1 )) , B i = L 0 , i + 2 1 log det K i .
The fluctuation-dependent part of the density contains
∑ i = 0 s − 1 [ L i ( c i + 1 , z i ) − L 0 , i ( c i + 1 ) + B i ( c i + 1 ) − B i ( c ^ i + 1 ) ] , (62) \sum_{i=0}^{s-1}\left[L_i(c_{i+1},z_i)-L_{0,i}(c_{i+1})+B_i(c_{i+1})-B_i(\widehat{c}_{i+1})\right], \tag*{(62)} i = 0 ∑ s − 1 [ L i ( c i + 1 , z i ) − L 0 , i ( c i + 1 ) + B i ( c i + 1 ) − B i ( c i + 1 ) ] , (62)
where c ^ i + 1 \widehat{c}_{i+1} c i + 1 is the reconstruction with every fluctuation increment equal to zero. These differences start at degree one in ζ \zeta ζ , whereas the classical remainders start at degree three. The comparison includes the correction log ( T I I ) − ( log T ) I I \log(T_{\mathrm{II}})-(\log T)_{\mathrm{II}} log ( T II ) − ( log T ) II that arises when the logarithm of a torus matrix T T T is restricted to the list I I I of unfixed labels. A torus diagonal density that is absent from the regional problem and has an unfixed external index has a network visiting Ω c \Omega^{c} Ω c and is summable from that index; its constant part remains an explicit sum over omitted labels.
All constants have the dependence described in item (3) of Definition 4.14, including independence of the later widths. They are evaluated on the fixed larger products before the final radius for the free coordinates is decreased. In particular they do not change at the renormalization group steps following the creation of a new component of Ω c \Omega^{c} Ω c .
Proof. Steps 1 and 2 verify that the incoming perturbation satisfies the hypotheses of Theorem 19.9 on all parameter domains. Steps 3 to 5 construct the conditional minimizers by a contraction argument that is uniform in the higher parameters, and check that every sequential substitution stays in the admissible domains. Steps 6 to 9 compare the regional and torus quantities, first without and then with the perturbation. Step 10 treats the matrix functions and the Haar densities, Step 11 the triangular determinant, and Step 12 the dependence of the constants on the frozen values.
We fix the real frozen value b = ( ξ , b r e c ) b=(\xi,b_{\mathrm{rec}}) b = ( ξ , b rec ) and write Q θ = F 0 + θ V Q_\theta=F_0+\theta V Q θ = F 0 + θ V , where V = P ‾ ∘ Φ 0 V=\overline{P}\circ\Phi_0 V = P ∘ Φ 0 and 0 ≤ θ ≤ 1 0\leq\theta\leq1 0 ≤ θ ≤ 1 ; here θ \theta θ is an interpolation parameter. Returning from u u u to z = z 0 + S 0 u z=z_0+S_0u z = z 0 + S 0 u is an invertible affine change of variables at fixed v v v . In the free logarithmic coordinates, the objective is
Q θ ( v , z ) = F 0 ( v , z ; ξ ) + θ P ‾ ( Log [ ( exp Ψ R ( v , z ; ξ ) ) G Ω ( v ) ] ; ξ , b r e c ) . Q_\theta(v,z)=F_0(v,z;\xi)+\theta\overline{P}\left(\operatorname{Log}\left[(\exp\Psi_R(v,z;\xi))^{G^\Omega(v)}\right];\xi,b_{\mathrm{rec}}\right). Q θ ( v , z ) = F 0 ( v , z ; ξ ) + θ P ( Log [ ( exp Ψ R ( v , z ; ξ ) ) G Ω ( v ) ] ; ξ , b rec ) .
where Ψ R \Psi_R Ψ R is the reconstruction before the gauge transformation. We use only this change of variables; we do not need the incoming perturbation to take the same value at the reconstruction before and after the gauge transformation.
Step 1: the incoming hypotheses. The incoming analytic labels form the set D i n D_{\mathrm{in}} D in , with their positions on the lattices where they were introduced and their volume factors. On this set the incoming profile is uniformly ρ c l \rho_{\mathrm{cl}} ρ cl ; the coordinates ξ \xi ξ are fixed parameters, and the spatially varying profiles are output profiles. The strict bounds on the image under the reconstruction majorant place Φ 0 \Phi_0 Φ 0 , on every admissible stage, inside this uniform input ball with a margin of half the radius. The same holds for every segment between two compared images, since coefficientwise a convex combination is bounded by the larger of the two absolute image bounds. In particular, the input jet bounds of Theorem 19.9 apply before any conditional minimization, on all its parameter neighborhoods.
Step 2: the incoming boundary comparison. For example, the inclusion of fixed complements used in Theorem 19.9 implies (20.8) as follows. Let h h h be an incoming fixed position nearest to i i i . A path from e e e to the current fixed complement can first go to i / B s i/B_s i / B s , then follow the scaled image of a shortest path from i i i to h h h , of length d i n ( i ) / B s d_{\mathrm{in}}(i)/B_s d in ( i ) / B s , and finally cross the fixed buffer between the incoming and the current fixed complements, whose length is included in C b u f , J / B s C_{\mathrm{buf},J}/B_s C buf , J / B s . Multiplying by B s B_s B s , we obtain (20.8). This argument uses no regularity or component assumption on the boundary. We keep (20.8) as a hypothesis on the geometry; choosing small radii does not establish it.
In particular, if α o u t ≤ B s α c l \alpha_{\mathrm{out}} \le B_s\alpha_{\mathrm{cl}} α out ≤ B s α cl and a chart branch has spatial rate b c h > α o u t b_{\mathrm{ch}} > \alpha_{\mathrm{out}} b ch > α out , then
e α o u t d c u r r e n t ( e ) e − α c l d i n ( i ) e − b c h dist ( e , i / B s ) ≤ e α o u t C b u f , J / B s e − ( α c l − α o u t / B s ) d i n ( i ) e − ( b c h − α o u t ) dist ( e , i / B s ) . e^{\alpha_{\mathrm{out}}d_{\mathrm{current}}(e)}e^{-\alpha_{\mathrm{cl}}d_{\mathrm{in}}(i)}e^{-b_{\mathrm{ch}}\operatorname{dist}(e,i/B_s)}
\le e^{\alpha_{\mathrm{out}}C_{\mathrm{buf},J}/B_s}e^{-(\alpha_{\mathrm{cl}}-\alpha_{\mathrm{out}}/B_s)d_{\mathrm{in}}(i)}e^{-(b_{\mathrm{ch}}-\alpha_{\mathrm{out}})\operatorname{dist}(e,i/B_s)}. e α out d current ( e ) e − α cl d in ( i ) e − b ch dist ( e , i / B s ) ≤ e α out C buf , J / B s e − ( α cl − α out / B s ) d in ( i ) e − ( b ch − α out ) dist ( e , i / B s ) .
This is the transport of the input boundary weight used in the perturbation bound. Input network lengths are divided by B s B_s B s when expressed in output units. Column normalizations remain the ratios of the cubed mesh sizes of their output and input labels. Local rates may be chosen below the finite minimum of these bounds over 1 ≤ s ≤ J 1 \le s \le J 1 ≤ s ≤ J ; the rates returned after a full step use s = J s = J s = J and B s = B B_s = B B s = B . If there is no incoming boundary, the incoming nonconstant boundary function is zero.
Step 3: conditional minimizers. For each stage I I I (an initial segment of the conditional integrations, with y = z I c y = z_{I^c} y = z I c fixed), let a I ( v , y ) a_I(v,y) a I ( v , y ) be the unperturbed conditional minimizer. By Lemma 19.3 , there exist a fixed solution neighborhood, a smaller neighborhood of the higher parameters, and an invertible matrix A I = F 0 , I I ( v , a I , y ) A_I = F_{0,II}(v,a_I,y) A I = F 0 , I I ( v , a I , y ) . Let I 0 I_0 I 0 be a common bound for the coefficient norms of the inverses A I − 1 A_I^{-1} A I − 1 on the fixed larger neighborhoods. Using the uniform incoming jet bound with j = 1 j = 1 j = 1 and the fixed rows of the chart derivatives, we obtain, at every admissible higher parameter y y y and not only at the joint minimizer,
∥ V I ( v , a I , y ; b ) ∥ c o e f f ≤ C P σ ∗ ρ c l , δ V ≤ I 0 C P σ ∗ ρ c l , (63) \lVert V_I(v,a_I,y;b)\rVert_{\mathrm{coeff}} \le C_P\sigma_*\rho_{\mathrm{cl}}, \qquad\delta_V \le I_0C_P\sigma_*\rho_{\mathrm{cl}}, \tag*{(63)} ∥ V I ( v , a I , y ; b ) ∥ coeff ≤ C P σ ∗ ρ cl , δ V ≤ I 0 C P σ ∗ ρ cl , (63)
where δ V \delta_V δ V is the bound on the preconditioned conditional forces of Lemma 19.3 . The constant C P C_P C P is evaluated on the larger products at the preliminary bound σ 0 \sigma_0 σ 0 and includes the conversion from z z z to u u u . The jet bound with j = 2 j = 2 j = 2 and the chain rule, including the term P ˉ ′ Φ 0 , z z \bar{P}'\Phi_{0,zz} P ˉ ′ Φ 0 , z z , give a uniform bound C P σ ∗ C_{P}\sigma_{*} C P σ ∗ on the mixed second derivatives. The real gap, the complex coefficient margins, the inequality δ V ≤ r I / 4 \delta_{V} \le r_{I}/4 δ V ≤ r I /4 , where r I r_{I} r I is the radius of the solution neighborhood, and the derivative inequality below are then the hypotheses of Lemma 19.3 . On that neighborhood the map
h ⟼ h − A I − 1 Q θ , I ( v , a I + h , y ) h \longmapsto h-A_{I}^{-1}Q_{\theta,I}(v,a_{I}+h,y) h ⟼ h − A I − 1 Q θ , I ( v , a I + h , y )
has derivative norm at most 1 / 2 1/2 1/2 and value at zero of norm at most r I / 4 r_{I}/4 r I /4 . Hence it maps the closed ball of radius r I / 2 r_{I}/2 r I /2 into itself, and its successive iterates converge: the difference of the iterates n + 1 n+1 n + 1 and n n n is at most 2 − n 2^{-n} 2 − n times the first difference. The same bounds on absolute coefficients hold for the analytic iterates, which gives normal convergence on every smaller product and a holomorphic limit. Uniqueness follows by subtracting two fixed point equations. Since the objective is holomorphic and real at real arguments, the minimizer is real at real data. The real Hessian and complex resolvent margins are those of Theorem 19.9 ; its weighted resolvent bounds rest on the exponential conjugation method of Combes and Thomas [61]. Taking the finite minimum over s ≤ J s \le J s ≤ J and over the stages fixes λ J \lambda_{J} λ J and Q Q Q , and taking the finite maximum fixes L K , J L_{K,J} L K , J and the bound on the inverses. The factors M i M^{i} M i belong to these matrices before the extrema are taken.
Step 4: sequential substitution. For the sequential construction we use the inequality
C s e q , J ( 2 δ V + Q ) < 1 2 ε t u b e , (64) C_{\mathrm{seq},J}(2\delta_{V}+Q)<\frac{1}{2}\varepsilon_{\mathrm{tube}}, \tag*{(64)} C seq , J ( 2 δ V + Q ) < 2 1 ε tube , (64)
where ε t u b e \varepsilon_{\mathrm{tube}} ε tube is the minimal margin of the parameter and solution neighborhoods of Lemma 19.3 . Let z θ ( v ) z_{\theta}(v) z θ ( v ) and z 0 ( v ) z_{0}(v) z 0 ( v ) be the perturbed and unperturbed joint stationary points. For the initial segment of all free coordinates there is no higher argument, so the contraction bound gives ∥ z θ − z 0 ∥ c o e f f ≤ 2 θ δ V ≤ 2 δ V \lVert z_{\theta}-z_{0}\rVert_{\mathrm{coeff}}\le2\theta\delta_{V}\le2\delta_{V} ∥ z θ − z 0 ∥ coeff ≤ 2 θ δ V ≤ 2 δ V . The tube gate places every projection of z θ z_{\theta} z θ inside the corresponding parameter and solution domains. Writing a θ , I a_{\theta,I} a θ , I for the perturbed conditional minimizer, joint stationarity and conditional uniqueness therefore give a θ , I ( v , z θ , I c ) = z θ , I a_{\theta,I}(v,z_{\theta,I^{c}})=z_{\theta,I} a θ , I ( v , z θ , I c ) = z θ , I . The joint mixed coefficient bounds and the bound on the free inverses give a finite common Lipschitz constant D J D_{J} D J of the conditional minimizers in y y y , and ∥ K i − 1 / 2 ∥ ≤ C S , J \lVert K_{i}^{-1/2}\rVert\le C_{S,J} ∥ K i − 1/2 ∥ ≤ C S , J . Let Δ i s e q \Delta^{\mathrm{seq}}_{i} Δ i seq bound the deviations of the already reconstructed coarser coordinates from z θ z_{\theta} z θ , in the joint coefficient algebra of the background and the remaining increments. Integrating D y a θ , I D_{y}a_{\theta,I} D y a θ , I from z θ , I c z_{\theta,I^{c}} z θ , I c to the moving higher argument uses its joint mixed-derivative majorant on the larger product. The coefficient displacement of z θ z_{\theta} z θ permits this recentering on a fixed smaller product; enlarge D J D_{J} D J once to include the composition. The next insertion then satisfies the majorant inequality
Δ i − 1 s e q ≤ ( 1 + D J ) Δ i s e q + C S , J Q , Δ s s e q = 0. (65) \Delta^{\mathrm{seq}}_{i-1}\le(1+D_{J})\Delta^{\mathrm{seq}}_{i}+C_{S,J}Q,\qquad\Delta^{\mathrm{seq}}_{s}=0. \tag*{(65)} Δ i − 1 seq ≤ ( 1 + D J ) Δ i seq + C S , J Q , Δ s seq = 0. (65)
Iterating and adding the joint-center displacement, we find that the total deviation from z 0 z_{0} z 0 is at most 2 δ V + C S , J Q ∑ k = 0 J − 1 ( 1 + D J ) k 2\delta_{V}+C_{S,J}Q\sum_{k=0}^{J-1}(1+D_{J})^{k} 2 δ V + C S , J Q ∑ k = 0 J − 1 ( 1 + D J ) k . We choose C s e q , J C_{\mathrm{seq},J} C seq , J larger than max { 1 , C S , J ∑ k = 0 J − 1 ( 1 + \max\{1,C_{S,J}\sum_{k=0}^{J-1}(1+ max { 1 , C S , J ∑ k = 0 J − 1 ( 1 + D J k D_J^k D J k and include every fixed embedding constant of the stages. Then (64) keeps every substituted coarser argument in the admissible domains, even after the deviation is doubled. The mixed bounds and the constant C s e q , J C_{\mathrm{seq},J} C seq , J are fixed before the later widths. In particular the preliminary bound σ 0 \sigma_0 σ 0 must satisfy the same inequality with δ V ≤ I 0 C P σ 0 ρ c l \delta_V \le I_0 C_P \sigma_0 \rho_{\mathrm{cl}} δ V ≤ I 0 C P σ 0 ρ cl .
Step 5: dependence on the coarser coordinates. We explain why this construction controls coefficients when the coarser coordinates move. Differentiating the stationary equation at fixed y y y , we obtain
∂ θ z I = − Q θ , I I − 1 V I . (66) \partial_\theta z_I=-Q_{\theta,II}^{-1}V_I. \tag*{(66)} ∂ θ z I = − Q θ , I I − 1 V I . (66)
If y y y moves, the right side has the additional term − Q θ , I I − 1 Q θ , I I c ∂ θ y -Q_{\theta,II}^{-1}Q_{\theta,II^c}\partial_\theta y − Q θ , I I − 1 Q θ , I I c ∂ θ y . A sequential insertion also has the term ( ∂ θ K − 1 / 2 ) ζ i (\partial_\theta K^{-1/2})\zeta_i ( ∂ θ K − 1/2 ) ζ i . All these terms are included in Theorem 19.9 ; the joint mixed-majorant argument leading to (65) , together with the joint-center displacement, keeps their evaluations at the moving coarser coordinates inside the same admissible products. At zero free input, conditional uniqueness identifies the sequential chain with z θ z_\theta z θ . Decreasing Q Q Q therefore does not change the admissible profile of the coarse background.
Step 6: the unperturbed boundary comparison. We use matched reconstruction formulas on the unfixed labels. Away from the fixed construction enlargement of Ω c \Omega^c Ω c , the formulas defining the common unfixed outputs, the block axial gauge equations, and the choices of eliminated edges coincide. Their unique admissible solutions on the unfixed outputs therefore coincide. This does not assert that the extended fields agree inside a component of Ω c \Omega^c Ω c . The prescribed values in Ω c \Omega^c Ω c are separate bounded sources; a source deep inside a component of Ω c \Omega^c Ω c need not be small at its own position. For the torus action composed with two reconstructions Ψ R = Ψ T + d \Psi_R=\Psi_T+d Ψ R = Ψ T + d , the exact difference is
F R − F T = ∫ 0 1 D E d 0 ( Ψ T + τ d ) [ d ] d τ (67) F_R-F_T=\int_0^1 DE_{d_0}(\Psi_T+\tau d)[d]\,d\tau \tag*{(67)} F R − F T = ∫ 0 1 D E d 0 ( Ψ T + τ d ) [ d ] d τ (67)
on the common unfixed variables. If auxiliary missing variables are kept, their separate auxiliary quadratic form is kept as a term rooted in Ω c \Omega^c Ω c . Every differentiated coefficient of the integral contains either a derivative of a changed formula or the factor d d d attached to an action tensor. The action and chart tensors on the larger domain are the tensors of Theorem 19.9 , and their contractions with a changed formula are networks covered by Lemma 18.15. Rooting at any unfixed external index connects that index to the change. For example, the second derivative is
D z 2 ( E d 0 ∘ Ψ ) = Ψ z T E d 0 ′ ′ ( Ψ ) Ψ z + E d 0 ′ ( Ψ ) Ψ z z . D_z^2(E_{d_0}\circ\Psi)=\Psi_z^{\mathsf T}E_{d_0}''(\Psi)\Psi_z+E_{d_0}'(\Psi)\Psi_{zz}. D z 2 ( E d 0 ∘ Ψ ) = Ψ z T E d 0 ′′ ( Ψ ) Ψ z + E d 0 ′ ( Ψ ) Ψ z z .
Both terms are present: stationarity in z z z does not make E d 0 ′ ( Ψ ) E_{d_0}'(\Psi) E d 0 ′ ( Ψ ) vanish. Differentiating (67) produces derivatives falling on E d 0 E_{d_0} E d 0 as well as derivatives falling on d d d ; the former remain connected to the inserted change. Summing a bounded source in Ω c \Omega^{c} Ω c from an unfixed root is justified by (53) and (54) . This proves factorial bounds with a visit to Ω c \Omega^{c} Ω c for the difference before minimization, rooted at every external index.
Step 7: minimization, Schur complements, and restriction. We pass to the conditional minimizers through their stationary equations. A force carrying a visit to Ω c \Omega^{c} Ω c is multiplied by the free inverse, as in (66) . The coefficientwise majorant series of the implicit equation converges on the admissible products: if Y Y Y denotes the majorant of the solution without the visit, the variation carrying the visit has the form H ( x , Y ) / ( 1 − A ( x , Y ) ) H(x,Y)/(1-A(x,Y)) H ( x , Y ) / ( 1 − A ( x , Y )) , where A ≤ 3 / 4 A \le3/4 A ≤ 3/4 after the auxiliary radius has been fixed. This follows by substituting the majorant of the variation into the differentiated fixed point equation and summing ∑ n ≥ 0 A n \sum_{n\ge0} A^n ∑ n ≥ 0 A n . The visit and all fixed external indices are kept. For a partial minimum the exact second derivative is the Schur complement
Q y y − Q y I Q I I − 1 Q I y . (68) Q_{yy}-Q_{yI}Q_{II}^{-1}Q_{Iy}. \tag*{(68)} Q y y − Q y I Q I I − 1 Q I y . (68)
We subtract two such expressions by replacing the three factors of the product one at a time. An inverse difference is A − 1 − B − 1 = − A − 1 ( A − B ) B − 1 A^{-1}-B^{-1}=-A^{-1}(A-B)B^{-1} A − 1 − B − 1 = − A − 1 ( A − B ) B − 1 , so each summand keeps a factor with a visit to Ω c \Omega^{c} Ω c . Only constrained free blocks are inverted. Positivity passes to the Schur complement by minimizing the positive quadratic form over the coordinates in I I I .
A separate correction compares restriction with elimination. Let A A A be a single torus free matrix, and let R R R and H H H be the lists of unfixed and of missing labels (in this paragraph R R R and H H H denote lists of labels). Block elimination gives
( A − 1 ) R R − A R R − 1 = A R R − 1 A R H ( A H H − A H R A R R − 1 A R H ) − 1 A H R A R R − 1 . (69) (A^{-1})_{RR}-A_{RR}^{-1}=A_{RR}^{-1}A_{RH}(A_{HH}-A_{HR}A_{RR}^{-1}A_{RH})^{-1}A_{HR}A_{RR}^{-1}. \tag*{(69)} ( A − 1 ) R R − A R R − 1 = A R R − 1 A R H ( A H H − A H R A R R − 1 A R H ) − 1 A H R A R R − 1 . (69)
To check this identity, we solve A x = ( f R , 0 ) Ax=(f_R,0) A x = ( f R , 0 ) , first for x R x_R x R in terms of x H x_H x H and then for x H x_H x H . Each term on the right of (69) passes through a missing index. Its coefficient network from an unfixed root therefore reaches Ω c \Omega^{c} Ω c , and the part of the exponential decay not used for the network weight sums over that index. Adding this correction to the regional comparison, we obtain the unperturbed defect at each stage. This operation compares restrictions of one torus matrix; it does not identify a regional problem with a torus problem whose complement is set to zero.
Step 8: from factorial bounds to coefficient bounds. The bounds just obtained for the tensors of order k + 2 k+2 k + 2 , at every external index, are C j k + 2 ( k + 2 ) ! C_{j}^{k+2}(k+2)! C j k + 2 ( k + 2 )! . Dividing by the Taylor factorials and using (57) , we find that on a fixed smaller auxiliary radius h h h the summed matrix coefficients are at most
C j 2 ∑ k ≥ 0 ( k + 1 ) ( k + 2 ) ( C j h ) k = 2 C j 2 ( 1 − C j h ) − 3 . C_j^2\sum_{k\ge0}(k+1)(k+2)(C_jh)^k=2C_j^2(1-C_jh)^{-3}. C j 2 k ≥ 0 ∑ ( k + 1 ) ( k + 2 ) ( C j h ) k = 2 C j 2 ( 1 − C j h ) − 3 .
We choose the outer background and free products inside this radius and do not change them later. Since any index may be fixed before the sums are taken, we may fix a background or free input, which proves the column versions.
Step 9: adding the perturbation. We insert the unperturbed regional matrix at its own chain of coarser arguments. The defect in (61) is the sum of three terms: the change of the objective and of the minimizer at a common coarser argument, the change of that argument, and the unperturbed regional–torus defect. The first term has the rooted perturbation bound of Theorem 19.9 , with its visit to Ω c \Omega^{c} Ω c . The second term is the integral of an unperturbed matrix derivative against the difference of the coarser reconstructions, which carries a visit to Ω c \Omega^{c} Ω c ; both are defined on the same products by the statement on the sequential product in Lemma 19.3 , and the network joins the new index to the visit. The third term was treated in Steps 6 to 8. The torus quantities follow their own chain throughout. Differences of the intermediate arguments are included by the reconstruction comparison of Theorem 19.6 and by the sequential chain argument of Steps 4–5, with zero missing increments and matched retained data. The sum satisfies (61) for fixed smaller a 1 , α 1 > 0 a_1,\alpha_1>0 a 1 , α 1 > 0 . The real matrices are symmetric, which gives the transposed bound with the same constants. Rooting each contraction at the specified background or free index gives the other bounds, with its volume factor. The auxiliary blocks are identical throughout and have zero difference; no matrix element rooted purely at an auxiliary label is assigned a boundary decay.
Step 10: matrix functions and Haar densities. For positive matrices, integrating the scalar resolvent and applying the resolvent identity, we obtain the ordered identities of holomorphic functional calculus (see [27])
A − 1 / 2 − B − 1 / 2 = − 1 π ∫ 0 ∞ t − 1 / 2 ( A + t ) − 1 ( A − B ) ( B + t ) − 1 d t (70) A^{-1/2}-B^{-1/2}=-\frac{1}{\pi}\int_{0}^{\infty}t^{-1/2}(A+t)^{-1}(A-B)(B+t)^{-1}\,dt \tag*{(70)} A − 1/2 − B − 1/2 = − π 1 ∫ 0 ∞ t − 1/2 ( A + t ) − 1 ( A − B ) ( B + t ) − 1 d t (70)
log A − log B = ∫ 0 ∞ ( A + t ) − 1 ( A − B ) ( B + t ) − 1 d t (71) \log A-\log B=\int_{0}^{\infty}(A+t)^{-1}(A-B)(B+t)^{-1}\,dt \tag*{(71)} log A − log B = ∫ 0 ∞ ( A + t ) − 1 ( A − B ) ( B + t ) − 1 d t (71)
Indeed A − 1 / 2 = π − 1 ∫ 0 ∞ t − 1 / 2 ( A + t ) − 1 d t A^{-1/2}=\pi^{-1}\int_{0}^{\infty}t^{-1/2}(A+t)^{-1}\,dt A − 1/2 = π − 1 ∫ 0 ∞ t − 1/2 ( A + t ) − 1 d t and log A = ∫ 0 ∞ [ ( 1 + t ) − 1 I − ( A + t ) − 1 ] d t \log A=\int_{0}^{\infty}[(1+t)^{-1}I-(A+t)^{-1}]\,dt log A = ∫ 0 ∞ [( 1 + t ) − 1 I − ( A + t ) − 1 ] d t hold for positive numbers, hence for real symmetric positive matrices by orthogonal diagonalization. The common complex resolvent margins extend these identities holomorphically. The coefficient majorants of the integrands are C J t − 1 / 2 ( λ J + t ) − 2 C_Jt^{-1/2}(\lambda_J+t)^{-2} C J t − 1/2 ( λ J + t ) − 2 and C J ( λ J + t ) − 2 C_J(\lambda_J+t)^{-2} C J ( λ J + t ) − 2 , which are integrable at both endpoints. We may therefore integrate all contractions and derivatives termwise. Each difference keeps its factor A − B A-B A − B .
The logarithm of the Haar disintegration density is a sum of logarithms of the non-vanishing Haar factors of finitely many local formulas and of the determinants of the constraints at the eliminated edges, as in (4.4) and Lemma 18.7. On the fixed chart we write each factor as a ( 1 + h ) a(1+h) a ( 1 + h ) , with its nonzero value a a a at the base point and a fixed strict coefficient margin ∣ h ∣ < 1 |h|<1 ∣ h ∣ < 1 ; the series log ( 1 + h ) = ∑ k ≥ 1 ( − 1 ) k + 1 h k / k \log(1+h)=\sum_{k\geq1}(-1)^{k+1}h^k/k log ( 1 + h ) = ∑ k ≥ 1 ( − 1 ) k + 1 h k / k gives its absolute derivatives on every smaller chart. The base factors are moved into Λ \Lambda Λ . Away from the boundary the local formulas agree. In a difference of compositions, a changed local formula, or a changed minimizer, inverse square root, or coarser argument, is a factor carrying a visit to Ω c \Omega^{c} Ω c . The preceding calculus therefore yields the comparison of the Haar densities at all orders.
A diagonal matrix-logarithm density keeps its index until an external unfixed index is chosen. Its nonconstant derivatives are chains of resolvents with a differentiated matrix inserted. Rooting at that insertion makes the trace sum summable by (53) . If a torus diagonal density is missing from the regional problem, its network with any unfixed external index must reach the missing index, and the same argument sums over the missing index from the unfixed index. A pure constant has no such index and remains its exact sum over omitted labels.
To compare with the restriction of the torus logarithm, we set T = K i T T = K_i^{\mathrm{T}} T = K i T and C = T I I ⊕ T H H C = T_{II} \oplus T_{HH} C = T I I ⊕ T H H , where I = I i I = I_i I = I i and H = H i H = H_i H = H i are the lists of unfixed and missing free labels at scale i i i . Using the integral representation of the logarithm, we obtain
log T I I − ( log T ) I I = ∫ 0 ∞ [ ( T + t ) − 1 − ( C + t ) − 1 ] I I d t . (72) \log T_{II} - (\log T)_{II} = \int_{0}^{\infty} \left[(T+t)^{-1} - (C+t)^{-1}\right]_{II}\,\mathrm{d}t. \tag*{(72)} log T I I − ( log T ) I I = ∫ 0 ∞ [ ( T + t ) − 1 − ( C + t ) − 1 ] I I d t . (72)
The integrand equals − [ ( T + t ) − 1 ( T − C ) ( C + t ) − 1 ] I I -\left[(T+t)^{-1}(T-C)(C+t)^{-1}\right]_{II} − [ ( T + t ) − 1 ( T − C ) ( C + t ) − 1 ] I I . Its middle factor connects unfixed and missing labels, so it has a visit to Ω c \Omega^{c} Ω c , and it has the same integrable bound as (71) . Thus this correction, composed with every value of the coarser fluctuation variables, is included in the comparison.
Step 11: the triangular determinant. The determinant of the triangular change of variables of Theorem 12.13, or of the identical finite triangular calculation for the present minimizers, is
∏ i g i 3 f i det K i ( c i + 1 ) − 1 / 2 , \prod_i g_i^{3f_i}\det K_i(c_{i+1})^{-1/2}, i ∏ g i 3 f i det K i ( c i + 1 ) − 1/2 ,
where f i f_i f i is the number of free blocks at scale i i i . Adding and subtracting the logarithm of its value at zero fluctuation, we combine the Haar terms into (62) . Only one determinant occurs. For an analytic function G G G on the common product,
G ( v , ζ ) − G ( v , 0 ) = ∫ 0 1 D ζ G ( v , τ ζ ) [ ζ ] d τ . (73) G(v,\zeta) - G(v,0) = \int_0^1 D_\zeta G(v,\tau\zeta)[\zeta]\,\mathrm{d}\tau. \tag*{(73)} G ( v , ζ ) − G ( v , 0 ) = ∫ 0 1 D ζ G ( v , τ ζ ) [ ζ ] d τ . (73)
This keeps every positive degree in ζ \zeta ζ . Stationarity and the exact normalization by K i − 1 / 2 K_i^{-1/2} K i − 1/2 cancel the terms of degrees one and two of each classical remainder, so its degree is at least three. Composition with the coarser scales does not lower that degree.
Step 12: constants and measurability. The proof uses only margins on fixed domains, finitely many operations on the stages, (53) , normally convergent coefficientwise majorant series, and the resolvent integrals above. These give the dependence stated in part (4). A new component of Ω c \Omega^{c} Ω c changes the set of indices carrying a visit and the frozen values, but no constant. If Ω c \Omega^{c} Ω c is empty, no formula changes and there is no input carrying a visit. This proves the stated special cases and uniformities.
We also record the regularity in the frozen values needed for finite iterations. If the incoming coefficients and the charts depend on the real value b b b in a Borel way, then so does every iterate of the contraction in Step 3: finite sums, products, and admissible inverses preserve this property, and a convergent coefficient composition is a limit of its polynomial truncations. The geometric contraction bound is uniform in b b b , so the coefficients of the limit are Borel. The resolvent integrals for square roots and logarithms are limits of finite Riemann sums with uniformly bounded tails, and preserve this property. With continuous incoming data on a fixed branch, the same uniform convergence gives continuity. None of these arguments differentiates in b b b . In the finite analytic recursion, the initial torus functions and the local formulas in the identity chart have this regularity, and substitution of a frozen value and the integrals over fixed products below preserve Borel dependence. Thus every coefficient family constructed by the recursion depends on the frozen values in a Borel way.
The conditional fluctuation potential
Lemma 20.3. Use the data and products of Theorem 20.2. Let the incoming effective interaction be
R = r d 0 ( ⋅ ⊕ 0 ) + f , D r d 0 ( 0 ) = 0 , B ρ , a T ( r d 0 ) ≤ E , X ( f ) ≤ σ ∗ ≤ σ 0 , R=r_{d_0}(\mathord\cdot\oplus0)+f,\qquad Dr_{d_0}(0)=0,\qquad B_{\rho,a}^{T}(r_{d_0})\le E,\qquad\mathcal{X}(f)\le\sigma_* \le\sigma_0, R = r d 0 ( ⋅ ⊕ 0 ) + f , D r d 0 ( 0 ) = 0 , B ρ , a T ( r d 0 ) ≤ E , X ( f ) ≤ σ ∗ ≤ σ 0 , where r d 0 r_{d_0} r d 0 is the torus effective interaction at depth d 0 d_0 d 0 , with its torus symmetries, evaluated with the fixed slots set to zero, and f f f is its boundary part. The value of r d 0 r_{d_0} r d 0 at zero is unrestricted. For 0 ≤ i < s ≤ J 0\le i<s\le J 0 ≤ i < s ≤ J set
t max = max i g i = γ M m , g i = t max λ i , λ i = M ( i − s + 1 ) / 2 ∈ [ λ ∗ , 1 ] , t_{\max}=\max_i g_i=\sqrt{\frac{\gamma}{Mm}},\qquad g_i=t_{\max}\lambda_i,\qquad\lambda_i=M^{(i-s+1)/2}\in[\lambda_*,1], t m a x = i max g i = M m γ , g i = t m a x λ i , λ i = M ( i − s + 1 ) /2 ∈ [ λ ∗ , 1 ] , where λ ∗ = M − ( J − 1 ) / 2 \lambda_*=M^{-(J-1)/2} λ ∗ = M − ( J − 1 ) /2 . Let F i F_i F i be the partially minimized objective at scale i i i , without the coupling factor and with its exact normalization M i M_i M i , and define
T i ( c i + 1 , ζ i ) = F i ( c i + 1 , z i ∗ + K i − 1 / 2 ζ i ) − F i ( c i + 1 , z i ∗ ) − 1 2 ∣ ζ i ∣ 2 . T_i(c_{i+1},\zeta_i)=F_i(c_{i+1},z_i^*+K_i^{-1/2}\zeta_i)-F_i(c_{i+1},z_i^*)-\frac{1}{2}|\zeta_i|^2. T i ( c i + 1 , ζ i ) = F i ( c i + 1 , z i ∗ + K i − 1/2 ζ i ) − F i ( c i + 1 , z i ∗ ) − 2 1 ∣ ζ i ∣ 2 . For a fixed admissible real coarse background and fixed real frozen values, the conditional fluctuation potential, with its values at zero fluctuation moved into the prefactor, is exactly
V ( w ) = 1 2 ∑ i ∣ w i ∣ 2 + ∑ i g i − 2 T i ( c i + 1 , g i w i ) + ∑ i [ L i ( c i + 1 , z i ) − L 0 , i ( c i + 1 ) + B i ( c i + 1 ) − B i ( c ^ i + 1 ) ] + R ( Log C ( v , ( g i w i ) i ; b r e c ) ) − R ( Log C ( v , 0 ; b r e c ) ) , (74) \begin{aligned}
V(w)={}&\frac{1}{2}\sum_i|w_i|^2+\sum_i g_i^{-2}T_i(c_{i+1},g_iw_i)\\
&+\sum_i\left[L_i(c_{i+1},z_i)-L_{0,i}(c_{i+1})+B_i(c_{i+1})-B_i(\widehat c_{i+1})\right]\\
&+R\left(\operatorname{Log}C(v,(g_iw_i)_i;b_{\mathrm{rec}})\right)-R\left(\operatorname{Log}C(v,0;b_{\mathrm{rec}})\right), \tag*{(74)}
\end{aligned} V ( w ) = 2 1 i ∑ ∣ w i ∣ 2 + i ∑ g i − 2 T i ( c i + 1 , g i w i ) + i ∑ [ L i ( c i + 1 , z i ) − L 0 , i ( c i + 1 ) + B i ( c i + 1 ) − B i ( c i + 1 ) ] + R ( Log C ( v , ( g i w i ) i ; b rec ) ) − R ( Log C ( v , 0 ; b rec ) ) , (74)
where C ( v , ζ ; b r e c ) C(v,\zeta;b_{\mathrm{rec}}) C ( v , ζ ; b rec ) denotes the reconstructed configuration on the unfixed labels. Only logarithms of unfixed labels in the identity chart are taken. The Haar disintegration density, the degree-one terms of the incoming interaction, and all determinant terms of the coarser scales occur in (74) . The separate prefactor contains all constants of the block averaging constraint and of the Haar measure, the Gaussian powers of the couplings, the values at the minimizers, and the external compact factors.
There are A ≥ 1 A \ge1 A ≥ 1 , a ′ > 0 a' > 0 a ′ > 0 , and C V , J , C 0 < ∞ C_{V,J}, C_0 < \infty C V , J , C 0 < ∞ with C 0 ≥ 2 A 3 C_0 \ge2A^3 C 0 ≥ 2 A 3 , and a fixed radius s W > 0 s_W > 0 s W > 0 for the increments ζ i \zeta_i ζ i , such that on the product ∣ g i w i , e ∣ ≤ s W |g_iw_{i,e}| \le s_W ∣ g i w i , e ∣ ≤ s W
V ( w ) = 1 2 ∣ w ∣ 2 + ∑ k ≥ 3 t max k − 2 P k ( w ) + ∑ k ≥ 1 t max k Q k ( w ) , ∥ P k ∥ a ′ , 1 ≤ A k , ∥ Q k ∥ a ′ , 1 ≤ L V A k , (75) V(w)=\frac{1}{2}|w|^2+\sum_{k\ge3}t_{\max}^{k-2}P_k(w)+\sum_{k\ge1}t_{\max}^kQ_k(w),\qquad\|P_k\|_{a',1}\le A^k,\qquad\|Q_k\|_{a',1}\le\mathcal{L}_V A^k, \tag*{(75)} V ( w ) = 2 1 ∣ w ∣ 2 + k ≥ 3 ∑ t m a x k − 2 P k ( w ) + k ≥ 1 ∑ t m a x k Q k ( w ) , ∥ P k ∥ a ′ , 1 ≤ A k , ∥ Q k ∥ a ′ , 1 ≤ L V A k , (75)
where P k , Q k P_k,Q_k P k , Q k are homogeneous polynomials of degree k k k in the scalar coordinates of w w w and L V = C V , J ( 1 + E + σ 0 ) ≥ 1 \mathcal{L}_V=C_{V,J}(1+\mathbb{E}+\sigma_0)\ge1 L V = C V , J ( 1 + E + σ 0 ) ≥ 1 . For T k ( w ) = ∑ ∣ η ∣ = k t η w η T_k(w)=\sum_{|\eta|=k}t_\eta w^\eta T k ( w ) = ∑ ∣ η ∣ = k t η w η the norm is
∥ T k ∥ a ′ , 1 = sup j ∑ η j > 0 η j ∣ t η ∣ e a ′ t ( supp η ) . \|T_k\|_{a',1}=\sup_j\sum_{\eta_j>0}\eta_j|t_\eta|e^{a't(\operatorname{supp}\eta)}. ∥ T k ∥ a ′ , 1 = j sup η j > 0 ∑ η j ∣ t η ∣ e a ′ t ( supp η ) .
The bounds (75) also hold before the background is evaluated, in the coefficient algebra at the outer profile, with the background positions included in the network length.
The following inequalities suffice for a common enclosing product. All constants below are computed at the preliminary bound σ 0 \sigma_0 σ 0 , before σ ∗ \sigma_* σ ∗ , β r e c \beta_{\mathrm{rec}} β rec , r r r , and the later widths are fixed. Let r q , U , r q , P , u 0 r_{q,U},r_{q,P},u_0 r q , U , r q , P , u 0 be the radii of the unperturbed and perturbed sequential free products and of the joint product of Lemma 19.3 , and c t u b e , J c_{\mathrm{tube},J} c tube , J its fixed conversion constant. For ρ ∙ ∈ { ρ , ρ c l } \rho_\bullet\in\{\rho,\rho_{\mathrm{cl}}\} ρ ∙ ∈ { ρ , ρ cl } let R r a w R_{\mathrm{raw}} R raw , C l i f t , J C_{\mathrm{lift},J} C lift , J , s a u x , ∙ s_{\mathrm{aux},\bullet} s aux , ∙ , C T , ∙ C_{T,\bullet} C T , ∙ , C Δ , ∙ C_{\Delta,\bullet} C Δ , ∙ , and A 0 , ∙ A_{0,\bullet} A 0 , ∙ be the radius of the larger construction domain, the bound on the lifts, the auxiliary radii, the constants of the image increments, and the bound on the image at zero fluctuation in Lemma 19.7 . Choose R W > 0 R_W>0 R W > 0 strictly below
min { s W 2 , r q , U 2 , r q , P 2 , c t u b e , J u 0 , R r a w 64 C l i f t , J , 1 4 A , 1 16 C 0 , min ρ ∙ ∈ { ρ , ρ c l } { s a u x , ∙ 2 , 1 4 C T , ∙ , ρ ∙ 8 ( 2 C T , ∙ + C Δ , ∙ ) } } . \min\left\{\frac{s_W}{2},\frac{r_{q,U}}{2},\frac{r_{q,P}}{2},c_{\mathrm{tube},J}u_0,\frac{R_{\mathrm{raw}}}{64C_{\mathrm{lift},J}},\frac{1}{4A},\frac{1}{16C_0},\min_{\rho_\bullet\in\{\rho,\rho_{\mathrm{cl}}\}}\left\{\frac{s_{\mathrm{aux},\bullet}}{2},\frac{1}{4C_{T,\bullet}},\frac{\rho_\bullet}{8(2C_{T,\bullet}+C_{\Delta,\bullet})}\right\}\right\}. min { 2 s W , 2 r q , U , 2 r q , P , c tube , J u 0 , 64 C lift , J R raw , 4 A 1 , 16 C 0 1 , ρ ∙ ∈ { ρ , ρ cl } min { 2 s aux , ∙ , 4 C T , ∙ 1 , 8 ( 2 C T , ∙ + C Δ , ∙ ) ρ ∙ } } .
Impose the margins of the larger construction domain and of the stages in Lemma 19.7 and, for both radii, A 0 , ∙ ≤ ρ ∙ / 4 A_{0,\bullet}\le\rho_\bullet/4 A 0 , ∙ ≤ ρ ∙ /4 and C T , ∙ K ∗ ρ ∙ ≤ 1 / 4 C_{T,\bullet}K_*\rho_\bullet\le1/4 C T , ∙ K ∗ ρ ∙ ≤ 1/4 . Let C e n , J ≥ 1 C_{\mathrm{en},J}\ge1 C en , J ≥ 1 be a fixed constant (in the application it is the constant of the enclosure estimate Theorem 21.9; nothing else from that section is used), and require
r , β r e c , σ ∗ ρ c l ≤ λ ∗ R W 6 C e n , J , δ e n = C e n , J ( r + β r e c + σ ∗ ρ c l ) , R e n = 2 δ e n λ ∗ , (76) r,\ \beta_{\mathrm{rec}},\ \sigma_*\rho_{\mathrm{cl}}\le\frac{\lambda_*R_W}{6C_{\mathrm{en},J}},\qquad\delta_{\mathrm{en}}=C_{\mathrm{en},J}(r+\beta_{\mathrm{rec}}+\sigma_*\rho_{\mathrm{cl}}),\qquad R_{\mathrm{en}}=\frac{2\delta_{\mathrm{en}}}{\lambda_*}, \tag*{(76)} r , β rec , σ ∗ ρ cl ≤ 6 C en , J λ ∗ R W , δ en = C en , J ( r + β rec + σ ∗ ρ cl ) , R en = λ ∗ 2 δ en , (76)
where r r r is the radius of the small-link domain B r B_r B r and β r e c \beta_{\mathrm{rec}} β rec is the bound on the frozen values. Then the whole product
K e n = ∏ i , e { ∣ w i , e ∣ ≤ R e n / t max } \mathcal{K}_{\mathrm{en}}=\prod_{i,e}\{|w_{i,e}|\le R_{\mathrm{en}}/t_{\max}\} K en = i , e ∏ { ∣ w i , e ∣ ≤ R en / t m a x }
is admissible. On it the absolute coefficient increment of either incoming image is bounded by
C T , ∙ R e n 1 − C T , ∙ ( K ∗ ρ ∙ + R e n ) + C Δ , ∙ R e n ≤ ρ ∙ / 8 , \frac{C_{T,\bullet}R_{\mathrm{en}}}{1-C_{T,\bullet}(K_{*}\rho_{\bullet}+R_{\mathrm{en}})}+C_{\Delta,\bullet}R_{\mathrm{en}}\leq\rho_{\bullet}/8, 1 − C T , ∙ ( K ∗ ρ ∙ + R en ) C T , ∙ R en + C Δ , ∙ R en ≤ ρ ∙ /8 ,
the denominator is at least 1 / 2 1/2 1/2 , and the image under the absolute reconstruction majorant at spatial exponent zero is at most 3 ρ ∙ / 8 3\rho_{\bullet}/8 3 ρ ∙ /8 . If these bounds are imposed in a weighted coefficient row, the same bound holds in that row. In the alternative provided by Lemma 19.7 , the unperturbed image row S c 0 ≤ C J ρ ∙ S^{0}_{c}\leq C_{J}\rho_{\bullet} S c 0 ≤ C J ρ ∙ at an exponent c > 0 c>0 c > 0 , larger than the working exponent, gives one fixed positive exponent ϑ \vartheta ϑ with
S ϑ c 0 ≤ ( 3 ρ ∙ / 8 ) 1 − ϑ ( C J ρ ∙ ) ϑ < ( 3 / 4 ) ρ ∙ . S^{0}_{\vartheta c}\leq(3\rho_{\bullet}/8)^{1-\vartheta}(C_{J}\rho_{\bullet})^{\vartheta}<(\sqrt{3}/4)\rho_{\bullet}. S ϑ c 0 ≤ ( 3 ρ ∙ /8 ) 1 − ϑ ( C J ρ ∙ ) ϑ < ( 3 /4 ) ρ ∙ .
For the unperturbed construction, this exponent is fixed before the regional rates. For the perturbed reconstruction, impose the output-rate conditions of Theorem 19.9 at the allocated working rates below this exponent and bound its coefficient difference from the unperturbed reconstruction there. Fix the larger-product difference-row constant C P , ∙ C_{P,\bullet} C P , ∙ before the final perturbation bound and require C P , ∙ σ ∗ ρ c l ≤ ρ ∙ / 64 C_{P,\bullet}\sigma_{*}\rho_{\mathrm{cl}}\leq\rho_{\bullet}/64 C P , ∙ σ ∗ ρ cl ≤ ρ ∙ /64 for both radii. The total image row is then less than ( 3 / 4 + 1 / 64 ) ρ ∙ < ρ ∙ / 2 (\sqrt{3}/4+1/64)\rho_{\bullet}<\rho_{\bullet}/2 ( 3 /4 + 1/64 ) ρ ∙ < ρ ∙ /2 at those rates; no bound for the perturbed image at the original higher exponent is required. Thus the margin of half the incoming radius is an absolute weighted coefficient bound, not only a bound on values.
Let C 1 C_{1} C 1 be the constant of Lemma 20.5 , which depends only on A A A , a ′ a' a ′ , and v J v_{J} v J . For 0 < t max ≤ t ∗ < e − 1 0<t_{\max}\leq t_{*}<e^{-1} 0 < t m a x ≤ t ∗ < e − 1 impose separately
C 0 t ∗ 2 L V ≤ 1 / 16 , C 1 t ∗ ( L V + 2 ) < 1 4 3 . (77) C_{0}t_{*}^{2}\mathcal{L}_{V}\leq1/16,\qquad C_{1}t_{*}(\mathcal{L}_{V}+2)<\frac{1}{4\sqrt{3}}. \tag*{(77)} C 0 t ∗ 2 L V ≤ 1/16 , C 1 t ∗ ( L V + 2 ) < 4 3 1 . (77)
Then on a neighborhood of the whole real product K e n \mathcal{K}_{\mathrm{en}} K en ,
∥ D w 2 ( V − ∣ w ∣ 2 / 2 ) ∥ r o w , c o l , a ′ ≤ C 0 ( R e n + t max 2 L V ) ( 1 − A R e n ) 3 ≤ 8 27 < 1 2 . (78) \left\|D_{\mathbf{w}}^{2}(V-\lvert w\rvert^{2}/2)\right\|_{\mathrm{row,col},a'}\leq\frac{C_{0}(R_{\mathrm{en}}+t_{\max}^{2}\mathcal{L}_{V})}{(1-AR_{\mathrm{en}})^{3}}\leq\frac{8}{27}<\frac{1}{2}. \tag*{(78)} D w 2 ( V − ∣ w ∣ 2 /2 ) row , col , a ′ ≤ ( 1 − A R en ) 3 C 0 ( R en + t m a x 2 L V ) ≤ 27 8 < 2 1 . (78)
If p ( t ) = A w log ( 1 / t ) p(t)=A_{w}\log(1/t) p ( t ) = A w log ( 1/ t ) , the products of balls with radii p i ∈ [ p ( t max ) , p ( t max ) + A w ( J − 1 ) ( log M ) / 2 ] p_{i}\in[p(t_{\max}),p(t_{\max})+A_{w}(J-1)(\log M)/2] p i ∈ [ p ( t m a x ) , p ( t m a x ) + A w ( J − 1 ) ( log M ) /2 ] lie in K e n \mathcal{K}_{\mathrm{en}} K en whenever
t max [ p ( t max ) + A w ( J − 1 ) ( log M ) / 2 ] < R e n . t_{\max}[p(t_{\max})+A_{w}(J-1)(\log M)/2]<R_{\mathrm{en}}. t m a x [ p ( t m a x ) + A w ( J − 1 ) ( log M ) /2 ] < R en .
The cutoff of the small-field event (4.6) is the case p i = p ( t max ) p_{i}=p(t_{\max}) p i = p ( t m a x ) for all i i i ; the cutoffs p i = p ( g i ) p_{i}=p(g_{i}) p i = p ( g i ) at the individual scales are a separate case. All constants have the dependence described in item (3) of Definition 4.14, and are also independent of the later widths, of the frozen values, and of the individual radii of the fluctuation balls.
Proof. We keep b = ( ξ , b r e c ) b=(\xi,b_{\mathrm{rec}}) b = ( ξ , b rec ) fixed and write t = t max t=t_{\max} t = t m a x . The data of Theorem 20.2 include the uniform incoming classical profile, (20.8), and (20.12) on all parameter neighborhoods. We choose the radius for the increments ζ \zeta ζ below Q Q Q ; then (20.13) holds at every stage. These are the hypotheses for the perturbed tensors used here, and no derivative in the frozen values is needed. Step 1: the exact formula. At scale i i i , the minimum at fixed coarser arguments is F i ( c i + 1 , z i ∗ ) F_i(c_{i+1},z_i^*) F i ( c i + 1 , z i ∗ ) . With S i = K i − 1 / 2 S_i=K_i^{-1/2} S i = K i − 1/2 , Taylor’s formula with integral remainder gives
g i − 2 T i ( c i + 1 , g i w i ) = g i 2 ∫ 0 1 ( 1 − τ ) 2 D i 3 F i ( c i + 1 , z i ∗ + τ g i S i w i ) [ S i w i , S i w i , S i w i ] d τ . (79) g_i^{-2}T_i(c_{i+1},g_iw_i)=\frac{g_i}{2}\int_0^1(1-\tau)^2D_i^3F_i(c_{i+1},z_i^*+\tau g_iS_iw_i)[S_iw_i,S_iw_i,S_iw_i]\,d\tau. \tag*{(79)} g i − 2 T i ( c i + 1 , g i w i ) = 2 g i ∫ 0 1 ( 1 − τ ) 2 D i 3 F i ( c i + 1 , z i ∗ + τ g i S i w i ) [ S i w i , S i w i , S i w i ] d τ . (79)
The terms of orders zero, one, and two have been, respectively, subtracted, annihilated by stationarity, and normalized to ∣ g i w i ∣ 2 / 2 |g_iw_i|^2/2 ∣ g i w i ∣ 2 /2 by S i T K i S i = I S_i^{\mathsf T}K_iS_i=I S i T K i S i = I . The successive partial minima satisfy F i + 1 = M min F i F_{i+1}=M\min F_i F i + 1 = M min F i , and the couplings satisfy g i + 1 2 = M g i 2 g_{i+1}^2=Mg_i^2 g i + 1 2 = M g i 2 , so the values at the minimizers telescope in the exponent. In decreasing order of the scales, the map from w w w to z z z is triangular with diagonal blocks g i S i g_iS_i g i S i . This proves the classical part of (74) and its Gaussian term, by the same finite algebra as in Theorem 12.13. The Haar disintegration density and the single triangular determinant split as in part (3) of Theorem 20.2. Splitting R ( Log C ) R(\operatorname{Log} C) R ( Log C ) at its value at zero fluctuation, we obtain the last two terms. All values removed at zero fluctuation, the Gaussian powers, and the constant Haar densities belong to the prefactor; none of them is set to one.
Step 2: the two series. We fix an auxiliary radius h 0 > 0 h_0>0 h 0 > 0 for the increments ζ \zeta ζ , strictly inside every admissible stage product and its joint neighborhood. Theorem 20.2 gives tensors with bounds at every external index on a larger product. Dividing by the Taylor factorials, composing coefficientwise majorants (Lemma 20.1), and using (57) , we bound the degree-k k k scalar coefficients, rooted at any free index, by D J H J k D_JH_J^k D J H J k for fixed D J , H J D_J,H_J D J , H J . An additional multiplicity η j ≤ k \eta_j\le k η j ≤ k is bounded by 2 k 2^k 2 k . For an incoming function, the strict image bound gives the jet bound
∥ D j R ∥ r o o t ≤ ( E + σ 0 ) ρ 2 − j j ! ( 1 − κ ) j + 1 , κ ≤ 1 2 , j ≥ 1 , (80) \lVert D^jR\rVert_{\mathrm{root}}\le(E+\sigma_0)\rho^{2-j}\frac{j!}{(1-\kappa)^{j+1}},\qquad\kappa\le\frac12,\quad j\ge1, \tag*{(80)} ∥ D j R ∥ root ≤ ( E + σ 0 ) ρ 2 − j ( 1 − κ ) j + 1 j ! , κ ≤ 2 1 , j ≥ 1 , (80)
at a point whose image under the absolute reconstruction majorant is at most κ ρ \kappa\rho κ ρ . To verify it, we fix the first differentiated input index. At original degree l l l , the fixed-index weight in the input norm and the sum over the other j − 1 j-1 j − 1 differentiated positions give the factor ( l ) j (l)_j ( l ) j , with the remaining factor κ l − j ρ 2 − j \kappa^{l-j}\rho^{2-j} κ l − j ρ 2 − j , and the second identity in (57) sums the result. The torus seminorm starts at degree two and the boundary norm at degree one; the same estimate bounds both. Constants of R R R disappear only in the fluctuation difference. The input networks are joined to the reconstruction networks, and a prescribed differentiated index stays fixed. This proves (80) in the coefficient algebra, before the background is evaluated.
Consider a monomial of T i T_i T i , and let η j \eta_j η j be its degree in the free coordinates of scale j j j . It has at least three indices at scale i i i , so η i ≥ 3 \eta_i\ge3 η i ≥ 3 and ∑ j ≥ i η j = k ≥ 3 \sum_{j\ge i}\eta_j=k\ge3 ∑ j ≥ i η j = k ≥ 3 . After division by g i 2 g_i^2 g i 2 its coupling factor is
t k − 2 λ i η i − 2 ∏ j > i λ j η j . t^{k-2}\lambda_i^{\eta_i-2}\prod_{j>i}\lambda_j^{\eta_j}. t k − 2 λ i η i − 2 j > i ∏ λ j η j .
The product of ratios is at most one; the weaker bound λ ∗ − 2 \lambda_*^{-2} λ ∗ − 2 would also be an admissible fixed constant. By (20.22), every Haar, determinant, and incoming difference has at least one free index. A term of degree k k k there has the factor t k ∏ j λ j η j t^k \prod_j \lambda_j^{\eta_j} t k ∏ j λ j η j , whose product of ratios is again at most one. The values B i B_i B i in (20.23) are evaluated at the coarser reconstructions, and their coefficients in the coarser free variables belong to the same series. Summing over at most J J J scales, increasing A ≥ max { 1 , 2 H J , 2 D J H J } A \ge\max\{1,2H_J,2D_JH_J\} A ≥ max { 1 , 2 H J , 2 D J H J } , and decreasing h 0 h_0 h 0 if necessary, we obtain
∥ P k ∥ a ′ , 1 ≤ A k , ∥ Q k ∥ a ′ , 1 ≤ C V , J ( 1 + E + σ 0 ) A k . \|P_k\|_{a',1} \le A^k,\qquad\|Q_k\|_{a',1} \le C_{V,J}(1+E+\sigma_0)A^k. ∥ P k ∥ a ′ , 1 ≤ A k , ∥ Q k ∥ a ′ , 1 ≤ C V , J ( 1 + E + σ 0 ) A k .
The first quantum coefficient Q 1 Q_1 Q 1 includes every degree-one term of the incoming interaction. Only powers of the fixed radii were absorbed in these constants. We choose s W < h 0 s_W<h_0 s W < h 0 with a strict margin in the larger domain. The series are then normally convergent on that radius, at the outer profile as well as after evaluation.
Step 3: the enclosing product is admissible. The three inequalities in (20.25) imply that
δ e n ≤ λ ∗ R W / 2 , R e n ≤ R W . \delta_{\mathrm{en}} \le\lambda_* R_W/2,\qquad R_{\mathrm{en}} \le R_W. δ en ≤ λ ∗ R W /2 , R en ≤ R W .
For w ∈ K e n w \in K_{\mathrm{en}} w ∈ K en we have ∣ ζ i ∣ = t λ i ∣ w i ∣ ≤ R e n |\zeta_i|=t\lambda_i|w_i|\le R_{\mathrm{en}} ∣ ζ i ∣ = t λ i ∣ w i ∣ ≤ R en . The minimum defining R W R_W R W places this whole product strictly inside the unperturbed and perturbed sequential products and the joint neighborhood; the constant c t u b e , J c_{\mathrm{tube},J} c tube , J includes every finite sequential embedding. Indeed, the joint-center displacement is still at most 2 δ V 2\delta_V 2 δ V at the preliminary bound σ 0 \sigma_0 σ 0 , and the new free radius is at most R e n ≤ Q R_{\mathrm{en}}\le Q R en ≤ Q . The gate (20.13) therefore gives C s e q , J ( 2 δ V + R e n ) < ε t u b e / 2 C_{\mathrm{seq},J}(2\delta_V+R_{\mathrm{en}})<\varepsilon_{\mathrm{tube}}/2 C seq , J ( 2 δ V + R en ) < ε tube /2 . Applying (20.14) with free radius R e n R_{\mathrm{en}} R en and adding the joint-center displacement keeps every moving coarser argument in its admissible domain. The separate margin of the larger construction domain in Lemma 19.7 includes the background, the frozen values, and the free radius; it places all reconstructions before the gauge transformation in the larger charts of the Haar densities and of the action.
Step 4: the image increments. For either radius, we abbreviate C T = C T , ∙ C_T=C_{T,\bullet} C T = C T , ∙ and C Δ = C Δ , ∙ C_\Delta=C_{\Delta,\bullet} C Δ = C Δ , ∙ . By hypothesis, we have C T ( K ∗ ρ ∙ + R e n ) ≤ 1 / 2 C_T(K_*\rho_\bullet+R_{\mathrm{en}})\le1/2 C T ( K ∗ ρ ∙ + R en ) ≤ 1/2 . The rational increment in the statement is therefore at most ( 2 C T + C Δ ) R e n ≤ ρ ∙ / 8 (2C_T+C_\Delta)R_{\mathrm{en}}\le\rho_\bullet/8 ( 2 C T + C Δ ) R en ≤ ρ ∙ /8 . Adding the bound A 0 , ∙ ≤ ρ ∙ / 4 A_{0,\bullet}\le\rho_\bullet/4 A 0 , ∙ ≤ ρ ∙ /4 at zero fluctuation, we obtain the bound 3 ρ ∙ / 8 3\rho_\bullet/8 3 ρ ∙ /8 at exponent zero. If these inequalities hold in a weighted coefficient row, each calculation holds in that row. Otherwise apply the following interpolation to the unperturbed image row at such an exponent c c c , with S c 0 ≤ C J ρ ∙ S_c^0\le C_J\rho_\bullet S c 0 ≤ C J ρ ∙ . For nonnegative coefficient weights b η b_\eta b η and network lengths l η l_\eta l η , Hölder’s inequality for finite sums gives
∑ η b η e ϑ c l η ≤ ( ∑ η b η ) 1 − ϑ ( ∑ η b η e c l η ) ϑ . \sum_\eta b_\eta e^{\vartheta c l_\eta}\le\left(\sum_\eta b_\eta\right)^{1-\vartheta}\left(\sum_\eta b_\eta e^{c l_\eta}\right)^\vartheta. η ∑ b η e ϑ c l η ≤ ( η ∑ b η ) 1 − ϑ ( η ∑ b η e c l η ) ϑ .
Passing to increasing finite sums proves the stated interpolation. Its right side, divided by ρ ∙ \rho_\bullet ρ ∙ , tends to 3 / 8 3/8 3/8 as ϑ ↓ 0 \vartheta\downarrow0 ϑ ↓ 0 . Hence one positive ϑ \vartheta ϑ works for both radii and the finitely many stages of the unperturbed construction. Its exponent is fixed before the regional rates are chosen. At allocated working rates below this exponent and satisfying Theorem 19.9 , that theorem bounds the coefficient row of the perturbed-minus-unperturbed reconstruction by C P , σ ∗ ϱ c l C_{P,\sigma^{\ast}}\varrho_{\mathrm{cl}} C P , σ ∗ ϱ cl on the same enclosing products. Fixing this constant before the final choice of σ ∗ \sigma_{\ast} σ ∗ and making that bound smaller than the remaining half-radius margin gives the asserted perturbed image bound. No interpolation at the unperturbed exponent is applied to inherited perturbative networks.
Step 5: curvature. For the curvature bound, we fix one derivative at the scalar coordinate j j j and sum over the second differentiated index. Differentiating a homogeneous monomial of degree k k k twice gives the factor k − 1 k-1 k − 1 relative to the norm ∥ ⋅ ∥ a ′ , 1 \|\cdot\|_{a',1} ∥ ⋅ ∥ a ′ , 1 . The matrix weight is bounded by the weight of the original network, since the distance between the two differentiated indices is at most the network length. At the radius ϱ = R e n \varrho=R_{\mathrm{en}} ϱ = R en in the variables ζ \zeta ζ , the two contributions are bounded by
∑ k ≥ 3 ( k − 1 ) A k ϱ k − 2 = A 2 [ ( 1 − A ϱ ) − 2 − 1 ] ≤ 2 A 3 ϱ ( 1 − A ϱ ) 2 , \sum_{k\geq3}(k-1)A^k\varrho^{k-2}
= A^2\left[(1-A\varrho)^{-2}-1\right]
\leq\frac{2A^3\varrho}{(1-A\varrho)^2}, k ≥ 3 ∑ ( k − 1 ) A k ϱ k − 2 = A 2 [ ( 1 − A ϱ ) − 2 − 1 ] ≤ ( 1 − A ϱ ) 2 2 A 3 ϱ ,
t 2 L V ∑ k ≥ 2 ( k − 1 ) A k ϱ k − 2 = A 2 t 2 L V ( 1 − A ϱ ) 2 . t^2\mathcal{L}_V\sum_{k\geq2}(k-1)A^k\varrho^{k-2}
= \frac{A^2t^2\mathcal{L}_V}{(1-A\varrho)^2}. t 2 L V k ≥ 2 ∑ ( k − 1 ) A k ϱ k − 2 = ( 1 − A ϱ ) 2 A 2 t 2 L V .
Here we used that on K e n \mathcal{K}_{\mathrm{en}} K en the coordinates satisfy ∣ w j ∣ ≤ R e n / t |w_j|\leq R_{\mathrm{en}}/t ∣ w j ∣ ≤ R en / t , so that t k − 2 ( R e n / t ) k − 2 = R e n k − 2 t^{k-2}(R_{\mathrm{en}}/t)^{k-2}=R_{\mathrm{en}}^{k-2} t k − 2 ( R en / t ) k − 2 = R en k − 2 for the classical terms and t k ( R e n / t ) k − 2 = t 2 R e n k − 2 t^k(R_{\mathrm{en}}/t)^{k-2}=t^2R_{\mathrm{en}}^{k-2} t k ( R en / t ) k − 2 = t 2 R en k − 2 for the others. We enlarge C 0 ≥ 2 A 3 C_0\geq2A^3 C 0 ≥ 2 A 3 to include the fixed conversion between scalar and three-component block norms. The column bound follows from the symmetry of the Hessian. The weaker denominator ( 1 − A R e n ) − 3 (1-AR_{\mathrm{en}})^{-3} ( 1 − A R en ) − 3 in (78) applies to both contributions. Since A R e n ≤ 1 / 4 AR_{\mathrm{en}}\leq1/4 A R en ≤ 1/4 , C 0 R e n ≤ 1 / 16 C_0R_{\mathrm{en}}\leq1/16 C 0 R en ≤ 1/16 , and C 0 t 2 L V ≤ 1 / 16 C_0t^2\mathcal{L}_V\leq1/16 C 0 t 2 L V ≤ 1/16 , the result is
1 / 16 + 1 / 16 ( 1 − 1 / 4 ) 3 = 8 27 < 1 2 . \frac{1/16+1/16}{(1-1/4)^3}=\frac{8}{27}<\frac{1}{2}. ( 1 − 1/4 ) 3 1/16 + 1/16 = 27 8 < 2 1 .
All inequalities have strict margins in the charts, so they hold on a neighborhood of the real product, including its faces and corners. The constant C 1 C_1 C 1 of Lemma 20.5 is determined from A A A , a ′ a' a ′ , the label density, and the curvature bound just proved, so the second inequality in (77) is a separate requirement compatible with the present choices.
Step 6: logarithmic cutoffs. A ball of radius p i p_i p i lies in K e n \mathcal{K}_{\mathrm{en}} K en whenever t p i < R e n tp_i<R_{\mathrm{en}} t p i < R en . The finite logarithmic offset in the statement bounds every such p i p_i p i , since p ( g i ) = p ( t ) + A w log ( 1 / λ i ) p(g_i)=p(t)+A_{\mathrm{w}}\log(1/\lambda_i) p ( g i ) = p ( t ) + A w log ( 1/ λ i ) and λ i ≥ λ ∗ \lambda_i\geq\lambda_{\ast} λ i ≥ λ ∗ . This proves the inclusion both for the common cutoff and for the cutoffs at the individual scales. All constants were computed at σ 0 \sigma_0 σ 0 on fixed larger products, before any final enclosure quantity or width. The proof involves no derivative in the frozen values, no sum over earlier renormalization group steps, and no dependence on the individual radii of the fluctuation balls.
Remark 20.4. The incoming boundary function f f f in Lemma 20.3 is measured in the norm X ρ \mathcal{X}^{\rho} X ρ with the uniform profile ρ \rho ρ , and (20.29) is stated for this case. For a function measured with a nonuniform profile ϕ \phi ϕ , the analogous estimate holds for the rescaled function R ~ ( x ) = R ( ( ϕ i x i ) i ) \widetilde{R}(x)=R((\phi_i x_i)_i) R ( x ) = R (( ϕ i x i ) i ) : if its degree-wise rooted sum at unit profile is at most H ρ 2 H\rho^2 H ρ 2 , then at a point whose image under the absolute reconstruction majorant is bounded by κ ϕ \kappa\phi κ ϕ , 0 ≤ κ < 1 0\leq\kappa<1 0 ≤ κ < 1 ,
∣ D j R ~ ∣ κ , r o o t ≤ H ρ 2 ∑ l ≥ j ( l j ) κ l − j = H ρ 2 j ! ( 1 − κ ) j + 1 , j ≥ 1. \left|D^j\widetilde{R}\right|_{\kappa,\mathrm{root}}\leq H\rho^2\sum_{l\geq j}\binom{l}{j}\kappa^{l-j}=H\rho^2\frac{j!}{(1-\kappa)^{j+1}},\qquad j\geq1. D j R κ , root ≤ H ρ 2 l ≥ j ∑ ( j l ) κ l − j = H ρ 2 ( 1 − κ ) j + 1 j ! , j ≥ 1.
Thus each differentiated slot carries its own radius ϕ i \phi_i ϕ i . The unscaled form (20.29) does not follow for a spatially varying radius profile. Indeed, fix an unfixed index i i i at distance d i d_i d i from Ω c \Omega^c Ω c . The function R ( v ) = H ρ 2 e − α b d i v i / ϕ i R(v)=H\rho^2 e^{-\alpha b d_i}v_i/\phi_i R ( v ) = H ρ 2 e − α b d i v i / ϕ i has norm H H H and derivative H ρ 2 e − α b d i / ϕ i H\rho^2e^{-\alpha b d_i}/\phi_i H ρ 2 e − α b d i / ϕ i , which is unbounded as the boundary radius ϕ i = ρ ∂ \phi_i=\rho_{\partial} ϕ i = ρ ∂ decreases with d i d_i d i fixed. In the coefficient sums of Section 20.4 every differentiated background index is multiplied by its own radius, so no inverse power of the small boundary radius ρ ∂ \rho_{\partial} ρ ∂ enters.
Lemma 20.5. The following statements hold.
Let K \mathcal{K} K be a finite product of three-dimensional balls, let S , W ⊂ K \mathcal{S},\mathcal{W}\subset\mathcal{K} S , W ⊂ K be Borel sets, let χ : K → [ 0 , 1 ] \chi:\mathcal{K}\to[0,1] χ : K → [ 0 , 1 ] be Borel, let e − V e^{-V} e − V be a positive Borel function with 0 < Z K = ∫ K e − V d w < ∞ 0<Z_{\mathcal{K}}=\int_{\mathcal{K}}e^{-V}\,dw<\infty 0 < Z K = ∫ K e − V d w < ∞ , and let P > 0 P>0 P > 0 be a constant prefactor (unrelated to the perturbation P P P of Theorem 20.2). Set
G = P ∫ S χ e − V d w , F ~ = P ∫ W e − V d w , d ν K = Z K − 1 χ e − V d w . G=P\int_{\mathcal{S}}\chi e^{-V}\,dw,\qquad\widetilde{F}=P\int_{\mathcal{W}}e^{-V}\,dw,\qquad d\nu_{\mathcal{K}}=Z_{\mathcal{K}}^{-1}\chi e^{-V}\,dw. G = P ∫ S χ e − V d w , F = P ∫ W e − V d w , d ν K = Z K − 1 χ e − V d w . If ν K ( W c ) ≤ δ < 1 \nu_{\mathcal{K}}(\mathcal{W}^{c})\leq\delta<1 ν K ( W c ) ≤ δ < 1 , then G ≤ ( 1 − δ ) − 1 F ~ G\leq(1-\delta)^{-1}\widetilde{F} G ≤ ( 1 − δ ) − 1 F . Nonnegative functions of the coarse background and external factors may multiply both sides. No inclusion between W \mathcal{W} W and S \mathcal{S} S is needed. A product with no factors has mass one.
Let K = K e n \mathcal{K}=\mathcal{K}_{\mathrm{en}} K = K en be the enclosing product of Lemma 20.3, under its hypotheses with t max ≤ t ∗ t_{\max}\leq t_* t m a x ≤ t ∗ , at an admissible real coarse background and fixed real frozen values, and let V V V be the potential (20.23). There is a constant C 1 C_1 C 1 , depending only on A A A , a ′ a' a ′ , and v J v_J v J , such that
D w 2 V ≥ I / 2 , sup j ∣ E ν K w j ∣ ≤ C 1 t max ( L V + 2 ) . D_w^2V\geq I/2,\qquad\sup_j\left|\mathbb{E}_{\nu_{\mathcal{K}}}w_j\right|\leq C_1t_{\max}(\mathcal{L}_V+2). D w 2 V ≥ I /2 , j sup ∣ E ν K w j ∣ ≤ C 1 t m a x ( L V + 2 ) . where j j j runs over the scalar coordinates. For every scalar function F F F that is 1-Lipschitz for the Euclidean norm and every $v\geq 0,
ν K ( ∣ F − E F ∣ ≥ v ) ≤ 2 e − v 2 / 4 . \nu_{\mathcal{K}}(|F-\mathbb{E}F|\geq v)\leq2e^{-v^2/4}. ν K ( ∣ F − E F ∣ ≥ v ) ≤ 2 e − v 2 /4 . Let p ( t ) = A w log ( 1 / t ) p(t)=A_w\log(1/t) p ( t ) = A w log ( 1/ t ) , and choose radii p i ≥ p ( t max ) p_i\geq p(t_{\max}) p i ≥ p ( t m a x ) with t max max i p i < R e n t_{\max}\max_i p_i<R_{\mathrm{en}} t m a x max i p i < R en , so that the product W = ∏ i , e { ∣ w i , e ∣ ≤ p i } \mathcal{W}=\prod_{i,e}\{|w_{i,e}|\leq p_i\} W = ∏ i , e { ∣ w i , e ∣ ≤ p i } is contained in K \mathcal{K} K . If C 1 t max ( L V + 2 ) ≤ p ( t max ) / ( 2 3 ) C_1t_{\max}(\mathcal{L}_V+2)\leq p(t_{\max})/(2\sqrt{3}) C 1 t m a x ( L V + 2 ) ≤ p ( t m a x ) / ( 2 3 ) , which holds for all sufficiently small t max t_{\max} t m a x and is one of the requirements on the coupling bound g ∗ g_* g ∗ in Theorem 20.14, then
ν K ( W c ) ≤ 6 b exp { − p ( t max ) 2 / 48 } , b ≤ C J m 3 , \nu_{\mathcal{K}}(\mathcal{W}^{c})\leq6b\exp\{-p(t_{\max})^2/48\},\qquad b\leq C_Jm^3, ν K ( W c ) ≤ 6 b exp { − p ( t m a x ) 2 /48 } , b ≤ C J m 3 ,
where b b b counts all fluctuation blocks of the step, jointly over all components of Ω \Omega Ω and all scales. These are means under ν K \nu_K ν K at fixed frozen values; no symmetry of the frozen values is assumed. The constants have the dependence described in item (3) of Definition 4.14 and are independent of the later widths.
Consider the stopping sides m t = m B t m_t = mB^t m t = m B t , t ≥ 0 t \ge0 t ≥ 0 , together with at most one additional side η m B j \eta mB^j η m B j with j ∈ N 0 j \in\mathbb{N}_0 j ∈ N 0 , where η \eta η ranges over a fixed finite set of positive ratios (arising from a final composition of fewer than J J J averaging steps). Suppose that nonnegative errors at these sides satisfy
δ t ≤ C J m t 3 e − a w log 2 M m t / γ , z t ≤ C J m t 3 e − b m t / γ , a w , b > 0 , \delta_t \le C_Jm_t^3 e^{-a_w \log^2 \sqrt{Mm_t/\gamma}}, \qquad z_t \le C_Jm_t^3 e^{-bm_t/\gamma}, \qquad a_w,b>0, δ t ≤ C J m t 3 e − a w l o g 2 M m t / γ , z t ≤ C J m t 3 e − b m t / γ , a w , b > 0 ,
and similarly at the additional side. By part (2), the first bound holds with a w = A W 2 / 48 a_w=A^2_W/48 a w = A W 2 /48 for the comparison errors of this lemma. Then for every fixed Γ ∈ ( 0 , ∞ ) \Gamma\in(0,\infty) Γ ∈ ( 0 , ∞ ) , in particular for every Γ = γ 0 ≤ 1 \Gamma=\gamma_0\le1 Γ = γ 0 ≤ 1 , there is a minimum side m 0 m_0 m 0 , depending only on Γ \Gamma Γ , the fixed data, and the finite set of ratios, such that, uniformly for m ≥ m 0 m\ge m_0 m ≥ m 0 and 0 < γ ≤ Γ 0<\gamma\le\Gamma 0 < γ ≤ Γ , all errors are at most 1 / 2 1/2 1/2 and
∑ t ( δ t + z t ) < ∞ , log ∏ t ( 1 − δ t ) − 1 ≤ 2 ∑ t δ t ≤ C δ . \sum_t(\delta_t+z_t)<\infty,\qquad\log\prod_t(1-\delta_t)^{-1}\le2\sum_t\delta_t\le C_\delta. t ∑ ( δ t + z t ) < ∞ , log t ∏ ( 1 − δ t ) − 1 ≤ 2 t ∑ δ t ≤ C δ .
The sums and the logarithm tend to zero as m → ∞ m\to\infty m → ∞ , uniformly in 0 < γ ≤ Γ 0<\gamma\le\Gamma 0 < γ ≤ Γ and in the finite or infinite number of sides. The same conclusion holds for ϵ t = δ t / ( 1 − z t ) \epsilon_t=\delta_t/(1-z_t) ϵ t = δ t / ( 1 − z t ) after increasing m 0 m_0 m 0 . There is one comparison for the whole product at each stopping side, not a separate factor for each component of Ω c \Omega^c Ω c . The minimum side and C δ C_\delta C δ have the dependence described in item (4) of Definition 4.14; none of them depends on the number of steps or on the large-field decomposition.
Proof. Step 1 proves part (1). Steps 2 to 6 prove part (2): we first check that the measure is well defined, then interpolate between the Gaussian product and the potential, bound the force of the interaction using fourth moments, integrate the derivative of the means along the interpolation, and finally bound the probability outside the chosen product. Step 7 reduces part (3) to Lemma 20.15.
Step 1: the abstract comparison. The Borel hypotheses make both integrands measurable, and since 0 ≤ χ ≤ 1 0\le\chi\le1 0 ≤ χ ≤ 1 and 0 < Z K < ∞ 0<Z_K<\infty 0 < Z K < ∞ , we have 0 ≤ G , F ~ ≤ P Z K < ∞ 0\le G,\widetilde F\le PZ_K<\infty 0 ≤ G , F ≤ P Z K < ∞ . Since S ⊂ K S\subset K S ⊂ K , we have G ≤ P Z K G\le PZ_K G ≤ P Z K . On the other hand
F ~ = P Z K ν K ( W ) ≥ P Z K ( 1 − δ ) . \widetilde F=PZ_K\nu_K(W)\ge PZ_K(1-\delta). F = P Z K ν K ( W ) ≥ P Z K ( 1 − δ ) .
Dividing by the positive finite number P Z K PZ_K P Z K gives part (1); the common prefactor cancels before any estimate. Multiplying by nonnegative functions of the coarse background preserves the inequality pointwise. For a product with no factors we use the unit point mass, and the same proof applies.
Step 2: normalizability on the enclosing product. The potential of Lemma 20.3 is real and smooth on a neighborhood of the fixed compact product K e n \mathcal{K}_{\mathrm{en}} K en . Its ball radii R e n / t max R_{\mathrm{en}}/t_{\max} R en / t m a x are positive and finite, so the product has positive finite Lebesgue volume. If V min V_{\min} V m i n and V max V_{\max} V m a x are the minimum and maximum of V V V on K e n \mathcal{K}_{\mathrm{en}} K en , then
0 < e − V max vol ( K e n ) ≤ Z K e n ≤ e − V min vol ( K e n ) < ∞ . 0 < e^{-V_{\max}}\operatorname{vol}(\mathcal{K}_{\mathrm{en}}) \le Z_{\mathcal{K}_{\mathrm{en}}} \le e^{-V_{\min}}\operatorname{vol}(\mathcal{K}_{\mathrm{en}}) < \infty. 0 < e − V m a x vol ( K en ) ≤ Z K en ≤ e − V m i n vol ( K en ) < ∞.
Thus the integrand is positive, Borel, and normalizable. The chosen product W \mathcal{W} W is closed, hence Borel; the comparison in part (1) keeps its Borel hypotheses on S S S and χ \chi χ .
Step 3: convexity and concentration. We write t = t max t=t_{\max} t = t m a x and H i n t ( w ) = V ( w ) − ∣ w ∣ 2 / 2 H_{\mathrm{int}}(w)=V(w)-|w|^2/2 H int ( w ) = V ( w ) − ∣ w ∣ 2 /2 , and interpolate on the same product
V θ ( w ) = ∣ w ∣ 2 / 2 + θ H i n t ( w ) , 0 ≤ θ ≤ 1. V_\theta(w)=|w|^2/2+\theta H_{\mathrm{int}}(w), \qquad0\le\theta\le1. V θ ( w ) = ∣ w ∣ 2 /2 + θ H int ( w ) , 0 ≤ θ ≤ 1.
Each V θ V_\theta V θ is real and smooth on the compact product, so the normalization bounds of Step 2 hold throughout the interpolation. By (78) the Hessian perturbation is at most κ = 8 / 27 \kappa=8/27 κ = 8/27 , so every V θ V_\theta V θ has Hessian at least ( 1 − κ ) I ≥ I / 2 (1-\kappa)I\ge I/2 ( 1 − κ ) I ≥ I /2 ; for θ = 1 \theta=1 θ = 1 this is the first assertion. The absolute coefficient bound in (78) bounds the exponentially weighted sums over y y y of the suprema over the product of ∣ ( V θ ) x y ∣ |(V_\theta)_{xy}| ∣ ( V θ ) x y ∣ , in rows and in columns, by the same fixed constant; the three-component block form follows by including the fixed block conversion in C 0 C_0 C 0 . Fix 0 < b c o v ≤ a ′ 0<b_{\mathrm{cov}}\le a' 0 < b cov ≤ a ′ so small that
max { sup x ∑ y ≠ x ( e b c o v dist ( x , y ) − 1 ) sup ∥ ( V θ ) x y ∥ , sup y ∑ x ≠ y ( e b c o v dist ( x , y ) − 1 ) sup ∥ ( V θ ) x y ∥ } ≤ 1 4 . \max\left\{\sup_x\sum_{y\ne x}\left(e^{b_{\mathrm{cov}}\operatorname{dist}(x,y)}-1\right)\sup\|(V_\theta)_{xy}\|,\ \sup_y\sum_{x\ne y}\left(e^{b_{\mathrm{cov}}\operatorname{dist}(x,y)}-1\right)\sup\|(V_\theta)_{xy}\|\right\}\le\frac14. max ⎩ ⎨ ⎧ x sup y = x ∑ ( e b cov dist ( x , y ) − 1 ) sup ∥ ( V θ ) x y ∥ , y sup x = y ∑ ( e b cov dist ( x , y ) − 1 ) sup ∥ ( V θ ) x y ∥ ⎭ ⎬ ⎫ ≤ 4 1 .
Such a choice depends only on the fixed weighted bound, since the convexity of u ↦ e u u\mapsto e^u u ↦ e u gives e b c o v d − 1 ≤ ( b c o v / a ′ ) ( e a ′ d − 1 ) e^{b_{\mathrm{cov}}d}-1\le(b_{\mathrm{cov}}/a')(e^{a'd}-1) e b cov d − 1 ≤ ( b cov / a ′ ) ( e a ′ d − 1 ) for b c o v ≤ a ′ b_{\mathrm{cov}}\le a' b cov ≤ a ′ . The off-diagonal blocks of the Gaussian part vanish. Thus all hypotheses of Theorem 5.4, which adapts the Bakry–Émery and Helffer–Sjöstrand methods [8, 110] to products of balls with reflecting boundary, hold with convexity constant 1 / 2 1/2 1/2 , on a neighborhood of the finite product; its ball faces and product corners are included in that theorem. It yields the scalar tail 2 e − v 2 / 4 2e^{-v^2/4} 2 e − v 2 /4 and the covariance bound
∣ Cov θ ( F , G ) ∣ ≤ 4 ∑ x , y e − b c o v dist ( x , y ) ∥ ∇ x F ∥ 2 ∥ ∇ y G ∥ 2 (81) \left|\operatorname{Cov}_{\theta}(F,G)\right|\le4\sum_{x,y}e^{-b_{\mathrm{cov}}\operatorname{dist}(x,y)}\|\nabla_xF\|_2\|\nabla_yG\|_2 \tag*{(81)} ∣ Cov θ ( F , G ) ∣ ≤ 4 x , y ∑ e − b cov dist ( x , y ) ∥ ∇ x F ∥ 2 ∥ ∇ y G ∥ 2 (81)
where Cov θ \operatorname{Cov}_{\theta} Cov θ and ∥ ⋅ ∥ 2 \|\cdot\|_2 ∥ ⋅ ∥ 2 refer to the normalized measure with density proportional to e − V θ e^{-V_\theta} e − V θ on K e n \mathcal{K}_{\mathrm{en}} K en . These constants do not depend on the number or the radii of the balls.
Step 4: the force. Let M ( θ ) = max j ∣ E θ w j ∣ M(\theta)=\max_j|\mathbb{E}_\theta w_j| M ( θ ) = max j ∣ E θ w j ∣ , where j j j is scalar (in this proof M ( θ ) M(\theta) M ( θ ) is unrelated to the blocking factor M M M ). The tail just obtained implies a universal bound on the fourth moment of w j − E θ w j w_j-\mathbb{E}_\theta w_j w j − E θ w j , by integrating ∫ 0 ∞ 4 u 3 P ( ∣ w j − E θ w j ∣ > u ) d u \int_0^\infty4u^3\mathbb{P}(|w_j-\mathbb{E}_\theta w_j|>u)\,du ∫ 0 ∞ 4 u 3 P ( ∣ w j − E θ w j ∣ > u ) d u against the Gaussian majorant. Hence
∥ w j ∥ 4 ≤ C 4 + M ( θ ) . \|w_j\|_4\le C_4+M(\theta). ∥ w j ∥ 4 ≤ C 4 + M ( θ ) .
Consider a differentiated classical monomial of degree k ≥ 3 k \ge3 k ≥ 3 . After fixing its differentiated coordinate, we place two of its remaining k − 1 k-1 k − 1 factors in L 4 L^4 L 4 and the other k − 3 k-3 k − 3 factors in the supremum norm R e n / t R_{\mathrm{en}}/t R en / t of the product. Hölder’s inequality, which applies also when the two factors are the same scalar coordinate, gives the factor
t k − 2 ( R e n / t ) k − 3 ( C 4 + M ( θ ) ) 2 = t R e n k − 3 ( C 4 + M ( θ ) ) 2 . t^{k-2}\left(R_{\mathrm{en}}/t\right)^{k-3}\left(C_4+M(\theta)\right)^2=tR_{\mathrm{en}}^{k-3}\left(C_4+M(\theta)\right)^2. t k − 2 ( R en / t ) k − 3 ( C 4 + M ( θ ) ) 2 = t R en k − 3 ( C 4 + M ( θ ) ) 2 .
The coefficient sum with the differentiated index fixed is bounded by A k A^k A k , and the geometric sum is at most C ( A ) t ( 1 + M ( θ ) 2 ) C(A)t(1+M(\theta)^2) C ( A ) t ( 1 + M ( θ ) 2 ) because A R e n ≤ 1 / 4 AR_{\mathrm{en}}\le1/4 A R en ≤ 1/4 . For the quantum terms, including degree one, a direct supremum bound gives
∑ k ≥ 1 L V A k t k ( R e n / t ) k − 1 = L V A t 1 − A R e n . \sum_{k\ge1}\mathcal{L}_{V}A^kt^k\left(R_{\mathrm{en}}/t\right)^{k-1}=\frac{\mathcal{L}_{V}At}{1-AR_{\mathrm{en}}}. k ≥ 1 ∑ L V A k t k ( R en / t ) k − 1 = 1 − A R en L V A t .
Combining the two bounds, we obtain
sup j ∥ ∂ j H i n t ∥ 2 ≤ C ( A ) t [ L V + 1 + M ( θ ) 2 ] . (82) \sup_j\left\|\partial_jH_{\mathrm{int}}\right\|_2\le C(A)t\left[\mathcal{L}_{V}+1+M(\theta)^2\right]. \tag*{(82)} j sup ∥ ∂ j H int ∥ 2 ≤ C ( A ) t [ L V + 1 + M ( θ ) 2 ] . (82)
Step 5: the means. Differentiating the finite integrals on the fixed compact product, we obtain
d d θ E θ w j = − Cov θ ( w j , H i n t ) . \frac{d}{d\theta}\mathbb{E}_{\theta}w_j=-\operatorname{Cov}_{\theta}(w_j,H_{\mathrm{int}}). d θ d E θ w j = − Cov θ ( w j , H int ) .
Indeed, the derivative of the numerator is − ∫ w j H i n t e − V θ -\int w_jH_{\mathrm{int}}e^{-V_\theta} − ∫ w j H int e − V θ and that of the denominator is − ∫ H i n t e − V θ -\int H_{\mathrm{int}}e^{-V_\theta} − ∫ H int e − V θ , and their quotient is the displayed centered expression. The integrands and their derivatives are bounded on this finite product, so the differentiation is justified directly. This is the calculation of Theorem 13.10 on a fixed product, carried out here for the present measure; we do not use the conclusions of that theorem, whose hypotheses concern the torus products. Combining (81) , (82) , and (53) , we obtain a fixed C 1 C_1 C 1 with
∣ d d θ E θ w j ∣ ≤ C 1 t [ L V + 1 + M ( θ ) 2 ] . (83) \left|\frac{d}{d\theta}\mathbb{E}_{\theta}w_j\right|\le C_1t\left[\mathcal{L}_{V}+1+M(\theta)^2\right]. \tag*{(83)} d θ d E θ w j ≤ C 1 t [ L V + 1 + M ( θ ) 2 ] . (83)
It depends only on A A A , a ′ a' a ′ , and the label density; the margin κ < 1 / 2 \kappa<1/2 κ < 1/2 has already been imposed. This is the constant in the hypothesis (77) of Lemma 20.3 .
At θ = 0 \theta=0 θ = 0 the measure is a product of centered truncated Gaussian measures on balls, so every coordinate has mean zero, for every frozen value. All means are continuous in θ \theta θ . Suppose that M M M reached the value one, and let θ 1 \theta_1 θ 1 be the first time at which it does. Integrating (83) for each coordinate on [ 0 , θ 1 ] [0,\theta_1] [ 0 , θ 1 ] , where M ≤ 1 M\le1 M ≤ 1 , would give M ( θ 1 ) ≤ C 1 t ( L V + 2 ) < 1 / ( 4 3 ) < 1 M(\theta_1)\le C_1t(\mathcal{L}_{V}+2)<1/(4\sqrt{3})<1 M ( θ 1 ) ≤ C 1 t ( L V + 2 ) < 1/ ( 4 3 ) < 1 , a contradiction. Inserting M < 1 M<1 M < 1 into (83) and integrating once more proves M ( 1 ) ≤ C 1 t ( L V + 2 ) M(1)\le C_1t(\mathcal{L}_{V}+2) M ( 1 ) ≤ C 1 t ( L V + 2 ) . This argument integrates the inequalities for the individual coordinates before taking the maximum; it does not use any differentiability of the maximum.
Step 6: the probability outside the chosen product. The explicit hypothesis t max i p i < R e n t\max_i p_i < R_{\mathrm{en}} t max i p i < R en puts the chosen product strictly inside K e n \mathcal{K}_{\mathrm{en}} K en . The later logarithmic gates (20.48) provide a sufficient choice of this inclusion in Theorem 20.14 ; the present local argument uses the inclusion itself. If a fluctuation block lies outside its ball of radius p i ≥ p ( t ) p_i \ge p(t) p i ≥ p ( t ) , then some scalar coordinate has absolute value larger than p ( t ) / 3 p(t)/\sqrt{3} p ( t ) / 3 . Since the mean of that coordinate is at most p ( t ) / ( 2 3 ) p(t)/(2\sqrt{3}) p ( t ) / ( 2 3 ) in absolute value, the coordinate deviates from its mean by at least p ( t ) / ( 2 3 ) p(t)/(2\sqrt{3}) p ( t ) / ( 2 3 ) . By the scalar tail and a union bound over the three coordinates, the probability that a given block lies outside its ball is at most 6 e − p ( t ) 2 / 48 6e^{-p(t)^2/48} 6 e − p ( t ) 2 /48 . The step has at most b ≤ C J m 3 b \le C_Jm^3 b ≤ C J m 3 blocks: each of the at most J J J scales has a fixed finite number of free labels per output block, and removing blocks from the region only decreases this number. This proves the bound on ν K ( W c ) \nu_{\mathcal{K}}(\mathcal{W}^c) ν K ( W c ) , with no factor depending on the number of components. The enclosure and curvature hypotheses are used on the whole product K e n \mathcal{K}_{\mathrm{en}} K en , which does not depend on the configuration.
Step 7: summation. Part (3) is the case C = C J C=C_J C = C J , c = b c=b c = b , a = a W a=a_W a = a W of Lemma 20.15 , with the ratio B B B and the finite set of additional ratios, since log 2 M m / γ = z 2 \log^2\sqrt{Mm/\gamma}=z^2 log 2 M m / γ = z 2 with z = ( 1 / 2 ) log ( M m / γ ) z=(1/2)\log(Mm/\gamma) z = ( 1/2 ) log ( M m / γ ) . With p ( t max ) = A W log M m / γ p(t_{\max})=A_W\log\sqrt{Mm/\gamma} p ( t m a x ) = A W log M m / γ , the bound of part (2) is of the stated form with a W = A W 2 / 48 a_W=A_W^2/48 a W = A W 2 /48 , since 6 b ≤ C J m 3 6b\le C_Jm^3 6 b ≤ C J m 3 .
Cluster expansion with boundary-localized coefficients
We now expand the logarithm of the fluctuation integral, keeping two structures apart. Probabilistic independence is determined by the set of fluctuation blocks on which a monomial depends: two monomials depending on disjoint sets of blocks are independent under a product reference measure, even if they depend on the same background coordinates. Spatial decay is carried instead by a connected network through all positions of the coefficient, background positions included, which may lie far from the fluctuation blocks. Overlapping monomials share a fluctuation block, so their networks meet there, and the union of their networks is connected without any additional connecting path.
The organization into connected polymer activities and a convergent logarithm follows the constructive cluster expansion; see Bałaban [18, 19] for its implementation in lattice gauge theories, Brydges [38] for a general exposition, and Kotecký and Preiss [120] for the abstract polymer criterion. The cancellation of connected graphs against rooted trees used in Theorem 12.6 is Penrose’s partition [136]; its construction by generations and parents, and the positive tree recursions used below, are described in [87, Sections 2 and 4.1]. The finite graph inequality is imported in the form of [68, Appendix B, (323), p. 42], through Theorem 12.6; no other result of that paper, which concerns a scalar field theory, is used. We prove directly the features needed here: the two kinds of roots, the exact fluctuation supports, and the visits to the fixed complement.
Complex coupling parameter. In this section we replace the coupling powers in (20.24) by powers of a complex parameter z z z : the classical powers t max q − 2 t_{\max}^{q-2} t m a x q − 2 become z q − 2 z^{q-2} z q − 2 and the other powers t max q t_{\max}^{q} t m a x q become z q z^{q} z q . The incoming functions, the couplings of earlier steps, the ratios λ i \lambda_i λ i , and the normalized symmetric reference measures on balls are kept fixed. The incoming insertion becomes
A R ( z ) = R ( Log C ( v , ( z λ i w i ) i ; b r e c ) ) − R ( Log C ( v , 0 ; b r e c ) ) . A_R(z)=R\bigl(\operatorname{Log} C(v,(z\lambda_iw_i)_i;b_{\mathrm{rec}})\bigr)-R\bigl(\operatorname{Log} C(v,0;b_{\mathrm{rec}})\bigr). A R ( z ) = R ( Log C ( v , ( z λ i w i ) i ; b rec ) ) − R ( Log C ( v , 0 ; b rec ) ) .
We call the interaction without insertion the sum of all terms of V − ∣ w ∣ 2 / 2 V-|w|^2/2 V − ∣ w ∣ 2 /2 in (20.23) except the incoming insertion, after this substitution. For p x > 0 p_x>0 p x > 0 attached to the fluctuation blocks, with p x = p i p_x=p_i p x = p i for a block at scale i i i , we write p = max i p i p=\max_i p_i p = max i p i , and we always assume p ≥ 1 p\geq1 p ≥ 1 .
Definition 20.6. Let F F F be an interaction, given with a resolution into scalar monomials
U λ = u λ v ν λ w η λ , ∣ η λ ∣ ≥ 1 , U_\lambda=u_\lambda v^{\nu_\lambda}w^{\eta_\lambda},\qquad|\eta_\lambda|\geq1, U λ = u λ v ν λ w η λ , ∣ η λ ∣ ≥ 1 , as in Lemma 20.1. A fluctuation index is a whole fluctuation block at its scale and position; a background index is an unfixed scalar background coordinate on the lattice where it was introduced. Write S λ S_\lambda S λ for the set of fluctuation blocks on which η λ \eta_\lambda η λ is positive (the fluctuation support of the monomial) and k λ = ∣ ν λ ∣ k_\lambda=|\nu_\lambda| k λ = ∣ ν λ ∣ for its background degree. Each monomial carries a connected network through all of its fluctuation and background positions; for a monomial whose coefficient has a visit to Ω L J c \Omega_{L_J}^{c} Ω L J c , the network contains this visit. We write ℓ λ \ell_\lambda ℓ λ for the total length of the network, including the visit. No enlargement of S λ S_\lambda S λ is made. Internal indices are summed, using a fixed part of the exponential decay rate, before the resolution into monomials, and repeated occurrences keep their own network refinements. For the exponent s 0 > 0 s_0>0 s 0 > 0 set
b λ ( F ) = ∣ u λ ( F ) ∣ ψ p r o f ν λ ∏ x p x ∣ η λ , x ∣ e s 0 ℓ λ , b_\lambda(F)=|u_\lambda(F)|\psi_{\mathrm{prof}}^{\nu_\lambda}\prod_x p_x^{|\eta_{\lambda,x}|}e^{s_0\ell_\lambda}, b λ ( F ) = ∣ u λ ( F ) ∣ ψ prof ν λ x ∏ p x ∣ η λ , x ∣ e s 0 ℓ λ , and for h ≥ 0 h\geq0 h ≥ 0
W h ( F ) = sup x ∑ λ : x ∈ S λ e h ∣ S λ ∣ b λ ( F ) , W_h(F)=\sup_x\sum_{\lambda:x\in S_\lambda}e^{h|S_\lambda|}b_\lambda(F), W h ( F ) = x sup λ : x ∈ S λ ∑ e h ∣ S λ ∣ b λ ( F ) , D h ( F ) = ∑ k ≥ 1 sup i ∑ λ : k λ = k ν λ , i e h ∣ S λ ∣ b λ ( F ) . D_h(F)=\sum_{k\geq1}\sup_i\sum_{\lambda:k_\lambda=k}\nu_{\lambda,i}e^{h|S_\lambda|}b_\lambda(F). D h ( F ) = k ≥ 1 ∑ i sup λ : k λ = k ∑ ν λ , i e h ∣ S λ ∣ b λ ( F ) . In W h W_h W h the root x x x is a fluctuation block, and in D h D_h D h the root i i i is a background index. In D h D_h D h the supremum over the root is taken before the sum over the background degree, as in the coefficient norm with boundary decay.
Lemma 20.7. Fix the admissible larger products and the preliminary perturbation ball of Theorem 20.2 and Lemma 20.3, including both strict margins for the images under the reconstruction
majorant and the inequalities on the conditional forces at every stage. Fix the ratio K ∗ K_\ast K ∗ and the profiles ϕ ^ = K ∗ ϕ \widehat{\phi}=K_\ast\phi ϕ = K ∗ ϕ and ψ p r o f = K ∗ ϕ \psi_{\mathrm{prof}}=\sqrt{K_\ast}\phi ψ prof = K ∗ ϕ . For a composition of s s s averaging steps put B s = M s B_s=M^s B s = M s , and write a i n , α i n a_{\mathrm{in}},\alpha_{\mathrm{in}} a in , α in for the incoming rates. Fix an allocation of the intrinsic coefficient exponent and impose the output-rate restrictions of Theorem 19.9 (4), with B s B_s B s in place of B B B : in particular, for the fixed geometric allocation constant require s 0 = a + α < min ( c J , B s a i n ) / C g e o m s_0=a+\alpha<\min(c_J,B_sa_{\mathrm{in}})/C_{\mathrm{geom}} s 0 = a + α < min ( c J , B s a in ) / C geom and α < B s α i n / C g e o m \alpha<B_s\alpha_{\mathrm{in}}/C_{\mathrm{geom}} α < B s α in / C geom . The output rates may be chosen smaller for a shorter composition; no return of the full-step rates is asserted for it. In the full-step iteration the incoming rates are a , α a,\alpha a , α , and the stronger allocation conditions of Theorem 20.9 apply. Consider the following families of interactions, resolved into monomials as in Definition 20.6 : the interaction without insertion, the incoming insertion, the differences between the regional families and the restricted torus families, and the mixed difference of the boundary change and the incoming insertion. Here X ρ ( f ) \mathcal{X}^{\rho}(f) X ρ ( f ) is evaluated at the incoming regional rates a i n , α i n a_{\mathrm{in}},\alpha_{\mathrm{in}} a in , α in , whereas E E E bounds the incoming torus interaction at its fixed torus rate; a i n a_{\mathrm{in}} a in is no larger than that available spatial rate after conversion to the graph metric. The resolved norms W h , D h W_h,D_h W h , D h use s 0 = a + α s_0=a+\alpha s 0 = a + α and network lengths in output units. Let H = E + X ρ ( f ) H=E+\mathcal{X}^{\rho}(f) H = E + X ρ ( f ) and set
s ( z ) = C J ∣ z ∣ ( p 3 + p + H p ) 1 − C J ∣ z ∣ p , \mathfrak{s}(z)=\frac{C_J|z|(p^3+p+Hp)}{1-C_J|z|p}, s ( z ) = 1 − C J ∣ z ∣ p C J ∣ z ∣ ( p 3 + p + H p ) ,
for z z z in the fixed larger radius with C J ∣ z ∣ p < 1 C_J|z|p<1 C J ∣ z ∣ p < 1 , inside both incoming image margins. Then, uniformly in the fixed real frozen values,
W 16 ( F ) + D 16 ( F ) ≤ ϵ F W_{16}(F)+D_{16}(F)\leq\epsilon_F W 16 ( F ) + D 16 ( F ) ≤ ϵ F
for each of these families F F F . One may take ϵ F ≤ C J s ( z ) \epsilon_F\leq C_J\mathfrak{s}(z) ϵ F ≤ C J s ( z ) for the family without insertion and for the boundary differences, and
ϵ A + ϵ ∂ A ≤ C J ∣ z ∣ H p 1 − C J ∣ z ∣ p (84) \epsilon_A+\epsilon_{\partial A}\leq\frac{C_J|z|Hp}{1-C_J|z|p} \tag*{(84)} ϵ A + ϵ ∂ A ≤ 1 − C J ∣ z ∣ p C J ∣ z ∣ H p (84)
for the incoming insertion A A A and its boundary difference ∂ A \partial A ∂ A . In particular the latter bound holds with a visit to Ω c \Omega^c Ω c and with an arbitrary background root. The constants have the dependence described in item (3) of Definition 4.14 and are independent of the later widths; the profiles and the input radii are not decreased at later steps.
Proof. Step 1 records the coefficient resolution of the terms of the potential and its basic bounds. Steps 2 and 3 write out, with every contraction, the resolution in the two typical cases of a classical monomial and of a monomial of the Haar density. Step 4 treats the incoming insertion and the boundary differences. Step 5 converts the coefficient bounds into the bounds on W 16 W_{16} W 16 and D 16 D_{16} D 16 , keeping the order of the supremum and the degree sum.
Step 1: the resolution. The tensors of Theorem 20.2 , with bounds at every external index, and the coefficient version of Lemma 20.3 are available on the outer profile, before the background is evaluated. By Lemma 20.1 , every term of the potential has an exact resolution into monomials carrying connected networks. Rooted at an arbitrary free index, this resolution has convergent majorants in the free degree q q q , namely
C q ∣ z ∣ q − 2 p q ( q ≥ 3 ) and C q ∣ z ∣ p q ( q ≥ 1 ) , C^{q}|z|^{q-2}p^{q} \qquad(q \ge3) \qquad\text{and}\qquad C^{q}|z|p^{q} \qquad(q \ge1), C q ∣ z ∣ q − 2 p q ( q ≥ 3 ) and C q ∣ z ∣ p q ( q ≥ 1 ) ,
for the classical terms and for the other terms, respectively. For the incoming insertion the second majorant is multiplied by H H H . The ratios λ j \lambda_j λ j have modulus at most one, and J J J is fixed. For the last assertion the input is (20.29) with E + σ 0 E+\sigma_0 E + σ 0 replaced by H H H : fixing its first differentiated input and summing over the other inputs gives
∥ D j R ∥ r o o t ≤ H ρ 2 − j j ! ( 1 − κ ) j + 1 , κ ≤ 1 2 , \lVert D^jR\rVert_{\mathrm{root}} \le H\rho^{2-j}\frac{j!}{(1-\kappa)^{j+1}},\qquad\kappa\le\frac{1}{2}, ∥ D j R ∥ root ≤ H ρ 2 − j ( 1 − κ ) j + 1 j ! , κ ≤ 2 1 ,
by applying the factor ( l ) j κ l − j (l)_{j}\kappa^{l-j} ( l ) j κ l − j to the input norm at degree l l l with that input fixed and summing the geometric series. The constants of R R R disappear in the fluctuation difference; they are not bounded by H H H . The composition with every index rooted in Theorem 19.9, with its volume-normalized input columns, keeps a specified external background index as well as a specified free index. At a background root i i i , the sum of the monomials containing that index is bounded by ϕ ^ ( i ) \hat{\phi}(i) ϕ ^ ( i ) times the absolute coefficient majorant of ∂ i F \partial_iF ∂ i F . This produces the factor ν λ , i \nu_{\lambda,i} ν λ , i in D h D_h D h . No division by the smaller boundary radius occurs.
Step 2: a classical monomial. Fix a scale i i i . By the definition of the partial minimization, F i ( v , c i + 1 , y ) F_i(v,c_{i+1},y) F i ( v , c i + 1 , y ) is the perturbed objective Q 1 Q_1 Q 1 of the proof of Theorem 20.2, in the free logarithmic coordinates and multiplied by the factor M i M^i M i , evaluated with the coordinates of the finer scales at their conditional minimizers, at fixed frozen values. We write this as F i ( v , c i + 1 , y ) = Q ( i ) ( Ψ i ( v , c i + 1 , y ) ) F_i(v,c_{i+1},y)=Q^{(i)}(\Psi_i(v,c_{i+1},y)) F i ( v , c i + 1 , y ) = Q ( i ) ( Ψ i ( v , c i + 1 , y )) , where Q ( i ) = M i Q 1 Q^{(i)}=M^iQ_1 Q ( i ) = M i Q 1 , y y y denotes the free logarithmic coordinates of scale i i i , c i + 1 = C i + 1 ( v , ζ > i ) c_{i+1}=C_{i+1}(v,\zeta_{>i}) c i + 1 = C i + 1 ( v , ζ > i ) is the coarser reconstruction, and Ψ i \Psi_i Ψ i maps ( v , c i + 1 , y ) (v,c_{i+1},y) ( v , c i + 1 , y ) to the full list of coordinates ( v , z < i ∗ , y , z > i ) (v,z^*_{<i},y,z_{>i}) ( v , z < i ∗ , y , z > i ) , with the finer coordinates z < i ∗ z^*_{<i} z < i ∗ at their conditional minimizers. The coefficient bounds for Q ( i ) Q^{(i)} Q ( i ) on the common product are those of Theorem 19.9, and those for Ψ i \Psi_i Ψ i , which contains the conditional minimizers of the finer stages, are those of Theorem 20.2. The factorization only displays the chain rule; the general term can equally be bounded through the tensors D y r F i D_y^rF_i D y r F i of that theorem. At fixed ( v , c i + 1 ) (v,c_{i+1}) ( v , c i + 1 ) the minimizer z i ∗ z_i^* z i ∗ satisfies F i , y ( v , c i + 1 , z i ∗ ) = 0 F_{i,y}(v,c_{i+1},z_i^*)=0 F i , y ( v , c i + 1 , z i ∗ ) = 0 , and K i = F i , y y ( v , c i + 1 , z i ∗ ) K_i=F_{i,yy}(v,c_{i+1},z_i^*) K i = F i , y y ( v , c i + 1 , z i ∗ ) , S i = K i − 1 / 2 S_i=K_i^{-1/2} S i = K i − 1/2 , z i = z i ∗ + g i S i w i z_i=z_i^*+g_iS_iw_i z i = z i ∗ + g i S i w i . Since S i T K i S i = I S_i^{\mathsf T}K_iS_i=I S i T K i S i = I , Taylor’s formula in y y y on the common coefficient product gives
F i ( v , c i + 1 , z i ) − F i ( v , c i + 1 , z i ∗ ) g i 2 − ∣ w i ∣ 2 2 = ∑ r ≥ 3 g i r − 2 r ! ∑ a 1 , … , a r ∑ β 1 , … , β r F i , a 1 ⋯ a r ∗ ∏ l = 1 r S i ( a l , β l ) w i , β l . \frac{F_i(v,c_{i+1},z_i)-F_i(v,c_{i+1},z_i^*)}{g_i^2}-\frac{|w_i|^2}{2}
=\sum_{r\ge3}\frac{g_i^{r-2}}{r!}\sum_{a_1,\ldots,a_r}\sum_{\beta_1,\ldots,\beta_r}F^*_{i,a_1\cdots a_r}\prod_{l=1}^{r}S_i(a_l,\beta_l)w_{i,\beta_l}. g i 2 F i ( v , c i + 1 , z i ) − F i ( v , c i + 1 , z i ∗ ) − 2 ∣ w i ∣ 2 = r ≥ 3 ∑ r ! g i r − 2 a 1 , … , a r ∑ β 1 , … , β r ∑ F i , a 1 ⋯ a r ∗ l = 1 ∏ r S i ( a l , β l ) w i , β l .
The star denotes evaluation at y = z i ∗ y=z_i^* y = z i ∗ (not at zero coarser input), and the β l \beta_l β l are scalar labels at scale i i i . This expression is analytic at g i = 0 g_i=0 g i = 0 : stationarity and the exact subtraction of the quadratic term remove the apparent singularity.
We display the term r = 3 r=3 r = 3 . For a multi-index τ = ( ν , η > i ) \tau=(\nu,\eta_{>i}) τ = ( ν , η > i ) in the background variables and the coarser free variables, we write [ ⋅ ] τ [\cdot]_{\tau} [ ⋅ ] τ for the coefficient of v ν ζ > i η > i v^{\nu}\zeta_{>i}^{\eta_{>i}} v ν ζ > i η > i after the substitution c i + 1 = C i + 1 ( v , ζ > i ) c_{i+1}=C_{i+1}(v,\zeta_{>i}) c i + 1 = C i + 1 ( v , ζ > i ) , and set
e x 1 ⋯ x m ; τ ( m ) = [ ( D m Q ( i ) ) x 1 ⋯ x m ( Ψ i ∗ ) ] τ , P x ; A ; τ = [ ( ∂ y A Ψ i , x ) ] τ , s a β ; τ = [ S i ( a , β ) ] τ , e^{(m)}_{x_1\cdots x_m;\tau}=\left[(D^mQ^{(i)})_{x_1\cdots x_m}(\Psi_i^*)\right]_{\tau},\qquad P_{x;A;\tau}=\left[(\partial_{y_A}\Psi_{i,x})\right]_{\tau},\qquad s_{a\beta;\tau}=[S_i(a,\beta)]_{\tau}, e x 1 ⋯ x m ; τ ( m ) = [ ( D m Q ( i ) ) x 1 ⋯ x m ( Ψ i ∗ ) ] τ , P x ; A ; τ = [ ( ∂ y A Ψ i , x ) ] τ , s a β ; τ = [ S i ( a , β ) ] τ ,
where ∂ y A = ∂ y a ∂ y b \partial_{y_A}=\partial_{y_a}\partial_{y_b} ∂ y A = ∂ y a ∂ y b for A = ( a , b ) A=(a,b) A = ( a , b ) , and similarly for longer A A A . The chain rule gives, for T a b c ; τ = [ F i , a b c ∗ ] τ T_{abc;\tau}=[F^*_{i,abc}]_{\tau} T ab c ; τ = [ F i , ab c ∗ ] τ ,
T a b c ; τ = ∑ x 1 , x 2 , x 3 ∑ τ 0 + τ 1 + τ 2 + τ 3 = τ e x 1 x 2 x 3 ; τ 0 ( 3 ) P x 1 ; a ; τ 1 P x 2 ; b ; τ 2 P x 3 ; c ; τ 3 + ∑ x 1 , x 2 ∑ τ 0 + τ 1 + τ 2 = τ e x 1 x 2 ; τ 0 ( 2 ) ( P x 1 ; a b ; τ 1 P x 2 ; c ; τ 2 + P x 1 ; a c ; τ 1 P x 2 ; b ; τ 2 + P x 1 ; b c ; τ 1 P x 2 ; a ; τ 2 ) + ∑ x ∑ τ 0 + τ 1 = τ e x ; τ 0 ( 1 ) P x ; a b c ; τ 1 . \begin{aligned}
T_{abc;\tau}
={}&\sum_{x_1,x_2,x_3}\sum_{\tau_0+\tau_1+\tau_2+\tau_3=\tau}
e^{(3)}_{x_1x_2x_3;\tau_0}P_{x_1;a;\tau_1}P_{x_2;b;\tau_2}P_{x_3;c;\tau_3}\\
&+\sum_{x_1,x_2}\sum_{\tau_0+\tau_1+\tau_2=\tau}
e^{(2)}_{x_1x_2;\tau_0}\left(P_{x_1;ab;\tau_1}P_{x_2;c;\tau_2}+P_{x_1;ac;\tau_1}P_{x_2;b;\tau_2}+P_{x_1;bc;\tau_1}P_{x_2;a;\tau_2}\right)\\
&+\sum_x\sum_{\tau_0+\tau_1=\tau}e^{(1)}_{x;\tau_0}P_{x;abc;\tau_1}.
\end{aligned} T ab c ; τ = x 1 , x 2 , x 3 ∑ τ 0 + τ 1 + τ 2 + τ 3 = τ ∑ e x 1 x 2 x 3 ; τ 0 ( 3 ) P x 1 ; a ; τ 1 P x 2 ; b ; τ 2 P x 3 ; c ; τ 3 + x 1 , x 2 ∑ τ 0 + τ 1 + τ 2 = τ ∑ e x 1 x 2 ; τ 0 ( 2 ) ( P x 1 ; ab ; τ 1 P x 2 ; c ; τ 2 + P x 1 ; a c ; τ 1 P x 2 ; b ; τ 2 + P x 1 ; b c ; τ 1 P x 2 ; a ; τ 2 ) + x ∑ τ 0 + τ 1 = τ ∑ e x ; τ 0 ( 1 ) P x ; ab c ; τ 1 .
In particular, the Hessian is
F i , a b ∗ = ∑ x , x ′ Q x x ′ ( i ) ( Ψ i ∗ ) Ψ i , x ′ , a ∗ Ψ i , x ′ , b ∗ + ∑ x Q x ( i ) ( Ψ i ∗ ) Ψ i , x , a b ∗ . F^*_{i,ab}=\sum_{x,x'}Q^{(i)}_{xx'}(\Psi_i^*)\Psi^*_{i,x',a}\Psi^*_{i,x',b}+\sum_xQ^{(i)}_x(\Psi_i^*)\Psi^*_{i,x,ab}. F i , ab ∗ = x , x ′ ∑ Q x x ′ ( i ) ( Ψ i ∗ ) Ψ i , x ′ , a ∗ Ψ i , x ′ , b ∗ + x ∑ Q x ( i ) ( Ψ i ∗ ) Ψ i , x , ab ∗ .
Stationarity in y y y does not imply ( Q ( i ) ) ′ ( Ψ i ∗ ) = 0 (Q^{(i)})'(\Psi_i^*)=0 ( Q ( i ) ) ′ ( Ψ i ∗ ) = 0 , so the term with ( Q ( i ) ) ′ (Q^{(i)})' ( Q ( i ) ) ′ and the last line of the formula for T a b c ; τ T_{abc;\tau} T ab c ; τ cannot be deleted. Let η i \eta_i η i be a multi-index of degree three in the fluctuation variables of scale i i i . After the substitutions g l = z λ l g_l=z\lambda_l g l = z λ l and ζ > i = z λ > i w > i \zeta_{>i}=z\lambda_{>i}w_{>i} ζ > i = z λ > i w > i , the coefficient of v ν w i η i w > i η > i v^\nu w_i^{\eta_i}w_{>i}^{\eta_{>i}} v ν w i η i w > i η > i in this cubic term is
z 1 + ∣ η > i ∣ λ i 6 ∏ j = i + 1 s − 1 λ j ∣ η j ∣ ∑ a , b , c ∑ β 1 , β 2 , β 3 e β 1 + e β 2 + e β 3 = η i ∑ τ 0 + τ 1 + τ 2 + τ 3 = ( ν , η > i ) T a b c ; τ 0 s a β 1 ; τ 1 s b β 2 ; τ 2 s c β 3 ; τ 3 . (85) \frac{z^{1+|\eta_{>i}|\lambda_i}}{6}\prod_{j=i+1}^{s-1}\lambda_j^{|\eta_j|}
\sum_{a,b,c}\sum_{\substack{\beta_1,\beta_2,\beta_3\\ \mathbf e_{\beta_1}+\mathbf e_{\beta_2}+\mathbf e_{\beta_3}=\eta_i}}
\sum_{\tau_0+\tau_1+\tau_2+\tau_3=(\nu,\eta_{>i})}
T_{abc;\tau_0}s_{a\beta_1;\tau_1}s_{b\beta_2;\tau_2}s_{c\beta_3;\tau_3}. \tag*{(85)} 6 z 1 + ∣ η > i ∣ λ i j = i + 1 ∏ s − 1 λ j ∣ η j ∣ a , b , c ∑ β 1 , β 2 , β 3 e β 1 + e β 2 + e β 3 = η i ∑ τ 0 + τ 1 + τ 2 + τ 3 = ( ν , η > i ) ∑ T ab c ; τ 0 s a β 1 ; τ 1 s b β 2 ; τ 2 s c β 3 ; τ 3 . (85)
where η j \eta_j η j is the part of the multi-index η > i \eta_{>i} η > i at scale j j j . The factor 1 / 6 1/6 1/6 also accounts for repeated β \beta β ’s; no further factorial occurs. For each summand we take the networks of its coefficients and join them at their common indices x x x and a a a , and at the coarser arguments. The network of the resulting monomial is this connected union, and its fluctuation support is
S λ = { blocks of β 1 , β 2 , β 3 } ∪ { coarser blocks used by η > i } , S_\lambda=\{\text{blocks of }\beta_1,\beta_2,\beta_3\}\cup\{\text{coarser blocks used by }\eta_{>i}\}, S λ = { blocks of β 1 , β 2 , β 3 } ∪ { coarser blocks used by η > i } ,
with repeated blocks counted once. The background positions in ν \nu ν belong to the network even if they lie outside S λ S_\lambda S λ . Let Q m + Q_m^+ Q m + , P m + P_m^+ P m + , and S + S^+ S + be the absolute scalar majorants of the jets and maps above on their common product, including the volume ratios of part (4) of Lemma 20.1. The formula for T a b c ; τ T_{abc;\tau} T ab c ; τ and Lemma 20.1 give the cubic majorant
1 6 ( Q 3 + ( P 1 + ) 3 + 3 Q 2 + P 2 + P 1 + + Q 1 + P 3 + ) ( S + ) 3 , \frac{1}{6}\left(Q_3^+(P_1^+)^3+3Q_2^+P_2^+P_1^++Q_1^+P_3^+\right)(S^+)^3, 6 1 ( Q 3 + ( P 1 + ) 3 + 3 Q 2 + P 2 + P 1 + + Q 1 + P 3 + ) ( S + ) 3 ,
and the same expression with a derivative falling on one factor bounds the sum rooted at a fluctuation or background index. Each product here corresponds to a connected contraction in the formula for T a b c ; τ T_{abc;\tau} T ab c ; τ , so no unrooted product over an extensive set appears. For general r r r the rule is the partition formula
∂ a 1 ⋯ ∂ a r ( Q ( i ) ∘ Ψ i ) = ∑ π ∑ ( x B ) B ∈ π ( D ∣ π ∣ Q ( i ) ) ( x B ) ( Ψ i ) ∏ B ∈ π ∂ a B Ψ i , x B \partial_{a_1}\cdots\partial_{a_r}\left(Q^{(i)}\circ\Psi_i\right)=\sum_{\pi}\sum_{(x_B)_{B\in\pi}}\left(D^{|\pi|}Q^{(i)}\right)_{(x_B)}\left(\Psi_i\right)\prod_{B\in\pi}\partial_{a_B}\Psi_{i,x_B} ∂ a 1 ⋯ ∂ a r ( Q ( i ) ∘ Ψ i ) = π ∑ ( x B ) B ∈ π ∑ ( D ∣ π ∣ Q ( i ) ) ( x B ) ( Ψ i ) B ∈ π ∏ ∂ a B Ψ i , x B
where π \pi π runs over the partitions of { 1 , … , r } \{1,\ldots,r\} { 1 , … , r } . Together with Lemma 20.1, it gives a convergent coefficientwise majorant on a common smaller free product. Since r ≥ 3 r\ge3 r ≥ 3 , a term of total free degree q = r + ∣ η > i ∣ q=r+|\eta_{>i}| q = r + ∣ η > i ∣ has the coupling power z q − 2 z^{q-2} z q − 2 , and the ratios of the scales have modulus at most one. This proves the cubic starting order before any expectation over the fluctuation variables is taken.
Step 3: a monomial of the Haar density. Let Γ ( v , c i + 1 , y ) \Gamma(v,c_{i+1},y) Γ ( v , c i + 1 , y ) be the chart of solved links at scale i i i , let E i ∘ \mathcal{E}_i^\circ E i ∘ be its links outside the block axial gauge trees, and let c c c run over the set C i \mathcal{C}_i C i of constraints solved at scale i i i , with eliminated edges e c ∗ e_c^* e c ∗ . We write
A c = D X e c ∗ c ( Γ ) , L i = − ∑ e ∈ E i ∘ log j H ( Γ e ) + ∑ c ∈ C i log det A c , A_c=D_{X_{e_c^*}}c(\Gamma),\qquad L_i=-\sum_{e\in\mathcal{E}_i^\circ}\log j_H(\Gamma_e)+\sum_{c\in\mathcal{C}_i}\log\det A_c, A c = D X e c ∗ c ( Γ ) , L i = − e ∈ E i ∘ ∑ log j H ( Γ e ) + c ∈ C i ∑ log det A c ,
with j H ( X ) = sinc 2 ( ∣ X ∣ / 2 ) j_H(X)=\operatorname{sinc}^2(|X|/2) j H ( X ) = sinc 2 ( ∣ X ∣/2 ) as in (4.2) and Lemma 5.1(2); only field-independent factors have been removed. The fluctuation-dependent part of L i L_i L i at scale i i i is
L i ( v , c i + 1 , z i ) − L i ( v , c i + 1 , z i ∗ ) = ∑ r ≥ 1 g i r r ! ∑ a 1 , … , a r ∑ β 1 , … , β r L i , a 1 ⋯ a r ∗ ∏ l = 1 r S i ( a l , β l ) w i , β l . L_i(v,c_{i+1},z_i)-L_i(v,c_{i+1},z_i^*)=\sum_{r\ge1}\frac{g_i^r}{r!}\sum_{a_1,\ldots,a_r}\sum_{\beta_1,\ldots,\beta_r}L^*_{i,a_1\cdots a_r}\prod_{l=1}^{r}S_i(a_l,\beta_l)w_{i,\beta_l}. L i ( v , c i + 1 , z i ) − L i ( v , c i + 1 , z i ∗ ) = r ≥ 1 ∑ r ! g i r a 1 , … , a r ∑ β 1 , … , β r ∑ L i , a 1 ⋯ a r ∗ l = 1 ∏ r S i ( a l , β l ) w i , β l .
Its linear coefficient is
L i , a ∗ = − ∑ e ∈ E i ∘ ∑ ς = 1 3 ( ∂ ς log j H ) ( Γ e ∗ ) Γ e ς , a ∗ + ∑ c ∈ C i ∑ A , B = 1 3 ( A c − 1 ) B A ∗ ( ∂ y a A c ) A B ∗ , L^*_{i,a}=-\sum_{e\in\mathcal{E}_i^\circ}\sum_{\varsigma=1}^{3}(\partial_{\varsigma}\log j_H)(\Gamma_e^*)\Gamma^*_{e\varsigma,a}+\sum_{c\in\mathcal{C}_i}\sum_{A,B=1}^{3}(A_c^{-1})^*_{BA}(\partial_{y_a}A_c)^*_{AB}, L i , a ∗ = − e ∈ E i ∘ ∑ ς = 1 ∑ 3 ( ∂ ς log j H ) ( Γ e ∗ ) Γ e ς , a ∗ + c ∈ C i ∑ A , B = 1 ∑ 3 ( A c − 1 ) B A ∗ ( ∂ y a A c ) A B ∗ ,
where ( ∂ y a A c ) A B = ∑ e , ς ( ∂ X e ς A c ) A B Γ e ς , a (\partial_{y_a}A_c)_{AB}=\sum_{e,\varsigma}(\partial_{X_{e\varsigma}}A_c)_{AB}\Gamma_{e\varsigma,a} ( ∂ y a A c ) A B = ∑ e , ς ( ∂ X e ς A c ) A B Γ e ς , a , and all factors are evaluated at the minimizer and at the coarser reconstruction C i + 1 ( v , ζ > i ) C_{i+1}(v,\zeta_{>i}) C i + 1 ( v , ζ > i ) . Taking the coefficient [ ⋅ ] τ [\cdot]_\tau [ ⋅ ] τ of each factor and summing over all allocations of τ \tau τ gives the coefficients l a ; τ = [ L i , a ∗ ] τ l_{a;\tau}=[L^*_{i,a}]_\tau l a ; τ = [ L i , a ∗ ] τ , with every link, constraint, color, integrated coordinate, and multi-index sum displayed. For ∣ η i ∣ = 1 |\eta_i|=1 ∣ η i ∣ = 1 the coefficient after the coupling substitution is
z 1 + ∣ η > i ∣ λ i ∏ j = i + 1 s − 1 λ j ∣ η j ∣ ∑ a ∑ β : e β = η i ∑ τ 0 + τ 1 = ( ν , η > i ) l a ; τ 0 s a β ; τ 1 . z^{1+|\eta_{>i}|}\lambda_i\prod_{j=i+1}^{s-1}\lambda_j^{|\eta_j|}\sum_a\sum_{\beta:e_\beta=\eta_i}\sum_{\tau_0+\tau_1=(\nu,\eta_{>i})}l_{a;\tau_0}s_{a\beta;\tau_1}. z 1 + ∣ η > i ∣ λ i j = i + 1 ∏ s − 1 λ j ∣ η j ∣ a ∑ β : e β = η i ∑ τ 0 + τ 1 = ( ν , η > i ) ∑ l a ; τ 0 s a β ; τ 1 .
The trace in L i , a ∗ L^*_{i,a} L i , a ∗ is rooted at the index a a a through the derivative ∂ y a A c \partial_{y_a}A_c ∂ y a A c ; its coefficient bound sums over c , e , A , B , ς c,e,A,B,\varsigma c , e , A , B , ς . It is not the unrooted norm of a sum of logarithmic determinants.
The network is the connected union of the networks of the Haar factors, the inverse matrices, and the chart, joined to the network of S i S_i S i at a a a , and the support is the block of β \beta β together with the coarser blocks used by η > i \eta_{>i} η > i . Lemma 20.1 bounds these coefficients by the products of the majorants of the factors and their derivatives. The series starts at total free degree one, with coupling power equal to the total free degree.
The remaining part of the measure fluctuation in (20.11) is the difference B i ( c i + 1 ) − B i ( c ^ i + 1 ) B_i(c_{i+1})-B_i(\widehat{c}_{i+1}) B i ( c i + 1 ) − B i ( c i + 1 ) , whose fluctuation support may lie entirely in coarser scales; it is not part of the field-independent prefactor. For a coordinate h γ h_\gamma h γ of the coarser argument, the total derivative is
∂ h γ B i = L i , h γ ∗ + ∑ a L i , a ∗ ∂ h γ z i , a ∗ + 1 2 ∑ a , b ( K i − 1 ) b a ∂ h γ K i , a b \partial_{h_\gamma} B_i=L^*_{i,h_\gamma}+\sum_a L^*_{i,a}\partial_{h_\gamma}z^*_{i,a}+\frac{1}{2}\sum_{a,b}(K_i^{-1})_{ba}\partial_{h_\gamma}K_{i,ab} ∂ h γ B i = L i , h γ ∗ + a ∑ L i , a ∗ ∂ h γ z i , a ∗ + 2 1 a , b ∑ ( K i − 1 ) ba ∂ h γ K i , ab
where ∂ h γ z i ∗ = − K i − 1 F i , ⋅ h γ ∗ \partial_{h_\gamma}z^*_i=-K_i^{-1}F^*_{i,\,\cdot\,h_\gamma} ∂ h γ z i ∗ = − K i − 1 F i , ⋅ h γ ∗ and ∂ h γ K i , a b = F i , a b h γ ∗ + ∑ a ′ F i , a b a ′ ∗ ∂ h γ z i , a ′ ∗ \partial_{h_\gamma}K_{i,ab}=F^*_{i,abh_\gamma}+\sum_{a'}F^*_{i,aba'}\partial_{h_\gamma}z^*_{i,a'} ∂ h γ K i , ab = F i , ab h γ ∗ + ∑ a ′ F i , ab a ′ ∗ ∂ h γ z i , a ′ ∗ . The difference equals the convergent coefficient expansion of
∫ 0 1 ∑ j = i + 1 s − 1 ∑ β ∈ I j ζ j , β ∑ γ ( ∂ h γ B i ) ( v , C i + 1 ( v , τ ζ > i ) ) ( ∂ ζ j , β C i + 1 , γ ) ( v , τ ζ > i ) d τ , \int_0^1\sum_{j=i+1}^{s-1}\sum_{\beta\in I_j}\zeta_{j,\beta}\sum_\gamma(\partial_{h_\gamma}B_i)(v,C_{i+1}(v,\tau\zeta_{>i}))(\partial_{\zeta_{j,\beta}}C_{i+1,\gamma})(v,\tau\zeta_{>i})\,d\tau, ∫ 0 1 j = i + 1 ∑ s − 1 β ∈ I j ∑ ζ j , β γ ∑ ( ∂ h γ B i ) ( v , C i + 1 ( v , τ ζ > i )) ( ∂ ζ j , β C i + 1 , γ ) ( v , τ ζ > i ) d τ ,
since the integrand is the derivative in τ \tau τ of B i ( v , C i + 1 ( v , τ ζ > i ) ) B_i(v,C_{i+1}(v,\tau\zeta_{>i})) B i ( v , C i + 1 ( v , τ ζ > i )) ; here I j I_j I j is the set of unfixed scalar free labels at scale j j j , and ∂ ζ j , β C i + 1 \partial_{\zeta_{j,\beta}}C_{i+1} ∂ ζ j , β C i + 1 is evaluated at τ ζ > i \tau\zeta_{>i} τ ζ > i without an extra factor τ \tau τ . Each Taylor coefficient of this integral carries the factor 1 / ( n + 1 ) 1/(n+1) 1/ ( n + 1 ) when the remaining integrand has degree n n n in τ \tau τ . Likewise, at fixed couplings,
d z i = ( ∂ v z i ∗ ) d v + ( ∂ c z i ∗ ) d c i + 1 + g i { ( ∂ v S i ) d v + ( ∂ c S i ) d c i + 1 } w i + g i S i d w i , dz_i=(\partial_vz_i^*)\,dv+(\partial_cz_i^*)\,dc_{i+1}+g_i\{(\partial_vS_i)\,dv+(\partial_cS_i)\,dc_{i+1}\}w_i+g_iS_i\,dw_i, d z i = ( ∂ v z i ∗ ) d v + ( ∂ c z i ∗ ) d c i + 1 + g i {( ∂ v S i ) d v + ( ∂ c S i ) d c i + 1 } w i + g i S i d w i ,
and none of these terms may be dropped. The inverse square root S i S_i S i is treated as a constituent tensor with the bounds of Theorem 20.2; if it is expanded, (20.19) and the Neumann series of the resolvent around K i ( 0 , 0 ) K_i(0,0) K i ( 0 , 0 ) give a chain of ordinary matrix products whose coefficients are its network refinements. The comparison of determinant matrices with the same indices also includes the correction (20.21), which is a further constituent tensor with a visit to Ω c \Omega^c Ω c .
Step 4: the incoming insertion and the boundary differences. We write A ( v , ζ ) = Log C ( v , ζ ; c b r e c ) \mathcal{A}(v,\zeta)=\operatorname{Log}C(v,\zeta;c_{\mathrm{brec}}) A ( v , ζ ) = Log C ( v , ζ ; c brec ) . Then
R ( A ( v , ζ ) ) − R ( A ( v , 0 ) ) = ∑ n ≥ 1 1 n ! D n R ( A ( v , 0 ) ) [ A ( v , ζ ) − A ( v , 0 ) , … , A ( v , ζ ) − A ( v , 0 ) ] , R(\mathcal{A}(v,\zeta))-R(\mathcal{A}(v,0))=\sum_{n\geq1}\frac{1}{n!}D^nR(\mathcal{A}(v,0))[\mathcal{A}(v,\zeta)-\mathcal{A}(v,0),\ldots,\mathcal{A}(v,\zeta)-\mathcal{A}(v,0)], R ( A ( v , ζ )) − R ( A ( v , 0 )) = n ≥ 1 ∑ n ! 1 D n R ( A ( v , 0 )) [ A ( v , ζ ) − A ( v , 0 ) , … , A ( v , ζ ) − A ( v , 0 )] ,
and each factor A ( v , ζ ) − A ( v , 0 ) \mathcal{A}(v,\zeta)-\mathcal{A}(v,0) A ( v , ζ ) − A ( v , 0 ) has positive degree in ζ \zeta ζ . The coefficient of v ν ζ η v^\nu\zeta^\eta v ν ζ η is the sum over n n n , over the contracted indices j 1 , … , j n j_1,\ldots,j_n j 1 , … , j n of the old lattice, and over the allocations of the multi-indices to the n n n factors with positive free degree in each, of the products of the jet coefficients of R R R and the coefficients of the factors. No additional multinomial factor occurs, since each coefficient is already a multi-index coefficient. Each such occurrence keeps the union of the old network of R R R , scaled by B s − 1 = M − s B_s^{-1}=M^{-s} B s − 1 = M − s in output units, and the attached reconstruction networks, and its fluctuation support is the set of blocks used by η \eta η . By the jet bound of Step 1 with j = 1 j=1 j = 1 at an image bounded by ρ / 2 \rho/2 ρ /2 , the derivative series of R R R at the image is at most 4 H ρ 4H\rho 4 H ρ . The column majorant of the reconstruction with a free index fixed vanishes at zero free radius and is linear in the free radius on the fixed product, so it is at most C J ∣ z ∣ p C_J|z|^p C J ∣ z ∣ p . Fixing an external input in the old counting convention contributes the fixed volume ratio Γ J = max i v i / v o l d ≤ M 3 J \Gamma_J=\max_i v_i/v_{\mathrm{old}}\le M^{3J} Γ J = max i v i / v old ≤ M 3 J between the input mesh and the old mesh. We obtain the incoming bound, linear in H H H and starting at ∣ z ∣ p |z|^p ∣ z ∣ p .
For a difference between the regional and the restricted torus constructions we compare factor by factor, before any expansion of logarithms. In each product of coefficients we use the ordered identity
∏ r = 1 m A r 1 − ∏ r = 1 m A r 0 = ∑ r = 1 m ( ∏ l < r A l 1 ) ( A r 1 − A r 0 ) ( ∏ l > r A l 0 ) , \prod_{r=1}^{m} A_r^1-\prod_{r=1}^{m} A_r^0=\sum_{r=1}^{m}\left(\prod_{l<r}A_l^1\right)(A_r^1-A_r^0)\left(\prod_{l>r}A_l^0\right), r = 1 ∏ m A r 1 − r = 1 ∏ m A r 0 = r = 1 ∑ m ( l < r ∏ A l 1 ) ( A r 1 − A r 0 ) ( l > r ∏ A l 0 ) ,
and for the incoming function at two reconstructed arguments C 0 , C 1 C_0,C_1 C 0 , C 1 in the identity chart (here C 0 , C 1 C_0,C_1 C 0 , C 1 denote arguments, not the constants of Lemma 20.3),
R ( C 1 ) − R ( C 0 ) = ∫ 0 1 D R ( C 0 + s ( C 1 − C 0 ) ) [ C 1 − C 0 ] d s . R(C_1)-R(C_0)=\int_0^1 DR\left(C_0+s(C_1-C_0)\right)[C_1-C_0]\,\mathrm{d}s. R ( C 1 ) − R ( C 0 ) = ∫ 0 1 D R ( C 0 + s ( C 1 − C 0 ) ) [ C 1 − C 0 ] d s .
Both images under the absolute reconstruction majorant lie in the common product with half the incoming radius, so the whole segment does. Terms involving the incoming boundary function use its incoming boundary weight; torus terms acquire the visit carried by the difference of the reconstructions. Fixed input slots are visits to Ω c \Omega^c Ω c . Changed formulas, minimizers, inverse square roots, and Haar densities are treated by the same identity, with − A − 1 ( A − B ) B − 1 -A^{-1}(A-B)B^{-1} − A − 1 ( A − B ) B − 1 for inverses and (70) for inverse square roots. Each term contains one changed factor with its visit to Ω c \Omega^c Ω c . The versions with every index rooted in Theorem 19.9 keep every differentiated background or free index, even when the changed factor is not differentiated; the degree-one terms are therefore included.
We join the incoming networks, scaled by B s − 1 B_s^{-1} B s − 1 , to the chart networks at their contracted indices. The transport of the route to the fixed complement is exactly (42) , and Lemma 18.15 sums the remaining internal indices. We append the resulting route to the network of the monomial. One fixed allocation of the available exponents bounds e s 0 ℓ λ e^{s_0\ell_\lambda} e s 0 ℓ λ , including this route and the fixed stencil offsets; the offsets contribute a fixed factor in the degree, absorbed on the margin already provided by the construction. The allocation is the same for every order and every contraction, and no exponent is lost at each vertex. The scaled incoming networks and routes are bounded by the inequalities against B s α i n B_s\alpha_{\mathrm{in}} B s α in and B s α i n B_s\alpha_{\mathrm{in}} B s α in imposed in the statement. Every term still contains an incoming factor, so its bound is linear in H H H and starts at ∣ z ∣ p |z|p ∣ z ∣ p . The arguments of Step 5 apply to this majorant with a visit to Ω c \Omega^{c} Ω c and give (20.33). Finally, a torus monomial deleted because it uses a missing fluctuation block already has a visit at that block; if it has an unfixed root, its network provides the route to Ω c \Omega^{c} Ω c .
Step 5: conversion to W 16 W_{16} W 16 and D 16 D_{16} D 16 . By Steps 1 to 4, there exist a free radius V 1 > 0 V_1>0 V 1 > 0 , strictly inside all image and stage margins, and a common bound A 0 A_0 A 0 for the coefficientwise majorants of the diagrams on the outer profile ϕ ^ \widehat{\phi} ϕ and at free radius V 1 V_1 V 1 , including each permitted fixed coefficient index and the volume ratios of part (4) of Lemma 20.1. For the incoming family the bound is A 0 H A_0H A 0 H , and for a difference it is linear in the bound of the changed factor. We write the resolved coefficients of free degree q q q and background degree k k k as u λ 0 v λ ν ζ λ η u_\lambda^0v_\lambda^\nu\zeta_\lambda^\eta u λ 0 v λ ν ζ λ η , after removing the factor z q − 2 z^{q-2} z q − 2 for the classical terms and z q z^q z q for the others; the ratios λ i \lambda_i λ i have modulus at most one. The mixed factorial identity is
1 k ! q ! ∑ other ordered indices ∣ D k ν D ζ q ( u 0 v ν ζ η ) ∣ = { ( η l / q ) ∣ u 0 ∣ , one free input l fixed , ( ν i / k ) ∣ u 0 ∣ , one background input i fixed , \frac{1}{k!q!}\sum_{\text{other ordered indices}}\left|D_k^\nu D_\zeta^q\left(u^0v^\nu\zeta^\eta\right)\right| =
\begin{cases}
(\eta_l/q)|u^0|, & \text{one free input }l\text{ fixed},\\
(\nu_i/k)|u^0|, & \text{one background input }i\text{ fixed},
\end{cases} k ! q ! 1 other ordered indices ∑ D k ν D ζ q ( u 0 v ν ζ η ) = { ( η l / q ) ∣ u 0 ∣ , ( ν i / k ) ∣ u 0 ∣ , one free input l fixed , one background input i fixed ,
for q ≥ 1 q\geq1 q ≥ 1 and k ≥ 1 k\geq1 k ≥ 1 respectively, and the radii of all inputs multiply these expressions. With C ~ = max { 1 , V 1 − 1 } \widetilde{C}=\max\{1,V_1^{-1}\} C = max { 1 , V 1 − 1 } , Lemma 20.1 and the finite majorant at V 1 V_1 V 1 give
sup l ∑ λ : ( k λ , q λ ) = ( k , q ) η λ , l q ∣ u λ 0 ∣ ϕ ^ ν λ e s 0 ℓ λ ≤ A 0 C ~ q , sup i ∑ λ : ( k λ , q λ ) = ( k , q ) ν λ , i k ∣ u λ 0 ∣ ϕ ^ ν λ e s 0 ℓ λ ≤ A 0 C ~ q . \sup_l\sum_{\lambda:(k_\lambda,q_\lambda)=(k,q)}\frac{\eta_{\lambda,l}}{q}|u_\lambda^0|\widehat{\phi}^{\nu_\lambda}e^{s_0\ell_\lambda}\leq A_0\widetilde{C}^q,\qquad
\sup_i\sum_{\lambda:(k_\lambda,q_\lambda)=(k,q)}\frac{\nu_{\lambda,i}}{k}|u_\lambda^0|\widehat{\phi}^{\nu_\lambda}e^{s_0\ell_\lambda}\leq A_0\widetilde{C}^q. l sup λ : ( k λ , q λ ) = ( k , q ) ∑ q η λ , l ∣ u λ 0 ∣ ϕ ν λ e s 0 ℓ λ ≤ A 0 C q , i sup λ : ( k λ , q λ ) = ( k , q ) ∑ k ν λ , i ∣ u λ 0 ∣ ϕ ν λ e s 0 ℓ λ ≤ A 0 C q .
These bounds concern the total variation of the resolution produced by composition of coefficientwise majorants.
Let ϑ = K ∗ − 1 / 2 < 1 \vartheta=K_*^{-1/2}<1 ϑ = K ∗ − 1/2 < 1 . Passing from ϕ ^ \widehat{\phi} ϕ to ψ p r o f \psi_{\mathrm{prof}} ψ prof multiplies a term of background degree k k k by exactly ϑ q k \vartheta^{qk} ϑ q k , independently of the positions of its background indices in the spatially varying radius profile. Each term has ∣ S λ ∣ ≤ q |S_\lambda|\leq q ∣ S λ ∣ ≤ q , and selecting a fluctuation block is bounded by the sum of its three scalar multiplicities. At fixed k , q k,q k , q , the preceding bounds therefore contribute at most 3 q A 0 C ~ q ϑ q k 3qA_0\widetilde{C}^q\vartheta^{qk} 3 q A 0 C q ϑ q k for a fluctuation root and k A 0 C ~ q ϑ q k kA_0\widetilde{C}^q\vartheta^{qk} k A 0 C q ϑ q k for a background root; the latter bound is uniform in the root at each degree. Taking the supremum over the root at each degree and only then summing over k k k , and using q ≤ 2 q q\leq2^q q ≤ 2 q , we find that the sums in the definitions of W h W_h W h and D h D_h D h , at fixed q q q , are at most
A 0 C ~ q e h q ( 3 q 1 − ϑ + ϑ ( 1 − ϑ ) 2 ) ≤ A B q , A = A 0 ( 3 1 − ϑ + ϑ ( 1 − ϑ ) 2 ) , B = 2 e h C ~ . A_0\widetilde{C}^q e^{hq}\left(\frac{3q}{1-\vartheta}+\frac{\vartheta}{(1-\vartheta)^2}\right)\leq AB^q,\qquad
A=A_0\left(\frac{3}{1-\vartheta}+\frac{\vartheta}{(1-\vartheta)^2}\right),\qquad
B=2e^h\widetilde{C}. A 0 C q e h q ( 1 − ϑ 3 q + ( 1 − ϑ ) 2 ϑ ) ≤ A B q , A = A 0 ( 1 − ϑ 3 + ( 1 − ϑ ) 2 ϑ ) , B = 2 e h C .
Multiplying by ∣ z ∣ q − 2 p q |z|^{q-2}p^q ∣ z ∣ q − 2 p q , q ≥ 3 q\geq3 q ≥ 3 , or by ∣ z ∣ q p q |z|^qp^q ∣ z ∣ q p q , q ≥ 1 q\geq1 q ≥ 1 , and summing the geometric series, we obtain on B ∣ z ∣ p < 1 B|z|p<1 B ∣ z ∣ p < 1
W h c l + D h c l ≤ A B 3 ∣ z ∣ p 3 1 − B ∣ z ∣ p , W h H a a r + D h H a a r ≤ A B ∣ z ∣ p 1 − B ∣ z ∣ p , W h i n c + D h i n c ≤ A B H ∣ z ∣ p 1 − B ∣ z ∣ p . W_h^{\mathrm{cl}}+D_h^{\mathrm{cl}}\leq\frac{AB^3|z|p^3}{1-B|z|p},\qquad
W_h^{\mathrm{Haar}}+D_h^{\mathrm{Haar}}\leq\frac{AB|z|p}{1-B|z|p},\qquad
W_h^{\mathrm{inc}}+D_h^{\mathrm{inc}}\leq\frac{ABH|z|p}{1-B|z|p}. W h cl + D h cl ≤ 1 − B ∣ z ∣ p A B 3 ∣ z ∣ p 3 , W h Haar + D h Haar ≤ 1 − B ∣ z ∣ p A B ∣ z ∣ p , W h inc + D h inc ≤ 1 − B ∣ z ∣ p A B H ∣ z ∣ p .
for the classical, Haar-density, and incoming families. The same bounds hold for the differences with a visit to Ω c \Omega^{c} Ω c , with their majorants. Note that the classical bound starts with ∣ z ∣ p |z|^{p} ∣ z ∣ p : the two coupling powers removed by the division by g i 2 g_{i}^{2} g i 2 are not restored. For h = 16 h=16 h = 16 , enlarging the constant to a common C J C_{J} C J and restricting the disc accordingly gives the assertions of the lemma. Choosing the auxiliary disc smaller by this fixed factor keeps it inside the same larger free radius, and neither an input radius nor the ratio K ∗ K_{*} K ∗ changes. Finitely many additional background marks contribute a polynomial factor in k k k , which is summable since ∑ k ( k + 1 ) m ϑ k < ∞ \sum_{k}(k+1)^{m}\vartheta^{k}<\infty ∑ k ( k + 1 ) m ϑ k < ∞ . The argument never interchanges sup i ∑ k \sup_{i}\sum_{k} sup i ∑ k with ∑ k sup i \sum_{k}\sup_{i} ∑ k sup i . The bounds at fixed real frozen values are proved first and are uniform, so taking the supremum over the frozen values afterwards preserves them. All coefficientwise majorant series converge on the same fixed larger products, and finite truncations and the absolute majorants justify the displayed sums.
Lemma 20.8. Let π = ⨂ x π x \pi=\bigotimes_{x}\pi_{x} π = ⨂ x π x be a finite product of probability measures on fluctuation blocks, each π x \pi_{x} π x supported in the ball { ∣ w x ∣ ≤ p x } \{|w_{x}|\le p_{x}\} { ∣ w x ∣ ≤ p x } with p x < ∞ p_{x}<\infty p x < ∞ ; these are the radii used in b λ b_{\lambda} b λ , W 16 W_{16} W 16 , and D 16 D_{16} D 16 . Use the same working rates a , α > 0 a,\alpha>0 a , α > 0 as in X ϕ X^{\phi} X ϕ , and assume s 0 ≥ a + α s_{0}\ge a+\alpha s 0 ≥ a + α . Consider monomial families as in Definition 20.6, with nonempty supports S λ S_{\lambda} S λ and networks through all of their positions, and put ϵ = W 16 ( U ) + D 16 ( U ) \epsilon=W_{16}(U)+D_{16}(U) ϵ = W 16 ( U ) + D 16 ( U ) . There is an absolute constant ϵ 0 > 0 \epsilon_{0}>0 ϵ 0 > 0 such that ϵ < ϵ 0 \epsilon<\epsilon_{0} ϵ < ϵ 0 implies the following.
The product expansion followed by the hard-core polymer expansion, with incompatibility S λ ∩ S μ ≠ ∅ S_{\lambda}\cap S_{\mu}\ne\emptyset S λ ∩ S μ = ∅ , gives a normally convergent logarithm of ∫ e − ∑ λ U λ d π \int e^{-\sum_{\lambda}U_{\lambda}}\,d\pi ∫ e − ∑ λ U λ d π , equal to zero at zero interaction, and this integral does not vanish in the coefficient ball. Neither spatial connectedness of S λ S_{\lambda} S λ nor membership of the background positions in S λ S_{\lambda} S λ is needed.
Every resolved term of the logarithm carries the connected union of the networks of its monomials, with no additional length. Its absolute sum rooted at a fluctuation block, with the weight e 4 ∣ S ∣ e^{4|S|} e 4∣ S ∣ where S S S is the union of the supports of its monomials, is at most C ϵ C\epsilon C ϵ . After passage from ψ p r o f \psi_{\mathrm{prof}} ψ prof to ϕ \phi ϕ , the sum over the background degrees of its rooted coefficient suprema is at most C ϵ C\epsilon C ϵ .
For one marked family A A A the same bounds hold with C ϵ A C\epsilon_{A} C ϵ A . Let U ( s , r ) = U + s B + r A + s r C U(s,r)=U+sB+rA+srC U ( s , r ) = U + s B + r A + sr C , 0 ≤ s , r ≤ 1 0\le s,r\le1 0 ≤ s , r ≤ 1 , where B B B carries the boundary change, A A A the incoming insertion, and C C C both (here s , r s,r s , r are interpolation parameters and B , C B,C B , C are families of monomials). Assume that every occurrence in B B B and C C C has an assigned network visiting Ω L J c \Omega^{c}_{L_{J}} Ω L J c , with that visit included in ℓ λ \ell_{\lambda} ℓ λ and in its weighted bound, and that every interpolated family lies in the same fixed small ball. Then the mixed derivative ∂ s ∂ r \partial_{s}\partial_{r} ∂ s ∂ r of the logarithm and its four-corner difference are bounded, in both rooted norms after passage of profile, by
C ( ϵ ∂ A + ϵ ∂ ϵ A ) , (86) C(\epsilon_{\partial A}+\epsilon_{\partial}\epsilon_{A}), \tag*{(86)} C ( ϵ ∂ A + ϵ ∂ ϵ A ) , (86) where ϵ ∂ , ϵ A , ϵ ∂ A \epsilon_{\partial},\epsilon_{A},\epsilon_{\partial A} ϵ ∂ , ϵ A , ϵ ∂ A are the values of W 16 + D 16 W_{16}+D_{16} W 16 + D 16 for B , A , C B,A,C B , A , C . A term containing the boundary change keeps a visit to Ω L J c \Omega^{c}_{L_{J}} Ω L J c . Its network can be extended to Ω c \Omega^{c} Ω c by a segment of length at
most the fixed enlargement constant L 0 L_0 L 0 of Lemma 20.1(5), whose cost is absorbed in the fixed constants. A term containing both changes keeps an incoming factor. These bounds keep the weight of the network length. In particular, for a difference F F F of logarithms containing the boundary change, on the working profile,
X ϕ ( F ) ≤ C J ( ϵ ∂ A + ϵ ∂ ϵ A ) \mathcal{X}^{\phi}(F) \le C_J(\epsilon_{\partial A}+\epsilon_{\partial}\epsilon_A) X ϕ ( F ) ≤ C J ( ϵ ∂ A + ϵ ∂ ϵ A )
in the case of two changes, and X ϕ ( F ) ≤ C J ϵ ∂ \mathcal{X}^{\phi}(F) \le C_J\epsilon_{\partial} X ϕ ( F ) ≤ C J ϵ ∂ in the case of one change.
The logarithm and its derivatives in the changes are holomorphic on this ball. The combinatorial constants are absolute, with the dependence described in item (1) of Definition 4.14; the constants of the passage of profile, of the density, and of the spatial conversion have the dependence described in item (3).
Proof. The proof follows the standard two-stage cluster expansion; we give the estimates so that the root and the support weights stay visible. We first truncate to finitely many monomials.
Step 1: the product expansion. Let f λ = e − U λ − 1 f_\lambda=e^{-U_\lambda}-1 f λ = e − U λ − 1 . Its absolute coefficient majorant is at most e b λ − 1 ≤ 2 b λ e^{b_\lambda}-1\le2b_\lambda e b λ − 1 ≤ 2 b λ for small ϵ \epsilon ϵ . Indeed ∣ w x ∣ ≤ p x |w_x|\le p_x ∣ w x ∣ ≤ p x holds π x \pi_x π x -almost surely, so the polynomial bounds defining b λ b_\lambda b λ bound the integrands on the support of the product measure. The normalized truncated Gaussian measures on balls used later satisfy this hypothesis. We expand the exact identity
e − ∑ λ U λ = ∏ λ ( 1 + f λ ) . e^{-\sum_\lambda U_\lambda}=\prod_\lambda(1+f_\lambda). e − ∑ λ U λ = λ ∏ ( 1 + f λ ) .
We group each selected family of distinct indices by the connected components of the graph in which λ , μ \lambda,\mu λ , μ are joined when S λ ∩ S μ ≠ ∅ S_\lambda\cap S_\mu\ne\varnothing S λ ∩ S μ = ∅ . Different components depend on disjoint sets of fluctuation blocks, so their expectations factor, even if their background parameters coincide. This is an exact hard-core gas whose polymers are unions of fluctuation supports. Spatially disconnected sets may serve as polymers, because compatibility uses only set intersection and the independence of the reference variables.
Step 2: the tree majorant at the first stage. Let n λ = ∣ S λ ∣ n_\lambda=|S_\lambda| n λ = ∣ S λ ∣ and a λ = 2 b λ e 10 n λ a_\lambda=2b_\lambda e^{10n_\lambda} a λ = 2 b λ e 10 n λ . The majorant by rooted trees is the increasing iteration from zero of
T λ = a λ exp ( ∑ μ : S μ ∩ S λ ≠ ∅ T μ ) (87) T_\lambda=a_\lambda\exp\left(\sum_{\mu:S_\mu\cap S_\lambda\ne\varnothing}T_\mu\right) \tag*{(87)} T λ = a λ exp μ : S μ ∩ S λ = ∅ ∑ T μ (87)
Every connected selected family has a spanning tree; expanding the exponential supplies the factorials for unordered children, and dropping the distinctness of the indices only increases this sum of nonnegative terms. If T μ ≤ a μ e n μ T_\mu\le a_\mu e^{n_\mu} T μ ≤ a μ e n μ for all μ \mu μ , then the exponent in (20.36) is at most
n λ sup x ∑ μ : x ∈ S μ a μ e n μ ≤ 2 ϵ n λ . n_\lambda\sup_x\sum_{\mu:x\in S_\mu}a_\mu e^{n_\mu}\le2\epsilon n_\lambda. n λ x sup μ : x ∈ S μ ∑ a μ e n μ ≤ 2 ϵ n λ .
For 2 ϵ ≤ 1 2\epsilon\le1 2 ϵ ≤ 1 the candidate a λ e n λ a_\lambda e^{n_\lambda} a λ e n λ therefore bounds every iterate. Its sum rooted at a fluctuation block is at most 2 ϵ 2\epsilon 2 ϵ . Thus the polymer activities are absolutely summable with the weight e 10 ∣ S ∣ e^{10|S|} e 10∣ S ∣ .
Step 3: the second stage. We apply the finite graph inequality of Theorem 12.6 to this hard-core gas; only its bound on alternating sums over connected graphs is imported. Let z γ z_\gamma z γ be the absolute polymer activities and n γ = ∣ S γ ∣ n_\gamma= |S_\gamma| n γ = ∣ S γ ∣ . Step 2 gives sup x ∑ γ : x ∈ S γ z γ e 10 n γ ≤ 2 ϵ \sup_x \sum_{\gamma:x\in S_\gamma} z_\gamma e^{10n_\gamma} \le2\epsilon sup x ∑ γ : x ∈ S γ z γ e 10 n γ ≤ 2 ϵ . The second rooted-tree majorant is the increasing iteration from zero of
U γ = z γ e 4 n γ exp ( ∑ η : S η ∩ S γ ≠ ∅ U η ) , U γ ≤ z γ e 5 n γ . U_\gamma= z_\gamma e^{4n_\gamma}\exp\left(\sum_{\eta:S_\eta\cap S_\gamma\ne\varnothing} U_\eta\right), \qquad U_\gamma\le z_\gamma e^{5n_\gamma}. U γ = z γ e 4 n γ exp η : S η ∩ S γ = ∅ ∑ U η , U γ ≤ z γ e 5 n γ .
Indeed, under the displayed candidate bound the child sum is at most
n γ sup x ∑ η : x ∈ S η z η e 5 n η ≤ 2 ϵ n γ ≤ n γ ( 2 ϵ ≤ 1 ) , n_\gamma\sup_x \sum_{\eta:x\in S_\eta} z_\eta e^{5n_\eta} \le2\epsilon n_\gamma\le n_\gamma\qquad(2\epsilon\le1), n γ x sup η : x ∈ S η ∑ z η e 5 n η ≤ 2 ϵ n γ ≤ n γ ( 2 ϵ ≤ 1 ) ,
so the candidate is a supersolution and its sum rooted at a fluctuation block is at most 2 ϵ 2\epsilon 2 ϵ . This gives the fluctuation-rooted bound of part (2) with weight e 4 ∣ S ∣ e^{4|S|} e 4∣ S ∣ . Repeated polymer occurrences are allowed, as in the finite graph theorem. Thus the input weight is e 16 ∣ S ∣ e^{16|S|} e 16∣ S ∣ , the polymer activities have weight e 10 ∣ S ∣ e^{10|S|} e 10∣ S ∣ , and the logarithm has weight e 4 ∣ S ∣ e^{4|S|} e 4∣ S ∣ .
Step 4: the networks. We keep the resolution into monomials inside each exponential and each polymer. Within a connected family of the first stage, the networks of adjacent factors share a fluctuation position. At the second stage, adjacent polymers share a fluctuation block used by a factor in each of them, so their networks also meet at that position. By induction along a spanning tree at each stage, the union is connected and
ℓ ( N u n i o n ) ≤ ∑ λ o c c u r r e n c e s ℓ λ . (88) \ell(\mathcal{N}_{\mathrm{union}}) \le\sum_{\lambda\ \mathrm{occurrences}} \ell_\lambda. \tag*{(88)} ℓ ( N union ) ≤ λ occurrences ∑ ℓ λ . (88)
Coincident positions of different color coordinates have distance zero. Copies of the same monomial in an exponential have the same support and network. A visit to Ω c \Omega^c Ω c remains a visit in the union. Thus the network weights multiply, with no extra connecting length at either stage.
Step 5: the background root. Fix a background coordinate i i i in a coefficient of a product. The identity ν i = ∑ λ ν λ , i \nu_i = \sum_\lambda\nu_{\lambda,i} ν i = ∑ λ ν λ , i assigns the derivative in i i i to one monomial occurrence, with its multiplicity. We root the spanning tree of the first stage at this factor and, at the second stage, at the polymer containing it. The estimates for the children just proved bound the remaining part of the expansion by C e c n λ b λ C e^{c n_\lambda} b_\lambda C e c n λ b λ with c < 16 c < 16 c < 16 , so that the total support weight used is less than e 16 n λ e^{16n_\lambda} e 16 n λ . At the distinguished factor we use the degree-k k k term in D 16 D_{16} D 16 instead of its norm rooted at a fluctuation block. This works even if the background position is not a fluctuation position. The other factors are summed through their intersections of fluctuation supports.
Fixing i i i and summing over all output degrees at the intermediate profile, we obtain the bound C ∑ k d k ≤ C ε C\sum_k d_k \le C\varepsilon C ∑ k d k ≤ C ε , where d k d_k d k is the degree-k k k supremum in D 16 D_{16} D 16 for the distinguished monomial. Indeed, the remaining part of the expansion is summed at its fixed fluctuation root, and its total coefficient majorant there is bounded by the estimates for the children. We thus obtain a bound on sup i ∑ K \sup_i\sum_K sup i ∑ K at the intermediate profile, which is not yet the coefficient norm with boundary decay. Let q ∗ = K ∗ − 1 / 2 < 1 q_* = K_*^{-1/2} < 1 q ∗ = K ∗ − 1/2 < 1 . At output degree K K K , passing to the working profile multiplies every coefficient by q ∗ K q_*^K q ∗ K . Thus, taking the supremum separately at each degree,
∑ K ≥ 1 sup i { output coefficients of degree K rooted at i , at ϕ } ≤ C ε ∑ K ≥ 1 q ∗ K . \sum_{K\ge1}\sup_i\{\text{output coefficients of degree }K\text{ rooted at }i,\text{ at }\phi\}\le C\varepsilon\sum_{K\ge1}q_*^K. K ≥ 1 ∑ i sup { output coefficients of degree K rooted at i , at ϕ } ≤ C ε K ≥ 1 ∑ q ∗ K .
This is the required order of operations; a bound on sup i ∑ K \sup_i\sum_K sup i ∑ K at the same radius would not imply a bound on ∑ K sup i \sum_K\sup_i ∑ K sup i . The second half of the ratio between the profiles is used here, once for each local map on the fixed profiles. The factor ν i / K \nu_i/K ν i / K in the coefficient norm with boundary decay only decreases the multiplicity bound. Repeated copies in an exponential obey the same identity for ν i \nu_i ν i , and differentiating the exponential accounts for all possible distinguished occurrences.
Step 6: one and two changes. The tree estimates also hold in the rooted norms with changes. Differentiating (20.36) once gives
T λ ′ = e ∑ μ ∼ λ T μ a λ ′ + T λ ∑ μ ∼ λ T μ ′ . T'_\lambda=e^{\sum_{\mu\sim\lambda}T_\mu}a'_\lambda+T_\lambda\sum_{\mu\sim\lambda}T'_\mu. T λ ′ = e ∑ μ ∼ λ T μ a λ ′ + T λ μ ∼ λ ∑ T μ ′ .
where μ ∼ λ \mu\sim\lambda μ ∼ λ means S μ ∩ S λ ≠ ∅ S_\mu\cap S_\lambda\ne\varnothing S μ ∩ S λ = ∅ . In the norm rooted at a fluctuation block, the last operator has norm at most C ε < 1 / 2 C\varepsilon<1/2 C ε < 1/2 , since its incidence factor n λ ≤ e n λ n_\lambda\le e^{n_\lambda} n λ ≤ e n λ is bounded by a factor e n λ e^{n_\lambda} e n λ of the support weight that is not used in the output bound. Its inverse is a Neumann series with nonnegative terms and norm at most two. Hence a first variation is bounded by the norm of the changed source. Writing h λ = ∑ μ ∼ λ T μ h_\lambda=\sum_{\mu\sim\lambda}T_\mu h λ = ∑ μ ∼ λ T μ , the mixed variation satisfies
T λ , s r = e h λ a λ , s r + e h λ ( a λ , s h λ , r + a λ , r h λ , s ) + T λ h λ , s h λ , r + T λ ∑ μ ∼ λ T μ , s r . T_{\lambda,sr}=e^{h_\lambda}a_{\lambda,sr}+e^{h_\lambda}(a_{\lambda,s}h_{\lambda,r}+a_{\lambda,r}h_{\lambda,s})+T_\lambda h_{\lambda,s}h_{\lambda,r}+T_\lambda\sum_{\mu\sim\lambda}T_{\mu,sr}. T λ , sr = e h λ a λ , sr + e h λ ( a λ , s h λ , r + a λ , r h λ , s ) + T λ h λ , s h λ , r + T λ μ ∼ λ ∑ T μ , sr .
Its inhomogeneous terms are a source with both changes, products of the two first variations, and a source with one change times a first variation. Their bounds are C ( ϵ ∂ A + ϵ ∂ ϵ A ) C(\epsilon_{\partial A}+\epsilon_{\partial}\epsilon_A) C ( ϵ ∂ A + ϵ ∂ ϵ A ) , and the same inverse bounds the mixed variation. The polynomial incidence factors have degree at most two, and sup n ≥ 1 n 2 e − n < ∞ \sup_{n\ge1}n^2e^{-n}<\infty sup n ≥ 1 n 2 e − n < ∞ . In the bilinear interpolation, terms containing the family C C C are bounded by a constant times ϵ ∂ A \epsilon_{\partial A} ϵ ∂ A , since the other monomial norms are bounded on the fixed small ball. Integrating the mixed derivative over the unit square, we obtain the four-corner difference, with the same bound. These calculations apply at both stages of the expansion. For a background root, we first sum the resulting rooted expressions at the distinguished factor at the intermediate profile, with the sequence d k d_k d k described above, and then apply the factor q K ∗ q^*_K q K ∗ at each output degree. If the background root lies on a factor different from both changes, the remaining expansion has at most one additional distinguished branch. Its cubic incidence factor is bounded by a further factor e n λ e^{n\lambda} e nλ of the support weight, since sup n ≥ 1 n 3 e − n < ∞ \sup_{n\geq1} n^3e^{-n}<\infty sup n ≥ 1 n 3 e − n < ∞ ; the passages from e 16 ∣ S ∣ e^{16|S|} e 16∣ S ∣ to e 10 ∣ S ∣ e^{10|S|} e 10∣ S ∣ and from e 10 ∣ S ∣ e^{10|S|} e 10∣ S ∣ to e 4 ∣ S ∣ e^{4|S|} e 4∣ S ∣ leave these factors available. The source with both changes keeps its incoming factor, and the product source contains the first variation in the incoming insertion. This proves (86) in both rooted norms, including the case of a background root on an unchanged factor.
Step 7: spatial conversion. We keep the joined networks. By the hypothesis on B B B and C C C , every term containing the boundary change carries a visit to Ω L J c \Omega^c_{L_J} Ω L J c . If a coefficient with unfixed background root i i i carries such a visit, then by part (5) of Lemma 20.1
a t ( ν ) + α d Ω ( i ) ≤ ( a + α ) ℓ ( N u n i o n ) + C J α . a t(\nu)+\alpha d_{\Omega}(i)\leq(a+\alpha)\ell(\mathcal{N}_{\mathrm{union}})+C_J\alpha. a t ( ν ) + α d Ω ( i ) ≤ ( a + α ) ℓ ( N union ) + C J α .
Since s 0 ≥ a + α s_0\geq a+\alpha s 0 ≥ a + α , the product of the weights e s 0 ℓ λ e^{s_0\ell_\lambda} e s 0 ℓ λ in b λ b_\lambda b λ dominates e ( a + α ) ℓ ( N u n i o n ) e^{(a+\alpha)\ell(\mathcal{N}_{\mathrm{union}})} e ( a + α ) ℓ ( N union ) , by (88) . The fixed buffer is used only at the final visit to Ω c \Omega^c Ω c . The factor ν i / K \nu_i/K ν i / K in (4.11) is at most the multiplicity of the distinguished index. The bound at the intermediate profile and the degree sum ∑ K ≥ 1 q K ∗ \sum_{K\geq1}q^*_K ∑ K ≥ 1 q K ∗ just proved, multiplied by the fixed factor ρ − 2 \rho^{-2} ρ − 2 and the buffer constant, give the stated bounds on X φ X^\varphi X φ .
Step 8: the logarithm. The bounds on the coefficientwise majorants prove the normal convergence of every resolved series and of the displayed derivatives. For a small multiple of the interaction, the two series are the Taylor expansion of the logarithm of the finite product integral. Their exponential agrees with the integral throughout the strict ball, by the identity theorem in that multiplier. Hence the integral does not vanish there. The uniform absolute bounds allow us to remove the truncations to finitely many monomials and to differentiate termwise in the parameters, and they prove holomorphy in the absolute coefficient norms. We have not used the hypothesis on background supports in the holomorphy theorem of Section 12. □
Theorem 20.9. Fix the torus rate a T > 0 a_T>0 a T > 0 and the ratio K ∗ K_* K ∗ between the profiles ϕ ^ = K ∗ ϕ \hat{\phi}=K_*\phi ϕ ^ = K ∗ ϕ and ψ p r o f = K ∗ ϕ \psi_{\mathrm{prof}}=\sqrt{K_*}\phi ψ prof = K ∗ ϕ . There are geometric constants C a l l o c ≥ C g e o m C_{\mathrm{alloc}}\geq C_{\mathrm{geom}} C alloc ≥ C geom and C g ≥ 2 C_g\geq2 C g ≥ 2 , independent of J J J , with the following property; here C g e o m C_{\mathrm{geom}} C geom includes the transport allocation of Theorem 19.9 (4). Choose C ∗ g e o m ≥ C g C_*^{\mathrm{geom}}\geq C_g C ∗ geom ≥ C g and choose J J J once with
B = M J ≥ B g e o m , where B g e o m = 256 C a l l o c ( C ∗ g e o m + 1 ) . B=M^J\geq B_{\mathrm{geom}},\qquad\text{where } B_{\mathrm{geom}}=256C_{\mathrm{alloc}}\left(C_*^{\mathrm{geom}}+1\right). B = M J ≥ B geom , where B geom = 256 C alloc ( C ∗ geom + 1 ) . After J J J has been fixed, choose positive rates
a = a c l = a s p , α = α c l = a / 16 , m r a t e = C ∗ g e o m ( a + α ) , b 1 = 8 C ∗ g e o m α a=a_{\mathrm{cl}}=a_{\mathrm{sp}},\qquad\alpha=\alpha_{\mathrm{cl}}=a/16,\qquad m_{\mathrm{rate}}=C_*^{\mathrm{geom}}(a+\alpha),\qquad b_1=8C_*^{\mathrm{geom}}\alpha a = a cl = a sp , α = α cl = a /16 , m rate = C ∗ geom ( a + α ) , b 1 = 8 C ∗ geom α
so that
m r a t e + b 1 < min ( c p r i m , J , B a ) / C a l l o c , b 1 < B α / C a l l o c , 3 ( a + α ) < a T , m_{\mathrm{rate}}+b_1<\min(c_{\mathrm{prim},J},B_a)/C_{\mathrm{alloc}},\qquad b_1<B_\alpha/C_{\mathrm{alloc}},\qquad\sqrt{3}(a+\alpha)<a_{\mathrm{T}}, m rate + b 1 < min ( c prim , J , B a ) / C alloc , b 1 < B α / C alloc , 3 ( a + α ) < a T ,
where c p r i m , J c_{\mathrm{prim},J} c prim , J is the fixed positive reserve for the intrinsic tensors and the unperturbed weighted reconstruction images, in the graph metric of § 2.2.1 , chosen no larger than the intrinsic exponent c J c_J c J of Theorem 19.9 . The inherited perturbative networks use the separate inequalities against B a B_a B a and B α B_\alpha B α from Theorem 19.9 ; their images satisfy the half-radius bound at the allocated working rates.
Consider a full renormalization group step (s = J s=J s = J ) on admissible data. Sum the internal indices first, keep networks through all positions of the monomials, and use the exact nonempty fluctuation supports of Definition 20.6 in both expansions. All factors and polymer occurrences that are joined share a fluctuation position; the union of their networks keeps every background index and every visit to Ω c \Omega^c Ω c , and adds no length. Set p = max i p i ≥ 1 p=\max_i p_i\geq1 p = max i p i ≥ 1 , x = X ρ ( f ) x=\mathcal{X}^{\rho}(f) x = X ρ ( f ) , and H = E + x H=E+x H = E + x , and substitute the complex parameter z z z for the coupling powers as described before Definition 20.6 , with the insertion A R ( z ) A_R(z) A R ( z ) . Let Ξ 0 \Xi_0 Ξ 0 and Ξ R \Xi_R Ξ R be the integrals of e − U e^{-U} e − U against the product of normalized symmetric reference measures on balls, for the interaction U U U without insertion and for U U U plus the insertion, respectively. For the matched torus potential, with the fixed background slots set to zero, define its restriction by deleting the monomials that use a missing fluctuation block, and adjoin the missing blocks with the identical normalized reference measures.
There are ϵ J , c z , J > 0 \epsilon_J,c_{z,J}>0 ϵ J , c z , J > 0 and C J < ∞ C_J<\infty C J < ∞ such that the following holds if
s ( z ) = C J ∣ z ∣ ( p 3 + p + H p ) 1 − C J ∣ z ∣ p ≤ ϵ J , C J ∣ z ∣ p < 1. s(z)=\frac{C_J|z|(p^3+p+Hp)}{1-C_J|z|^p}\leq\epsilon_J,\qquad C_J|z|^p<1. s ( z ) = 1 − C J ∣ z ∣ p C J ∣ z ∣ ( p 3 + p + H p ) ≤ ϵ J , C J ∣ z ∣ p < 1.
The integrals without and with insertion, the restricted torus integrals, and the linearly interpolated integrals do not vanish, and their logarithms are the normally convergent expansions of Lemma 20.8 . The regional and restricted families satisfy W 16 + D 16 ≤ C J s ( z ) W_{16}+D_{16}\leq C_Js(z) W 16 + D 16 ≤ C J s ( z ) ; their boundary-difference families satisfy the same bound with networks visiting Ω c \Omega^c Ω c . The comparisons of the nonconstant parts of the logarithms, including the deleted torus terms with an unfixed background index, satisfy
X ϕ ( log Ξ ∙ Ω − ( log Ξ ∙ T ) ) ≤ C J s ( z ) , ∙ ∈ { 0 , R } , \mathcal{X}^{\phi}\left(\log\Xi^\Omega_{\bullet}-\left(\log\Xi^{\mathbb{T}}_{\bullet}\right)\right)\leq C_Js(z),\qquad\bullet\in\{0,R\}, X ϕ ( log Ξ ∙ Ω − ( log Ξ ∙ T ) ) ≤ C J s ( z ) , ∙ ∈ { 0 , R } ,
where the bar denotes zero extension of the fixed background slots; constants supported in Ω c \Omega^c Ω c remain separate. The mixed difference of the boundary change and the insertion is bounded by C J ( ϵ ∂ A + ϵ ∂ ϵ A ) C_J(\epsilon_{\partial A}+\epsilon_{\partial}\epsilon_A) C J ( ϵ ∂ A + ϵ ∂ ϵ A ) in the norms of Definition 20.6 , and every term keeps an incoming factor. The family extends holomorphically to z = 0 z=0 z = 0 , and its compared insertion logarithm F F F satisfies F ( 0 ) = F ′ ( 0 ) = 0 F(0)=F'(0)=0 F ( 0 ) = F ′ ( 0 ) = 0 . In particular, if
r z = c z , J p 3 + p + H p , 0 < t max ≤ r z / 2 , r_z=\frac{c_{z,J}}{p^3+p+Hp},\qquad0<t_{\max}\leq r_z/2, r z = p 3 + p + H p c z , J , 0 < t m a x ≤ r z /2 ,
then
X ϕ ( D Ω R − ( D T r d 0 ) ) ≤ C J t max 2 H ( p 4 + H p 2 ) , D R = − log Ξ R + log Ξ 0 . \mathcal{X}^{\phi}\left(D^\Omega R-\left(D^{\mathbb{T}}r_{d_0}\right)\right)\leq C_Jt_{\max}^2H(p^4+Hp^2),\qquad DR=-\log\Xi_R+\log\Xi_0. X ϕ ( D Ω R − ( D T r d 0 ) ) ≤ C J t m a x 2 H ( p 4 + H p 2 ) , D R = − log Ξ R + log Ξ 0 .
After the whole boundary layer with reduced radius has been frozen as in Lemma 18.9, the classical and quantum comparisons return the same rates a , α a,\alpha a , α and the same input radii; the torus rate remains a T a_{\mathrm{T}} a T . The constants have the dependence described in item (3) of Definition 4.14 and are independent of the later widths. The return of the same profile is asserted only for full steps with B ≥ B g e o m B \ge B_{\mathrm{geom}} B ≥ B geom , not for compositions of fewer than J J J averaging steps.
Proof. The constants C a l l o c C_{\mathrm{alloc}} C alloc and C g ≥ 2 C_{\mathrm{g}} \ge2 C g ≥ 2 include the fixed contraction and transport allocation needed to retain a strict positive reserve above s 0 = a + α s_0 = a + \alpha s 0 = a + α in the constituent difference rows; in particular m r a t e ≥ 2 s 0 m_{\mathrm{rate}} \ge2s_0 m rate ≥ 2 s 0 . They bound the contraction with every index rooted and the transport of the input boundary weight in Theorem 19.9 and Lemma 18.15; they include the factor 3 \sqrt{3} 3 of the conversion between Euclidean and graph exponents in § 2.2.1 , and hence do not depend on J J J . The torus coefficients, bounded in B ρ , a T \mathcal{B}_{\rho,a_{\mathrm{T}}} B ρ , a T with Euclidean network lengths, have graph exponent a T / 3 > a + α a_{\mathrm{T}}/\sqrt{3} > a + \alpha a T / 3 > a + α by the last inequality in the statement. With α = a / 16 \alpha= a/16 α = a /16 we have m r a t e + b 1 = ( 25 / 16 ) C g e o m ∗ a m_{\mathrm{rate}} + b_1 = (25/16)C_{\mathrm{geom}}^{*}a m rate + b 1 = ( 25/16 ) C geom ∗ a , and both restrictions against B a B a B a and B α B \alpha B α follow from the lower bound on B B B . At this fixed J J J , we decrease a > 0 a > 0 a > 0 to satisfy the bounds on the exponents of the coefficients of the potential, of the weighted images, and of the torus, including 3 ( a + α ) < a T \sqrt{3}(a+\alpha) < a_{\mathrm{T}} 3 ( a + α ) < a T . No constant depending on J J J is used to choose J J J . The comparison of the classical boundary networks uses the same allocation of the decay rate and no expansion with respect to a reference product. At zero fluctuation it works as follows. Consider a classical comparison, or the comparison of the pullback terms, with the boundary tensors rooted at every index. At background degree k ≥ 1 k \ge1 k ≥ 1 , its rooted absolute sum on ϕ ^ \hat{\phi} ϕ ^ is bounded by its tensor majorant for fixed J J J , and that allocation already accounts for the weights of its output network and of its distance to Ω c \Omega^{\mathrm{c}} Ω c . Passing to ϕ \phi ϕ gives the factor K ∗ − k K_*^{-k} K ∗ − k . Summing over the degree, including any fixed polynomial factor for the choice of a distinguished index, we obtain the bound in the coefficient norm with boundary decay. For a pulled-back incoming boundary term, we use its jet majorant at half the radius and its distinguished input columns before this calculation. Thus the classical coefficients are converted without assigning a fluctuation support to a term at zero fluctuation, and these terms never enter either expansion.
We apply Lemma 20.7 at these rates and then Lemma 20.8 . The hypotheses of the latter on the rates and on the boundary-change families hold, since Lemma 20.7 uses s 0 = a + α s_0 = a + \alpha s 0 = a + α and gives the boundary differences with their networks visiting Ω c \Omega^{\mathrm{c}} Ω c . Restricting a torus family decreases its norms, which are sums of absolute values. For its difference from the full family, every deleted coefficient with an unfixed background root visits a missing fluctuation position and hence Ω c \Omega^{\mathrm{c}} Ω c , and the same lemma applies. Pure background constants without an unfixed index have zero norm with boundary decay and stay outside this comparison. The estimate with two changes implies the comparison of the logarithms with and without insertion. The logarithms and their derivatives in the parameters are determined by the finite identities and normal convergence proved above, not by a choice of a complex logarithm after integration.
We choose c z , J c_{z,J} c z , J so that the disc ∣ z ∣ ≤ r z |z| \le r_z ∣ z ∣ ≤ r z lies strictly inside the margins of Lemma 20.7, of the free product, and of the logarithm. On this disc, (20.33) and the bounds with one and two changes give
∥ F ( z ) ∥ X ϕ ≤ C J ∣ z ∣ H p . \|F(z)\|_{\mathcal{X}^{\phi}} \le C_J|z|Hp. ∥ F ( z ) ∥ X ϕ ≤ C J ∣ z ∣ H p .
At z = 0 z = 0 z = 0 every fluctuating monomial vanishes. The coefficient of z z z is cubic in the fluctuation variables for the classical terms and linear for the quantum, Haar density, determinant, and incoming terms. Each of these is odd under the simultaneous reversal w ↦ − w w \mapsto-w w ↦ − w , so its expectation under the symmetric reference measures on balls vanishes. At first order the logarithm contains a single monomial occurrence, hence F ( 0 ) = F ′ ( 0 ) = 0 F(0) = F'(0) = 0 F ( 0 ) = F ′ ( 0 ) = 0 coefficientwise, at every real frozen value.
The normally convergent series F ( z ) / z F(z)/z F ( z ) / z is holomorphic, vanishes at zero, and has norm at most C J H p C_JHp C J H p . Cauchy’s coefficient bound on ∣ z ∣ = r < r z |z| = r < r_z ∣ z ∣ = r < r z , followed by r ↑ r z r \uparrow r_z r ↑ r z , gives for t ≤ r z / 2 t \le r_z/2 t ≤ r z /2
∥ F ( t ) ∥ X ϕ ≤ 2 C J t 2 H p / r z ≤ C J t 2 H ( p 4 + H p 2 ) . \|F(t)\|_{\mathcal{X}^{\phi}} \le2C_Jt^2Hp/r_z \le C_Jt^2H(p^4+Hp^2). ∥ F ( t ) ∥ X ϕ ≤ 2 C J t 2 H p / r z ≤ C J t 2 H ( p 4 + H p 2 ) .
where in the last step we used p ≥ 1 p \ge1 p ≥ 1 . The integral proof of the coefficient bound applies to the supremum over the fixed index at every background degree, and then to their sum, by the triangle inequality.
Finally, Lemma 18.9 restores the incoming radii when the whole zone with reduced radius is frozen with the stated bounds on the frozen values. All conversions of coefficients were made on the fixed outer and intermediate profiles, and none is iterated at later steps. This proves the return of the same profile and the stated independence of the constants.
Effective interactions and normalization constants
Theorem 20.10. Consider a full renormalization group step, s = J s = J s = J , on admissible data as in Theorem 20.9. Set
P Ω R ( v ) = R ( Log C Ω ( v , 0 ; b r e c ) ) , Q Ω = B r e f Ω − ∑ i f i log z p i − log Ξ 0 Ω , P^\Omega R(v) = R\left(\operatorname{Log} C^\Omega(v,0;b_{\mathrm{rec}})\right), \qquad Q^\Omega= B^\Omega_{\mathrm{ref}}-\sum_i f_i\log z_{p_i}-\log\Xi^\Omega_0, P Ω R ( v ) = R ( Log C Ω ( v , 0 ; b rec ) ) , Q Ω = B ref Ω − i ∑ f i log z p i − log Ξ 0 Ω , B r e f = ∑ i B i ( c ^ i + 1 ) , B_{\mathrm{ref}} = \sum_i B_i(\widehat{c}_{i+1}), B ref = i ∑ B i ( c i + 1 ) , where f i f_i f i is the number of free blocks at scale i i i and z p = ∫ ∣ w ∣ ≤ p e − ∣ w ∣ 2 / 2 d 3 w z_p = \int_{|w|\le p} e^{-|w|^2/2}\,d^3w z p = ∫ ∣ w ∣ ≤ p e − ∣ w ∣ 2 /2 d 3 w , in the convention of Theorem 12.13. The field-independent factors of the block averaging constraint and of the Haar measure extracted from L i L_i L i are kept in Λ \Lambda Λ . Up to the exact constant kept in Λ \Lambda Λ , the new effective interaction M Ω R M^\Omega R M Ω R produced by the step satisfies
M Ω R = Q Ω + P Ω R + D Ω R . M^\Omega R = Q^\Omega+ P^\Omega R + D^\Omega R. M Ω R = Q Ω + P Ω R + D Ω R .
Here Q Ω Q^{\Omega} Q Ω is the newly generated term, P Ω R P^{\Omega}R P Ω R is the pullback of the incoming interaction by the zero-fluctuation reconstruction, and D Ω R D^{\Omega}R D Ω R is the fluctuation correction of Theorem 20.9 . For the matched torus reference with zero extension of the fixed slots, there are constants A J A_J A J , K J K_J K J , and C J C_J C J , independent of the later widths, such that
X ϕ ( Q Ω − Q T ∣ Ω ′ + P Ω ( r d 0 ∣ Ω ) − ( P T r d 0 ) ∣ Ω ′ ) ≤ A J ( 1 + E ) , \mathcal{X}^{\phi}\left(Q^{\Omega}-Q^{\mathsf{T}}|_{\Omega'}+P^{\Omega}(r_{d_0}|\Omega)-(P^{\mathsf{T}}r_{d_0})|_{\Omega'}\right)\le A_J(1+E), X ϕ ( Q Ω − Q T ∣ Ω ′ + P Ω ( r d 0 ∣Ω ) − ( P T r d 0 ) ∣ Ω ′ ) ≤ A J ( 1 + E ) ,
X ϕ ( P Ω f ) ≤ K J X ϕ ( f ) . \mathcal{X}^{\phi}(P^{\Omega}f)\le K_J\mathcal{X}^{\phi}(f). X ϕ ( P Ω f ) ≤ K J X ϕ ( f ) .
Here Ω ′ \Omega' Ω ′ is the set of unfixed scalar output labels. Assume that the column bounds of Theorem 19.6 and their perturbed versions hold on the product with doubled output analytic radii, and that the image under the reconstruction majorant on this product is at most half the incoming radius, as in Lemma 19.7 . Then one may take K J = 4 ( log 2 ) K c o l , J K_J=4(\log2)K_{\mathrm{col},J} K J = 4 ( log 2 ) K col , J . Here K c o l , J K_{\mathrm{col},J} K col , J bounds, for one fixed output, the sum over all old input labels, weighted by their boundary decay, of the distinguished derivative series of the perturbed zero-fluctuation reconstruction; it is chosen uniformly on the preliminary perturbation ball by Theorem 19.9 . The degree-one terms and every nonconstant dependence of an omitted torus density on the unfixed labels are included.
At every admissible complex value of the coarse background and every fixed real admissible frozen value, let I i I_i I i and H i H_i H i be the sets of unfixed and of missing scalar free labels at scale i i i . With matched torus arguments extended by zero,
log det K i Ω = ∑ x ∈ I i ( log K i T ) x x + B i , ∣ ( log K i T ) x x ∣ ≤ C J , ∣ B i ∣ ≤ C J ∣ ∂ b l k Ω ∣ . \log\det K_i^{\Omega}=\sum_{x\in I_i}(\log K_i^{\mathsf{T}})_{xx}+B_i,\qquad|(\log K_i^{\mathsf{T}})_{xx}|\le C_J,\qquad|B_i|\le C_J|\partial_{\mathrm{blk}}\Omega|. log det K i Ω = x ∈ I i ∑ ( log K i T ) xx + B i , ∣ ( log K i T ) xx ∣ ≤ C J , ∣ B i ∣ ≤ C J ∣ ∂ blk Ω∣.
These determinant statements also hold in the coefficient norm before evaluation. Scalar determinant labels and fluctuation blocks are counted separately. All constants have the dependence described in item (3) of Definition 4.14 and are independent of the later widths.
Proof. Every step here is a full step, s = J s=J s = J ; a final composition of fewer than J J J averaging steps that is evaluated deterministically adds no fluctuation integral.
Step 1: the exact factorization. The triangular calculation in the proof of Lemma 20.3 gives, at a fixed coarse background and fixed frozen values, the value of the integral
( ∏ i J i , r e f g i 3 f i ) e − F o u t / g s 2 − B r e f − R ( log C ( v , 0 ; b r e c ) ) ∏ i Z p i f i Ξ R ( v ) , (89) \left(\prod_i \mathcal{J}_{i,\mathrm{ref}}\mathfrak{g}_i^{3f_i}\right)e^{-F_{\mathrm{out}}/g_s^2-B_{\mathrm{ref}}-R(\log C(v,0;b_{\mathrm{rec}}))}\prod_i \mathcal{Z}_{p_i}^{f_i}\Xi_R(v), \tag*{(89)} ( i ∏ J i , ref g i 3 f i ) e − F out / g s 2 − B ref − R ( l o g C ( v , 0 ; b rec )) i ∏ Z p i f i Ξ R ( v ) , (89)
where F o u t / g s 2 F_{\mathrm{out}}/g_s^2 F out / g s 2 is the classical value of the step with the output coupling, and the J i , r e f \mathcal{J}_{i,\mathrm{ref}} J i , ref are the extracted constant Haar disintegration factors. Inserting the definitions of Q Q Q , P P P , and D D D into the exponent of Eq. (89) , and writing log Ξ R = log Ξ 0 − D R \log\Xi_R=\log\Xi_0-DR log Ξ R = log Ξ 0 − D R , we obtain M R = Q + P R + D R MR=Q+PR+DR M R = Q + P R + D R . This proves the recursion, including its signs and its Gaussian convention. Centering moves the value at zero into Λ \Lambda Λ and changes no factor of this identity.
Step 2: the generated term and the torus pullback. The deterministic part of Q Ω − Q T Q^{\Omega}-Q^{\mathrm{T}} Q Ω − Q T is the difference of the Haar densities at the minimizers and of the diagonal matrix-logarithm densities along the two zero-fluctuation chains. Their rooted comparisons with boundary decay are part (3) of Theorem 20.2 . A density based at a missing torus label is not discarded: a nonconstant coefficient with an unfixed index has a network to that missing label, and its sum over the missing label is bounded from the unfixed index by (53) . Its constant is left in Λ \Lambda Λ . The terms log z p i \log z_{p_i} log z p i and the constant Haar factors have no nonconstant background coefficient. The difference of the logarithms without insertion has the same fixed rates by Theorem 20.9 , with a bounded norm in the cluster expansion. We obtain a finite bound C J C_J C J for the comparison of the nonconstant generated terms.
For the torus pullback, denote by a Ω a_{\Omega} a Ω the zero extension of Log C Ω ( v , 0 ; b r e c ) \operatorname{Log} C^{\Omega}(v,0;b_{\mathrm{rec}}) Log C Ω ( v , 0 ; b rec ) and by a T a_{\mathrm{T}} a T the matched torus argument; in this paragraph a T a_{\mathrm{T}} a T is an argument, not the torus rate. We have
r d 0 ( a Ω ) − r d 0 ( a T ) = ∫ 0 1 D r d 0 ( a T + τ ( a Ω − a T ) ) [ a Ω − a T ] d τ . (90) r_{d_0}(a_{\Omega})-r_{d_0}(a_{\mathrm{T}})=\int_{0}^{1}D r_{d_0}\left(a_{\mathrm{T}}+\tau(a_{\Omega}-a_{\mathrm{T}})\right)[a_{\Omega}-a_{\mathrm{T}}]\,\mathrm{d}\tau. \tag*{(90)} r d 0 ( a Ω ) − r d 0 ( a T ) = ∫ 0 1 D r d 0 ( a T + τ ( a Ω − a T ) ) [ a Ω − a T ] d τ . (90)
Both absolute images are below half the incoming radius, so the whole segment admits the coefficient expansion of the integrand. On an unfixed old input, the difference of the arguments is the unperturbed defect of the reconstruction followed by the gauge transformation (Theorem 19.6 ) plus the defect of the perturbation (Theorem 19.9 ). On a fixed old input, the difference is a visit to Ω c \Omega^{c} Ω c . The action and interaction tensors join every output index to that visit, even if the source does not decay inside its own component of Ω c \Omega^{c} Ω c . The jet bound (80) with σ 0 = 0 \sigma_{0}=0 σ 0 = 0 bounds the integrand by C J E C_J E C J E , with its visit. The conversion of networks and degrees in the proof of Theorem 20.9 therefore bounds its X φ \mathcal{X}^{\varphi} X φ norm by C J E C_J E C J E . Together these bounds prove the first inequality, with A J A_J A J fixed on the larger products. No first derivative of the output was canceled.
Step 3: the pullback of the boundary function. For the zero-fluctuation reconstruction Φ e ( v ) = Log C e ( v , 0 ; b r e c ) \Phi_e(v)=\operatorname{Log} C_e(v,0;b_{\mathrm{rec}}) Φ e ( v ) = Log C e ( v , 0 ; b rec ) , we define
C e , x ( 2 ϕ ) = ∑ ν ∣ ν x ∣ ∣ Φ e , ν ∣ ( 2 ϕ ) e a 1 t ( { e } ∪ supp ν ) . C_{e,x}(2\phi)=\sum_{\nu}\lvert\nu_x\rvert\lvert\Phi_{e,\nu}\rvert(2\phi)e^{a_1t(\{e\}\cup\operatorname{supp}\nu)}. C e , x ( 2 ϕ ) = ν ∑ ∣ ν x ∣ ∣ Φ e , ν ∣ ( 2 ϕ ) e a 1 t ({ e } ∪ supp ν ) .
All constants in this step refer to the perturbed map. Let D J 0 D_J^0 D J 0 bound the unperturbed distinguished series on the larger product with doubled output analytic radius. The difference of the reconstructions in Theorem 19.9 , rooted at every index, with the uniform incoming norm and the inequalities on the conditional forces checked above, bounds the additional distinguished series by σ 0 D J 1 \sigma_0D_J^1 σ 0 D J 1 there. We fix a spatial decay rate b > α b>\alpha b > α and define
D J a c t = D J 0 + σ 0 D J 1 , C e , x ( 2 ϕ ) ≤ D J a c t ( 2 ϕ x ) e − b dist ( e , x ) . (91) D_J^{\mathrm{act}}=D_J^0+\sigma_0D_J^1,\qquad C_{e,x}(2\phi)\le D_J^{\mathrm{act}}(2\phi_x)e^{-b\operatorname{dist}(e,x)}. \tag*{(91)} D J act = D J 0 + σ 0 D J 1 , C e , x ( 2 ϕ ) ≤ D J act ( 2 ϕ x ) e − b dist ( e , x ) . (91)
The factor 2 ϕ x 2\phi_x 2 ϕ x appears because the input index is distinguished before the sum of absolute coefficients is taken. The rates are below the common exponent available in the coefficient bounds of the potential, and both D J 0 D^0_J D J 0 and D J 1 D^1_J D J 1 are computed at the preliminary bound on fixed larger products, before the later widths. Here and below, for an old input label e e e of R R R (a label of the incoming set D i n D_{\mathrm{in}} D in , on the input lattice of the step), we write d o l d ( e ) = d D i n ( e ) d_{\mathrm{old}}(e)=d_{D_{\mathrm{in}}}(e) d old ( e ) = d D in ( e ) for its distance to the incoming fixed complement, measured in units of the input lattice as in the incoming norm. For a current unfixed background input x x x , we write d o u t ( x ) = d Ω ( x ) d_{\mathrm{out}}(x)=d_{\Omega}(x) d out ( x ) = d Ω ( x ) for its distance to Ω c \Omega^c Ω c , in units of the output lattice. Finally, dist ( e , x ) \operatorname{dist}(e,x) dist ( e , x ) is the distance between x x x and the position e / B S e/B_S e / B S of e e e , in units of the output lattice. Applying (59) on the old input set of R R R , which is the same set of unfixed labels, with quantum radius ρ \rho ρ , and dividing by B S B_S B S , we obtain
d o u t ( x ) ≤ d o l d ( e ) / B S + dist ( e , x ) + C b u f , J / B S . d_{\mathrm{out}}(x) \le d_{\mathrm{old}}(e)/B_S+\operatorname{dist}(e,x)+C_{\mathrm{buf},J}/B_S. d out ( x ) ≤ d old ( e ) / B S + dist ( e , x ) + C buf , J / B S .
In Theorem 19.6 the distance d o l d d_{\mathrm{old}} d old is measured in units of the output lattice, and the inequality d o u t ( x ) ≤ d o l d ( e ) + dist ( e , x ) + C J d_{\mathrm{out}}(x)\le d_{\mathrm{old}}(e)+\operatorname{dist}(e,x)+C_J d out ( x ) ≤ d old ( e ) + dist ( e , x ) + C J is an explicit hypothesis of that theorem. With d o l d d_{\mathrm{old}} d old there equal to our d o l d ( e ) / B S d_{\mathrm{old}}(e)/B_S d old ( e ) / B S , the inequality just displayed verifies this hypothesis, with buffer C b u f , J / B S C_{\mathrm{buf},J}/B_S C buf , J / B S ; it holds because the incoming fixed complement lies on the current fixed side. Since B S ≥ 1 B_S\ge1 B S ≥ 1 , this also gives the weaker comparison with d o l d ( e ) d_{\mathrm{old}}(e) d old ( e ) in place of d o l d ( e ) / B S d_{\mathrm{old}}(e)/B_S d old ( e ) / B S . The density of the old labels then implies
∑ e o l d u n f i x e d e − α d o l d ( e ) C e , x ( 2 ϕ ) ≤ K c o l , J ρ e − α d o u t ( x ) . (92) \sum_{e\ \mathrm{old\ unfixed}} e^{-\alpha d_{\mathrm{old}}(e)}C_{e,x}(2\phi)\le K_{\mathrm{col},J}\rho e^{-\alpha d_{\mathrm{out}}(x)}. \tag*{(92)} e old unfixed ∑ e − α d old ( e ) C e , x ( 2 ϕ ) ≤ K col , J ρ e − α d out ( x ) . (92)
We define K c o l , J K_{\mathrm{col},J} K col , J as a common bound for this inequality, uniformly on the preliminary perturbation ball. One admissible choice is any number at least 2 D J a c t e α C b u f , J / B S S J ( b − α ) 2D^{\mathrm{act}}_J e^{\alpha C_{\mathrm{buf},J}/B_S}S_J(b-\alpha) 2 D J act e α C buf , J / B S S J ( b − α ) , with the fixed volume factors included in S J S_J S J , and at least the unperturbed column bound; we take the finite maximum over the label types and the stages. The absolute image on the product with doubled output analytic radius is at most ρ / 2 \rho/2 ρ /2 . The construction used a ratio larger than two between the profiles, so no repeated reduction of radii is involved. This is the column calculation of Theorem 19.6 , with its constant enlarged by (91) . The hypotheses on the doubled product and on the incoming image at half the radius are used here.
Let x f = X ρ ( f ) x_f=\mathcal{X}^{\rho}(f) x f = X ρ ( f ) . At an old input root e e e , the distinguished derivative sum evaluated at ρ / 2 \rho/2 ρ /2 satisfies
∑ η : η e > 0 η e ∣ f η ∣ ( ρ / 2 ) ∣ η ∣ − 1 e α t ( supp η ) ≤ 4 ρ x f e − α d o l d ( e ) . (93) \sum_{\eta:\eta_e>0}\eta_e|f_\eta|(\rho/2)^{|\eta|-1}e^{\alpha t(\operatorname{supp}\eta)}\le4\rho x_f e^{-\alpha d_{\mathrm{old}}(e)}. \tag*{(93)} η : η e > 0 ∑ η e ∣ f η ∣ ( ρ /2 ) ∣ η ∣ − 1 e α t ( supp η ) ≤ 4 ρ x f e − α d old ( e ) . (93)
For degree l l l this follows by multiplying the normalized coefficient with this input fixed by 1 2 1 − l ρ 1^{2^{1-l}\rho} 1 2 1 − l ρ , which is at most 4 ρ 4\rho 4 ρ ; the case l = 1 l=1 l = 1 is included. The chain rule for coefficientwise majorants, the union of networks, and (92) bound the distinguished coefficient sum of the output at 2 ϕ 2\phi 2 ϕ by 4 K c o l , J ρ 2 x f e − α d o u t ( x ) 4K_{\mathrm{col},J}\rho^2x_f e^{-\alpha d_{\mathrm{out}}(x)} 4 K col , J ρ 2 x f e − α d out ( x ) . At degree k k k the fixed-index weight of the coefficient norm with boundary decay inserts the factor 1 / k 1/k 1/ k , and evaluation at ϕ \phi ϕ instead of 2 ϕ 2\phi 2 ϕ inserts 2 − k 2^{-k} 2 − k . We take the supremum over the fixed index at each degree and then sum. Since integrating ∑ j ≥ 0 u j = ( 1 − u ) − 1 \sum_{j\ge0}u^j=(1-u)^{-1} ∑ j ≥ 0 u j = ( 1 − u ) − 1 over [ 0 , 1 / 2 ] [0,1/2] [ 0 , 1/2 ] gives ∑ k ≥ 1 2 − k / k = log 2 \sum_{k\ge1}2^{-k}/k=\log2 ∑ k ≥ 1 2 − k / k = log 2 , we obtain
X ϕ ( D Ω f ) ≤ 4 ( log 2 ) K c o l , J X f . \mathcal{X}^{\phi}(\mathbb{D}^{\Omega}f)\le4(\log2)K_{\mathrm{col},J}\mathcal{X}_f. X ϕ ( D Ω f ) ≤ 4 ( log 2 ) K col , J X f .
This also shows that a finite K J K_J K J exists when no particular column constant is specified.
Step 4: the logarithmic determinants. The diagonal coefficient series of log K i T \log K_i^{\mathsf{T}} log K i T is bounded uniformly for each scalar label. To bound the constant as well as the derivatives, we write
log K = ∫ 0 ∞ [ ( 1 + t ) − 1 I − ( K + t ) − 1 ] d t . \log K=\int_{0}^{\infty}\left[(1+t)^{-1}I-(K+t)^{-1}\right]\,dt. log K = ∫ 0 ∞ [ ( 1 + t ) − 1 I − ( K + t ) − 1 ] d t .
Near t = 0 t=0 t = 0 the gap bounds the integrand. For large t t t , the resolvent identity with I I I bounds it by C J ( 1 + t ) − 2 C_J(1+t)^{-2} C J ( 1 + t ) − 2 , since K − I K-I K − I has a bounded coefficient row on the larger product. This integrable bound and the derivative bounds with every index rooted allow us to integrate coefficient by coefficient. The exact normalization of the scales adds only the same term i log M i\log M i log M for each scalar label, with i < J i<J i < J . Thus ∣ ( log K i T ) x x ∣ ≤ C J \left|(\log K_i^{\mathsf{T}})_{xx}\right|\le C_J ( log K i T ) xx ≤ C J on the stated complex profile, not only at zero background.
The comparison of a regional logarithm with the restriction of the torus logarithm is the sum of (71) and (72) . Its diagonal density at an unfixed root u u u is bounded in coefficient norm by C J e − c J d Ω ( u ) C_J e^{-c_Jd_{\Omega}(u)} C J e − c J d Ω ( u ) . Summing these diagonal defects and using (55) , we obtain the bound on B i \mathcal{B}_i B i . If Ω c \Omega^c Ω c is empty, the defect is zero. □
To state the recursion for Λ \Lambda Λ we fix the counts that enter it.
Definition 20.11. Consider a full step on the block region Ω \Omega Ω and the matched torus step. Every scalar label is counted on the lattice on which it is introduced.
For each scale i i i , ∣ H i ∣ |H_i| ∣ H i ∣ is the number of missing scalar free labels at scale i i i , that is, free labels of the torus step that are not labels of the regional step.
N m i s s i n g N_{\mathrm{missing}} N missing is the number of scalar labels of the torus step (background labels on the output lattice and free labels at the intermediate scales) that are not labels of the regional step. We write M b l o c k \mathcal{M}_{\mathrm{block}} M block for the set of missing torus fluctuation blocks; since each such block consists of three missing scalar labels, ∣ M b l o c k ∣ ≤ N m i s s i n g |\mathcal{M}_{\mathrm{block}}|\le N_{\mathrm{missing}} ∣ M block ∣ ≤ N missing .
N b d N_{\mathrm{bd}} N bd is the number of scalar labels of the regional step whose positions lie within the fixed distance w b a n d ≥ 1 w_{\mathrm{band}}\ge1 w band ≥ 1 of W W W , where w b a n d w_{\mathrm{band}} w band is the width of the interface band determined by the construction stencils; it does not depend on the later widths. This is N b a n d N_{\mathrm{band}} N band of (56) , so ∣ ∂ b l k Ω ∣ ≤ N b d ≤ C J ∣ ∂ b l k Ω ∣ |\partial_{\mathrm{blk}}\Omega|\le N_{\mathrm{bd}}\le C_J|\partial_{\mathrm{blk}}\Omega| ∣ ∂ blk Ω∣ ≤ N bd ≤ C J ∣ ∂ blk Ω∣ , including for isolated final blocks. Both counts are zero when W W W is empty.
N f r e e z e N_{\mathrm{freeze}} N freeze is the number of analytic scalar labels that are newly frozen at the end of the step, including, at the first step, the labels frozen at initialization.
When the steps are indexed by t t t , these counts are written ∣ H i ∣ |H_i| ∣ H i ∣ for the scales i i i of step t t t , and N m i s s i n g , t N_{\mathrm{missing},t} N missing , t , N b d , t N_{\mathrm{bd},t} N bd , t , N f r e e z e , t N_{\mathrm{freeze},t} N freeze , t .
Lemma 20.12. Assume the hypotheses on the products and the small-interaction hypothesis of Theorem 20.9 , in a full step, and fix a real admissible coarse background, a real admissible frozen value, and a real coupling t = t max t=t_{\max} t = t m a x . Use the potential without or with the insertion, from Lemma 20.3 , and remove its entire zero-fluctuation part before forming the interaction U ∙ T U_{\bullet}^{\mathrm T} U ∙ T , ∙ ∈ { 0 , R } \,\bullet\in\{0,\mathrm R\} ∙ ∈ { 0 , R } . Let π \pi π be the product of the identical normalized three-dimensional Gaussian reference measures on balls, including the reference measures of all missing torus blocks, and write
Ξ ∙ T = ∫ e − U ∙ T d π , Ξ ∙ Ω = ∫ e − U ∙ Ω d π , ∙ ∈ { 0 , R } . \Xi_{\bullet}^{\mathrm T}=\int e^{-U_{\bullet}^{\mathrm T}}\,d\pi,\qquad\Xi_{\bullet}^{\Omega}=\int e^{-U_{\bullet}^{\Omega}}\,d\pi,\qquad\bullet\in\{0,\mathrm R\}. Ξ ∙ T = ∫ e − U ∙ T d π , Ξ ∙ Ω = ∫ e − U ∙ Ω d π , ∙ ∈ { 0 , R } . The regional interaction does not depend on the adjoined variables. With N m i s s i n g N_{\mathrm{missing}} N missing and N b d N_{\mathrm{bd}} N bd as in Definition 20.11 , there is a restricted interaction U ∙ c U_{\bullet}^{c} U ∙ c on the same product such that
∥ U ∙ T − U ∙ c ∥ ∞ ≤ C J S ( t ) ( N m i s s i n g + N b d ) , \left\|U_{\bullet}^{\mathrm T}-U_{\bullet}^{c}\right\|_{\infty}\leq C_{J\mathcal{S}}(t)\left(N_{\mathrm{missing}}+N_{\mathrm{bd}}\right), U ∙ T − U ∙ c ∞ ≤ C J S ( t ) ( N missing + N bd ) , ∥ U ∙ Ω − U ∙ c ∥ ∞ ≤ C J S ( t ) N b d . \left\|U_{\bullet}^{\Omega}-U_{\bullet}^{c}\right\|_{\infty}\leq C_{J\mathcal{S}}(t)N_{\mathrm{bd}}. U ∙ Ω − U ∙ c ∞ ≤ C J S ( t ) N bd . Consequently, for the positive real logarithms,
∣ log Ξ ∙ T − log Ξ ∙ Ω ∣ ≤ C J S ( t ) ( N m i s s i n g + N b d ) . (94) \left|\log\Xi_{\bullet}^{\mathrm T}-\log\Xi_{\bullet}^{\Omega}\right|\leq C_{J\mathcal{S}}(t)\left(N_{\mathrm{missing}}+N_{\mathrm{bd}}\right). \tag*{(94)} log Ξ ∙ T − log Ξ ∙ Ω ≤ C J S ( t ) ( N missing + N bd ) . (94) This includes evaluation at coarse background zero, with the frozen values unchanged. Empty products have integral one.
Proof. The reference measure on a block ( i , e ) (i,e) ( i , e ) is z p i − 1 1 ∣ w ∣ ≤ p i e − ∣ w ∣ 2 / 2 d 3 w z_{p_i}^{-1}\mathbf{1}_{|w|\leq p_i}e^{-|w|^2/2}\,d^3w z p i − 1 1 ∣ w ∣ ≤ p i e − ∣ w ∣ 2 /2 d 3 w . We take the product of these measures over the union of the two sets of blocks. Adjoining blocks multiplies the regional integral by one, even for blocks at different scales with different radii.
For this comparison of constants only, we resolve the torus monomials as in Definition 20.6 , and let D m o n = { λ : S λ ∩ M b l o c k ≠ ∅ } \mathcal{D}_{\mathrm{mon}}=\{\lambda:S_{\lambda}\cap\mathcal{M}_{\mathrm{block}}\neq\varnothing\} D mon = { λ : S λ ∩ M block = ∅ } be the set of deleted monomials. We call the remaining interaction U ∙ c U_{\bullet}^{c} U ∙ c . The bound on W 16 W_{16} W 16 gives
∥ U T − U c ∥ ∞ ≤ ∑ x ∈ M b l o c k ∑ λ : x ∈ S λ b λ ≤ C J S ∣ M b l o c k ∣ . \left\|U^{\mathrm T}-U^{c}\right\|_{\infty}\leq\sum_{x\in\mathcal{M}_{\mathrm{block}}}\sum_{\lambda:x\in S_{\lambda}}b_{\lambda}\leq C_{J\mathcal{S}}\left|\mathcal{M}_{\mathrm{block}}\right|. U T − U c ∞ ≤ x ∈ M block ∑ λ : x ∈ S λ ∑ b λ ≤ C J S ∣ M block ∣ .
The set of monomials may be infinite; no bound on its cardinality is used. By Definition 20.11 , ∣ M b l o c k ∣ ≤ N m i s s i n g ≤ N m i s s i n g + N b d \left|\mathcal{M}_{\mathrm{block}}\right|\leq N_{\mathrm{missing}}\leq N_{\mathrm{missing}}+N_{\mathrm{bd}} ∣ M block ∣ ≤ N missing ≤ N missing + N bd . This proves the first inequality.
For the regional and the restricted monomials, we compare coefficients on the common set of unfixed fluctuation labels, at the same background argument extended by zero. We use the comparison before the logarithm in Theorem 20.2 and Lemma 20.3 : a reconstruction formula, a minimizer, an inverse square root, or a Haar density is replaced one factor at a time. A change of the coarser arguments is the integral of their derivative against the difference of the reconstructions, which carries a visit to Ω c \Omega^{c} Ω c . For an incoming torus term we integrate D r d 0 Dr_{d_0} D r d 0 on the segment between the two absolute images, which lie below half the radius; fixed old input slots are visits to Ω c \Omega^c Ω c , and unfixed slots use the defect of the reconstruction in Theorem 19.9 . The boundary function f f f enters with its own incoming boundary weight. By the versions with every index rooted, these comparisons keep a visit to Ω c \Omega^c Ω c and every specified unfixed fluctuation index, including all degree-one terms. They are estimates on monomials before either logarithm is expanded, so the present restriction need not coincide with the one used in the earlier expansions.
Every remaining fluctuation monomial has an unfixed fluctuation index. We assign each compared monomial to its first such index y y y . The comparison of coefficientwise majorants with every index rooted bounds its absolute sum there by C J S ( t ) e − α d Ω ( y ) C J_{\mathfrak{S}}(t)e^{-\alpha d_{\Omega}(y)} C J S ( t ) e − α d Ω ( y ) . Summing over these indices, we obtain
∥ U Ω − U c ∥ ∞ ≤ C J S ( t ) ∑ y unfixed e − α d Ω ( y ) ≤ C J S ′ ( t ) ∣ ∂ b l k Ω ∣ ≤ C S ( t ) N b d . \left\|U^{\Omega}-U^{c}\right\|_{\infty}\leq C J_{\mathfrak{S}}(t)\sum_{\substack{y\ \text{unfixed}}}e^{-\alpha d_{\Omega}(y)}\leq C J_{\mathfrak{S}}^{\prime}(t)\lvert\partial_{\mathrm{blk}}\Omega\rvert\leq C_{\mathfrak{S}}(t)N_{\mathrm{bd}}. U Ω − U c ∞ ≤ C J S ( t ) y unfixed ∑ e − α d Ω ( y ) ≤ C J S ′ ( t ) ∣ ∂ blk Ω ∣ ≤ C S ( t ) N bd .
by (55) and Definition 20.11. This proves the second inequality, at every fixed real frozen value.
The zero-fluctuation values of F i F_i F i , L i L_i L i , B i B_i B i , and R R R in Lemma 20.3 were subtracted before this sum and stay in the prefactor. In particular, a constant depending on the frozen values does not enter the boundary estimate. The convergent coefficientwise majorants justify all assignments and sums, by finite truncation and monotone convergence of absolute sums.
Finally, if real bounded U U U , V V V satisfy ∣ U − V ∣ ≤ D \lvert U-V\rvert\leq D ∣ U − V ∣ ≤ D pointwise, then e − D e − V ≤ e − U ≤ e D e − V e^{-D}e^{-V}\leq e^{-U}\leq e^{D}e^{-V} e − D e − V ≤ e − U ≤ e D e − V . Integrating against a probability measure and taking logarithms gives
∣ log ∫ e − U d π − log ∫ e − V d π ∣ ≤ D . \left\lvert\log\int e^{-U}\,\mathrm{d}\pi-\log\int e^{-V}\,\mathrm{d}\pi\right\rvert\leq D. log ∫ e − U d π − log ∫ e − V d π ≤ D .
All integrals here are finite and strictly positive because the potentials have bounded majorants. We apply this twice, through U c U^c U c , which proves (94) . For a real interaction, the integral with the interaction multiplied by 0 ≤ θ ≤ 1 0\leq\theta\leq1 0 ≤ θ ≤ 1 is positive and equals one at θ = 0 \theta=0 θ = 0 . The logarithm given by the cluster expansion vanishes at θ = 0 \theta=0 θ = 0 and therefore agrees with the positive real logarithm along this path. The comparison therefore also bounds the constant used in the analytic recursion.
Theorem 20.13. Use full steps on admissible data with p i ≥ 1 p_i\geq1 p i ≥ 1 at every intermediate scale, the exact extraction of field-independent factors, and the centered torus recursion of Theorem 20.10 . At every full step assume the product and rate hypotheses of Theorem 20.9 , together with its small-interaction and auxiliary-disc gates for the actual potentials without and, whenever used, with insertion. The bounds in those gates include the incoming torus and boundary interactions; admissibility alone does not impose cluster smallness. Restrict the coarse backgrounds, the couplings, and the frozen values to real data on the chosen products of fluctuation balls. Index the full steps by t = 0 , 1 , … t=0,1,\ldots t = 0 , 1 , … , keep the exact incoming value of Λ \Lambda Λ , and write Δ Λ t = Λ t − Λ t T \Delta\Lambda_t=\Lambda_t-\Lambda_t^{\mathrm{T}} Δ Λ t = Λ t − Λ t T for the
difference between the regional and the torus normalization constants before step t t t . At the end of step t t t , the newly frozen analytic scalar labels form a set F F F , frozen simultaneously at the values c j ♯ c_j^\sharp c j ♯ .
Assume the following.
(a0) The incoming centered torus interaction and boundary part have fixed bounds B ρ , a T ( r i n ) ≤ E i n \mathcal{B}_{\rho,a_{\mathsf{T}}}(r_{\mathrm{in}}) \le E_{\mathrm{in}} B ρ , a T ( r in ) ≤ E in and X ρ ( f i n ) ≤ X i n \mathcal{X}^{\rho}(f_{\mathrm{in}}) \le X_{\mathrm{in}} X ρ ( f in ) ≤ X in . In applications of Theorem 20.14 , one has E i n ≤ E E_{\mathrm{in}} \le E E in ≤ E and X i n ≤ σ ∗ X_{\mathrm{in}} \le\sigma_* X in ≤ σ ∗ .
(a) The outgoing centered effective interaction of step t t t has the form R o u t = r T + f R^{\mathrm{out}}=r_{\mathsf{T}}+f R out = r T + f with R o u t ( 0 ) = 0 R^{\mathrm{out}}(0)=0 R out ( 0 ) = 0 , D r T ( 0 ) = 0 D r_{\mathsf{T}}(0)=0 D r T ( 0 ) = 0 , B ρ , a ( r T ) ≤ E \mathcal{B}_{\rho,a}(r_{\mathsf{T}})\le E B ρ , a ( r T ) ≤ E , and X Ω ϕ ( f ) ≤ Y \mathcal{X}^{\phi}_{\Omega}(f)\le Y X Ω ϕ ( f ) ≤ Y , where Y Y Y is a fixed bound.
(b) Every newly frozen value satisfies ∣ c j ♯ ∣ ≤ ϕ ( j ) / 4 \lvert c_j^\sharp\rvert\le\phi(j)/4 ∣ c j ♯ ∣ ≤ ϕ ( j ) /4 for j ∈ F j\in F j ∈ F .
Then at each step, including the simultaneous restriction to fixed values at its end,
Δ Λ t + 1 ≤ Δ Λ t + ∑ i in step t ∣ H i ∣ log ( 1 / g i ) + C J ( N m i s s i n g , t + N b d , t + N f r e e z e , t ) , (95) \Delta\Lambda_{t+1}\le\Delta\Lambda_t+\sum_{i\text{ in step }t}\lvert H_i\rvert\log(1/g_i)+C_J(N_{\mathrm{missing},t}+N_{\mathrm{bd},t}+N_{\mathrm{freeze},t}), \tag*{(95)} Δ Λ t + 1 ≤ Δ Λ t + i in step t ∑ ∣ H i ∣ log ( 1/ g i ) + C J ( N missing , t + N bd , t + N freeze , t ) , (95)
with the counts of Definition 20.11, where C J C_J C J may also depend on E E E , Y Y Y , E i n E_{\mathrm{in}} E in , and X i n X_{\mathrm{in}} X in . Each Gaussian dimension occurs once, and the normalized auxiliary variables and the Haar measure of the block axial gauge trees have mass one. The restriction to fixed values is exactly
R n e w ( v ) = R o u t ( v , c ♯ ) − R o u t ( 0 , c ♯ ) , Λ n e w = Λ o u t − R o u t ( 0 , c ♯ ) , R_{\mathrm{new}}(v)=R^{\mathrm{out}}(v,c^\sharp)-R^{\mathrm{out}}(0,c^\sharp),\qquad\Lambda_{\mathrm{new}}=\Lambda^{\mathrm{out}}-R^{\mathrm{out}}(0,c^\sharp), R new ( v ) = R out ( v , c ♯ ) − R out ( 0 , c ♯ ) , Λ new = Λ out − R out ( 0 , c ♯ ) ,
and the frozen constant satisfies
∣ R o u t ( 0 , c ♯ ) ∣ ≤ ρ 2 ∣ F ∣ ( E / 16 + Y / 4 ) . (96) \left\lvert R^{\mathrm{out}}(0,c^\sharp)\right\rvert\le\rho^2\lvert F\rvert(E/16+Y/4). \tag*{(96)} R out ( 0 , c ♯ ) ≤ ρ 2 ∣ F ∣ ( E /16 + Y /4 ) . (96)
The classical constants remain in the full classical action. A constant of the incoming interaction is first moved into Λ \Lambda Λ exactly, using e Λ − R = e Λ − R ( 0 ) e − ( R − R ( 0 ) ) e^{\Lambda-R}=e^{\Lambda-R(0)}e^{-(R-R(0))} e Λ − R = e Λ − R ( 0 ) e − ( R − R ( 0 )) ; no nonconstant norm bounds this constant. No additional term proportional to ∣ Ω ∣ log ( 1 / g ) \lvert\Omega\rvert\log(1/g) ∣ Ω ∣ log ( 1/ g ) appears. The restriction at initialization is included, with R o u t R^{\mathrm{out}} R out the torus reference and f = 0 f=0 f = 0 . The constant C J C_J C J has the dependence described in item (3) of Definition 4.14 and is independent of the later widths, of the preceding depth, and of the number of steps. No inequality for complex values and no recursion for compositions of fewer than J J J averaging steps is asserted.
Proof. We compare the logarithms of the regional and the torus factorizations term by term: first the deterministic factors (Step 1), then the fluctuation integrals (Step 2), then the pullback at zero fluctuation (Step 3), and finally the constant produced by fixing coordinates (Step 4).
Step 1: coupling powers, determinants, and Haar factors. At fixed real frozen values, the factorization (20.38) contains exactly one factor g i g_i g i for each scalar free coordinate and one triangular determinant. Subtracting the logarithms of the regional and the torus factorizations, we find for the coupling powers
∑ i [ ∑ x ∈ I i log g i − ∑ x ∈ I i ∪ H i log g i ] = ∑ i ∣ H i ∣ log ( 1 / g i ) . \sum_i\left[\sum_{x\in I_i}\log g_i-\sum_{x\in I_i\cup H_i}\log g_i\right]=\sum_i\lvert H_i\rvert\log(1/g_i). i ∑ [ x ∈ I i ∑ log g i − x ∈ I i ∪ H i ∑ log g i ] = i ∑ ∣ H i ∣ log ( 1/ g i ) .
The determinant statement of Theorem 20.10 bounds the difference of the logarithmic determinants by C J C_J C J for each missing scalar label plus C J ∣ ∂ b l k Ω ∣ ≤ C J N b d C_J|\partial_{\mathrm{blk}}\Omega| \le C_JN_{\mathrm{bd}} C J ∣ ∂ blk Ω∣ ≤ C J N bd , by (20.5). The same argument applies to the Haar densities at zero fluctuation, by the comparison in part (3) of Theorem 20.2. Their constant normalizations depend only on M M M and J J J and contribute C J C_J C J for each omitted or interface coordinate. Since p i ≥ 1 p_i \ge1 p i ≥ 1 we have 0 < z 1 ≤ z p i ≤ ( 2 π ) 3 / 2 0 < z_1 \le z_{p_i} \le(2\pi)^{3/2} 0 < z 1 ≤ z p i ≤ ( 2 π ) 3/2 , so the terms log z p i \log z_{p_i} log z p i contribute a bounded amount for each missing block. The Haar measure of the block axial gauge trees and the adjoined reference measures contribute no dimension.
Step 2: the fluctuation integrals. Under the expressly assumed product, rate, and small-interaction hypotheses at this step, Lemma 20.12, at coarse background zero, bounds the difference of the fluctuation logarithms without insertion, and of those with insertion whenever the factorization uses them, by C J s ( N m i s s i n g + N b d ) C_Js(N_{\mathrm{missing}} + N_{\mathrm{bd}}) C J s ( N missing + N bd ) . In that proof, each missing fluctuation block is counted among the missing labels and each comparison at unfixed roots in the interface count; no geometric crossing argument is needed. Using s ≤ ϵ J s \le\epsilon_J s ≤ ϵ J , we find that this step contributes C J ( N m i s s i n g + N b d ) C_J(N_{\mathrm{missing}} + N_{\mathrm{bd}}) C J ( N missing + N bd ) .
Step 3: the pullback at zero fluctuation. The zero-fluctuation value of the pullback has not entered Step 2. For its centered torus and boundary parts we assign each monomial of degree l ≥ 1 l \ge1 l ≥ 1 to its input roots with the shares η e / l \eta_e/l η e / l , whose sum is one. The input norms in hypothesis (a0) and Definition 4.6 give C J E i n C_JE_{\mathrm{in}} C J E in or C J X i n C_JX_{\mathrm{in}} C J X in for each input root, after the common image at half the radius has summed the degree series. Missing roots are counted as missing labels. Differences rooted at unfixed labels carry the visit in (20.39) and are summed by (20.4). The incoming constant was moved into Λ \Lambda Λ exactly before this calculation.
Step 4: fixing coordinates. Finally we apply Lemma 18.9. To prove (20.45), note that when the unfixed variables are set to zero, only monomials supported in F F F survive. On such a monomial of degree l l l we have ∑ j ∈ F ν j / l = 1 \sum_{j\in F}\nu_j/l = 1 ∑ j ∈ F ν j / l = 1 , and by hypothesis (b) its frozen value has modulus at most 4 − l ∣ f ν ∣ ϕ l 4^{-l}|f_\nu|\phi^l 4 − l ∣ f ν ∣ ϕ l . Summing over j ∈ F j \in F j ∈ F and using the definition of X Ω ϕ \mathcal{X}_{\Omega}^{\phi} X Ω ϕ with the weights larger than one dropped, we find that the boundary part contributes at most ∣ F ∣ ρ 2 Y / 4 |F|\rho^2Y/4 ∣ F ∣ ρ 2 Y /4 , since l ≥ 1 l \ge1 l ≥ 1 . The torus part has l ≥ 2 l \ge2 l ≥ 2 and ∣ c j ♯ ∣ ≤ ρ / 4 |c_j^{\sharp}| \le\rho/4 ∣ c j ♯ ∣ ≤ ρ /4 , so the seminorm bounds its contribution by ∣ F ∣ ρ 2 E / 16 |F|\rho^2E/16 ∣ F ∣ ρ 2 E /16 . This proves (20.45); the bound contributes C J C_J C J for each newly frozen analytic label, degree one included. The value is subtracted from Λ \Lambda Λ with the sign in the statement, and its absolute value gives the upper bound. The same identity applies at initialization. Data deep inside the fixed complement stay in the external factors, and the classical constants stay in the classical action. Adding the finitely many contributions proves (20.44). Every estimate used a coefficient bound for each label or the boundary sum (20.4), which does not depend on the shape of the region; none of them estimates the incoming constant or introduces a factor depending on the number of components or of steps.
Preservation of the analytic bounds under iteration
The following theorem is the main result of this section.
Theorem 20.14. Fix the torus analytic radii and rates, the torus support weights and spatial margins, and M M M , and fix A w ≥ 1 A_{\mathrm{w}} \ge1 A w ≥ 1 . Choose J J J once so that the conditions on J J J of the torus construction used in Lemma 19.7 hold and B ≥ B g e o m B \ge B_{\mathrm{geom}} B ≥ B geom , as in Theorem 20.9 . There exist positive fixed base radii ρ \rho ρ , ρ c l \rho_{\mathrm{cl}} ρ cl , R r a w R_{\mathrm{raw}} R raw and rates a a a , α \alpha α as in Theorem 20.9 . At these radii and rates, fix a finite torus ceiling E ≥ 0 E \ge0 E ≥ 0 as specified below. For each such fixed E E E , there exist positive radii R W R_{\mathrm{W}} R W , R e n R_{\mathrm{en}} R en , a perturbation bound σ ∗ \sigma_* σ ∗ , the classical and quantum radius profiles defined in § 2.2.1 , a width s e r s_{\mathrm{er}} s er of the frozen layer, a bound β r e c \beta_{\mathrm{rec}} β rec on the frozen values, a support radius r r r , and a coupling bound g ∗ < e − 1 g_* < e^{-1} g ∗ < e − 1 , with the following property. One may take ρ = ρ c l \rho= \rho_{\mathrm{cl}} ρ = ρ cl ; other fixed ratios are allowed if both image conditions are imposed. The dependent constants and g ∗ g_* g ∗ may depend on E E E , and g ∗ g_* g ∗ may also depend on A w A_{\mathrm{w}} A w . More generally, after fixing the larger construction data, base radii and rates, and torus ceilings, the conclusions hold for every choice of the remaining parameters made in the order of Table 20.1 and satisfying all requirements in Steps 4–10 of the proof, as well as the explicit hypotheses of the theorem.
The fixed E E E is required to be a common ceiling for every incoming and outgoing centered torus reference interaction on the quantum radius and rate used below, including the reference at initialization and the terminal output. It is fixed before all parameter choices whose bounds depend on it, in particular the local output and mean constants and the final coupling ceiling. Consider a full step on an admissible block region satisfying the protected-neighborhood conditions of Lemma 19.5 , at any preceding depth and after any number of earlier steps, with a real admissible frozen value and an input satisfying
X c l ( P ‾ ) ≤ 2 σ ∗ , B ρ , a Π T ( r d 0 ) ≤ E , X ρ ( f ) ≤ σ ∗ , max i g i ≤ g ∗ . \mathcal{X}^{\mathrm{cl}}(\overline{P}) \le2\sigma_*, \qquad\mathcal{B}_{\rho,a\Pi T}(r_{d_0}) \le E, \qquad\mathcal{X}^{\rho}(f) \le\sigma_*, \qquad\max_i g_i \le g_*. X cl ( P ) ≤ 2 σ ∗ , B ρ , a Π T ( r d 0 ) ≤ E , X ρ ( f ) ≤ σ ∗ , i max g i ≤ g ∗ . Assume the uniform incoming classical norm on D i n D_{\mathrm{in}} D in , the boundary comparison (20.8), and the inequality on the conditional forces of Theorem 19.9 and Lemma 19.3 . Set p ( t ) = A w log ( 1 / t ) p(t) = A_{\mathrm{w}}\log(1/t) p ( t ) = A w log ( 1/ t ) and require at the t t t -th step, with t t = max i g t , i t_t = \max_i g_{t,i} t t = max i g t , i ,
p ( t t ) ≤ p t , i ≤ p ( t t ) + 1 2 A w ( J − 1 ) log M . p(t_t) \le p_{t,i} \le p(t_t) + \frac{1}{2} A_{\mathrm{w}}(J-1)\log M. p ( t t ) ≤ p t , i ≤ p ( t t ) + 2 1 A w ( J − 1 ) log M . Assume the inclusion of the products, the positivity of the rational image denominators, the small-interaction condition, and the conditions on the auxiliary half-disc of Lemma 20.3 and Theorem 20.9 , with the bound chosen for this whole family of radii. The small-field event uses p t , i = p ( t t ) p_{t,i}=p(t_t) p t , i = p ( t t ) . All conclusions formulated relative to the fixed torus recursion r d r_d r d —the recursion for Λ \Lambda Λ , the iterated boundary norm bounds, and the resulting oscillation bound—use this common
Table 20.1. The order of the parameter choices in the proof of Theorem 20.14. The requirements are stated in Steps 4 to 10.
choice of fluctuation radii. For any other radii in the displayed range, the local single-step comparisons are with the torus update computed on exactly the same fluctuation product; no identification of that update with the fixed r d 0 + J r_{d_0+J} r d 0 + J is asserted. Under these hypotheses the perturbed objective has all of the following properties: the coefficient bounds of the potential with each index fixed in turn, the gap of the constrained Hessians at every stage, and the strict bounds on the image under the reconstruction majorant, on the outer profile times the whole enclosing product; the comparisons with boundary decay of Theorem 20.2 ; the potential of Lemma 20.3 ; the expansions without and with insertion of Theorem 20.9 , with input support weight e 16 ∣ S ∣ e^{16|S|} e 16∣ S ∣ and output support weight e 4 ∣ S ∣ e^{4|S|} e 4∣ S ∣ as in Lemma 20.8 , and with the same returned rates and radii; the bounds on the generated term, the pullback, and the determinants of Theorem 20.10 ; and the recursion for Λ \Lambda Λ of Theorem 20.13 . These properties are conclusions of the earlier results.
All output constants are evaluated on the preliminary ball σ 0 \sigma_0 σ 0 and on the fixed boundary-completion ceiling X c l ( S ζ c l ) ≤ σ 0 / 4 \mathcal{X}^{\mathrm{cl}}(S^{\mathrm{cl}}_\zeta) \le\sigma_0/4 X cl ( S ζ cl ) ≤ σ 0 /4 . Thus one may take A J c l = A J c l , 0 + K J c l σ 0 / 4 A^{\mathrm{cl}}_J=A^{\mathrm{cl},0}_J+K^{\mathrm{cl}}_J\sigma_0/4 A J cl = A J cl , 0 + K J cl σ 0 /4 when passing from the adjusted perturbation to P t P_t P t ; this uses no later choice of β r e c \beta_{\mathrm{rec}} β rec or σ ∗ \sigma_* σ ∗ . The final freezing gate implies the stronger ceiling σ ∗ / 4 \sigma_*/4 σ ∗ /4 .
More explicitly, let x t = X ( f t ) x_t=\mathcal{X}(f_t) x t = X ( f t ) and x t c l = X c l ( P t ) x_t^{\mathrm{cl}}=\mathcal{X}^{\mathrm{cl}}(P_t) x t cl = X cl ( P t ) be the incoming boundary norms of the quantum and the classical parts at step t t t , the boundary-completion adjustment of the classical part being understood. Before coordinates are fixed, the outgoing boundary norms y t y_t y t and y t c l y_t^{\mathrm{cl}} y t cl satisfy
y t ≤ A J ( 1 + E ) + K J x t + C J t t 2 ( E + x t ) [ p t 4 + ( E + x t ) p t 2 ] , y_t \le A_J(1+E)+K_Jx_t+C_Jt_t^2(E+x_t)\left[p_t^4+(E+x_t)p_t^2\right], y t ≤ A J ( 1 + E ) + K J x t + C J t t 2 ( E + x t ) [ p t 4 + ( E + x t ) p t 2 ] ,
y t c l ≤ A J c l + K J c l x t c l , t t = max i g t , i , p t = max i p t , i . y_t^{\mathrm{cl}}\le A_J^{\mathrm{cl}}+K_J^{\mathrm{cl}}x_t^{\mathrm{cl}},\qquad
t_t=\max_i g_{t,i},\qquad p_t=\max_i p_{t,i}. y t cl ≤ A J cl + K J cl x t cl , t t = i max g t , i , p t = i max p t , i .
Under the hypotheses on the doubled output analytic radius and on the incoming image at half the radius, K J = 4 ( log 2 ) K c o l , J K_J=4(\log2)K_{\mathrm{col},J} K J = 4 ( log 2 ) K col , J , where K c o l , J K_{\mathrm{col},J} K col , J is the column bound of Theorem 20.10 for the perturbed zero-fluctuation reconstruction, uniform on the preliminary perturbation ball. The recursion for Λ \Lambda Λ is Theorem 20.13 .
Whenever coordinates are fixed, fix every label with reduced radius and keep the remaining unfixed labels at distance at least s e r s_{\mathrm{er}} s er from the previous boundary. For ∙ ∈ { c l , q } \bullet\in\{\mathrm{cl},q\} ∙ ∈ { cl , q } , with ( ρ q , E q ) = ( ρ , E ) (\rho_q,E_q)=(\rho,E) ( ρ q , E q ) = ( ρ , E ) and classical radius ρ c l \rho_{\mathrm{cl}} ρ cl , choose
E c l ≥ sup d ≥ 0 B ρ c l , a T ( E d ) . E_{\mathrm{cl}}\ge\sup_{d\ge0}B_{\rho_{\mathrm{cl}},a_T}(E_d). E cl ≥ d ≥ 0 sup B ρ cl , a T ( E d ) .
where a T a_T a T is no larger than the classical coefficient rate of Theorem 10.11. The factorial all-root bounds give this finite ceiling after the analytic radius is chosen inside their convergence radius. For both values of ∙ \bullet ∙ , require
C b c , ∙ E ∙ β r e c / ρ ∙ ≤ σ ∗ / 4 , ∣ c j ∙ ∣ ≤ min { β r e c , ϕ ∙ ( j ) / 4 } . C_{\mathrm{bc},\bullet}E_\bullet\beta_{\mathrm{rec}}/\rho_\bullet\le\sigma_*/4,\qquad
|c^\bullet_j|\le\min\{\beta_{\mathrm{rec}},\phi_\bullet(j)/4\}. C bc , ∙ E ∙ β rec / ρ ∙ ≤ σ ∗ /4 , ∣ c j ∙ ∣ ≤ min { β rec , ϕ ∙ ( j ) /4 } .
Here C b c , ∙ ≥ 8 C_{\mathrm{bc},\bullet}\ge8 C bc , ∙ ≥ 8 is fixed before β r e c \beta_{\mathrm{rec}} β rec and the later widths, and includes any finite sum over types of boundary-completion differences and over conversions between coordinates; a single simultaneous restriction of the union of all newly frozen labels to their fixed values uses C b c , ∙ = 8 C_{\mathrm{bc},\bullet}=8 C bc , ∙ = 8 in the coordinates in which the labels were introduced. Every frozen label, whether in the boundary layer, in a protected neighborhood, or among the boundary-crossing coordinates, and including the labels frozen at initialization, is covered by this bound. Here c ♯ c^{\sharp} c ♯ is the value in the analytic coordinates in which the label was introduced; any conversion from group values or transported coordinates is included in the bound by β r e c \beta_{\mathrm{rec}} β rec beforehand. Assume x 0 , x 0 c l ≤ σ ∗ x_0,x_0^{\mathrm{cl}}\leq\sigma_\ast x 0 , x 0 cl ≤ σ ∗ . Then, including the new boundary-completion differences of the torus parts,
x t + 1 ≤ z t + 2 e − α ( s e r − 1 ) y t , x t + 1 c l ≤ z t c l + 2 e − α ( s e r − 1 ) y t c l x_{t+1}\leq z_t+2e^{-\alpha(s_{\mathrm{er}}-1)}y_t,\qquad x_{t+1}^{\mathrm{cl}}\leq z_t^{\mathrm{cl}}+2e^{-\alpha(s_{\mathrm{er}}-1)}y_t^{\mathrm{cl}} x t + 1 ≤ z t + 2 e − α ( s er − 1 ) y t , x t + 1 cl ≤ z t cl + 2 e − α ( s er − 1 ) y t cl
where z t z_t z t and z t c l z_t^{\mathrm{cl}} z t cl are the norms of the boundary-completion differences of the torus parts created when coordinates are fixed at the end of step t t t . The choices can be made with
x 0 , x 0 c l ≤ σ ∗ , z t , z t c l ≤ σ ∗ / 4 , x_0,x_0^{\mathrm{cl}}\leq\sigma_\ast,\qquad z_t,z_t^{\mathrm{cl}}\leq\sigma_\ast/4, x 0 , x 0 cl ≤ σ ∗ , z t , z t cl ≤ σ ∗ /4 ,
and with the last terms of both recursions at most σ ∗ / 2 \sigma_\ast/2 σ ∗ /2 , uniformly on the input ball. In particular
x t , x t c l ≤ 3 σ ∗ / 4 ( t ≥ 1 ) x_t,x_t^{\mathrm{cl}}\leq3\sigma_\ast/4\qquad(t\geq1) x t , x t cl ≤ 3 σ ∗ /4 ( t ≥ 1 )
after any finite number of steps, with no further loss of radius or rate. For ∣ v i ∣ ≤ ϕ ( i ) |v_i|\leq\phi(i) ∣ v i ∣ ≤ ϕ ( i ) the oscillation of the centered effective interaction satisfies
∣ R t ( v ; b r e c ) − R t ( 0 ; b r e c ) ∣ ≤ ρ 2 [ E ∣ Ω t ∣ + σ ∗ ∑ j ∈ Ω t e − α d Ω t ( j ) ] |R_t(v;b_{\mathrm{rec}})-R_t(0;b_{\mathrm{rec}})|\leq\rho^2\left[E|\Omega_t|+\sigma_\ast\sum_{j\in\Omega_t}e^{-\alpha d_{\Omega_t}(j)}\right] ∣ R t ( v ; b rec ) − R t ( 0 ; b rec ) ∣ ≤ ρ 2 E ∣ Ω t ∣ + σ ∗ j ∈ Ω t ∑ e − α d Ω t ( j )
where, for this count and this sum, Ω t \Omega_t Ω t denotes the set of unfixed scalar labels, directions and colors included. At the output of a step that is not followed by fixing coordinates, σ ∗ \sigma_\ast σ ∗ in this formula is replaced by the fixed output bound for y t y_t y t .
All constants have the dependence described in item (3) of Definition 4.14 and exclude the later widths. The output bounds are computed first and s e r s_{\mathrm{er}} s er afterwards; J J J is not enlarged by means of these constants for fixed J J J . The theorem concerns admissible data; that the configurations of the large-field decomposition are covered by them is proved separately in Section 21. For a composition of s < J s<J s < J averaging steps the estimates at each stage hold on their admissible products, but a conclusion about the two images or about the enclosing product also requires the separate image condition for that shorter composition; the preservation of the profile is asserted only for full steps. An empty set of unfixed labels has nonconstant norm zero, determinant one, logarithmic determinant zero, and fluctuation integral one.
Proof. The proof has three parts. Steps 1 to 3 prove the bounds on y t y_t y t and y t c l y_t^{\mathrm{cl}} y t cl before coordinates are fixed. Steps 4 to 10 make the parameter choices in the order summarized in Table 20.1 and check that each choice uses only quantities fixed before it. Steps 11 and 12 prove the iteration and the bound on the oscillation.
Step 1: the hypotheses at each step. We fix A W ≥ 1 A_W \ge1 A W ≥ 1 before the coupling bound. The coefficient constants of the potential and the torus and geometric conditions used to choose J J J do not depend on A W A_W A W ; only the final coupling bound may depend on it. At each step, the incoming classical norm is the uniform norm on D i n D_{\mathrm{in}} D in , the boundary comparison (20.8) holds by assumption, with all label positions and volume factors on the lattices where the labels were introduced, and the conditional forces satisfy (20.12) and (20.13). The strict bounds on the two images place all segments of evaluation inside the uniform incoming products. Thus, at the fixed frozen value b = ( ξ , b r e c ) b=(\xi,b_{\mathrm{rec}}) b = ( ξ , b rec ) , both the perturbation and its comparison with zero satisfy the hypotheses of Theorem 19.9.
Step 2: the classical comparison before coordinates are fixed. Let a R a_R a R and a T a_T a T be the two admissible zero-fluctuation chains before the gauge transformation, the regional one with its missing slots set to zero. The comparison of values, before the common factor M J M^J M J , is
F R ( v , a R ) − F T ( v , a T ) = F R ( v , a R ) − F T ( v , a R ) + ∫ 0 1 D z F T ( v , a T + τ ( a R − a T ) ) [ a R − a T ] d τ . F_R(v,a_R)-F_T(v,a_T)=F_R(v,a_R)-F_T(v,a_R)
+\int_0^1 D_zF_T\bigl(v,a_T+\tau(a_R-a_T)\bigr)[a_R-a_T]\,\mathrm{d}\tau. F R ( v , a R ) − F T ( v , a T ) = F R ( v , a R ) − F T ( v , a R ) + ∫ 0 1 D z F T ( v , a T + τ ( a R − a T ) ) [ a R − a T ] d τ .
The fixed larger charts contain this segment, since the smaller solution neighborhoods have strict room in them. The first difference has the representation (20.16) through changed formulas. On unfixed free slots, the difference of the chains in the integral is the comparison of minimizers in Theorem 20.2 , here with perturbation zero; on missing free slots it is a bounded source in Ω c \Omega^c Ω c . Every nonconstant coefficient with an unfixed output index therefore has a network to Ω c \Omega^c Ω c , by the action tensors rooted at every index and Lemma 18.15. This includes derivatives falling on the action tensor while the source stays undifferentiated. We sum over the indices in Ω c \Omega^c Ω c from the unfixed index using (53) . Pure constants supported in Ω c \Omega^c Ω c stay in the classical value. The conversion of coefficients and profiles in the proof of Theorem 20.9 , without its expansions, now yields a finite bound A J c l A_J^{\mathrm{cl}} A J cl for the centered comparison. The factor M J M^J M J depends only on J J J and is included in this constant. No interpolation between different fibers is used in this calculation of values.
Step 3: the perturbation and the quantum comparison. Next we interpolate in the adjusted input perturbation. By the envelope formula, the derivative of the minimum value is the evaluation of the perturbation at the minimizer. The strict bound on the image under the reconstruction majorant bounds its input jets, and joining its input boundary network to the reconstruction networks proves the output bound K J c l x c l K_J^{\mathrm{cl}}x^{\mathrm{cl}} K J cl x cl . Equivalently, this is the conclusion of Theorem 19.9 on the minimum values for two perturbations, with its exact boundary-completion adjustment. The fixed small boundary-completion difference is included with the unperturbed term in A J c l A_J^{\mathrm{cl}} A J cl . This proves y c l ≤ A J c l + K J c l x c l y^{\mathrm{cl}} \le A_J^{\mathrm{cl}} + K_J^{\mathrm{cl}}x^{\mathrm{cl}} y cl ≤ A J cl + K J cl x cl . The coefficient networks in this argument are the classical ones, with the common rates chosen in Theorem 20.9 .
For the common cutoff p t , i = p ( t t ) p_{t,i}=p(t_t) p t , i = p ( t t ) , we subtract the matched torus recursion, which is the fixed recursion r d r_d r d , from the recursion of Theorem 20.10 . The difference is the sum of the comparison of the generated terms and torus pullbacks, the term P Ω f P^\Omega f P Ω f , and the compared fluctuation correction. The first two are bounded in Theorem 20.10 by A J ( 1 + E ) A_J(1+E) A J ( 1 + E ) and K J x K_Jx K J x , and the third by Theorem 20.9 . By the triangle inequality, we obtain the stated bound on y y y .
Step 4: radii, rates, and the perturbation bound. The fixed products include the auxiliary reserve h J h_J h J in the definition of admissibility. Choose the analytic radii within K ∗ max ( ρ , ρ c l ) < h J K_* \max(\rho,\rho_{\mathrm{cl}}) < h_J K ∗ max ( ρ , ρ cl ) < h J , and the free radii below h J h_J h J , before fixing the preliminary perturbation or any output coefficient bound. In particular every factorial series and implicit or matrix-function series used for Theorem 20.2 is summed on a strictly smaller product.
We fix K ∗ K_* K ∗ before J J J and choose J J J for the torus conditions and the geometric condition. We choose the base radii inside the fixed larger construction domains; one may choose ρ = ρ c l \rho=\rho_{\mathrm{cl}} ρ = ρ cl , and different fixed ratios simply keep both image conditions. We choose the common rates below the intrinsic coefficient and unperturbed image exponents at fixed J J J , and impose the separate output-rate conditions against B α B\alpha B α and B α B\alpha B α for the inherited perturbation. Its controlled coefficient difference preserves the half-radius image at these working rates after the final perturbation bound is decreased. We decrease R r a w R_{\mathrm{raw}} R raw and the base radii if necessary, for the fixed margin of the larger construction domain and for C T , ∙ K ∗ ρ ∙ ≤ 1 / 4 C_{\mathrm{T},\bullet}K_{*}\rho_{\bullet} \le1/4 C T , ∙ K ∗ ρ ∙ ≤ 1/4 . The torus conditions are used with their stronger torus rate, not through an estimate involving an uncontrolled regional constant.
At the chosen base radii and rates we fix the torus ceiling E E E in the statement and the classical ceiling E c l E_{\mathrm{cl}} E cl furnished by the all-root bounds. Subsequent local constants and the final coupling ceiling may depend on these fixed ceilings. We choose a preliminary bound σ 0 \sigma_{0} σ 0 below the admissible inverse, conditional force, and real gap margins, and a free radius r q r_{q} r q with
C s e q , J ( 2 I 0 C P σ 0 ρ c l + r q ) < ε t u b e / 2. C_{\mathrm{seq},J}(2I_{0}C_{\mathrm{P}}\sigma_{0}\rho_{\mathrm{cl}}+r_{q})<\varepsilon_{\mathrm{tube}}/2. C seq , J ( 2 I 0 C P σ 0 ρ cl + r q ) < ε tube /2.
The constants are those of the larger construction neighborhoods in (63) , whose uniformity in all higher parameters was checked there. We compute all output, tensor, image, sequential, and enclosure constants uniformly for this bound on the fixed larger products. In particular we fix A A A , C 0 C_{0} C 0 , C 1 C_{1} C 1 , C e n , J C_{\mathrm{en},J} C en , J , and L V L_{V} L V , and the difference-row constants C P , ∙ C_{\mathrm{P},\bullet} C P , ∙ at the allocated working rates, and choose R W R_{W} R W from the positive minimum in Lemma 20.3 . Then we fix σ ∗ ≤ σ 0 \sigma_{*}\leq\sigma_{0} σ ∗ ≤ σ 0 so that, for both radii,
σ ∗ ρ c l ≤ λ ∗ R W / ( 6 C e n , J ) , C P , ∙ σ ∗ ρ c l ≤ ρ ∙ / 64. \sigma_{*}\rho_{\mathrm{cl}}\leq\lambda_{*}R_{W}/(6C_{\mathrm{en},J}),\qquad C_{\mathrm{P},\bullet}\sigma_{*}\rho_{\mathrm{cl}}\leq\rho_{\bullet}/64. σ ∗ ρ cl ≤ λ ∗ R W / ( 6 C en , J ) , C P , ∙ σ ∗ ρ cl ≤ ρ ∙ /64.
This order is necessary because this summand of the enclosure remains when r r r and β r e c \beta_{\mathrm{rec}} β rec are decreased.
Step 5: the boundary radii. We choose the two reduced boundary radii so that the bounds on the image at zero fluctuation near the boundary in Lemma 19.7 are at most ρ ∙ / 16 \rho_{\bullet}/16 ρ ∙ /16 . Then we choose a fixed width w b d w_{\mathrm{bd}} w bd of the boundary layer so that the bounds on the tails far from the boundary are at most ρ ∙ / 16 \rho_{\bullet}/16 ρ ∙ /16 . The remaining bounds at zero fluctuation are the torus bulk bound and the bound for the frozen values. The first is a torus condition, and the second will be imposed on β r e c \beta_{\mathrm{rec}} β rec . All constants in these choices were computed independently of these widths and reduced radii.
Step 6: output bounds and the width of the frozen layer. We set
c J log = 1 2 A W ( J − 1 ) log M , p + ( t ) = A W log ( 1 / t ) + c J log , H ∗ = E + σ ∗ . (97) c_{J}^{\log}=\frac{1}{2}A_{W}(J-1)\log M,\qquad p^{+}(t)=A_{W}\log(1/t)+c_{J}^{\log},\qquad H_{*}=E+\sigma_{*}. \tag*{(97)} c J l o g = 2 1 A W ( J − 1 ) log M , p + ( t ) = A W log ( 1/ t ) + c J l o g , H ∗ = E + σ ∗ . (97)
At a step indexed by k k k , write t = t k t=t_{k} t = t k . By the hypothesis on the radii, 1 < p ( t ) ≤ p k , i ≤ p + ( t ) 1<p(t)\leq p_{k,i}\leq p^{+}(t) 1 < p ( t ) ≤ p k , i ≤ p + ( t ) , since t ≤ g ∗ < e − 1 t\leq g_{*}<e^{-1} t ≤ g ∗ < e − 1 . On the input ball we impose provisionally the requirement
C J t 2 H ∗ [ ( p + ( t ) ) 4 + H ∗ ( p + ( t ) ) 2 ] ≤ 1 , C_{J}t^{2}H_{*}\left[(p^{+}(t))^{4}+H_{*}(p^{+}(t))^{2}\right]\leq1, C J t 2 H ∗ [ ( p + ( t ) ) 4 + H ∗ ( p + ( t ) ) 2 ] ≤ 1 ,
which holds for small couplings. We then take the fixed output bounds
Y = A J ( 1 + E ) + K J σ ∗ + 1 , Y c l = A J c l + K J c l σ ∗ . Y=A_{J}(1+E)+K_{J}\sigma_{*}+1,\qquad Y^{\mathrm{cl}}=A_{J}^{\mathrm{cl}}+K_{J}^{\mathrm{cl}}\sigma_{*}. Y = A J ( 1 + E ) + K J σ ∗ + 1 , Y cl = A J cl + K J cl σ ∗ .
We choose the integer width of the frozen layer with
s e r ≥ max { 2 , w b d , 1 + α − 1 log ( 4 max { γ , γ c l , 1 } / σ ∗ ) } . s_{\mathrm{er}} \ge\max\left\{2,\,w_{\mathrm{bd}},\,1+\alpha^{-1}\log\left(4\max\{\gamma,\gamma^{\mathrm{cl}},1\}/\sigma_*\right)\right\}. s er ≥ max { 2 , w bd , 1 + α − 1 log ( 4 max { γ , γ cl , 1 } / σ ∗ ) } .
The bound may be enlarged to account for the fixed buffer between label positions in Lemma 18.9. Each output restricted to the interior of the frozen layer is then at most σ ∗ / 2 \sigma_*/2 σ ∗ /2 , since 2 e − α ( s e r − 1 ) max { γ , γ c l } ≤ σ ∗ / 2 2e^{-\alpha(s_{\mathrm{er}}-1)}\max\{\gamma,\gamma^{\mathrm{cl}}\}\le\sigma_*/2 2 e − α ( s er − 1 ) max { γ , γ cl } ≤ σ ∗ /2 .
Step 7: the frozen values and the new boundary-completion differences. Whenever coordinates are fixed, let F F F be the union of all newly frozen analytic labels: the whole zone with reduced radius, the protected neighborhoods, and the boundary-crossing coordinates. The same convention covers the restriction at initialization. In the coordinates in which the labels were introduced, a single application of Lemma 18.9 to this union bounds the torus boundary-completion difference by 8 E ∙ β r e c / ρ ∙ 8E_\bullet\beta_{\mathrm{rec}}/\rho_\bullet 8 E ∙ β rec / ρ ∙ . The rate hypothesis of part (2) of that lemma holds with the torus rate a T a_{\mathrm{T}} a T and the regional rates a , α a,\alpha a , α , since 3 ( a + α ) < a T \sqrt{3}(a+\alpha)<a_{\mathrm{T}} 3 ( a + α ) < a T by Theorem 20.9 . No separate factor arises for each component of Ω c \Omega^{\mathrm{c}} Ω c . If a fixed finite list of types of boundary-completion differences is kept separately, with conversion factors L k , ∙ L_{k,\bullet} L k , ∙ , the triangle inequality yields instead the bound with C b c , ∙ = 8 ∑ k L k , ∙ C_{\mathrm{bc},\bullet}=8\sum_k L_{k,\bullet} C bc , ∙ = 8 ∑ k L k , ∙ . We include the classical boundary-completion adjustment among these types whenever it is recorded separately. These constants are fixed on the larger charts, before β r e c \beta_{\mathrm{rec}} β rec and all later widths.
The bound concerns the analytic values c j ♯ c_j^{\sharp} c j ♯ in the coordinates in which the labels were introduced. For a frozen group element in the identity chart, ∣ Log U ∣ = ρ G ( U ) \lvert\operatorname{Log}U\rvert=\rho_G(U) ∣ Log U ∣ = ρ G ( U ) in the Euclidean norm of the Lie algebra, so this conversion has constant one. A longer fixed chart formula, in the coordinates before the gauge transformation or in transported coordinates, has a finite bound C w o r d , J C_{\mathrm{word},J} C word , J on its derivative rows on the larger construction domain. Integrating along the input segment, we bound its logarithm in the identity chart by C w o r d , J C_{\mathrm{word},J} C word , J times the maximal input norm. These conversion constants are fixed before the bounds on the frozen values are chosen. If a separate estimate for the initial frozen values is provided with a constant C a d , J C_{\mathrm{ad},J} C ad , J , we choose C c o n v , J ≥ max { 1 , C a d , J , C w o r d , J } C_{\mathrm{conv},J}\ge\max\{1,C_{\mathrm{ad},J},C_{\mathrm{word},J}\} C conv , J ≥ max { 1 , C ad , J , C word , J } before using a condition relating the support radius to the frozen values. The theorem assumes the bound on every c j ♯ c_j^{\sharp} c j ♯ . It does not deduce this bound from a small support radius in an arbitrary gauge representative, and it transports a norm bound between gauge frames for the frozen coordinates only when the change of frame is the norm-preserving adjoint action of a block.
We choose β r e c > 0 \beta_{\mathrm{rec}}>0 β rec > 0 below the bound for the frozen values in the image, below the enclosure bound λ ∗ R W / ( 6 C e n , J ) \lambda_*R_{\mathrm{W}}/(6C_{\mathrm{en},J}) λ ∗ R W / ( 6 C en , J ) , and, for both ∙ ∈ { c l , q } \bullet\in\{\mathrm{cl},\mathrm{q}\} ∙ ∈ { cl , q } , below
min { ρ ∙ / 4 , ρ ∙ / 4 , σ ∗ ρ ∙ 4 C b c , ∙ ( E ∙ + 1 ) } (98) \min\left\{\rho_{\bullet}/4,\ \rho_{\bullet}/4,\ \frac{\sigma_*\rho_{\bullet}}{4C_{\mathrm{bc},\bullet}(E_\bullet+1)}\right\} \tag*{(98)} min { ρ ∙ /4 , ρ ∙ /4 , 4 C bc , ∙ ( E ∙ + 1 ) σ ∗ ρ ∙ } (98)
These are upper bounds on a quantity chosen after their constants, and they remain positive when a bound E ∙ E_\bullet E ∙ is zero. They imply the two inequalities $C_{\mathrm{bc},\bullet}E_\bullet\beta_{\mathrm{rec}}/\rho_\bullet\le\sigma_*/4 in the statement. The hypothesis ∣ c j ♯ ∣ ≤ min { β r e c , ϕ ∙ ( j ) / 4 } |c_j^\sharp| \le\min\{\beta_{\mathrm{rec}},\phi_{\bullet}(j)/4\} ∣ c j ♯ ∣ ≤ min { β rec , ϕ ∙ ( j ) /4 } is imposed on all of F F F , not only on the old boundary zone.
Step 8: initialization. The assumption x 0 , x 0 c l ≤ σ ∗ x_0,x_0^{\mathrm{cl}} \le\sigma_\ast x 0 , x 0 cl ≤ σ ∗ of the iteration can be met: starting from the torus reference, both boundary parts are zero before initialization, and the same restriction of the union to fixed values, together with (98) , gives initial boundary parts of norm at most σ ∗ / 4 \sigma_\ast/4 σ ∗ /4 . For other prescribed inputs the bounds on the initial norms are kept as hypotheses. The coverage of the configurations and the estimate for the initial frozen values are separate inputs of the later applications.
Step 9: the support radius. We choose r > 0 r>0 r > 0 inside all margins of the small-link domain and of the larger construction domain, below λ ∗ R W / ( 6 C e n , J ) \lambda_\ast R_W/(6C_{\mathrm{en},J}) λ ∗ R W / ( 6 C en , J ) , and, when an estimate for the frozen values in terms of the support radius is provided, with C c o n v , J r ≤ β r e c C_{\mathrm{conv},J}r\le\beta_{\mathrm{rec}} C conv , J r ≤ β rec , imposed after the conversion constant has been fixed. Then R e n = 2 C e n , J ( r + β r e c + σ ∗ ρ c l ) / λ ∗ R_{\mathrm{en}}=2C_{\mathrm{en},J}(r+\beta_{\mathrm{rec}}+\sigma_\ast\rho_{\mathrm{cl}})/\lambda_\ast R en = 2 C en , J ( r + β rec + σ ∗ ρ cl ) / λ ∗ is positive and at most R W R_W R W . The image denominators now have their fixed margins, and Lemma 20.3 supplies the radius-dependent curvature budget on the whole product. The full curvature conclusion also requires the coupling condition C 0 t 2 L V ≤ 1 / 16 C_0t^2\mathcal{L}_V\le1/16 C 0 t 2 L V ≤ 1/16 , imposed in Step 10.
Step 10: the coupling bound. Fix C g a t e , J C_{\mathrm{gate},J} C gate , J dominating the primitive, small-interaction, and fluctuation-remainder constants of Lemma 20.7 and Theorems 20.9 and 20.10, including their fixed multipliers, while keeping the established ϵ J , c z , J \epsilon_J,c_{z,J} ϵ J , c z , J . Finally, choose g ∗ < e − 1 g_\ast<e^{-1} g ∗ < e − 1 so that the following inequalities hold for every 0 < t ≤ g ∗ 0<t\le g_\ast 0 < t ≤ g ∗ :
t p + ( t ) < R e n , C g a t e , J t p + ( t ) ≤ 1 2 , 2 C g a t e , J t D + ( t ) ≤ ϵ J , t D + ( t ) ≤ c z , J / 2 , D + ( t ) = ( p + ( t ) ) 3 + ( 1 + H ∗ ) p + ( t ) , C 0 t 2 L V ≤ 1 / 16 , C 1 t ( L V + 2 ) < 1 / ( 4 3 ) , C g a t e , J t 2 H ∗ [ ( p + ( t ) ) 4 + H ∗ ( p + ( t ) ) 2 ] ≤ 1. (99) \begin{aligned}
t p^+(t)&<R_{\mathrm{en}}, & C_{\mathrm{gate},J}t p^+(t)&\le\frac{1}{2},\\
2C_{\mathrm{gate},J}tD^+(t)&\le\epsilon_J, & tD^+(t)&\le c_{z,J}/2, & D^+(t)&=(p^+(t))^3+(1+H_\ast)p^+(t),\\
C_0t^2\mathcal{L}_V&\le1/16, & C_1t(\mathcal{L}_V+2)&<1/(4\sqrt{3}),\\
C_{\mathrm{gate},J}t^2H_\ast\left[(p^+(t))^4+H_\ast(p^+(t))^2\right]&\le1. \tag*{(99)}
\end{aligned} t p + ( t ) 2 C gate , J t D + ( t ) C 0 t 2 L V C gate , J t 2 H ∗ [ ( p + ( t ) ) 4 + H ∗ ( p + ( t ) ) 2 ] < R en , ≤ ϵ J , ≤ 1/16 , ≤ 1. C gate , J t p + ( t ) t D + ( t ) C 1 t ( L V + 2 ) ≤ 2 1 , ≤ c z , J /2 , < 1/ ( 4 3 ) , D + ( t ) = ( p + ( t ) ) 3 + ( 1 + H ∗ ) p + ( t ) , (99)
The image denominators are at least 1 / 2 1/2 1/2 on the whole enclosing product by the earlier image bounds, which do not involve the coupling. At a step indexed by k k k , apply these inequalities with t = t k ≤ g ∗ t=t_k\le g_\ast t = t k ≤ g ∗ . For any radii in the stated logarithmic range we have p k ≤ p + ( t ) p_k\le p^+(t) p k ≤ p + ( t ) and E + x k ≤ H ∗ E+x_k\le H_\ast E + x k ≤ H ∗ . The first line implies the inclusion of the products. The next line implies s ( t ) ≤ ϵ J s(t)\le\epsilon_J s ( t ) ≤ ϵ J and t ≤ c z , J / ( 2 [ p k 3 + p k + ( E + x k ) p k ] ) = r z / 2 t\le c_{z,J}/(2[p_k^3+p_k+(E+x_k)p_k])=r_z/2 t ≤ c z , J / ( 2 [ p k 3 + p k + ( E + x k ) p k ]) = r z /2 . The third line implies (77) ; since p ( t ) > 1 p(t)>1 p ( t ) > 1 , it also implies the condition C 1 t ( L V + 2 ) ≤ p ( t ) / ( 2 3 ) C_1t(\mathcal{L}_V+2)\le p(t)/(2\sqrt{3}) C 1 t ( L V + 2 ) ≤ p ( t ) / ( 2 3 ) on the means in Lemma 20.5. The last line is the provisional requirement of Step 6. The choice is simultaneous, because every left side that has to be small is a fixed combination of t log q ( 1 / t ) t\log^q(1/t) t log q ( 1/ t ) and t 2 log q ( 1 / t ) t^2\log^q(1/t) t 2 log q ( 1/ t ) for finitely many integers q q q . Writing t = e − u t=e^{-u} t = e − u , we have e − u u q → 0 e^{-u}u^q\to0 e − u u q → 0 ; for example the power series of e u / 2 e^{u/2} e u /2 bounds u q e − u u^qe^{-u} u q e − u by a constant times e − u / 2 e^{-u/2} e − u /2 . These upper bounds hold for all sufficiently small positive t t t , so a single bound works for all scales of every full step in the prescribed family of radii. The common cutoff has p k , i = p ( t ) p_{k,i}=p(t) p k , i = p ( t ) , and the cutoffs at the individual scales, p ( g k , i ) = p ( t ) + A W log ( 1 / λ i ) p(g_{k,i})=p(t)+A_W\log(1/\lambda_i) p ( g k , i ) = p ( t ) + A W log ( 1/ λ i ) , lie in the same range. The expansions are used on these chosen products, not on the enclosing product K e n \mathcal{K}_{\mathrm{en}} K en . This final choice does not change the output bounds or the width s e r s_{\mathrm{er}} s er , which were computed before.
Step 11: the iteration. Here and in the oscillation estimate we use the common fluctuation products, so every torus reference is the fixed recursion r d r_d r d . By Lemma 18.9 with η = 1 / 4 \eta=1/4 η = 1/4 , fixing the coordinates of the union F F F bounds the old boundary part by 2 e − α ( s e r − 1 ) y t 2e^{-\alpha(s_{\mathrm{er}}-1)}y_t 2 e − α ( s er − 1 ) y t , since ∣ c j ♭ ∣ ≤ ϕ ∙ ( j ) / 4 \lvert c_j^{\flat}\rvert\leq\phi_{\bullet}(j)/4 ∣ c j ♭ ∣ ≤ ϕ ∙ ( j ) /4 and all labels with reduced radius are frozen. Independently, the boundary-completion differences of the torus parts satisfy, by the hypothesis on the frozen values,
z t q ≤ C b c , ∙ E ∙ β r e c / ρ ∙ ≤ σ ∗ / 4 , (100) z_t^{\mathrm{q}}\leq C_{\mathrm{bc},\bullet}E_{\bullet}\beta_{\mathrm{rec}}/\rho_{\bullet}\leq\sigma_*/4, \tag*{(100)} z t q ≤ C bc , ∙ E ∙ β rec / ρ ∙ ≤ σ ∗ /4 , (100)
where z t q = z t z_t^{\mathrm{q}}=z_t z t q = z t . Fixing coordinates at an additional interior component of Ω c \Omega^{c} Ω c satisfies the same estimate, because its values are also bounded by β r e c \beta_{\mathrm{rec}} β rec . Combining the two bounds, we obtain
x t + 1 ≤ z t + 2 e − α ( s e r − 1 ) y t , x t + 1 c l ≤ z t c l + 2 e − α ( s e r − 1 ) y t c l . x_{t+1}\leq z_t+2e^{-\alpha(s_{\mathrm{er}}-1)}y_t,\qquad x^{\mathrm{cl}}_{t+1}\leq z^{\mathrm{cl}}_t+2e^{-\alpha(s_{\mathrm{er}}-1)}y^{\mathrm{cl}}_t. x t + 1 ≤ z t + 2 e − α ( s er − 1 ) y t , x t + 1 cl ≤ z t cl + 2 e − α ( s er − 1 ) y t cl .
These estimates hold for arbitrary additional components of Ω c \Omega^{c} Ω c , by Lemma 18.9. Its torus term uses the outgoing reference interaction, so the common ceiling E E E in the hypotheses is needed also at the output of this step, including the terminal step. If both input norms are at most σ ∗ \sigma_* σ ∗ , then y t ≤ Y y_t\leq Y y t ≤ Y and y t c l ≤ Y c l y^{\mathrm{cl}}_t\leq Y^{\mathrm{cl}} y t cl ≤ Y cl , and by Step 6 the right sides are at most 3 σ ∗ / 4 3\sigma_*/4 3 σ ∗ /4 . Starting from the initial bounds, finite induction proves the bounds on x t , x t c l x_t,x^{\mathrm{cl}}_t x t , x t cl after every number of steps. The older frozen values are simply fixed in the next application. The remaining labels keep their input radii because the whole zone with reduced radius was frozen: the conclusion of Lemma 18.9 is the norm with the uniform radius on the remaining labels. This verifies the hypothesis of a uniform incoming profile at the next step. The classical boundary-completion adjustment satisfies
X c l ( P ˉ ) ≤ X c l ( P ) + X c l ( S ζ c l ) ≤ σ ∗ + σ ∗ / 4 < 2 σ ∗ . \mathcal{X}^{\mathrm{cl}}(\bar{P})\leq\mathcal{X}^{\mathrm{cl}}(P)+\mathcal{X}^{\mathrm{cl}}(S^{\mathrm{cl}}_{\zeta})\leq\sigma_*+\sigma_*/4<2\sigma_*. X cl ( P ˉ ) ≤ X cl ( P ) + X cl ( S ζ cl ) ≤ σ ∗ + σ ∗ /4 < 2 σ ∗ .
by the included bound on the boundary-completion differences. The inclusion of the old fixed complement and the positions of the labels in the next geometry are hypotheses of that step, not conclusions of this norm inequality. The inequality on the conditional forces then holds at every coarser parameter of the next step, by (20.12). Every operation preserves the Borel dependence on the frozen values, as proved in Theorem 20.2. No analytic loss is summed over earlier steps.
Step 12: the oscillation. We write the centered effective interaction as its torus part plus its boundary part. At each degree we assign each monomial to its roots with the shares ν i / ∣ ν ∣ \nu_i/|\nu| ν i /∣ ν ∣ , whose sum is one. Evaluating at ∣ v i ∣ ≤ ϕ ( i ) |v_i| \le\phi(i) ∣ v i ∣ ≤ ϕ ( i ) and dropping the network weight, we obtain for the boundary part
∣ f t ( v ) − f t ( 0 ) ∣ ≤ ρ 2 X ( f t ) ∑ i ∈ Ω t e − α d Ω t ( i ) . |f_t(v)-f_t(0)| \le\rho^2 \mathcal{X}(f_t) \sum_{i\in\Omega_t} e^{-\alpha d_{\Omega_t}(i)}. ∣ f t ( v ) − f t ( 0 ) ∣ ≤ ρ 2 X ( f t ) i ∈ Ω t ∑ e − α d Ω t ( i ) .
The degree-one term is included in this assignment. For the torus part the same assignment and its seminorm give ρ 2 E ∣ Ω t ∣ \rho^2 E|\Omega_t| ρ 2 E ∣ Ω t ∣ . These are the stated bounds on the oscillation; before coordinates are fixed we use Y \mathcal{Y} Y instead of σ ∗ \sigma_* σ ∗ . The preceding results, with the choices just made, imply every assertion of the theorem. All local constants were evaluated before the later widths, and none of them was used to choose J J J . A composition of fewer than J J J averaging steps keeps its estimates at each stage on its admissible products, but its conclusions about the full radius require the separate image condition. The iteration for Λ \Lambda Λ and for the norms uses only full steps. A final composition of fewer than J J J averaging steps may remain as a deterministic function; no recursion for Λ \Lambda Λ is asserted for it here. The assertions for an empty set of unfixed labels are the usual conventions for empty sums, products, and determinants. This completes the proof.
Summation of truncation errors
Lemma 20.15. Fix 0 < Γ < ∞ 0 < \Gamma< \infty 0 < Γ < ∞ , M > 1 M > 1 M > 1 , a finite set R \mathcal{R} R of ratios R > 1 R > 1 R > 1 , a finite set of positive numbers η \eta η , and finitely many positive constants a a a , c c c , C C C , and ε \varepsilon ε .
For every finite list of nonnegative integers k k k , there is a minimum side m 0 m_0 m 0 , independent of 0 < γ ≤ Γ 0 < \gamma\le\Gamma 0 < γ ≤ Γ , such that
C [ 1 + log ( m / γ ) ] k ≤ ε m / γ ( m ≥ m 0 ) . C[1+\log(m/\gamma)]^k \le\varepsilon m/\gamma\qquad(m\ge m_0). C [ 1 + log ( m / γ ) ] k ≤ ε m / γ ( m ≥ m 0 ) . With z = ( 1 / 2 ) log ( M m / γ ) z=(1/2)\log(Mm/\gamma) z = ( 1/2 ) log ( M m / γ ) , the functions
X c ( m , γ ) = m 3 e − c m / γ , W a ( m , γ ) = m 3 e − a z 2 X_c(m,\gamma)=m^3e^{-cm/\gamma}, \qquad W_a(m,\gamma)=m^3e^{-az^2} X c ( m , γ ) = m 3 e − c m / γ , W a ( m , γ ) = m 3 e − a z 2 are decreasing in m m m and increasing in γ \gamma γ for m ≥ m 0 m \ge m_0 m ≥ m 0 and 0 < γ ≤ Γ 0 < \gamma\le\Gamma 0 < γ ≤ Γ , after increasing m 0 m_0 m 0 , and they satisfy X c ( R m , γ ) ≤ X c ( m , γ ) / 2 X_c(Rm,\gamma)\le X_c(m,\gamma)/2 X c ( R m , γ ) ≤ X c ( m , γ ) /2 and W a ( R m , γ ) ≤ W a ( m , γ ) / 2 W_a(Rm,\gamma)\le W_a(m,\gamma)/2 W a ( R m , γ ) ≤ W a ( m , γ ) /2 for R ∈ R R\in\mathcal{R} R ∈ R . Their sums over the sides m R k mR^k m R k , k ≥ 0 k\ge0 k ≥ 0 , are at most twice their first terms, and tend to zero as m → ∞ m\to\infty m → ∞ , uniformly on ( 0 , Γ ] (0,\Gamma] ( 0 , Γ ] .
Let m k = m R k m_k=mR^k m k = m R k , k ≥ 0 k\ge0 k ≥ 0 , for some R ∈ R R\in\mathcal{R} R ∈ R , together with at most one additional side η m R j \eta mR^j η m R j with j ∈ N 0 j\in\mathbb{N}_0 j ∈ N 0 , and let nonnegative errors δ \delta δ and z e x t z_{\mathrm{ext}} z ext at these sides satisfy δ ≤ C W a \delta\le CW_a δ ≤ C W a and z e x t ≤ C X c z_{\mathrm{ext}}\le CX_c z ext ≤ C X c . After increasing m 0 m_0 m 0 , depending only on Γ \Gamma Γ , M M M , R R R , the set of η \eta η , and the constants, the following holds uniformly for m ≥ m 0 m\ge m_0 m ≥ m 0 , $0<\gamma\le\Gamma, and the number of sides, finite or infinite. Every \text{, and the number of sides, finite or infinite. Every} , and the number of sides, finite or infinite. Every \delta
is at most 1 / 2 1/2 1/2 , every z e x t z_{\mathrm{ext}} z ext is at most 1 / 4 1/4 1/4 , and ϵ = δ / ( 1 − z e x t ) \epsilon= \delta/(1-z_{\mathrm{ext}}) ϵ = δ / ( 1 − z ext ) satisfies ϵ ≤ ( 4 / 3 ) δ ≤ 1 / 2 \epsilon\le(4/3)\delta\le1/2 ϵ ≤ ( 4/3 ) δ ≤ 1/2 . The sums of the δ \delta δ , of the z e x t z_{\mathrm{ext}} z ext , and of the ϵ \epsilon ϵ are finite, and
log ∏ ( 1 − δ ) − 1 ≤ 2 ∑ δ , log ∏ ( 1 − ϵ ) − 1 ≤ 2 ∑ ϵ , \log\prod(1-\delta)^{-1} \le2\sum\delta,\qquad\log\prod(1-\epsilon)^{-1} \le2\sum\epsilon, log ∏ ( 1 − δ ) − 1 ≤ 2 ∑ δ , log ∏ ( 1 − ϵ ) − 1 ≤ 2 ∑ ϵ ,
also for every finite subproduct. These sums are bounded by any prescribed positive number after increasing m 0 m_0 m 0 , and they tend to zero as m → ∞ m \to\infty m → ∞ , uniformly on ( 0 , Γ ] (0,\Gamma] ( 0 , Γ ] .
In particular, part (3) of Lemma 20.5 holds. All constants in this lemma have the dependence described in item (4) of Definition 4.14; the local constants of Lemma 20.5 keep their own dependence.
Proof. Let x = m / γ ≥ m 0 / Γ x = m/\gamma\ge m_0/\Gamma x = m / γ ≥ m 0 /Γ . For every fixed k k k we have ( 1 + log x ) k / x → 0 (1+\log x)^k/x \to0 ( 1 + log x ) k / x → 0 as x → ∞ x \to\infty x → ∞ : with y = log x y=\log x y = log x , the exponential series bounds y k e − y / 2 y^k e^{-y/2} y k e − y /2 on a half-line y ≥ y 0 y \ge y_0 y ≥ y 0 . This proves part (1), for finitely many k k k simultaneously, and allows us to assume x ≥ e x \ge e x ≥ e throughout.
For part (2), let Δ R = ( 1 / 2 ) log R > 0 \Delta_R = (1/2)\log R > 0 Δ R = ( 1/2 ) log R > 0 . Since z ( R m ) = z ( m ) + Δ R z(Rm) = z(m) + \Delta_R z ( R m ) = z ( m ) + Δ R , the ratios are
W a ( R m , γ ) W a ( m , γ ) = R 3 e − a ( 2 Δ R z + Δ R 2 ) , (101) \frac{W_a(Rm,\gamma)}{W_a(m,\gamma)} = R^3 e^{-a(2\Delta_R z+\Delta_R^2)}, \tag*{(101)} W a ( m , γ ) W a ( R m , γ ) = R 3 e − a ( 2 Δ R z + Δ R 2 ) , (101)
X c ( R m , γ ) X c ( m , γ ) = R 3 e − c ( R − 1 ) m / γ . (102) \frac{X_c(Rm,\gamma)}{X_c(m,\gamma)} = R^3 e^{-c(R-1)m/\gamma}. \tag*{(102)} X c ( m , γ ) X c ( R m , γ ) = R 3 e − c ( R − 1 ) m / γ . (102)
Their maxima over m ≥ m 0 m \ge m_0 m ≥ m 0 and 0 < γ ≤ Γ 0 < \gamma\le\Gamma 0 < γ ≤ Γ occur at the smallest m m m and the largest γ \gamma γ , and they tend to zero as m 0 m_0 m 0 increases, since z ≥ ( 1 / 2 ) log ( M m 0 / Γ ) → ∞ z \ge(1/2)\log(Mm_0/\Gamma) \to\infty z ≥ ( 1/2 ) log ( M m 0 /Γ ) → ∞ and m / γ ≥ m 0 / Γ → ∞ m/\gamma\ge m_0/\Gamma\to\infty m / γ ≥ m 0 /Γ → ∞ . We choose m 0 m_0 m 0 so that both ratios are at most 1 / 2 1/2 1/2 for every R ∈ R R \in\mathcal{R} R ∈ R . The logarithmic derivatives in log m \log m log m are 3 − c m / γ 3-cm/\gamma 3 − c m / γ for X c X_c X c and 3 − a z 3-az 3 − a z for W a W_a W a , and those in log γ \log\gamma log γ are c m / γ cm/\gamma c m / γ and a z az a z . Hence X c X_c X c and W a W_a W a are increasing in γ \gamma γ , and they are decreasing in m m m once c m 0 / Γ > 3 cm_0/\Gamma> 3 c m 0 /Γ > 3 and ( a / 2 ) log ( M m 0 / Γ ) > 3 (a/2)\log(Mm_0/\Gamma) > 3 ( a /2 ) log ( M m 0 /Γ ) > 3 . The logarithms of the first terms at γ = Γ \gamma=\Gamma γ = Γ are 3 log m − c m / Γ 3\log m-cm/\Gamma 3 log m − c m /Γ and 3 log m − a log 2 ( M m / Γ ) / 4 3\log m-a\log^2(Mm/\Gamma)/4 3 log m − a log 2 ( M m /Γ ) /4 , and both tend to − ∞ -\infty − ∞ . By monotonicity in γ \gamma γ , the first terms tend to zero uniformly on ( 0 , Γ ] (0,\Gamma] ( 0 , Γ ] . Summing the series with ratio at most 1 / 2 1/2 1/2 shows that the sums over m R k mR^k m R k are at most twice their first terms, which proves part (2).
For part (3), the additional side η m R j \eta mR^j η m R j is handled in the same way. Since j ≥ 0 j \ge0 j ≥ 0 and R > 1 R > 1 R > 1 , it is at least η m \eta m η m . Let m l o w = min { 1 , min η } m m_{\mathrm{low}}=\min\{1,\min\eta\}m m low = min { 1 , min η } m . For sufficiently large m m m , monotonicity in m m m bounds the term at the additional side by the corresponding function at m l o w m_{\mathrm{low}} m low , whether or not η < 1 \eta< 1 η < 1 . Since m l o w m_{\mathrm{low}} m low is a fixed positive multiple of m m m , the sums over all sides are at most three times the first terms evaluated at m l o w m_{\mathrm{low}} m low , and these tend to zero uniformly on ( 0 , Γ ] (0,\Gamma] ( 0 , Γ ] as m → ∞ m \to\infty m → ∞ , for finitely or infinitely many sides. We take the maximum of the finitely many minimum sides so obtained. Increasing m 0 m_0 m 0 once more, we make every δ \delta δ at most 1 / 2 1/2 1/2 , every z e x t z_{\mathrm{ext}} z ext at most 1 / 4 1/4 1/4 , and the sums smaller than any prescribed positive number. Then ϵ = δ / ( 1 − z e x t ) ≤ ( 4 / 3 ) δ \epsilon=\delta/(1-z_{\mathrm{ext}})\le(4/3)\delta ϵ = δ / ( 1 − z ext ) ≤ ( 4/3 ) δ , so the sum and the maximum of the ε \varepsilon ε can be made as small as desired as well, in particular every ε \varepsilon ε at most 1 / 2 1/2 1/2 . For 0 ≤ u ≤ 1 / 2 0 \le u \le1/2 0 ≤ u ≤ 1/2 ,
− log ( 1 − u ) = ∫ 0 u ( 1 − v ) − 1 d v ≤ 2 u . -\log(1-u) = \int_{0}^{u} (1-v)^{-1}\,\mathrm{d}v \le2u. − log ( 1 − u ) = ∫ 0 u ( 1 − v ) − 1 d v ≤ 2 u .
Summing this inequality bounds the logarithms of all reciprocal products, including every finite subproduct; the infinite products are defined by the increasing finite sums of logarithms. This proves part (3).
Part (3) of Lemma 20.5 is the case R = B R = B R = B , with C = C J C = C_J C = C J , a = a W a = a_W a = a W , c = b c = b c = b , and the finite set of additional ratios, since log 2 M m / γ = z 2 \log^2 \sqrt{Mm}/\gamma= z^2 log 2 M m / γ = z 2 ; its bound ε t ≤ 2 δ t \varepsilon_t \le2\delta_t ε t ≤ 2 δ t follows from ε t ≤ ( 4 / 3 ) δ t \varepsilon_t \le(4/3)\delta_t ε t ≤ ( 4/3 ) δ t . The local concentration statements of Lemma 20.5 use only its hypotheses on the products and on the coupling at the given step; the restriction to a coupling bound enters only in the summation just proved, so those local statements and their prefactors are unchanged.
Large-field decompositions and conditional densities
This section connects the renormalization group on unions of blocks, constructed in Sections 18 to § 2.2 , with the two measures to which it is applied in Sections 22 and 23: the Wilson measure and the stopped measure. All statements concern a fixed lattice and are finite-dimensional.
The analytic theorems of § 2.2 integrate the normalized fluctuation variables of one RG step over a fixed product of balls. They work in local coordinates near the identity, on a block region from which a fixed complement has been removed, with the link variables of that complement held fixed as frozen coordinates. To apply them to the Wilson measure, we must solve three problems. The first is coverage. A fiber of the block averaging map consists of all fine configurations with a given average. It is a Borel subset of the compact configuration space, can contain many gauge-equivalent representatives, and need not lie in a coordinate chart. We show that every point of every fiber carrying mass becomes, after a gauge transformation that fixes all prescribed vertices, a point of the domain of the local inverse coordinates, and that its normalized fluctuation coordinates lie in an enclosing product domain on which the conditional fluctuation potential is uniformly convex. The second problem is the choice of a density. The later estimates evaluate densities at particular gauge representatives. These form sets of Haar measure zero, so a Radon–Nikodym derivative defined only almost everywhere does not suffice. We construct an explicit Borel density from the Haar disintegration formulas and prove that it is invariant at every point under the relevant gauge transformations. The third problem is the preservation of frozen coordinates. When the integration stops near a region of large curvature, the link variables there are kept on the lattice where they were frozen. The gauge transformations used afterward must not change them, and the analytic estimates require their logarithms to be small.
Section 21.1 proves locality of the iterated block averaging map (Lemma 21.1 ) and a gauge-fixing theorem on arbitrary unions of large closed cells (Theorem 21.3 ). The last part of that theorem is a rule that protects the endpoints of all frozen coordinates. Section 21.2 defines a deterministic sequence of curvature tests, partitions the configuration space according to their outcomes into large-field decompositions (Theorem 21.5 ), and verifies the hypotheses of the analytic theorems of § 2.1 and § 2.2 on each decomposition (Lemma 21.7 ). Section 21.3 proves the coverage of every supported fiber point and constructs, at all stages, the representatives satisfying the reconstruction identities (Theorem 21.9): they are obtained from the coarsest scale downward by the reconstruction maps, so that each reproduces the output field above it and contains the frozen values. At the first RG step under the Wilson measure, this section uses the lower bound for the constrained Hessian of Lemma 18.14, which is uniform in the region and in the frozen variables. Section 21.4 constructs the invariant densities (Theorem 21.10 ), and Section 21.5 proves the identity relating the densities before and after freezing (Lemma 21.11 ). Section 21.6 first compares an integral over a fiber with an integral over the fluctuation cutoff domain (Lemma 21.12 ), and then combines these results in the induction of Theorem 21.13 .
The inputs are the quantitative extension of maps into SU(2) in Lemma 7.5; the stopped measure and its coordinates on the torus (Theorems 14.3 and 14.6); the Haar disintegration on block regions (Lemma 18.7); the gauge transformation fixed at prescribed endpoints and the perturbation theorem of Lemma 19.5 and Theorem 19.9 ; and the analytic integration, the comparison of normalization constants, and the concentration estimate of Theorems 20.13 and 20.14 and Lemma 20.5 . Theorems 21.5 , 21.10 , and 21.13 are used in Section 22 for the bounds on Wilson densities and the exponential moment bounds, and Theorems 21.3 , 21.5 , and 21.9 are used in the choice of parameters in Section 23.
The organization by nested regions, fields frozen on their complements, and effective densities on the remaining regions follows Bałaban’s large-field renormalization [18, 20, 21]. Dimock’s account of Bałaban’s method for the scalar ϕ 3 4 \phi^4_3 ϕ 3 4 model describes the partition into small-field and large-field regions, the unintegrated complement, and localized block averaging [67, §§1–2]; the summation of the resulting expansion is in [69]. These works concern a different model and different coordinates. The deterministic tests for the block averaging map T M T_M T M , the gauge construction on arbitrary unions of cells, the protection of frozen endpoints, the coverage of every fiber point, and the invariant density versions below are proved here.
Conventions. Throughout, M ≥ 3 M \ge3 M ≥ 3 is odd, J ≥ 1 J \ge1 J ≥ 1 , and B = M J B = M^J B = M J . One application of T M T_M T M (Definition 4.3), including its Borel extension, is an averaging step, and an RG step consists of J J J consecutive averaging steps. A group of 1 ≤ s ≤ J 1 \le s \le J 1 ≤ s ≤ J consecutive averaging steps with output side m m m has intermediate scales 0 ≤ i < s 0 \le i < s 0 ≤ i < s and couplings
g i = t max M ( i − s + 1 ) / 2 , t max = γ M m , λ ∗ = M − ( J − 1 ) / 2 ≤ g i t max ≤ 1. g_i = t_{\max}M^{(i-s+1)/2}, \qquad t_{\max} = \sqrt{\frac{\gamma}{Mm}}, \qquad\lambda_* = M^{-(J-1)/2} \le\frac{g_i}{t_{\max}} \le1. g i = t m a x M ( i − s + 1 ) /2 , t m a x = M m γ , λ ∗ = M − ( J − 1 ) /2 ≤ t m a x g i ≤ 1.
The angle of U ∈ S U ( 2 ) U \in\mathrm{SU}(2) U ∈ SU ( 2 ) is ρ G ( U ) \rho_G(U) ρ G ( U ) , its distance from I I I in the bi-invariant metric of Section 4; we have ρ G ( U 1 ⋯ U l ) ≤ ∑ j ρ G ( U j ) \rho_G(U_1 \cdots U_l) \le\sum_j \rho_G(U_j) ρ G ( U 1 ⋯ U l ) ≤ ∑ j ρ G ( U j ) and ∣ Log U ∣ = ρ G ( U ) |\operatorname{Log} U| = \rho_G(U) ∣ Log U ∣ = ρ G ( U ) when ρ G ( U ) < π \rho_G(U) < \pi ρ G ( U ) < π . We write c c c for the output field. In analytic functions and in the coordinates of Definition 18.6, the coarse argument is its identity-chart logarithm Log c \operatorname{Log} c Log c , which we also denote by c c c when no confusion can arise; since ∣ Log c e ∣ = ρ G ( c e ) |\operatorname{Log} c_e| = \rho_G(c_e) ∣ Log c e ∣ = ρ G ( c e ) for every output link e e e , bounds on the two agree. A block region Ω \Omega Ω is a union of blocks of the output lattice of the current group of averaging steps; its complement is the fixed complement. The frozen coordinates are written b r e c b_{\mathrm{rec}} b rec . They consist of frozen links and of based transports of frozen links across the boundary of the block region (defined in § 2.3.2 ), each kept on the lattice on which it was frozen. A link is across the boundary of the block region if exactly one of its endpoints lies in it.
Fix the cutoff amplitude A W ≥ 1 A_{\mathrm{W}} \ge1 A W ≥ 1 at the outset. Every application of Theorem 20.14 below uses the coupling ceiling g ∗ g_* g ∗ chosen after this amplitude, and g ∗ g_* g ∗ may depend on A W A_{\mathrm{W}} A W . The analytic constants and the conditions used to choose J J J do not depend on A W A_{\mathrm{W}} A W .
Unless stated otherwise, constants written C J C_J C J , c J c_J c J , u J u_J u J , ϵ J \epsilon_J ϵ J , and similar depend only on M M M , J J J , the fixed larger construction and analytic domains and their radius ratios, the spatial and boundary decay rates and their margins, the bounded density of labels, and the fixed ceilings on actions and perturbations. They are independent of L L L , γ \gamma γ , all lattice sides, the total and the preceding depth, the number of RG steps since a coordinate was frozen, the shape, topology, and number of components of the regions, and the values of the frozen coordinates. This is the dependence described in item (3) of Definition 4.14. Where stated, these constants are also independent of the widths of boundary bands, removed bands, protected neighborhoods, and cells chosen later; such widths enter only through explicit geometric counts. No uniformity as J → ∞ J \to\infty J → ∞ is asserted. The constant C δ C_\delta C δ in Theorem 21.13 has instead the dependence described in item (4) of Definition 4.14.
Local path products and closed cell complexes
The full dependence set of a link of an averaged configuration is the set of links read by the block averaging map, including its Borel extension at exceptional values, by its reference path, and by every path product used to check that the logarithms involved lie in the correct branch. For an iterate we replace every link recursively by its full dependence set, and for a plaquette we take the union of the sets of its four links. These lists are treated as sets when we define excluded label sets; repetitions are kept only to estimate the lengths of path products. Geometric neighborhoods use the periodic maximum coordinate distance. Replacing it by the lattice distance used in the analytic norms changes only dimensional comparison constants.
Lemma 21.1. Fix M M M and J J J . There exist finite constants D J D_J D J , L J L_J L J , v J v_J v J , depending only on M M M and J J J , and constants u J > 0 u_J > 0 u J > 0 and C J < ∞ C_J < \infty C J < ∞ , with the following properties for every group of s ≤ J s \le J s ≤ J averaging steps.
The full dependence set of a link, of a plaquette, or of a prescribed based transport of an averaged configuration is contained in the neighborhood of radius D J D_J D J of its output label, in output lattice units, and contains at most v J v_J v J input labels per output unit cube. The expanded hierarchical tree paths have length at most L J L_J L J .
Suppose that every input link of such a full dependence set has angle at most u ≤ u J u \le u_J u ≤ u J . Then every link appearing in the recursive computation of the selected output object has angle at most ( 10 M ) i u (10M)^i u ( 10 M ) i u at level i ≤ s i \le s i ≤ s . The corresponding relative logarithmic coordinates of the block averaging constraint, including the eliminated coordinates, have size at most C J u C_Ju C J u . After one fixed reduction of u J u_J u J , they belong to the domain of the local inverse of the block averaging constraint, with a fixed positive margin.
These conclusions do not depend on the region or on its components.
Proof. An ordered path from the center of an M M M -block to a vertex of the block has length at most 3 ( M − 1 ) / 2 3(M-1)/2 3 ( M − 1 ) /2 . The path F c , x F_{c,x} F c , x of Definition 4.3 has length at most 4 M − 3 4M-3 4 M − 3 , and F c , x C c − 1 F_{c,x}C_c^{-1} F c , x C c − 1 has length at most 5 M − 3 5M-3 5 M − 3 . If ( 5 M − 3 ) u < 2 π (5M-3)u < 2\pi ( 5 M − 3 ) u < 2 π , every relative transport U ( F c , x ) U ( C c ) − 1 U(F_{c,x})U(C_c)^{-1} U ( F c , x ) U ( C c ) − 1 is in the identity chart, and the exponential of the average of their logarithms has angle at most ( 5 M − 3 ) u (5M-3)u ( 5 M − 3 ) u . Multiplying by U ( C c ) U(C_c) U ( C c ) , we obtain the bound ( 6 M − 3 ) u ≤ 10 M u (6M-3)u \le10Mu ( 6 M − 3 ) u ≤ 10 M u . Iteration is permitted when ( 10 M ) J u (10M)^J u ( 10 M ) J u is smaller than the fixed positive margin of one averaging step. The Borel extension reads the same finite list of links, so the locality of the list does not require this smallness.
The based tree gauge on a block is obtained by multiplying transports along its ordered tree paths. It sets each tree link to I I I , equals I I I at the root, and is unique with these properties, since the equation on an oriented tree edge determines the gauge at the child from the gauge at the parent. At the next intermediate scale we make the same construction and lift the new gauge as a constant on each of the previous blocks. Such a lift conjugates each previously fixed tree link and hence preserves the value I I I . The conditions at coarser scales are fixed before the finer variables are expressed. The resulting hierarchical transports are products of these tree factors at the various scales; no eliminated constraint edge is inserted into a tree. Each factor equals I I I on the hierarchical tree gauge section. There are at most J J J scales, each with finitely many paths in a fixed block, and by substitution we obtain a finite maximal path length L J L_J L J and a finite maximal radius D J D_J D J of the dependence neighborhoods. Equivalently, we may take these constants to be the maxima over the finitely many oriented dependence neighborhoods for 1 ≤ s ≤ J 1 \le s \le J 1 ≤ s ≤ J on the infinite cubic lattice: restriction to a region only deletes path products, and passage to a periodic quotient only identifies labels. The maximal number of input labels in the same dependence neighborhoods gives ν J \nu_J ν J .
After the tree gauge transformation, every link outside the trees is a product of a number, bounded in terms of M M M and J J J , of the small links just bounded, and the coarse-field lift of the target is another such product. The relative vector is therefore bounded by C J u C_Ju C J u . The constraints have the eliminated-edge inverse and the uniqueness neighborhood, with its positive margin, used in Lemma 18.7 and Theorem 20.14 . We choose u J u_J u J so that the preceding bound is less than half the radius of that neighborhood, taking the minimum over s ≤ J s \le J s ≤ J . Then the relative solution, including its eliminated entries, is the unique solution given by that inverse. This identifies the branch of a given solution; it does not solve the constraints anew. Every path used above stays in a block of at most J J J scales, and none follows a tree spanning a whole component.
The next corollary gives the initial bounds for a stopped measure; it is used in the proofs of Theorems 21.5 and 21.9.
Corollary 21.2. Before the bound β r e c \beta_{\mathrm{rec}} β rec for frozen coordinates and the support radius r r r are fixed, one can fix C a d , J < ∞ C_{\mathrm{ad},J}<\infty C ad , J < ∞ , depending only on M M M , J J J , and the fixed tree conventions, with the following property. Let U ∈ D r U\in D_r U ∈ D r be an initial configuration of a stopped measure on a lattice of side M s m M^s m M s m , with s ≤ J s\le J s ≤ J , m ≥ 4 m\ge4 m ≥ 4 , and no previously frozen coordinates, and assume that r r r is below the fixed margin of the finite path-product charts. Then, for every union of blocks, there exists a gauge representative in which the output field and all frozen links and based transports at the input level are bounded by C a d , J r C_{\mathrm{ad},J}r C ad , J r . This is a geometric existence statement, uniform in the preceding depth, the side, and the region.
Proof. A gauge transformation taking U U U into B r B_r B r gives the bound r r r , not r / 2 r/2 r /2 , at the input. By Lemma 21.1 , all intermediate internal links have angle at most ( 10 M ) J r (10M)^Jr ( 10 M ) J r . The expanded hierarchical tree paths have at most L J L_J L J factors of these links. A link after the tree gauge transformation is a product of a path, a link, and an inverse path, and has angle at most ( 2 L J + 1 ) ( 10 M ) J r (2L_J+1)(10M)^Jr ( 2 L J + 1 ) ( 10 M ) J r . A based transport is a path times its link across the boundary of the block region and has the smaller bound ( L J + 1 ) ( 10 M ) J r (L_J+1)(10M)^Jr ( L J + 1 ) ( 10 M ) J r . Links lying entirely in the fixed complement remain below r r r , since the regional tree gauge is the identity there. The output roots are fixed by the tree gauge, and the output bound is ( 10 M ) J r (10M)^Jr ( 10 M ) J r . We conclude that C a d , J = 2 ( 2 L J + 3 ) ( 10 M ) J C_{\mathrm{ad},J}=2(2L_J+3)(10M)^J C ad , J = 2 ( 2 L J + 3 ) ( 10 M ) J exceeds twice each of these bounds. We fix it now, and only afterward require C a d , J r C_{\mathrm{ad},J}r C ad , J r to be below the fixed identity and inverse chart radii.
We now turn to gauge fixing on large closed cells. Small curvature does not force small links: a flat configuration with nontrivial holonomy around a noncontractible cycle cannot be gauge transformed to the identity. The following theorem shows that on a union of closed cells of side about R R R , plaquette angles of order R − 2 R^{-2} R − 2 allow a gauge in which all links have angle of order R − 1 R^{-1} R − 1 . The term R − 1 R^{-1} R − 1 comes from distributing compact gauge values over the edges of the cell grid, which have length at least R / 2 R/2 R /2 ; the curvature contributes O ( δ R ) O(\delta R) O ( δ R ) through an axial gauge in each cell. Transitions between the frames of neighboring cells are compared only on common edges and are never multiplied along chains of cells, so the constants do not depend on the topology of the union. We combine standard axial frames and the extension of maps into S U ( 2 ) ≃ S 3 \mathrm{SU}(2)\simeq S^3 SU ( 2 ) ≃ S 3 with the quantitative patching estimate of Lemma 7.5, preserving all prescribed vertices. The gauge fixing of lattice configurations with small plaquettes goes back to Bałaban [14]; the version below, on arbitrary unions of cells and with prescribed vertex values, is proved here.
Theorem 21.3. There exist universal constants C , C ′ , R 0 < ∞ C,C',R_0<\infty C , C ′ , R 0 < ∞ and c > 0 c>0 c > 0 with the following properties. Work in lattice units, let R ≥ R 0 R\ge R_0 R ≥ R 0 be an integer and n ≥ 8 R n\ge8R n ≥ 8 R , and partition each
coordinate circle of the periodic lattice of side n n n into intervals of lengths in [ R / 2 , 2 R ] [R/2,2R] [ R /2 , 2 R ] . Let P P P be any union of the resulting closed three-dimensional cells, and let U U U be a configuration on the links of P P P such that
ρ G ( U p ) ≤ δ for every plaquette of every cell of P , δ R 2 ≤ 1 , \rho_G(U_p) \le\delta\quad\text{for every plaquette of every cell of }P,\qquad\delta R^2 \le1, ρ G ( U p ) ≤ δ for every plaquette of every cell of P , δ R 2 ≤ 1 ,
including the plaquettes in the boundaries of the cells. We take δ ≥ 0 \delta\ge0 δ ≥ 0 and, in part (2), η ≥ 0 \eta\ge0 η ≥ 0 . A maximum of link angles over an empty set of links is defined to be zero. These hypotheses are in force in all parts below.
There exists a gauge transformation k k k on the vertices of P P P with
max e ∈ P ρ G ( U e k ) ≤ C ( δ R + R − 1 ) ≤ C ′ / R . \max_{e\in P}\rho_G(U_e^k) \le C(\delta R+R^{-1}) \le C'/R. e ∈ P max ρ G ( U e k ) ≤ C ( δ R + R − 1 ) ≤ C ′ / R .
This holds for disconnected P P P , for cells meeting only at vertices or edges, and for arbitrary holes, handles, periodic wrapping, and holonomy; the constants depend on none of these features, nor on n n n , L L L , the coupling, the depth, or J J J . If q ( U p ) ≤ d ∗ q(U_p)\le d_* q ( U p ) ≤ d ∗ for every such plaquette, one may take δ = π 2 d ∗ \delta=\pi\sqrt{2d_*} δ = π 2 d ∗ . For a fixed sufficiently large A 8 A_8 A 8 , the choice R = ⌈ A 8 / r ⌉ R=\lceil A_8/r\rceil R = ⌈ A 8 / r ⌉ together with π 2 d ∗ R 2 ≤ 1 \pi\sqrt{2d_*}R^2\le1 π 2 d ∗ R 2 ≤ 1 gives the bound ρ G ( U e k ) < r / 2 \rho_G(U_e^k)<r/2 ρ G ( U e k ) < r /2 for every link.
Let Γ ⊂ P \Gamma\subset P Γ ⊂ P be a union of closed cells of the grid, and let k Γ k_\Gamma k Γ be a gauge transformation on the vertices of Γ \Gamma Γ with
ρ G ( U e k Γ ) ≤ η ( e ⊂ Γ ) , K R = 1 + η R + δ R 2 , K R 2 / R ≤ c . \rho_G(U_e^{k_\Gamma})\le\eta\quad(e\subset\Gamma),\qquad K_R=1+\eta R+\delta R^2,\qquad K_R^2/R\le c. ρ G ( U e k Γ ) ≤ η ( e ⊂ Γ ) , K R = 1 + η R + δ R 2 , K R 2 / R ≤ c .
Then there exists an extension k k k equal to k Γ k_\Gamma k Γ at every vertex of Γ \Gamma Γ with
max e ∈ P ρ G ( U e k ) ≤ C K R 10 / R . \max_{e\in P}\rho_G(U_e^k)\le CK_R^{10}/R. e ∈ P max ρ G ( U e k ) ≤ C K R 10 / R .
Every link contained in a cell with no vertex in Γ \Gamma Γ satisfies the stronger bound C ′ / R C'/R C ′ / R . In particular, this bound holds after removing two layers of cells around Γ \Gamma Γ .
For fixed M M M and J J J , let C B ≥ ( 10 M ) J C_B\ge(10M)^J C B ≥ ( 10 M ) J also bound the lengths of the finite tree path products. Assume η ≤ C B r \eta\le C_Br η ≤ C B r , r ≤ 1 r\le1 r ≤ 1 , and R = ⌈ A 8 / r ⌉ R=\lceil A_8/r\rceil R = ⌈ A 8 / r ⌉ , and set
K B = 2 + C B ( A 8 + 1 ) , C c o l = C K B 10 / A 8 . K_B=2+C_B(A_8+1),\qquad C_{\mathrm{col}}=CK_B^{10}/A_8. K B = 2 + C B ( A 8 + 1 ) , C col = C K B 10 / A 8 .
If K B 2 r / A 8 ≤ c K_B^2r/A_8\le c K B 2 r / A 8 ≤ c and C c o l r < β r e c / 2 C_{\mathrm{col}}r<\beta_{\mathrm{rec}}/2 C col r < β rec /2 , then all links of P P P are smaller than β r e c / 2 \beta_{\mathrm{rec}}/2 β rec /2 , and, after one fixed increase of A 8 A_8 A 8 , the links in cells with no vertex in Γ \Gamma Γ are smaller than r / 2 r/2 r /2 . The constants in this part have the dependence described in item (3) of Definition 4.14. These statements assert the existence of gauge transformations; no configuration-dependent gauge transformation is used as a change of variables in an integral.
At later scales, use the following endpoint protection rule. On the lattice on which each frozen coordinate was frozen, mark both endpoints of every frozen link, both endpoints of
every based transport, and every prescribed basepoint or root. A block of the current output lattice is excluded whenever its full centered preimage on the lattice of some mark contains that mark. Then every gauge transformation supported on the remaining blocks has a lift, constant on blocks, which equals I I I at every mark, and this lift preserves every frozen link and every based transport exactly. This holds at every finite depth, for coordinates frozen at any earlier step and at periodic seams, with no sum over the steps at which the coordinates were frozen. A newly chosen root of the Haar disintegration is fixed separately by the current based gauge. For the gauge transformation of Lemma 19.5, the current endpoints of the links across the boundary of the block region are included among the fixed vertices as well.
Proof. We give the absolute and the relative constructions on the same skeleton.
Step 1: axial frames. In each cell Q Q Q of P P P , we fix a corner o Q o_Q o Q , let τ x \tau_x τ x be the coordinate-ordered path in Q Q Q from o Q o_Q o Q to x x x , and define q Q ( x ) = U ( τ x ) q_Q(x)=U(\tau_x) q Q ( x ) = U ( τ x ) . Then U x y Q U^{Q}_{xy} U x y Q is the transport around the loop τ x ( x , y ) τ y − 1 \tau_x(x,y)\tau_y^{-1} τ x ( x , y ) τ y − 1 . To bound it, we compare the ordered paths to the adjacent vertices x x x and y y y . Moving the last elementary step through the ordered path inserts at most two rectangular strips, each of width one and length at most 2 R 2R 2 R . After canceling consecutive inverse edges, the resulting loop is a product of conjugated elementary plaquettes. There are at most C R CR C R of them, all in Q Q Q , including its boundary. By the product inequality for ρ G \rho_G ρ G , we have
ρ G ( U x y Q ) ≤ C 0 δ R . (103) \rho_G(U^{Q}_{xy}) \le C_0\delta R. \tag*{(103)} ρ G ( U x y Q ) ≤ C 0 δ R . (103)
Here C 0 C_0 C 0 is absolute, since the aspect ratio of a cell is at most four. This local estimate follows from Lemma 7.1: the cell has an injective periodic lift because 2 R ≤ n / 4 2R\le n/4 2 R ≤ n /4 , and the sum of its three side lengths is at most 6 R 6R 6 R .
Step 2: transitions. For two cells Q , Q ′ Q,Q' Q , Q ′ containing a common lattice edge ( x , y ) (x,y) ( x , y ) , we define t Q Q ′ ( x ) = q Q ( x ) q Q ′ ( x ) − 1 t_{QQ'}(x)=q_Q(x)q_{Q'}(x)^{-1} t Q Q ′ ( x ) = q Q ( x ) q Q ′ ( x ) − 1 . Direct multiplication gives
t Q Q ′ ( y ) = ( U x y Q ) − 1 t Q Q ′ ( x ) U x y Q ′ , d ( t Q Q ′ ( x ) , t Q Q ′ ( y ) ) ≤ 2 C 0 δ R , (104) t_{QQ'}(y)=(U^{Q}_{xy})^{-1}t_{QQ'}(x)U^{Q'}_{xy},\qquad d(t_{QQ'}(x),t_{QQ'}(y))\le2C_0\delta R, \tag*{(104)} t Q Q ′ ( y ) = ( U x y Q ) − 1 t Q Q ′ ( x ) U x y Q ′ , d ( t Q Q ′ ( x ) , t Q Q ′ ( y )) ≤ 2 C 0 δ R , (104)
where we used (21.1) and bi-invariance. This estimate holds whether the cells share a face or only an edge. When two cells share only a vertex, only the equality of the assigned value at that vertex is required.
Step 3: the edges of the grid. We write the desired gauge in the frame of Q Q Q as s Q ( x ) = k ( x ) q Q ( x ) − 1 s_Q(x)=k(x)q_Q(x)^{-1} s Q ( x ) = k ( x ) q Q ( x ) − 1 and set k = I k=I k = I at every vertex of the cell grid. On each edge of the grid, we fix one incident cell frame, interpolate the two endpoint values of s Q s_Q s Q by a shortest geodesic in S U ( 2 ) \mathrm{SU}(2) SU ( 2 ) , and sample it at the lattice vertices. Since the group has diameter 2 π 2\pi 2 π and the edge has length at least R / 2 R/2 R /2 , the speed is at most 4 π / R 4\pi/R 4 π / R . Each edge of the grid is assigned once, regardless of the number of incident cells. By (21.2), changing to another incident frame adds at most C δ R C\delta R C δ R per lattice edge. Since δ R 2 ≤ 1 \delta R^2\le1 δ R 2 ≤ 1 , the assigned frame values on every edge of the grid have lattice Lipschitz constant at most C 1 / R C_1/R C 1 / R . Step 4: the faces of the grid. On a face of the grid, we convert the assigned boundary values into the frame of one incident cell, and we use a single consistent fine triangulation on all faces of the grid. Since the vertices of each fine simplex can be joined by a bounded number of lattice edges in the boundary, their distances are at most C 2 / R C_2/R C 2 / R . We take R 0 R_0 R 0 so large that the normalized affine interpolation clause of Lemma 7.5 applies, with denominator at least 1 / 2 1/2 1/2 . The continuous boundary map then has Lipschitz constant C 3 / R C_3/R C 3 / R . The disk extension clause, rescaled to this rectangle, bounds the Lipschitz constant of its extension by C R − 1 ( 1 + C 3 ) 2 CR^{-1}(1+C_3)^2 C R − 1 ( 1 + C 3 ) 2 . Sampling the extension at all lattice vertices of the face, we recover k = s Q q Q k=s_Qq_Q k = s Q q Q . This is done once for each face, and previously assigned vertices are preserved exactly.
Step 5: the cells. For a cell Q Q Q of P P P , we convert the assigned boundary values into the frame of Q Q Q . On every lattice edge in the boundary their difference is at most C 4 / R C_4/R C 4 / R , by the face bound and (21.2). The same consistent triangulation and normalized interpolation give a boundary map from the sphere with Lipschitz constant C 5 / R C_5/R C 5 / R ; we increase R 0 R_0 R 0 once to satisfy its simplex margin. By the ball extension clause, we obtain a map with Lipschitz constant C R − 1 ( 1 + C 5 ) 5 CR^{-1}(1+C_5)^5 C R − 1 ( 1 + C 5 ) 5 , and sampling it gives k k k at the remaining vertices. For an edge of Q Q Q , the final comparison is the sum
ρ G ( k x U x y k y − 1 ) = ρ G ( s Q ( x ) U x y Q s Q ( y ) − 1 ) ≤ ρ G ( U x y Q ) + d ( s Q ( x ) , s Q ( y ) ) ≤ C ( δ R + R − 1 ) . (105) \begin{aligned}
\rho_G(k_xU_{xy}k_y^{-1})=\rho_G(s_Q(x)U_{xy}^Qs_Q(y)^{-1}) \\
&\le\rho_G(U_{xy}^Q)+d(s_Q(x),s_Q(y))\le C(\delta R+R^{-1}). \tag*{(105)}
\end{aligned} ρ G ( k x U x y k y − 1 ) = ρ G ( s Q ( x ) U x y Q s Q ( y ) − 1 ) ≤ ρ G ( U x y Q ) + d ( s Q ( x ) , s Q ( y )) ≤ C ( δ R + R − 1 ) . (105)
In the last inequality, we used (21.1) and the Lipschitz bound of the extension. Since δ R ≤ R − 1 \delta R\le R^{-1} δ R ≤ R − 1 , this proves the absolute estimate with C ′ = 2 C C'=2C C ′ = 2 C , after enlarging C C C to include the finitely many aspect and simplex constants. Since the assignments on the skeleton agree at every common vertex, they can be made directly on the periodic quotient and on any union of cells. Handles, the number of components, and the holonomy have not entered any constant.
Step 6: the relative construction. We keep every lattice vertex of Γ \Gamma Γ at its prescribed value. In the frame of a cell Q Q Q , the prescribed values satisfy, on each prescribed lattice edge,
d ( k Γ ( x ) q Q ( x ) − 1 , k Γ ( y ) q Q ( y ) − 1 ) ≤ η + C 0 δ R ≤ C K R / R . (106) d(k_\Gamma(x)q_Q(x)^{-1},k_\Gamma(y)q_Q(y)^{-1})\le\eta+C_0\delta R\le CK_R/R. \tag*{(106)} d ( k Γ ( x ) q Q ( x ) − 1 , k Γ ( y ) q Q ( y ) − 1 ) ≤ η + C 0 δ R ≤ C K R / R . (106)
This follows by applying the first inequality in (21.3) in reverse to the two transformed link matrices, whose angles are both known; it imposes no bound on the increments of k Γ k_\Gamma k Γ in the original frame. At the vertices of the grid not in Γ \Gamma Γ , we set k = I k=I k = I . An edge of the grid in Γ \Gamma Γ is kept, and every other edge of the grid is a geodesic in one incident frame between its assigned endpoint values. Since Γ \Gamma Γ is a union of closed cells of the grid, its intersection with an edge of the grid is either the whole edge or a subset of its endpoints, so this rule respects all prescribed values. Its lattice increments in any cell frame are bounded by C K R / R CK_R/R C K R / R .
We keep each face of the grid lying in Γ \Gamma Γ . On every other face, the boundary interpolation and the disk filling give the bound C K R 2 / R CK_R^2/R C K R 2 / R ; the interpolation is legitimate when C K R / R CK_R/R C K R / R is below its fixed simplex margin. On the boundary of a cell not in Γ \Gamma Γ , all assigned frame increments, including those on a face kept from Γ \Gamma Γ , are now at most C K R 2 / R CK_R^2/R C K R 2 / R . The universal c c c in the hypothesis is taken smaller than both required interpolation margins divided by these fixed constants. Since K R ≥ 1 K_R \ge1 K R ≥ 1 , the condition K R 2 / R ≤ c K_R^2/R \le c K R 2 / R ≤ c implies the first margin as well. The boundary interpolation on the sphere and the ball filling then give
Lip ( s Q ) ≤ C R − 1 ( 1 + C K R 2 ) 5 ≤ C K R 10 / R . \operatorname{Lip}(s_Q) \le CR^{-1}(1+CK_R^2)^5 \le CK_R^{10}/R. Lip ( s Q ) ≤ C R − 1 ( 1 + C K R 2 ) 5 ≤ C K R 10 / R .
All cells of Γ \Gamma Γ are kept unchanged. Using (21.3) and δ R ≤ R − 1 ≤ K R 10 / R \delta R \le R^{-1} \le K_R^{10}/R δ R ≤ R − 1 ≤ K R 10 / R , we obtain the relative bound.
Step 7: cells away from Γ \Gamma Γ . If Q Q Q has no vertex in Γ \Gamma Γ , no face or edge of Q Q Q is prescribed. Each edge of the grid in the boundary of Q Q Q is then a shortest geodesic of speed at most C / R C/R C / R , even if its endpoint values were assigned in the frame of another incident cell, and frame changes add at most C / R C/R C / R . The disk extensions on the faces of Q Q Q then have Lipschitz constant C / R C/R C / R , and so does the ball extension in Q Q Q . By (21.3), we obtain the stronger conclusion for every edge of Q Q Q . Removing two layers of cells around Γ \Gamma Γ removes every cell with a vertex in Γ \Gamma Γ .
Step 8: the choice of δ \delta δ and of A 8 A_8 A 8 . If U = e X U=e^X U = e X and x = ∣ X ∣ ≤ 2 π x=|X|\le2\pi x = ∣ X ∣ ≤ 2 π , then q ( U ) = 2 sin 2 ( x / 4 ) q(U)=2\sin^2(x/4) q ( U ) = 2 sin 2 ( x /4 ) and sin ( x / 4 ) ≥ x / ( 2 π ) \sin(x/4)\ge x/(2\pi) sin ( x /4 ) ≥ x / ( 2 π ) . Hence x 2 ≤ 2 π 2 q ( U ) x^2\le2\pi^2q(U) x 2 ≤ 2 π 2 q ( U ) , and the stated value of δ \delta δ is valid. Let A 8 A_8 A 8 be larger than 4 C ′ 4C' 4 C ′ and R 0 R_0 R 0 , increased by the fixed geometric factors used later in the excluded sets. The standing hypothesis gives R ≥ R 0 R\ge R_0 R ≥ R 0 . Also R ≥ A 8 / r R\ge A_8/r R ≥ A 8 / r , hence C ′ / R ≤ ( C ′ / A 8 ) r < r / 2 C'/R\le(C'/A_8)r<r/2 C ′ / R ≤ ( C ′ / A 8 ) r < r /2 . For part (3) we have A 8 / r ≤ R ≤ ( A 8 + 1 ) / r A_8/r\le R\le(A_8+1)/r A 8 / r ≤ R ≤ ( A 8 + 1 ) / r , so
K R ≤ 2 + C B ( A 8 + 1 ) = K B , C K R 10 / R ≤ C c o l r , K R 2 / R ≤ K B 2 r / A 8 . K_R \le2+C_B(A_8+1)=K_B,\qquad CK_R^{10}/R\le C_{\mathrm{col}}r,\qquad K_R^2/R\le K_B^2r/A_8. K R ≤ 2 + C B ( A 8 + 1 ) = K B , C K R 10 / R ≤ C col r , K R 2 / R ≤ K B 2 r / A 8 .
The two smallness conditions in part (3) then imply the stated bounds. Since the constant C B C_B C B is fixed before r r r , these conditions are upper bounds on r r r and a lower bound on the ratio β r e c / r \beta_{\mathrm{rec}}/r β rec / r , and they are not circular.
Step 9: endpoint protection. In one coordinate, the balanced-digit identity
{ ∑ j = 0 s − 1 M j a j : a j ∈ { − ( M − 1 ) / 2 , … , ( M − 1 ) / 2 } } = { − ( M s − 1 ) / 2 , … , ( M s − 1 ) / 2 } \left\{\sum_{j=0}^{s-1}M^j a_j:a_j\in\{-(M-1)/2,\ldots,(M-1)/2\}\right\}=\{-(M^s-1)/2,\ldots,(M^s-1)/2\} { j = 0 ∑ s − 1 M j a j : a j ∈ { − ( M − 1 ) /2 , … , ( M − 1 ) /2 } } = { − ( M s − 1 ) /2 , … , ( M s − 1 ) /2 }
holds: existence follows by successively taking the balanced residue modulo M M M , and uniqueness by reducing two such expressions modulo M M M and dividing their difference by M M M . Taking the product over the three coordinates and then the periodic quotient, we find that each vertex of the lattice of a mark has exactly one ancestor in any current centered partition, also at a seam. A mark in the preimage of a block would exclude that block. A gauge transformation supported on the remaining blocks is I I I at every such ancestor, and its lift, constant on blocks, is I I I at the mark. For any path, the gauge factors at the interior vertices cancel in order, and only the two endpoint factors remain. This proves that every marked frozen link and every based transport is preserved exactly, for all marks at once and with no sum over earlier steps. Finally, the cutoff in Lemma 19.5 vanishes at the prescribed endpoints of the links across the boundary of the block region as well, which proves the last statement.
Deterministic large-field decompositions
In this section, we partition the support of the initial measure according to the outcomes of a deterministic sequence of curvature tests, and we show that on each piece every block region admits a representative with small links and small newly frozen coordinates.
Based transports and frozen coordinates. Let e = ( u e , v e ) e=(u_e,v_e) e = ( u e , v e ) be a link across the boundary of the block region, oriented from its endpoint u e u_e u e in the block region to its endpoint v e v_e v e in the fixed complement, and let σ e \sigma_e σ e be the hierarchical tree path of Lemma 21.1 from the root α e \alpha_e α e of the block containing u e u_e u e to u e u_e u e . The based transport of e e e is b ^ e = U ( σ e ) U e \widehat{b}_e=U(\sigma_e)U_e b e = U ( σ e ) U e , the transport from α e \alpha_e α e to v e v_e v e ; its endpoints are α e \alpha_e α e and v e v_e v e , while the endpoints of the link itself are u e u_e u e and v e v_e v e . When integration stops at a boundary, such links are frozen through their based transports, and the other links leaving the integration are frozen as links. We write c j ♯ c_j^\sharp c j ♯ for the analytic coordinates of newly frozen coordinates: in the coordinates we use, the logarithm Log b \operatorname{Log} b Log b of the frozen group value b b b , or a fixed analytic function of finitely many such logarithms (see (107) ). Group values of frozen coordinates are never compared with analytic radii directly.
Parameters. We fix M M M , J J J , the construction domains of Theorem 20.14 , and A 8 A_8 A 8 sufficiently large as in Theorem 21.3 . Then we fix the frozen-coordinate conversion constant C c o n v , J C_{\mathrm{conv},J} C conv , J of (109) below. It is computed from the relative extension of Theorem 21.3 , the finite path products of the constraints and trees, the constant C a d , J C_{\mathrm{ad},J} C ad , J of Corollary 21.2 , the constant converting the near values at the first step under the Wilson measure, and the constants converting frozen group values into analytic coordinates. It is fixed before the two radii below. Then r > 0 r>0 r > 0 and the bound β r e c > C c o n v , J r \beta_{\mathrm{rec}}>C_{\mathrm{conv},J}r β rec > C conv , J r for frozen coordinates are taken so small that every newly frozen analytic coordinate satisfies
∣ c j ♯ ∣ ≤ min { β r e c , ϕ ∙ ( j ) 4 } , C b c , ∙ E ∙ β r e c ρ ∙ ≤ σ ∗ 4 , ∙ ∈ { c l , q } , \left|c_j^\sharp\right| \le\min\left\{\beta_{\mathrm{rec}},\frac{\phi_{\bullet}(j)}{4}\right\}, \qquad\frac{C_{\mathrm{bc},\bullet}E_{\bullet}\beta_{\mathrm{rec}}}{\rho_{\bullet}} \le\frac{\sigma_*}{4}, \qquad\bullet\in\{\mathrm{cl},\mathrm{q}\}, c j ♯ ≤ min { β rec , 4 ϕ ∙ ( j ) } , ρ ∙ C bc , ∙ E ∙ β rec ≤ 4 σ ∗ , ∙ ∈ { cl , q } ,
for the classical and the quantum spatially varying analytic radii ϕ c l , ϕ q \phi_{\mathrm{cl}},\phi_{\mathrm{q}} ϕ cl , ϕ q , with the radii ρ ∙ \rho_{\bullet} ρ ∙ , ceilings E ∙ E_{\bullet} E ∙ , and boundary-completion constants C b c , ∙ C_{\mathrm{bc},\bullet} C bc , ∙ of Theorem 20.14 . These are conditions on deterministic parameters, imposed before the coupling. The deep frozen group variables at the first step under the Wilson measure are arbitrary elements of SU(2), not small analytic logarithms; they are kept as fixed parameters and are never inserted into this analytic freezing. We set
R c e l l = ⌈ A 8 / r ⌉ , W = 8 B R c e l l , π 2 d ∗ R c e l l 2 ≤ 1. R_{\mathrm{cell}}=\left\lceil A_8/r\right\rceil,\qquad W=8BR_{\mathrm{cell}},\qquad\pi\sqrt{2d_*}R_{\mathrm{cell}}^2\leq1. R cell = ⌈ A 8 / r ⌉ , W = 8 B R cell , π 2 d ∗ R cell 2 ≤ 1.
The width W W W is measured in output units of the current RG step; in output units of the next RG step it is W / B = 8 R c e l l W/B=8R_{\mathrm{cell}} W / B = 8 R cell . The excluded sets below also contain the band of width s e r s_{\mathrm{er}} s er removed at each boundary (Theorem 20.14 ) and the protected neighborhoods of the finite path products. We may decrease r r r initially so that W ≥ s e r W\geq s_{\mathrm{er}} W ≥ s er .
At each boundary between RG steps, the previous fixed complement and the earlier frozen basepoints lie outside the next block region, up to the fixed buffer of the dependence neighborhoods of the path products. Equivalently, their incoming and current boundary distances satisfy the comparison required in Theorem 19.9 ; this is proved as (21.10). The endpoint protection rule of Theorem 21.3 contains these basepoints and the newly frozen labels, and additional excluded regions preserve the inclusion. At the first step, there is no previous fixed complement, and the condition is vacuous.
Cell grids. At each lattice side n ≥ 8 R c e l l n\geq8R_{\mathrm{cell}} n ≥ 8 R cell , we fix the cell grid deterministically. For example, we divide each coordinate circle into ⌈ n / R c e l l ⌉ \lceil n/R_{\mathrm{cell}}\rceil ⌈ n / R cell ⌉ intervals, distributing the remainder among intervals of the two nearest integer lengths. For R c e l l ≥ 2 R_{\mathrm{cell}}\geq2 R cell ≥ 2 these lengths lie in [ R c e l l / 2 , 2 R c e l l ] [R_{\mathrm{cell}}/2,2R_{\mathrm{cell}}] [ R cell /2 , 2 R cell ] . When an input grid must be aligned with the blocks, we perform the same construction in units of the blocking grid and round its scale up to a multiple of the block side. Since B B B is fixed and R c e l l R_{\mathrm{cell}} R cell may be taken larger than B B B initially, this changes the scale by at most a fixed factor.
Definition 21.4. Fix a finite range of scales consisting of complete RG steps and an initial configuration U U U . The ordered tests are the following finite procedure.
Before any gauge transformation, compute the iterates V ℓ = T M ℓ U V^\ell=T_M^\ell U V ℓ = T M ℓ U at all levels of the current RG step, with the Borel extension of T M T_M T M . Go through these levels in increasing order and, at each level, through the plaquettes in a fixed order.
A plaquette f f f fails if q f = q ( V f ℓ ) ≥ d ∗ q_f=q(V_f^\ell)\geq d_* q f = q ( V f ℓ ) ≥ d ∗ and its full dependence set does not meet the current excluded label set. When f f f fails, all variables of its full dependence set that are not yet frozen are frozen at the input of its RG step, together with those of a fixed neighborhood of the dependence set. This neighborhood is aligned with the blocks of the output lattice and has output radius 2 W 2W 2 W . It includes the fixed margins of the dependence neighborhoods and the protected neighborhood of the endpoint protection rule. Previously frozen variables stay where they are. The neighborhood is added to the excluded label set. A plaquette whose dependence set meets the excluded label set is not tested; it is excluded, not declared good.
After the tests of an RG step, retain only the closed cells of the output cell grid whose full dependence sets and protected neighborhoods avoid the excluded label set, and
add every touching excluded cell to the fixed complement. Freeze the full output boundary band: the layer of retained cells next to the excluded set, its incident cells (all retained grid cells meeting that layer, even at a vertex), all labels of reduced analytic radius, the removed band of width s e r s_{\mathrm{er}} s er , and the margins of the dependence neighborhoods. Enlarge the previous fixed complement by an output neighborhood of width W W W , and then by the protected neighborhood of the endpoint protection rule of Theorem 21.3 , using every endpoint and basepoint of every frozen coordinate on the lattice where it was frozen. A cell is retained only if its closure avoids all these exclusions, so a plaquette in the boundary of a cell is tested on the same terms as an interior plaquette.
For the next RG step, keep only the blocks that lie, together with their full dependence neighborhoods, in the retained interior. The discarded band is frozen. Excluded regions are combined by taking unions. If the block region becomes empty, its frozen coordinates are kept and the procedure stops.
When the initial measure is the Wilson measure, the first RG step is modified as in Theorem 21.5 (4). The outcome of the procedure, consisting of the failing plaquettes, the excluded regions, and the label lists of the coordinates frozen at all steps, is a large-field decomposition, denoted X \mathcal{X} X . The values of these coordinates are variables of the stage measures and are not part of the finite index X \mathcal{X} X . We write A X A_{\mathcal{X}} A X for the set of initial configurations for which the procedure produces X \mathcal{X} X .
Theorem 21.5. Make the parameter choices above, including the initial smallness reductions and fixed mesh factors specified in the proof below.
For every finite range of scales, every initial side, every preceding depth d 0 = a J d_0=aJ d 0 = a J with a ≥ 0 a\geq0 a ≥ 0 , and every normalized stopped measure
d ν ^ = Z ^ − 1 1 D r e − E d 0 , s a t / g m b a s e 2 − [ R d 0 , s a t − R d 0 ( I ) ] d H m b a s e , \mathrm{d}\hat{\nu}=\hat{Z}^{-1}1_{D_r}e^{-E_{d_0,\mathrm{sat}}/g^2_{m_{\mathrm{base}}}-[R_{d_0,\mathrm{sat}}-R_{d_0}(I)]}\mathrm{d}H_{m_{\mathrm{base}}}, d ν ^ = Z ^ − 1 1 D r e − E d 0 , sat / g m base 2 − [ R d 0 , sat − R d 0 ( I )] d H m base , the sets A X A_{\mathcal{X}} A X form a finite deterministic Borel partition of the support of ν ^ \hat{\nu} ν ^ , provided that the output side of every complete RG step is at least 8 R c e l l 8R_{\mathrm{cell}} 8 R cell . The same holds when the initial measure is the Wilson measure, with the first step described in part (4). Every tested plaquette that is not excluded satisfies q f < d ∗ q_f<d_* q f < d ∗ . The gauge transformations used below to exhibit small representatives do not enter the tests.
At every boundary between RG steps there exists a gauge representative in which the links in the interior of the block region are smaller than r / 2 r/2 r /2 , and the newly frozen analytic links and based transports at that output boundary are smaller than β R e c \beta_{\mathrm{Rec}} β Rec on the lattice where they are frozen. In the same representative, the ordinary links of the prescribed closed-cell collar Γ \Gamma Γ used in the relative output extension have the stronger native bound ρ G ( b e ) ≤ C B r \rho_G(b_e)\leq C_Br ρ G ( b e ) ≤ C B r on that output lattice. These collar links belong to the full frozen output band, and both of their endpoints are protected at subsequent steps. The first-input deep Wilson group variables of
part (4) are not among these analytic output coordinates. All previously frozen coordinates are preserved together with their endpoints; no smallness of their later averaged images is asserted. A block region that becomes empty keeps its frozen coordinates, and the procedure stops there. The same representative bounds hold on the closed supports of the stage measures in part (5), by the closed-support argument in § 2.3.3 .
Every added cell and every frozen label can be assigned to a failing plaquette by a local rule. For a failing plaquette f f f , the radius R t ( f ) R_t(f) R t ( f ) of the neighborhood assigned to f f f , measured at the outputs of successive RG steps, satisfies
R t + 1 ( f ) ≤ B − 1 R t ( f ) + C J ( W + R c e l l + 1 ) , R f i r s t ( f ) ≤ C J ( W + R c e l l + 1 ) . R_{t+1}(f) \le B^{-1}R_t(f) + C_J(\mathrm{W} + R_{\mathrm{cell}} + 1), \qquad R_{\mathrm{first}}(f) \le C_J(\mathrm{W} + R_{\mathrm{cell}} + 1). R t + 1 ( f ) ≤ B − 1 R t ( f ) + C J ( W + R cell + 1 ) , R first ( f ) ≤ C J ( W + R cell + 1 ) .
where R f i r s t ( f ) R_{\mathrm{first}}(f) R first ( f ) refers to the RG step in which f f f fails. The number of scalar labels assigned to f f f is at most C J ( B W + R c e l l + 1 ) 3 C_J(B\mathrm{W} + R_{\mathrm{cell}} + 1)^3 C J ( B W + R cell + 1 ) 3 in that RG step and at most C J ( W + R c e l l + 1 ) 3 C_J(\mathrm{W} + R_{\mathrm{cell}} + 1)^3 C J ( W + R cell + 1 ) 3 in each later RG step. The final plaquettes whose full dependence sets meet a neighborhood assigned to f f f are at most
F J = C J ( W + R c e l l + 1 ) 3 F_J = C_J(\mathrm{W} + R_{\mathrm{cell}} + 1)^3 F J = C J ( W + R cell + 1 ) 3
in number, with a fixed enlargement for each of the finitely many remaining numbers of averaging steps after the last RG step (Theorem 21.13 ). These estimates are subadditive under unions. A set of failing plaquettes that can be produced by the ordered tests determines at most one large-field decomposition. The constants C J C_J C J have the dependence stated at the beginning of this section and are independent of the later widths; the displayed polynomials exhibit the dependence of the counts on the widths.
For the first RG step under the Wilson measure, the ordered tests also test the plaquettes of the input configuration, before any representative is chosen. On a fixed input cell grid of scale S S S , aligned with the blocks, with C J − 1 R c e l l ≤ S ≤ C J R c e l l C_J^{-1}R_{\mathrm{cell}} \le S \le C_JR_{\mathrm{cell}} C J − 1 R cell ≤ S ≤ C J R cell , retain only the closed cells whose full dependence sets avoid the final excluded set. Move the block region inward by one full layer of cells and by the full dependence neighborhood of the constraint, tree, and basing path products, and require
n i n p u t ≥ 8 S , π 2 d ∗ S 2 ≤ 1. n_{\mathrm{input}} \ge8S, \qquad\pi\sqrt{2d_*}S^2 \le1. n input ≥ 8 S , π 2 d ∗ S 2 ≤ 1.
Then every plaquette of the resulting closed cell complex, including its boundary plaquettes, is a tested plaquette with q < d ∗ q < d_* q < d ∗ . The complex contains every plaquette whose Wilson term depends on an integration variable and every constraint and tree path product of the block region. There are no earlier frozen vertices at this step. The gauge transformation of Theorem 21.3 (1) on this complex makes the integration variables and the near values small; it extends to the remaining vertices, whose links are kept as arbitrary elements of S U ( 2 ) \mathrm{SU}(2) SU ( 2 ) (the deep frozen group variables). The relative extension at the output gives the bounds of part (2) for the newly frozen coordinates. All additional excluded sets are assigned to the same failing plaquettes and satisfy the bounds of part (3) after one enlargement of C J C_J C J .
For either initial measure ν \nu ν define the stage measures
η t X = ( O t X ) # ( 1 A X ν ) \eta_t^{\mathcal{X}} = \left(\mathcal{O}_t^{\mathcal{X}}\right)_{\#}\left(\mathbf{1}_{A_{\mathcal{X}}}\nu\right) η t X = ( O t X ) # ( 1 A X ν )
where O t X \mathcal{O}_t^{\mathcal{X}} O t X maps a configuration to the field on the current block region together with all frozen links and based transports, each on the lattice where it was frozen. All the tests remain in these measures. For every Borel set S S S (an event, not the grid scale of part (4)), in particular for every small-field event of the form (4.6) on the same finite hierarchy,
1 S = ∑ X 1 S 1 A X \mathbf{1}_S = \sum_{\mathcal{X}} \mathbf{1}_S\mathbf{1}_{A_{\mathcal{X}}} 1 S = X ∑ 1 S 1 A X
ν \nu ν -almost surely.
The partition does not change S S S and imposes no additional restriction on the Wilson measure. All statements are uniform in the holonomy, the number of RG steps since coordinates were frozen, the topology of the regions, and the deep frozen group variables. A remaining group of fewer than J J J averaging steps at the end of the range stops before the geometric conditions on the minimal side fail.
The following lemma records which changes of representative preserve the bounds on frozen coordinates; it is used in the proof of Theorem 21.5 and afterward.
Lemma 21.6. Consider a representative on a block region of a large-field decomposition of Theorem 21.5, with each frozen coordinate kept on the lattice where it was frozen. In part (3), assume that the analytic frozen coordinates satisfy (21.8).
Every gauge transformation supported on the current block region and lifted under the endpoint protection rule equals I I I at the endpoints of all earlier frozen links and based transports. It fixes their group values and their logarithms exactly.
A gauge transformation equal to I I I at the roots of the block region and at the outer endpoints v e v_e v e of the links across the boundary of the block region fixes every based transport b ^ e \widehat b_e b e , although it may move the inner endpoint u e u_e u e . The gauge transformation of Lemma 19.5 fixes u e u_e u e as well.
A simultaneous conjugation b ↦ k b k − 1 b \mapsto kbk^{-1} b ↦ k b k − 1 of the frozen coordinates preserves (21.8).
A change b x y ↦ k x b x y k y − 1 b_{xy} \mapsto k_xb_{xy}k_y^{-1} b x y ↦ k x b x y k y − 1 with independent k x , k y k_x,k_y k x , k y need not preserve (21.8).
In the following, analytic estimates are transported between representatives only by the changes in parts (1)–(3). Changes as in part (4) are used only for the invariance of densities in Theorem 21.10. None of these conventions changes the tests or restricts the initial measure.
Proof. Part (1) is Theorem 21.3(4). For part (2), the gauge factors at the interior vertices of the path σ e \sigma_e σ e and of the link e e e cancel, so b ^ e \widehat b_e b e changes only by the factors at α e \alpha_e α e and v e v_e v e , which are I I I ; the factor at u e u_e u e cancels. The gauge transformation of Lemma 19.5 is I I I at the protected endpoints of the links across the boundary of the block region, including u e u_e u e . For part (3), we have Log ( k b k − 1 ) = Ad k Log b \operatorname{Log}(kbk^{-1})=\operatorname{Ad}_k\operatorname{Log}b Log ( k b k − 1 ) = Ad k Log b on the identity chart, and Ad k \operatorname{Ad}_k Ad k is an isometry of s u ( 2 ) \mathfrak{su}(2) su ( 2 ) . Longer path-product coordinates have the same equivariance because the interior gauge factors cancel. Relative eliminated-coordinate maps are equivariant by covariance of their defining constraint and uniqueness of its local inverse. Thus the fixed analytic coordinates used here transform by the same global adjoint action; mere analyticity of an arbitrary coordinate map would not suffice. For part (4), take b = I b=I b = I and k y = I k_y=I k y = I ; then ρ G ( k x b k y − 1 ) = ρ G ( k x ) \rho_G(k_x b k_y^{-1})=\rho_G(k_x) ρ G ( k x b k y − 1 ) = ρ G ( k x ) is arbitrary.
Proof of Theorem 21.5 . We first fix all margins and constants that do not depend on a configuration (Steps 1 and 2). We then show that the tested plaquettes that are not excluded are good (Step 3), construct the small representatives by induction over the RG steps (Step 4), and compare boundary distances (Step 5). Steps 6 and 7 prove the counts and the measurability, Step 8 treats the first step under the Wilson measure, and Step 9 identifies the stage measures.
Step 1: fixed margins. All excluded sets in this proof are sets of labels, chosen by their locations and dependence sets; their construction uses no link representative. We fix D J D_J D J , L J L_J L J , ν J \nu_J ν J from Lemma 21.1 and C a d , J C_{\mathrm{ad},J} C ad , J from Corollary 21.2 . The protected neighborhoods include one full copy of each dependence neighborhood, its incident cells, and the removed band of width s e r s_{\mathrm{er}} s er , and A 8 A_8 A 8 includes the factors required in Theorem 21.3 . The support radius will be required to give R c e l l ≥ R 0 R_{\mathrm{cell}}\ge R_0 R cell ≥ R 0 and W W W larger than s e r s_{\mathrm{er}} s er and than the radii of these finitely many dependence neighborhoods. The number of copies and their enlargement factors depend only on M M M , J J J , and the fixed construction domains; in particular, they are fixed before W W W or R c e l l R_{\mathrm{cell}} R cell is evaluated.
Step 2: constants for frozen coordinates. We specify the margins for frozen coordinates before using any representative. On the lattice where it is frozen, the analytic coordinate of a frozen link b e b_e b e or based transport b ^ e \widehat b_e b e is its identity-chart logarithm. If its angle is below π \pi π , then
∣ Log b e ∣ = ρ G ( b e ) , ∣ Log b ^ e ∣ = ρ G ( b ^ e ) . (107) \lvert\operatorname{Log} b_e\rvert=\rho_G(b_e), \qquad\lvert\operatorname{Log}\widehat b_e\rvert=\rho_G(\widehat b_e). \tag*{(107)} ∣ Log b e ∣ = ρ G ( b e ) , ∣ Log b e ∣ = ρ G ( b e ) . (107)
Here ∣ ⋅ ∣ \lvert\cdot\rvert ∣ ⋅ ∣ is the adjoint-invariant Euclidean norm on s u ( 2 ) \mathfrak{su}(2) su ( 2 ) , and the conversion constant is one. If the fixed coordinate convention also uses a longer path product or a relative eliminated coordinate, let C c r , J ≥ 1 C_{\mathrm{cr},J}\ge1 C cr , J ≥ 1 bound its path length and the row sums of the first derivative of its fixed inverse chart on the larger construction domain. Such maps vanish at the configuration with all links equal to I I I . By the product inequality for angles, or by integration of the first derivative along the straight segment, their size is at most C c r , J C_{\mathrm{cr},J} C cr , J times the bound on their inputs. This constant depends only on the fixed charts and on M M M , J J J , not on any width or number of steps.
We fix C c o n v , J C_{\mathrm{conv},J} C conv , J by (109) below, after A 8 A_8 A 8 and before either radius; it includes C a d , J C_{\mathrm{ad},J} C ad , J and C c r , J C_{\mathrm{cr},J} C cr , J . Let ρ ∂ , c l \rho_{\partial,\mathrm{cl}} ρ ∂ , cl and ρ ∂ , q \rho_{\partial,q} ρ ∂ , q be the boundary values of the two spatially varying radii, and define
b min = 1 4 min { ρ c l , ρ , ρ ∂ , c l , ρ ∂ , q } . b_{\min}=\frac{1}{4}\min\{\rho_{\mathrm{cl}},\rho,\rho_{\partial,\mathrm{cl}},\rho_{\partial,q}\}. b m i n = 4 1 min { ρ cl , ρ , ρ ∂ , cl , ρ ∂ , q } .
We take β r e c < min { π , ρ / 4 , ρ c l / 4 } \beta_{\mathrm{rec}} < \min\{\pi,\rho/4,\rho_{\mathrm{cl}}/4\} β rec < min { π , ρ /4 , ρ cl /4 } satisfying the two conditions C b c , ∙ E ∙ β r e c / ρ ∙ ≤ σ ∗ / 4 C_{\mathrm{bc},\bullet}E\bullet\beta_{\mathrm{rec}}/\rho_{\bullet} \le\sigma_*/4 C bc , ∙ E ∙ β rec / ρ ∙ ≤ σ ∗ /4 , and then require
C c o n v , J r < min { β r e c , b min } . (108) C_{\mathrm{conv},J}r < \min\{\beta_{\mathrm{rec}},b_{\min}\}. \tag*{(108)} C conv , J r < min { β rec , b m i n } . (108)
These are the conditions on frozen coordinates in the choice of parameters. All quantities on the right side are fixed positive parameters, so this upper bound on r r r is compatible with its other upper bounds.
Step 3: the tests and the goodness of retained plaquettes. We first consider a stopped measure. At the input of the first RG step, a gauge transformation into B r B_r B r gives all links smaller than r r r . By Corollary 21.2, which precedes this construction, the initial frozen coordinates at the input are bounded by C a d , J r C_{\mathrm{ad},J}r C ad , J r in a common representative. Since C c o n v , J ≥ C a d , J C_{\mathrm{conv},J} \ge C_{\mathrm{ad},J} C conv , J ≥ C ad , J , their analytic coordinates satisfy (21.6) as well. At later inputs, by the induction in Step 4, the links in the interior of the block region are small, with all earlier frozen coordinates fixed. Before any gauge transformation, we compute the iterates at every level of the current RG step and perform the ordered tests of Definition 21.4. The tests always use the values of the iterates themselves, also when a path product that is not used has an exceptional value of the Borel extension.
After the tests, the retained cells are those of Definition 21.4(3). A dependence set avoiding the final excluded set also avoided the smaller excluded set at the time of its test. If that plaquette had value at least d ∗ d_* d ∗ , its failure would have put the dependence set into the excluded set. This contradiction shows that every tested plaquette that is not excluded is good, that is, q f < d ∗ q_f < d_* q f < d ∗ . Adding excluded sets after the tests can only reduce this set of plaquettes.
Step 4: small representatives. We prove part (2) by induction over the boundaries. Let P P P be the retained closed cell complex. Every link read for an output link of P P P comes from links of the current interior, before the new freezing at the input. By Lemma 21.1, its angle in the incoming representative is at most C B r C_Br C B r , where C B C_B C B includes the finite tree path lengths. By Step 3 above and Step 8 of the proof of Theorem 21.3, all plaquettes of P P P have angle at most δ = π 2 d ∗ \delta= \frac{\pi}{2\sqrt{d_*}} δ = 2 d ∗ π . Let Γ \Gamma Γ consist of one layer of retained cells next to the excluded set, enlarged to contain all prescribed boundary vertices. We prescribe the identity gauge on Γ \Gamma Γ in this representative; its link bound is η ≤ C B r \eta\le C_Br η ≤ C B r . After the initial reduction K B 2 r / A 8 ≤ c K_B^2r/A_8 \le c K B 2 r / A 8 ≤ c , parts (2) and (3) of Theorem 21.3 apply. The resulting gauge transformation is exactly I I I at every prescribed lattice vertex, has link bound C c o l r C_{\mathrm{col}}r C col r on P P P , and has the bound < r / 2 < r/2 < r /2 in cells without prescribed vertices. It extends by I I I outside P P P , because every boundary vertex of P P P is prescribed. We lift it as a constant through the current RG step. By the endpoint protection rule, the lift is I I I at all earlier marks and at the variables already frozen at the input in this RG step. Hence every earlier frozen coordinate, its endpoints, and every newly frozen input variable keep their values.
We freeze the full output boundary band, including Γ \Gamma Γ , its incident cells, all labels of reduced analytic radius, the removed band, and the margins of the dependence neighborhoods. Every ordinary link in Γ \Gamma Γ is recorded with its pre-extension value, since the extension gauge is I I I at both endpoints. Its bound C B r C_B r C B r is therefore native to its creation lattice, not an estimate obtained by averaging an earlier frozen value. Every remaining interior link lies in a cell with no vertex in Γ \Gamma Γ , because every retained cell meeting Γ \Gamma Γ belongs to the frozen band. Hence its angle is smaller than r / 2 r/2 r /2 . The newly frozen links have angle at most max ( C B , C c o l ) r \max(C_B,C_{\mathrm{col}})r max ( C B , C col ) r . A based transport multiplies such a link by a hierarchical transport with at most L J L_J L J factors, and its angle is at most the sum of their angles. As announced before the radii were chosen, we may therefore fix
C c o n v , J > 2 C c r , J ( 1 + L J ) max { C B , C c o l , C ′ / A 8 , C a d , J } , (109) C_{\mathrm{conv},J} > 2C_{\mathrm{cr},J}(1+L_J)\max\{C_B,C_{\mathrm{col}},C'/A_8,C_{\mathrm{ad},J}\}, \tag*{(109)} C conv , J > 2 C cr , J ( 1 + L J ) max { C B , C col , C ′ / A 8 , C ad , J } , (109)
with any additional fixed factors of dependence neighborhoods absorbed once. Every analytic coordinate of the union of the newly frozen boundary band, boundary layer, and based transports then has size at most C c o n v , J r C_{\mathrm{conv},J}r C conv , J r . By (21.5) and (21.6), we have simultaneously
∣ c j ♯ ∣ < β r e c , ∣ c j ♯ ∣ ≤ ϕ ∙ ( j ) / 4 ( ∙ ∈ { c l , q } ) . (110) \lvert c_j^\sharp\rvert< \beta_{\mathrm{rec}}, \qquad\lvert c_j^\sharp\rvert\le\phi_{\bullet}(j)/4 \quad(\bullet\in\{\mathrm{cl},\mathrm{q}\}). \tag*{(110)} ∣ c j ♯ ∣ < β rec , ∣ c j ♯ ∣ ≤ ϕ ∙ ( j ) /4 ( ∙ ∈ { cl , q }) . (110)
This includes every label of reduced radius, not only the boundary links of a layer. The constant is computed from A 8 A_8 A 8 , C B C_B C B , and the universal relative extension before r r r and β r e c \beta_{\mathrm{rec}} β rec are chosen.
For the next RG step, we retain only the blocks that lie, together with their full dependence neighborhoods, in the remaining interior. The discarded band has thickness at most C J C_J C J in that input lattice and consists of links that were small before discarding. The fixed factor coming from its dependence neighborhood was included in C c o n v , J C_{\mathrm{conv},J} C conv , J before the radii were fixed, so the band may be frozen with the same bound. The previous excluded region was enlarged by W W W in the current output lattice. After the next RG step, this is W / B = 8 R c e l l W/B=8R_{\mathrm{cell}} W / B = 8 R cell units, which suffices for the required layer of cells and the fixed incidences; the full dependence neighborhood is included in addition. None of these operations requires a later averaged image of a frozen coordinate to be small, since the coordinate stays on its own lattice. This completes the induction.
Step 5: incoming and current boundary distances. For the next RG step, let D i n D_{\mathrm{in}} D in be its list of remaining analytic input labels after the whole freezing, and let Ω ′ \Omega' Ω ′ be the set of its current retained blocks. Each incoming label is kept at its location on the lattice where it was created, with its original volume factor. A current block is retained only when its full input dependence neighborhood is contained in D i n D_{\mathrm{in}} D in . An omitted incoming label lies outside the block region, apart from the fixed displacement between the location of a chart label and its path product. A newly frozen boundary label, or a label of a link across the boundary of the block region, is omitted from D i n D_{\mathrm{in}} D in and satisfies the same statement. By the endpoint protection rule, an earlier frozen endpoint or basepoint has its current centered ancestor in an excluded block. By the balanced-digit identity, the difference between its rescaled position and that ancestor is less than one current unit, uniformly in the number of steps. Including the finite dependence neighborhoods of the charts, we may take L b u f , J = 3 D J + 6 L_{\mathrm{buf},J}=3D_J+6 L buf , J = 3 D J + 6 , so that every omitted input label x x x satisfies
dist ( x / B s , ( Ω ′ ) c ) ≤ L b u f , J , B s = M s . (111) \operatorname{dist}\left(x/B_s,(\Omega')^c\right)\le L_{\mathrm{buf},J}, \qquad B_s=M^s. \tag*{(111)} dist ( x / B s , ( Ω ′ ) c ) ≤ L buf , J , B s = M s . (111)
In the induction s = J s=J s = J . The same inclusion holds for a single group of s ≤ J s\le J s ≤ J averaging steps when its own domain conditions are imposed. Enlarging the excluded regions can only decrease the distance to ( Ω ′ ) c (\Omega')^c ( Ω ′ ) c , so it preserves the inclusion.
If the incoming fixed complement is nonempty, let i i i be an incoming retained label and x x x a nearest omitted input label. For a current retained label e e e we use the same distance as in the two boundary-weighted norms; the factor 3 in L b u f , J L_{\mathrm{buf},J} L buf , J covers the comparison with the maximum-distance convention for dependence neighborhoods. By the triangle inequality and (21.9), we have
d c u r r e n t ( e ) ≤ dist ( e , i / B s ) + B s − 1 d D i n ( i ) + L b u f , J + 1 , d_{\mathrm{current}}(e)\le\operatorname{dist}(e,i/B_s)+B_s^{-1}d_{D_{\mathrm{in}}}(i)+L_{\mathrm{buf},J}+1, d current ( e ) ≤ dist ( e , i / B s ) + B s − 1 d D in ( i ) + L buf , J + 1 ,
where the last term 1 accounts for the choice of a cell rather than a vertex as location. Rearranging, we obtain
d D i n ( i ) ≥ B s d c u r r e n t ( e ) − B s dist ( e , i / B s ) − C b u f , J , C b u f , J = B s ( L b u f , J + 1 ) . (112) d_{D_{\mathrm{in}}}(i)\ge B_s d_{\mathrm{current}}(e)-B_s\operatorname{dist}(e,i/B_s)-C_{\mathrm{buf},J}, \qquad C_{\mathrm{buf},J}=B_s(L_{\mathrm{buf},J}+1). \tag*{(112)} d D in ( i ) ≥ B s d current ( e ) − B s dist ( e , i / B s ) − C buf , J , C buf , J = B s ( L buf , J + 1 ) . (112)
Rescaling coordinates in this calculation does not rescale the counting measures of the labels or discard their volume factors in a column norm; these are the same as in the incoming analytic norm. If there is no incoming fixed complement, the boundary-weighted part of the input is zero, and no nearest omitted label is needed. A new excluded region satisfies the inclusion just proved once the initial split has been made. The buffer in (21.10) depends only on M M M , J J J , and the fixed charts, not on W W W , s e r \mathop{\mathrm{ser}} ser , the number of steps, or the number of excluded components.
Step 6: locality and counts. We verify locality and the counts without counting connected components. We label each initially added neighborhood by its failing plaquette; when neighborhoods overlap, a label in their intersection is assigned to the first incident failing plaquette in the fixed order. The enlargement of a union by a fixed neighborhood is contained in the union of the corresponding enlargements of its members. A cell meeting such an enlargement lies within a further 4 R c e l l 4R_{\mathrm{cell}} 4 R cell of it, and a full dependence neighborhood costs at most D J + 1 D_J+1 D J + 1 more. Projection of a mark to its current ancestor contracts distances by the blocking factor, with a rounding error of at most one current unit for the whole projection, not one for each earlier step. This follows from the balanced-digit identity in the proof of Theorem 21.3, since the whole preimage has radius less than half a block side.
If R t ( f ) R_t(f) R t ( f ) bounds all current neighborhoods and projected marks assigned to f f f , the next RG step gives, with a fixed a J < ∞ a_J < \infty a J < ∞ ,
R t + 1 ( f ) ≤ B − 1 R t ( f ) + a J ( W + R c e l l + 1 ) . (113) R_{t+1}(f) \le B^{-1}R_t(f) + a_J(W + R_{\mathrm{cell}} + 1). \tag*{(113)} R t + 1 ( f ) ≤ B − 1 R t ( f ) + a J ( W + R cell + 1 ) . (113)
The neighborhood created when f f f fails satisfies the same bound in the output units of its RG step. We take C J ≥ a J / ( 1 − B − 1 ) C_J \ge a_J/(1-B^{-1}) C J ≥ a J / ( 1 − B − 1 ) and at least the corresponding constant for the step in which f f f fails. Iterating (113) , we obtain R t ( f ) ≤ C J ( W + R c e l l + 1 ) R_t(f) \le C_J(W+R_{\mathrm{cell}}+1) R t ( f ) ≤ C J ( W + R cell + 1 ) at every later boundary between RG steps. A remaining group of fewer than J J J averaging steps at the end is read as a fixed finite dependence neighborhood at the last boundary; it creates no additional integrated step and no update of the normalization constant. All endpoint marks added when boundary layers are frozen lie in the same neighborhoods, so the protected neighborhoods of the endpoint protection rule obey (113) too, even if the frozen coordinates contain marks on many lattices.
A ball of radius a a a in the maximum distance contains at most ( 2 a + 3 ) 3 (2a+3)^3 ( 2 a + 3 ) 3 vertices. The link direction, the color component, the intermediate scale, and the type of frozen coordinate multiply this by a fixed C J C_J C J . In the RG step in which f f f fails we convert the radius into input units, which costs a factor B B B ; at later RG steps we use the bound in current units just proved. We obtain the counts
C J ( B W + R c e l l + 1 ) 3 , C J ( W + R c e l l + 1 ) 3 C_J(BW+R_{\mathrm{cell}}+1)^3,\qquad C_J(W+R_{\mathrm{cell}}+1)^3 C J ( B W + R cell + 1 ) 3 , C J ( W + R cell + 1 ) 3
of scalar labels. The lists of missing labels, boundary-band labels, and frozen labels used in Theorem 20.13 are subsets of these lists of bounded density; deterministic images of earlier frozen coordinates are not new independent labels. A final plaquette whose full dependence set meets an excluded region lies in a further fixed dependence neighborhood of one of the assigned sets. Counting the three plaquette directions there, we obtain the bound F J F_J F J . The finitely many remaining numbers of averaging steps after the last RG step only enlarge the dependence neighborhood. On a periodic quotient these counts can only decrease. Since we use unions rather than hulls of components, all the bounds are subadditive, including for excluded regions that touch or wrap around the torus.
Step 7: Borel measurability. The outcome of the ordered tests is determined by its set of failing plaquettes: replaying the ordered list and adding the prescribed excluded sets at exactly those failures reproduces it. Hence there is at most one decomposition for each set of failing plaquettes that the procedure can produce. Each event A X A_X A X is Borel, since it is a finite Boolean combination of tests of Borel iterates and of membership in deterministic dependence lists. These events are disjoint and exhaust the support. The small representatives were used only after these events had been specified.
Step 8: the first step under the Wilson measure. For the Wilson measure, we perform the same tests, including the plaquettes of the input configuration, without assuming a small input representative. Let F F F be the final excluded dependence set. On the fixed input cell grid of scale S S S , aligned with the blocks, we form the initial good-cell complex P i n P_{\mathrm{in}} P in , consisting of all closed cells whose full dependence sets avoid F F F . The blocks of the block region are chosen only after moving inward by one full layer of cells and by the full dependence neighborhood of the constraint, tree, basing, and plaquette path products. This choice implies that every input plaquette whose Wilson term depends on an integration variable, and every path product of the block region, is contained in P i n P_{\mathrm{in}} P in . For example, a point of a retained block together with its dependence neighborhood still lies in a retained cell or in its incident layer, and that layer was retained before moving inward. The near links are the links outside the block region read by these path products and plaquettes, as in (18.32), and their values are the near values. All remaining input links are deep links; their values are the deep frozen group variables, that is, the frozen group variables in the fixed complement that are not near values.
An input plaquette of P i n P_{\mathrm{in}} P in , including a boundary plaquette, avoids the final set F F F ; by the argument of Step 3, it is good. The absolute part of Theorem 21.3 applies with n i n p u t ≥ 8 S n_{\mathrm{input}} \ge8S n input ≥ 8 S and δ S 2 ≤ 1 \delta S^2 \le1 δ S 2 ≤ 1 , and we obtain the bound C ( δ S + S − 1 ) ≤ C ′ / S C(\delta S+S^{-1}) \le C'/S C ( δ S + S − 1 ) ≤ C ′ / S . Fixing the mesh factors when A 8 A_8 A 8 is chosen makes this smaller than the required fixed multiple of r r r . At this first input there are no earlier fixed vertices, and we extend the resulting gauge transformation arbitrarily to the other vertices. All transformed deep links are kept as arbitrary elements of S U ( 2 ) \mathrm{SU}(2) SU ( 2 ) in the frozen data. The finite path-product estimate and the relative construction at the output of Step 4 give the claimed bounds for the new boundary layers and based transports. The near values are small in this representative, not in the original configuration. Every additional excluded set lies within a neighborhood of radius C J S C_J S C J S of an excluded neighborhood, and we assign it to the same failing plaquette. Since S ≤ C J R c e l l S \le C_J R_{\mathrm{cell}} S ≤ C J R cell , the radius and label counts keep their displayed forms after one fixed enlargement. After this first step, the induction of Step 4 applies.
All analytically frozen adjacent variables of this first step lie in the complex P i n P_{\mathrm{in}} P in , and hence obey the same conversion and (110) . A dependence set frozen because its averaged plaquette fails can contain arbitrary elements of S U ( 2 ) \mathrm{SU}(2) SU ( 2 ) . Such deep frozen group variables are fixed parameters, not variables of the analytic freezing, and their plaquettes involving only frozen variables remain in the external Wilson factor. For a stopped measure, the corresponding initial variables lie in its gauge-invariant small-field domain, by Corollary 21.2 .
Step 9: stage measures. By Lemma 21.6, every gauge transformation used in this construction fixes the earlier frozen coordinates and their logarithms; this is the form in which the representatives of part (2) are used later. All stage maps are Borel, since they are compositions of restrictions, the block averaging map, and finite transport products; their pushforwards are the measures in part (5). Each frozen coordinate is recorded once, as an independent link or based transport, and carries all its dependence through the density; no statement about independence under the probability measure is needed. Since the events A X A_{\mathcal X} A X partition the support of ν \nu ν , multiplying their pointwise sum by 1 S 1_S 1 S gives the identity in part (5), and no restriction is added to the Wilson measure. All geometric arguments were finite local constructions; no step involves L L L , γ \gamma γ , d 0 d_0 d 0 , the depth, the number of steps since freezing, or the number of components. A final group of fewer than J J J averaging steps is stopped before its minimal-side condition fails, as stated.
The next lemma verifies, on every large-field decomposition, the hypotheses of Theorems 19.9 and 20.14 that concern the incoming radius, the boundary distances, the conditional forces, and the cutoff domains. We use the notation of Lemma 19.3. For each initial segment I I I of the integrated intermediate scales, y = z I c y=z_{I^c} y = z I c denotes the remaining free coordinates, T I \mathcal T_I T I their domain, a I ( c , y ; ξ ) a_I(c,y;\xi) a I ( c , y ; ξ ) the unperturbed conditional minimizer, and A I A_I A I the unperturbed conditional Hessian at a I a_I a I . Further, r I r_I r I is the radius of the conditional coordinate neighborhood, r q r_q r q the bound on the scaled fluctuations g i w i g_i w_i g i w i , ϵ t u b e \epsilon_{\mathrm{tube}} ϵ tube the fixed margin, and C s e q , J C_{\mathrm{seq},J} C seq , J the constant of the sequential construction. As in Theorem 19.9, Φ 0 \Phi_0 Φ 0 is the reconstruction in the target-dependent gauge and P ‾ \overline P P is the classical perturbation adjusted by the boundary-completion difference. Here Φ 0 ( c , u ; ξ ) \Phi_0(c,u;\xi) Φ 0 ( c , u ; ξ ) has the joint normalized argument u u u used in that theorem: if z 0 z_0 z 0 is the unperturbed joint stationary point, K 0 K_0 K 0 the unperturbed Hessian there, and S 0 = K 0 − 1 / 2 S_0=K_0^{-1/2} S 0 = K 0 − 1/2 , then z = z 0 + S 0 u z=z_0+S_0u z = z 0 + S 0 u . In the lemma below, the perturbing function in the original free coordinates is
V ( c , z ; b ) = P ‾ ( Φ 0 ( c , S 0 − 1 ( z − z 0 ) ; ξ ) ; b ) . (114) V(c,z;b)=\overline P\bigl(\Phi_0(c,S_0^{-1}(z-z_0);\xi);b\bigr). \tag*{(114)} V ( c , z ; b ) = P ( Φ 0 ( c , S 0 − 1 ( z − z 0 ) ; ξ ) ; b ) . (114)
Thus V I V_I V I and V I I V_{II} V I I below are derivatives in the original free coordinates z z z , at fixed c c c and real b b b , and the constants C P C_P C P and C P ′ C'_P C P ′ include the fixed coefficient bounds of this affine change of coordinates. For a function f f f of the labels D D D and frozen values c ♯ c^\sharp c ♯ on F ⊂ D F\subset D F ⊂ D , we write T c ♯ f T_{c^\sharp}f T c ♯ f for the function of the labels D ∖ F D\setminus F D ∖ F obtained by inserting these values, and X D ϕ \mathcal X_D^\phi X D ϕ for the coefficient norm with boundary decay of Definition 18.1 with spatially varying radius ϕ \phi ϕ .
Lemma 21.7. Assume the conditions on frozen coordinates and the construction conditions of Theorem 21.5.
Freezing at the boundary of a complete RG step removes every label of reduced analytic radius and leaves the incoming perturbation on the remaining labels with a uniform analytic input radius, separately for the classical and the quantum radius. Its incoming and current boundary distances satisfy (21.10).
Fix real b = ( ξ , b r e c ) b=(\xi,b_{\mathrm{rec}}) b = ( ξ , b rec ) , and suppose that the adjusted classical perturbation P ‾ \overline P P has uniform norm at most 2 σ ∗ 2\sigma_* 2 σ ∗ . Then the preconditioned conditional force δ V \delta_V δ V , taken over every initial segment I I I and every y ∈ T I y\in\mathcal T_I y ∈ T I and defined by the first expression below with V V V as in (21.12), satisfies
δ V = max I sup y ∈ T I ∥ A I − 1 V I ( c , a I ( c , y ; ξ ) , y ; b ) ∥ c o e f f ≤ I 0 C P σ ∗ ρ c l , (115) \delta_V=\max_I\sup_{y\in\mathcal T_I}\left\|A_I^{-1}V_I\bigl(c,a_I(c,y;\xi),y;b\bigr)\right\|_{\mathrm{coeff}}\leq I_0 C_P\sigma_*\rho_{\mathrm{cl}}, \tag*{(115)} δ V = I max y ∈ T I sup A I − 1 V I ( c , a I ( c , y ; ξ ) , y ; b ) coeff ≤ I 0 C P σ ∗ ρ cl , (115)
where I 0 I_0 I 0 and C P C_P C P are fixed bounds on the larger domains, independent of the later widths.
The enclosing product domain and every conditional coordinate neighborhood lie in the domains of Lemma 19.3 when the construction conditions include
δ V < 1 4 min I r I , R e n < r q , C s e q , J ( 2 I 0 C P σ ∗ ρ c l + r q ) < 1 2 ε t u b e , (116) \delta_V < \frac{1}{4}\min_I r_I,\qquad R_{\mathrm{en}} < r_q,\qquad C_{\mathrm{seq},J}(2I_0 C_P \sigma_* \rho_{\mathrm{cl}} + r_q) < \frac{1}{2}\varepsilon_{\mathrm{tube}}, \tag*{(116)} δ V < 4 1 I min r I , R en < r q , C seq , J ( 2 I 0 C P σ ∗ ρ cl + r q ) < 2 1 ε tube , (116)
together with the fixed Hessian conditions and the two reconstruction-image conditions (119) .
The fluctuation cutoff domain with common radius p ( t ) = A w log ( 1 / t ) p(t)=A_{\mathrm{w}}\log(1/t) p ( t ) = A w log ( 1/ t ) is one of the cutoff domains allowed in Theorem 20.14 . Its containment and expansion conditions hold below a single coupling ceiling, chosen after these fixed positive radii.
None of these statements requires a derivative in ζ \zeta ζ or in b r e c b_{\mathrm{rec}} b rec .
Proof. Steps 1 and 2 prove part (1) and the bound on the new boundary-completion differences, Step 3 proves part (2), Steps 4 and 5 verify the image and convexity conditions behind part (3), and Step 6 proves part (4).
Step 1: uniform radius on the remaining labels. The uniform radii are used in the freezing part of Theorem 20.14 . We write D D D for the old output list, I I I for the remaining labels, and F = D ∖ I F=D\setminus I F = D ∖ I . All labels of reduced radius belong to F F F , and d D ( i ) ≥ s e r d_D(i)\ge s_{\mathrm{er}} d D ( i ) ≥ s er for i ∈ I i\in I i ∈ I . If there is no old boundary, the boundary-weighted part of the incoming norm is zero, and the estimate below is immediate. Otherwise, since the full neighborhood of width s e r − 1 s_{\mathrm{er}}-1 s er − 1 of the old boundary was removed, a shortest route from i i i to the old boundary meets F F F at least s e r − 1 s_{\mathrm{er}}-1 s er − 1 steps earlier, and additional excluded regions can only shorten the distance to F F F . Hence
d D ( i ) ≥ d I ( i ) + s e r − 1. d_D(i)\ge d_I(i)+s_{\mathrm{er}}-1. d D ( i ) ≥ d I ( i ) + s er − 1.
For a monomial of degree k ≥ 1 k\ge1 k ≥ 1 in the remaining labels and degree j ≥ 0 j\ge0 j ≥ 0 in the frozen labels, inserting the frozen values multiplies its absolute coefficient contribution by at most η j \eta^j η j with η ≤ 1 / 4 \eta\le1/4 η ≤ 1/4 , by (110) . The new fraction ν i / k \nu_i/k ν i / k contributes at most a factor ( k + j ) / k ≤ 1 + j (k+j)/k\le1+j ( k + j ) / k ≤ 1 + j , and the remaining connecting network is no longer than the original one. Since every remaining label had the full radius ρ ∘ \rho_\circ ρ ∘ , the resulting series is estimated at the uniform radius ρ ∘ \rho_\circ ρ ∘ on I I I , not at a reduced boundary radius. Summing ∑ j ≥ 0 ( 1 + j ) η j = ( 1 − η ) − 2 < 2 \sum_{j\ge0}(1+j)\eta^j=(1-\eta)^{-2}<2 ∑ j ≥ 0 ( 1 + j ) η j = ( 1 − η ) − 2 < 2 and using the distance inequality, we obtain, including degree one,
χ I ρ ∘ ( T b f ) ≤ 2 e − α ∘ ( s e r − 1 ) χ D ρ ∘ ( f ) . (117) \chi_I^{\rho_\circ}(T_{\mathrm{b}}f)\le2e^{-\alpha_\circ(s_{\mathrm{er}}-1)}\chi_D^{\rho_\circ}(f). \tag*{(117)} χ I ρ ∘ ( T b f ) ≤ 2 e − α ∘ ( s er − 1 ) χ D ρ ∘ ( f ) . (117)
Replacing the radius and rates by the classical radius and rates in this monomial calculation, we obtain the classical statement, with the same factor ( 1 − η ) − 2 (1-\eta)^{-2} ( 1 − η ) − 2 .
Step 2: new boundary-completion differences. The bound ∣ c j ∙ ∣ ≤ ϕ ∙ ( j ) / 4 |c_j^\bullet| \le\phi_\bullet(j)/4 ∣ c j ∙ ∣ ≤ ϕ ∙ ( j ) /4 controls the inherited boundary term. Freezing also inserts the frozen values into the torus functions, and this produces a new boundary-completion difference. For it we use the bound ∣ c ˉ j ∙ ∣ ≤ β r e c |\bar{c}_j^\bullet| \le\beta_{\mathrm{rec}} ∣ c ˉ j ∙ ∣ ≤ β rec from (110) . Every contributing monomial has frozen degree j ≥ 1 j \ge1 j ≥ 1 and degree k ≥ 1 k \ge1 k ≥ 1 in the remaining labels. For a fixed remaining index i i i , the graph connecting network joins i i i to a frozen label, so d I ( i ) ≤ t ( ν ) d_I(i) \le\mathrm{t}(\nu) d I ( i ) ≤ t ( ν ) and t ( ν I ) ≤ t ( ν ) \mathrm{t}(\nu_I) \le\mathrm{t}(\nu) t ( ν I ) ≤ t ( ν ) . Let t 2 \mathrm{t}_2 t 2 denote the Euclidean network length in the torus coefficient norm, and let a ˉ ∙ \bar{a}_\bullet a ˉ ∙ be its Euclidean rate. The rate choices of Theorem 20.9 give a ˉ ∙ ≥ 3 ( a ∙ + α ∙ ) \bar{a}_\bullet\ge\sqrt{3}(a_\bullet+\alpha_\bullet) a ˉ ∙ ≥ 3 ( a ∙ + α ∙ ) . Therefore
a ∙ t ( ν I ) + α ∙ d I ( i ) ≤ ( a ∙ + α ∙ ) t ( ν ) ≤ 3 ( a ∙ + α ∙ ) t 2 ( ν ) ≤ a ˉ ∙ t 2 ( ν ) . a_\bullet\mathrm{t}(\nu_I)+\alpha_\bullet d_I(i)\le(a_\bullet+\alpha_\bullet)\mathrm{t}(\nu)\le\sqrt{3}(a_\bullet+\alpha_\bullet)\mathrm{t}_2(\nu)\le\bar{a}_\bullet\mathrm{t}_2(\nu). a ∙ t ( ν I ) + α ∙ d I ( i ) ≤ ( a ∙ + α ∙ ) t ( ν ) ≤ 3 ( a ∙ + α ∙ ) t 2 ( ν ) ≤ a ˉ ∙ t 2 ( ν ) .
With η ∙ = β r e c / ρ ∙ ≤ 1 / 4 \eta_\bullet=\beta_{\mathrm{rec}}/\rho_\bullet\le1/4 η ∙ = β rec / ρ ∙ ≤ 1/4 , the ratio of the new and old fractions ν i / k \nu_i/k ν i / k is at most 1 + j 1+j 1 + j . For each fixed original degree we split its absolute coefficient sum according to the frozen degree j j j , which is uniquely determined. Taking the supremum over the remaining index before the sum over degrees, we lose at most the factor
∑ j ≥ 1 ( 1 + j ) η ∙ j = η ∙ ( 2 − η ∙ ) ( 1 − η ∙ ) 2 ≤ 4 η ∙ . \sum_{j\ge1}(1+j)\eta_\bullet^j=\frac{\eta_\bullet(2-\eta_\bullet)}{(1-\eta_\bullet)^2}\le4\eta_\bullet. j ≥ 1 ∑ ( 1 + j ) η ∙ j = ( 1 − η ∙ ) 2 η ∙ ( 2 − η ∙ ) ≤ 4 η ∙ .
The normalization is ρ ∙ − 2 \rho_\bullet^{-2} ρ ∙ − 2 in both norms, and the powers of ρ ∙ \rho_\bullet ρ ∙ cancel against the factors of the frozen labels. This includes the new degree-one terms, and gives the bound 8 E ∙ β r e c / ρ ∙ 8E_\bullet\beta_{\mathrm{rec}}/\rho_\bullet 8 E ∙ β rec / ρ ∙ for one simultaneous freezing of the union of the frozen labels. The fixed constant C b c , ∙ C_{\mathrm{bc},\bullet} C bc , ∙ includes every required conversion and any finite sum of such terms. The total contribution therefore satisfies
z t ∙ ≤ C b c , ∙ E ∙ β r e c / ρ ∙ ≤ σ ∗ / 4. (118) z_t^\bullet\le C_{\mathrm{bc},\bullet}E_\bullet\beta_{\mathrm{rec}}/\rho_\bullet\le\sigma_\ast/4. \tag*{(118)} z t ∙ ≤ C bc , ∙ E ∙ β rec / ρ ∙ ≤ σ ∗ /4. (118)
The boundary band, the new boundary layers, the based transports, and the initialization are all included in this single bound. For a new excluded region in the interior, we do not claim that the removed band alone makes this difference small. Equations (117) and (118) imply the uniform incoming norms in the recursion of Theorem 20.14 . The frozen coordinates ζ \zeta ζ are parameters of those norms and need not have the full incoming analytic radius.
Step 3: conditional forces. We use the absolute coefficient bounds for the derivatives of P ˉ \bar{P} P ˉ in Theorem 19.9 , on its uniform input image, which lies strictly inside the input ball. By the fixed construction hypotheses, every conditional coordinate neighborhood has the same image, together with its connecting segments. Composing with the row-sum bounds of the first derivative of Φ 0 \Phi_0 Φ 0 and of the affine change of coordinates in (114) , we find ∥ V I ∥ c o e f f ≤ C P σ ∗ ρ c l \lVert V_I\rVert_{\mathrm{coeff}}\le C_P\sigma_\ast\rho_{\mathrm{cl}} ∥ V I ∥ coeff ≤ C P σ ∗ ρ cl , uniformly in y ∈ T I y\in\mathcal{T}_I y ∈ T I . Multiplying by the conditional inverse bound I 0 I_0 I 0 , we obtain (21.13). The second derivative also contains the term P ˉ ′ Φ 0 , I I \bar{P}'\Phi_{0,II} P ˉ ′ Φ 0 , I I , where Φ 0 , I I \Phi_{0,II} Φ 0 , I I is the second z I z_I z I -derivative of z ↦ Φ 0 ( c , S 0 − 1 ( z − z 0 ) ; ζ ) z\mapsto\Phi_0(c,S_0^{-1}(z-z_0);\zeta) z ↦ Φ 0 ( c , S 0 − 1 ( z − z 0 ) ; ζ ) ; keeping it, we obtain ∥ V I ∥ c o e f f ≤ C p ′ σ ∗ \lVert V_I\rVert_{\mathrm{coeff}} \le C'_p\sigma_\ast ∥ V I ∥ coeff ≤ C p ′ σ ∗ . The unperturbed neighborhoods are fixed first, so that their own Hessian variation uses at most half of the preconditioned margin. Reducing the fixed ceiling for σ ∗ \sigma_\ast σ ∗ makes I 0 C p ′ σ ∗ I_0C'_p\sigma_\ast I 0 C p ′ σ ∗ fit the remaining margin and preserves the real half-gap and the complex inverse margin. At an unperturbed conditional minimizer, we have F 0 , I = 0 F_{0,I}=0 F 0 , I = 0 , so the condition on inverse times force is precisely λ ∥ A I − 1 V I ∥ ≤ δ V < r I / 4 \lambda\lVert A_I^{-1}V_I\rVert\le\delta_V < r_I/4 λ ∥ A I − 1 V I ∥ ≤ δ V < r I /4 for 0 ≤ λ ≤ 1 0\le\lambda\le1 0 ≤ λ ≤ 1 . This verifies the hypothesis of the fixed-parameter contraction at every y y y , and its solution moves by at most 2 λ δ V 2\lambda\delta_V 2 λ δ V . The uniform mixed coefficient bounds of the same theorem control its dependence on the higher parameters. The sequential conclusion of Lemma 19.3 then applies under (116) to all intermediate scales. In particular, a product with ∣ g i w i , e ∣ ≤ R e n < r q \lvert g_iw_{i,e}\rvert\le R_{\mathrm{en}}<r_q ∣ g i w i , e ∣ ≤ R en < r q stays inside every parameter and solution neighborhood with margin. This is a check of the conditional force at moving parameters, not only at the joint minimizer.
Step 4: the image conditions. Both inherited image conditions are checked with the bound s q = R e n s_q=R_{\mathrm{en}} s q = R en for the scaled fluctuations. For each radius, the four zero-fluctuation terms of Lemma 19.7 satisfy
B − 1 Z ∙ ( K ∗ ρ ∙ ) ≤ ρ ∙ / 16 , C Δ , ∙ C a , ∙ R ∂ , ∙ ≤ ρ ∙ / 16 , B^{-1}\mathcal{Z}_{\bullet}(K_\ast\rho_{\bullet})\le\rho_{\bullet}/16,\qquad C_{\Delta,\bullet}C_{a,\bullet}R_{\partial,\bullet}\le\rho_{\bullet}/16, B − 1 Z ∙ ( K ∗ ρ ∙ ) ≤ ρ ∙ /16 , C Δ , ∙ C a , ∙ R ∂ , ∙ ≤ ρ ∙ /16 ,
C Δ , ∙ C a , ∙ e − a W w b d / 2 K ∗ ρ ∙ ≤ ρ ∙ / 16 , C F , ∙ β r e c ≤ ρ ∙ / 16. C_{\Delta,\bullet}C_{a,\bullet}e^{-aWw_{\mathrm{bd}}/2}K_\ast\rho_{\bullet}\le\rho_{\bullet}/16,\qquad C_{F,\bullet}\beta_{\mathrm{rec}}\le\rho_{\bullet}/16. C Δ , ∙ C a , ∙ e − aW w bd /2 K ∗ ρ ∙ ≤ ρ ∙ /16 , C F , ∙ β rec ≤ ρ ∙ /16.
Their sum is A 0 , ∙ A_{0,\bullet} A 0 , ∙ . The initial good-cell complex at the first step under the Wilson measure and the endpoint protection rule change neither these analytic constants nor the bound on the current data. In the notation of the fixed construction constants of Lemma 19.7 , the image conditions are
A 0 , ∙ ≤ ρ ∙ / 4 , 1 − C T , ∙ ( K ∗ ρ ∙ + R e n ) ≥ 1 2 , C T , ∙ R e n 1 − C T , ∙ ( K ∗ ρ ∙ + R e n ) + C Δ , ∙ R e n ≤ ρ ∙ / 8. (119) A_{0,\bullet}\le\rho_{\bullet}/4,\qquad1-C_{T,\bullet}(K_\ast\rho_{\bullet}+R_{\mathrm{en}})\ge\frac{1}{2},\qquad\frac{C_{T,\bullet}R_{\mathrm{en}}}{1-C_{T,\bullet}(K_\ast\rho_{\bullet}+R_{\mathrm{en}})}+C_{\Delta,\bullet}R_{\mathrm{en}}\le\rho_{\bullet}/8. \tag*{(119)} A 0 , ∙ ≤ ρ ∙ /4 , 1 − C T , ∙ ( K ∗ ρ ∙ + R en ) ≥ 2 1 , 1 − C T , ∙ ( K ∗ ρ ∙ + R en ) C T , ∙ R en + C Δ , ∙ R en ≤ ρ ∙ /8. (119)
Here R ∂ , ∙ = K ∗ ρ ∂ , ∙ R_{\partial,\bullet}=K_\ast\rho_{\partial,\bullet} R ∂ , ∙ = K ∗ ρ ∂ , ∙ . The margin in the larger construction domain of radius R r a w R_{\mathrm{raw}} R raw is, explicitly,
C l i f t , J [ K ∗ max ( ρ , ρ c l ) + r + β r e c + q r a w ] < R r a w / 8 , C_{\mathrm{lift},J}\left[K_\ast\max(\rho,\rho_{\mathrm{cl}})+r+\beta_{\mathrm{rec}}+q_{\mathrm{raw}}\right]<R_{\mathrm{raw}}/8, C lift , J [ K ∗ max ( ρ , ρ cl ) + r + β rec + q raw ] < R raw /8 ,
where q r a w q_{\mathrm{raw}} q raw is the fixed upper bound given by the row sums of the first derivative for the sequential product of scaled fluctuations and for all its initial segments; this bound is fixed before the eventual enclosing radius. Together with this margin, the image conditions give the image 3 ρ ∙ / 8 3\rho_{\bullet}/8 3 ρ ∙ /8 without spatial weights. The fixed interpolation of spatial rates in Theorem 20.14 then gives an absolute weighted image strictly below ρ ∙ / 2 \rho_{\bullet}/2 ρ ∙ /2 , so all evaluations and segments used for the force estimates above lie strictly in the uniform input ball, for both the classical and the quantum radius. These conditions on the parameters are already required by Theorem 21.5 and the analytic theorem, and they involve no test on the support of the measure.
Step 5: convexity on the enclosing product domain. The radius R W R_{\mathrm{W}} R W of the scaled fluctuations is fixed independently of the logarithmic cutoff. We first compute its constants on the fixed larger perturbation domain. Then R W R_{\mathrm{W}} R W is taken below the joint and sequential bounds, the bound of the larger construction domain, and the two inherited image bounds of the analytic theorem, as well as below 1 / ( 4 A ) 1/(4A) 1/ ( 4 A ) and 1 / ( 16 C 0 ) 1/(16C_0) 1/ ( 16 C 0 ) , where A A A and C 0 C_0 C 0 are the constants of Lemma 20.3. The remaining deterministic choices satisfy
r , β r e c , σ ∗ ρ c l ≤ λ ∗ R W 6 C e n , J , R e n = 2 C e n , J ( r + β r e c + σ ∗ ρ c l ) λ ∗ ≤ R W . (120) r,\ \beta_{\mathrm{rec}},\ \sigma_*\rho_{\mathrm{cl}} \le\frac{\lambda_*R_{\mathrm{W}}}{6C_{\mathrm{en},J}}, \qquad R_{\mathrm{en}} = \frac{2C_{\mathrm{en},J}(r+\beta_{\mathrm{rec}}+\sigma_*\rho_{\mathrm{cl}})}{\lambda_*} \le R_{\mathrm{W}}. \tag*{(120)} r , β rec , σ ∗ ρ cl ≤ 6 C en , J λ ∗ R W , R en = λ ∗ 2 C en , J ( r + β rec + σ ∗ ρ cl ) ≤ R W . (120)
Here C e n , J C_{\mathrm{en},J} C en , J is evaluated on the fixed larger products, as in the proof of Theorem 21.9, before these radii are reduced. Then A R e n ≤ 1 / 4 AR_{\mathrm{en}} \le1/4 A R en ≤ 1/4 and C 0 R e n ≤ 1 / 16 C_0R_{\mathrm{en}} \le1/16 C 0 R en ≤ 1/16 . When C 0 t 2 L V ≤ 1 / 16 C_0t^2\mathcal{L}_{\mathrm{V}} \le1/16 C 0 t 2 L V ≤ 1/16 , the potential estimate of Theorem 20.14 yields
∥ D w 2 ( V − ∣ w ∣ 2 2 ) ∥ r o w , c o l , a ′ ≤ C 0 ( R e n + t 2 L V ) ( 1 − A R e n ) 3 ≤ 1 / 8 ( 3 / 4 ) 3 = 8 27 < 1 2 . \left\|D^2_{\mathbf{w}}\left(V-\frac{|\mathbf{w}|^2}{2}\right)\right\|_{\mathrm{row,col},a'} \le\frac{C_0(R_{\mathrm{en}}+t^2\mathcal{L}_{\mathrm{V}})}{(1-AR_{\mathrm{en}})^3} \le\frac{1/8}{(3/4)^3} = \frac{8}{27} < \frac{1}{2}. D w 2 ( V − 2 ∣ w ∣ 2 ) row , col , a ′ ≤ ( 1 − A R en ) 3 C 0 ( R en + t 2 L V ) ≤ ( 3/4 ) 3 1/8 = 27 8 < 2 1 .
This verifies the Hessian margin on the whole enclosing product domain, separately from any expansion on the smaller logarithmic cutoff domain.
Step 6: the logarithmic cutoff domain. Let p ( t ) = A W log ( 1 / t ) p(t)=A_{\mathrm{W}}\log(1/t) p ( t ) = A W log ( 1/ t ) , c J log = A W ( J − 1 ) ( log M ) / 2 c_J^{\log}=A_{\mathrm{W}}(J-1)(\log M)/2 c J l o g = A W ( J − 1 ) ( log M ) /2 , and p + ( t ) = p ( t ) + c J log p^+(t)=p(t)+c_J^{\log} p + ( t ) = p ( t ) + c J l o g . The common choice p i = p ( t ) p_i=p(t) p i = p ( t ) lies in [ p ( t ) , p + ( t ) ] [p(t),p^+(t)] [ p ( t ) , p + ( t )] . Since g i / t = M ( i − s + 1 ) / 2 g_i/t=M^{(i-s+1)/2} g i / t = M ( i − s + 1 ) /2 , the choice p ( g i ) p(g_i) p ( g i ) at each intermediate scale lies in that interval as well; only the common choice is used here. With the fixed ceiling H = 1 + E + σ ∗ H=1+E+\sigma_* H = 1 + E + σ ∗ and the construction constants of the analytic theorem, the relevant conditions on the coupling can be checked in the following sufficient form:
t p + ( t ) < R e n , C J t p + ( t ) < 1 2 , C J t [ ( p + ( t ) ) 3 + p + ( t ) + H p + ( t ) ] 1 − C J t p + ( t ) ≤ ϵ J , 2 t [ ( p + ( t ) ) 3 + p + ( t ) + H p + ( t ) ] ≤ c z , J , C 0 t 2 L V ≤ 1 16 , C 1 t ( L V + 2 ) < min { 1 4 3 , p ( t ) 2 3 } . (121) \begin{aligned}
tp^+(t)&<R_{\mathrm{en}}, \qquad C_Jtp^+(t)<\frac{1}{2},\\
\frac{C_Jt[(p^+(t))^3+p^+(t)+Hp^+(t)]}{1-C_Jtp^+(t)}&\le\epsilon_J,\qquad2t[(p^+(t))^3+p^+(t)+Hp^+(t)]\le c_{z,J},\\
C_0t^2\mathcal{L}_{\mathrm{V}}&\le\frac{1}{16},\qquad C_1t(\mathcal{L}_{\mathrm{V}}+2)<\min\left\{\frac{1}{4\sqrt{3}},\frac{p(t)}{2\sqrt{3}}\right\}. \tag*{(121)}
\end{aligned} t p + ( t ) 1 − C J t p + ( t ) C J t [( p + ( t ) ) 3 + p + ( t ) + H p + ( t )] C 0 t 2 L V < R en , C J t p + ( t ) < 2 1 , ≤ ϵ J , 2 t [( p + ( t ) ) 3 + p + ( t ) + H p + ( t )] ≤ c z , J , ≤ 16 1 , C 1 t ( L V + 2 ) < min { 4 3 1 , 2 3 p ( t ) } . (121)
The first line ensures that the chosen radii are enclosed, and the second checks the conditions on the interaction and on the auxiliary half-disc of Theorem 20.14. The last line checks the curvature and mean conditions of Lemma 20.5, in addition to the already fixed small radius R W R_{\mathrm{W}} R W . Any needed provisional output remainder has the form C J t 2 ( E + σ ∗ ) [ ( p + ) 4 + ( E + σ ∗ ) ( p + ) 2 ] C_Jt^2(E+\sigma_*)[(p^+)^4+(E+\sigma_*)(p^+)^2] C J t 2 ( E + σ ∗ ) [( p + ) 4 + ( E + σ ∗ ) ( p + ) 2 ] and fits its fixed bound by the same reduction. Indeed, with u = log ( 1 / t ) u=\log(1/t) u = log ( 1/ t ) , each left side required to be small is bounded by a finite sum of terms e − k u ( A W u + c J log ) l e^{-ku}(A_{\mathrm{W}}u+c_J^{\log})^l e − k u ( A W u + c J l o g ) l with k > 0 k>0 k > 0 , or by such a sum divided by a denominator at least 1 / 2 1/2 1/2 , and these tend to zero as u → ∞ u\to\infty u → ∞ . With the amplitude A W A_{\mathrm{W}} A W fixed, we choose a finite U U U beyond all these thresholds and g ∗ ≤ e − U < e − 1 g_*\le e^{-U}<e^{-1} g ∗ ≤ e − U < e − 1 . Every condition then holds uniformly for 0 < t ≤ g ∗ 0 < t \le g_{*} 0 < t ≤ g ∗ , and p ( t ) ≥ 1 p(t) \ge1 p ( t ) ≥ 1 . The ceiling may depend on A W A_W A W ; it is independent of the side, the preceding depth, the number of steps, and the frozen coordinates. No expansion is applied on the larger radii R e n / t R_{\mathrm{en}}/t R en / t : that product is used only for the comparison of Lemma 21.12 .
Coordinate coverage and representatives satisfying the reconstruction identities
We show that every supported point of every relevant fiber has small relative coordinates in a suitable gauge, and that its normalized fluctuation coordinates lie in the enclosing product domain. We also construct, from the coarsest scale downward, the representatives satisfying the reconstruction identities, at which the densities will be evaluated. The comparison of the corresponding integrals with integrals over the fluctuation cutoff domain is Lemma 21.12 in § 2.3.6 . That comparison requires an incoming density bound, which is proved there by induction. The coverage statement proved here uses no density bound.
For a group of averaging steps, the conditional minimizers, Hessians K i K_i K i , and reconstruction maps C t C_t C t are those of Theorems 19.9 and 20.14 , with the gauge transformation of Lemma 19.5 , which depends only on the target. For a supported fiber point, z a c t z^{\mathrm{act}} z act denotes its free logarithmic coordinates and w a c t w^{\mathrm{act}} w act its normalized fluctuation coordinates, obtained by inverting the reconstruction. The enclosing product domain is
K t = ∏ i , e { ∣ w i , e ∣ ≤ R e n / t max } , \mathcal{K}_t = \prod_{i,e}\{|w_{i,e}| \le R_{\mathrm{en}}/t_{\max}\}, K t = i , e ∏ { ∣ w i , e ∣ ≤ R en / t m a x } ,
with R e n R_{\mathrm{en}} R en as in (120) .
Closed supports and retained transitions. Fix the finite lattice, range of scales, and decomposition X \mathcal{X} X . Before constructing densities, let η t X \eta_t^{\mathcal{X}} η t X be the actual stage measure obtained from the initial measure restricted to A X \mathcal{A}_{\mathcal{X}} A X by the integrations and freezings up to stage t t t . Its variables form the finite compact product X t X_t X t of current links and native frozen group coordinates, and put K t = supp η t X K_t = \operatorname{supp}\eta_t^{\mathcal{X}} K t = supp η t X . If F t F_t F t is the retained output map before freezing, a supported fiber over y y y means exactly K t ∩ F t − 1 ( y ) K_t \cap F_t^{-1}(y) K t ∩ F t − 1 ( y ) ; its free-coordinate set is the inverse image of this intersection under the local reconstruction. At a freezing boundary we use this pre-freeze convention and then the freezing bijection. This definition uses no version of a conditional measure.
The strict passing inequalities of the algorithm become q ( c p ) ≤ d ∗ q(c_p) \le d_{*} q ( c p ) ≤ d ∗ on these closed supports. Here is the local continuity justification. At each fixed finite stage, approach a support point by data actually produced by the restricted measure. The dependence exclusions ensure that every retained transition reads only the good-cell variables, the prescribed near data, and the protected native records; it never inserts exceptional deep frozen values into a local analytic map. The gauge tuples exhibiting the geometric bounds form a finite product of compact copies of S U ( 2 ) \mathrm{SU}(2) SU ( 2 ) , with closed endpoint compatibility conditions. A subsequence of such tuples converges. All group path products and native saving maps are continuous, so their bounds pass to the limit, and the plaquette bounds become weak inequalities. The resulting geometric bounds still place every relative logarithm and local constraint inverse strictly inside its fixed larger chart: the final radii were chosen below those margins by fixed factors. In this limiting gauge the retained transition is therefore the ordinary continuous analytic branch, on a neighborhood of the support point. Gauge equivariance identifies it with F t F_t F t in the original coordinates. This proves continuity of the retained transitions on the full K t K_t K t , by induction over the finitely many stages, starting with the continuous input plaquettes. It asserts no continuity of the globally Borel-extended averaging map on excluded data and chooses no continuous gauge map.
For completeness, this also propagates support through the actual restricted stage measures. If ν t \nu_t ν t is the measure at a transition after its retained indicators, then supp ν t ⊂ K t \operatorname{supp}\nu_t \subset K_t supp ν t ⊂ K t . Continuity on this compact set gives
supp ( F t ) # ν t = F t ( supp ν t ) . \operatorname{supp}(F_t)_\#\nu_t = F_t(\operatorname{supp}\nu_t). supp ( F t ) # ν t = F t ( supp ν t ) .
Indeed, the image is compact and carries the pushforward, while every neighborhood of an image point pulls back to a relative neighborhood of a support point of positive measure. Freezing and its inverse are continuous triangular group path-product maps, so the same assertion applies after freezing. Thus every supported output has supported predecessors, and the geometric and native collar bounds just proved apply to every supported fiber point, including the boundary. The ordered tests themselves retain their strict passing rule.
Lemma 21.8. There is a fixed ϵ r e l , J > 0 \epsilon_{\mathrm{rel},J} > 0 ϵ rel , J > 0 , depending only on the finite path and local inverse-chart margins fixed before the final radii, with the following property. Consider an inter-step output before its freezing map, with old records b r e c o l d b_{\mathrm{rec}}^{\mathrm{old}} b rec old and retained closed-cell complex P P P . Suppose that q ( c p ) ≤ d ∗ q(c_p) \le d_* q ( c p ) ≤ d ∗ for every plaquette of P P P , including its boundary, that the ordinary links of its prescribed collar Γ \Gamma Γ have their native creation-lattice bound ρ G ( c e ) ≤ C B r \rho_G(c_e) \le C_B r ρ G ( c e ) ≤ C B r , and that the other analytic output-band and near data have the bounds of Theorem 21.5 (2). Assume also r + β r e c < ϵ r e l , J r + \beta_{\mathrm{rec}} < \epsilon_{\mathrm{rel},J} r + β rec < ϵ rel , J . Then a gauge transformation K K K , prescribed to be I I I on Γ \Gamma Γ and extended by I I I outside P P P , puts the retained interior links below r / 2 r/2 r /2 and all links of P P P below β r e c / 2 \beta_{\mathrm{rec}}/2 β rec /2 . It is I I I at every protected endpoint and basepoint of an old record, including variables already frozen at the input of the current group, preserves b r e c o l d b_{\mathrm{rec}}^{\mathrm{old}} b rec old and the zero-completed earlier coordinates exactly, and gives a small output datum for the relative-coordinate argument of Theorem 21.9(1). The non-collar links of the current output band may move under K K K ; at this pre-freeze comparison they belong to the output field, not to b r e c o l d b_{\mathrm{rec}}^{\mathrm{old}} b rec old .
These hypotheses hold at every datum obtained by inverse freezing from a history-produced representative at the next stage, including every supported fiber point over that datum. Indeed, the ordinary Γ \Gamma Γ links are recorded literally in the full frozen band, and the inverse freezing map restores them independently of the retained free links. No claim is made that an arbitrary small representative whose saved coordinates are merely below β r e c \beta_{\mathrm{rec}} β rec has this native collar property.
Proof. The hypothesis q ( c p ) ≤ d ∗ q(c_p) \le d_* q ( c p ) ≤ d ∗ and Step 8 of the proof of Theorem 21.3 give ρ G ( c p ) ≤ δ = π 2 d ∗ \rho_G(c_p) \le\delta= \frac{\pi}{2d_*} ρ G ( c p ) ≤ δ = 2 d ∗ π on every plaquette of P P P ; on stage supports this weak hypothesis is supplied by the preceding closed-support argument. In Theorem 21.3 (2)–(3), take k Γ = I k_\Gamma= I k Γ = I , η = C B r \eta= C_B r η = C B r , and R = ⌈ A 8 / r ⌉ R = \lceil A_8/r \rceil R = ⌈ A 8 / r ⌉ . The already imposed conditions δ R 2 ≤ 1 \delta R^2 \le1 δ R 2 ≤ 1 and K B 2 r / A 8 ≤ c K_B^2r/A_8 \le c K B 2 r / A 8 ≤ c give the relative extension. It is I I I on Γ \Gamma Γ and on the boundary vertices of P P P , so extension by I I I outside P P P is consistent. Its link bounds on P P P and in cells without a vertex of Γ \Gamma Γ are those stated; every retained cell meeting Γ \Gamma Γ belongs to the frozen band, so every retained interior link lies in a cell with no vertex in Γ \Gamma Γ . Analytic near values outside P P P keep their original bounds, while the finitely many path products crossing the collar have the same fixed L J L_J L J conversion as in Step 4 of the proof of Theorem 21.5 . The path products and coordinate conversions have fixed length or fixed derivative-row bounds. Their values are at most C J ( r + β r e c ) C_J(r+\beta_{\mathrm{rec}}) C J ( r + β rec ) in the local chart, so choosing ϵ r e l , J \epsilon_{\mathrm{rel},J} ϵ rel , J below the minimum of the finitely many positive inverse-chart margins divided by this fixed C J C_J C J places every coordinate used in the small relative argument inside its chart. This threshold is fixed before the final radii. The endpoint protection rule ensures that the lift of K K K is I I I at every mark of b r e c o l d b_{\mathrm{rec}}^{\mathrm{old}} b rec old , including the input variables already frozen in this group. In particular, replacing an earlier frozen value by the identity in the regional completion commutes with K K K : both endpoints of that completed coordinate are fixed.
For the last assertion, at fixed retained output v v v the freezing map of Lemma 21.11 records ordinary band links directly. Only a crossing link is replaced by the based expression U v ( σ e ) a e U_v(\sigma_e)a_e U v ( σ e ) a e . The closed-cell collar lies in the full ordinary frozen band, whose incident retained cells all belong to the frozen band, so no collar link has an endpoint in the active region and its links use the former rule. Its recorded values and both endpoints were protected by Theorem 21.5 (2). Consequently every inverse freezing over the same next-stage datum restores the identical collar values, regardless of the supported free coordinates. The comparison gauge K K K is applied only to this restored pre-freeze output; it does not overwrite the recorded values in the actual downward family.
Theorem 21.9. There exist 0 < ϵ e n , J ≤ ϵ r e l , J 0 < \epsilon_{\mathrm{en},J} \le\epsilon_{\mathrm{rel},J} 0 < ϵ en , J ≤ ϵ rel , J and C e n , J < ∞ C_{\mathrm{en},J} < \infty C en , J < ∞ , with the dependence stated at the beginning of this section and independent of all later widths of boundary bands, removed bands, protected neighborhoods, and cells, such that the following holds for every large-field decomposition of Theorem 21.5 . Assume its construction conditions, the two reconstruction-image conditions (119) , and r + β r e c < ϵ e n , J r+\beta_{\mathrm{rec}} < \epsilon_{\mathrm{en},J} r + β rec < ϵ en , J .
Let 1 ≤ s ≤ J 1 \le s \le J 1 ≤ s ≤ J . Let ( c , b r e c ) (c,b_{\mathrm{rec}}) ( c , b rec ) be the output field and frozen coordinates of a configuration in the support of the stage measure, given either by a small representative of Theorem 21.5 (2) (at the first step for a stopped measure, of Corollary 21.2 ), by the small comparison datum of Lemma 21.8 , or by a representative obtained from either such datum by an output gauge change acting on the old frozen coordinates as in Lemma 21.6(1)–(3). Assume that this representative and the associated fields before the target-dependent gauge transformation belong to the common larger construction domain of Theorem 20.14 , and that its analytic frozen coordinates satisfy (110) . Let U U U be a configuration in the support of the fiber over ( c , b r e c ) (c,b_{\mathrm{rec}}) ( c , b rec ) . Then there exists a correcting gauge transformation, equal to I I I at the roots of the block region, at the vertices of the frozen boundary layer on the side of the fixed complement (in particular at the endpoints v e v_e v e of the links across the boundary of the block region), and at all protected earlier endpoints and basepoints, with the following property. After the hierarchical tree gauge fixing, every relative constraint vector X i X_i X i of U U U , including its eliminated entries, and its free part z i z_i z i satisfy
max 0 ≤ i ≤ s ( ∥ X i ∥ ∞ + ∥ z i ∥ ∞ ) ≤ C J ( r + β r e c ) . \max_{0 \le i \le s}\left(\lVert X_i\rVert_{\infty}+\lVert z_i\rVert_{\infty}\right)\le C_J(r+\beta_{\mathrm{rec}}). 0 ≤ i ≤ s max ( ∥ X i ∥ ∞ + ∥ z i ∥ ∞ ) ≤ C J ( r + β rec ) .
They lie in the common domain of the local inverse of the block averaging constraint, with a fixed positive margin. The correcting gauge transformation is the composition of a smallness gauge and the lift, constant on blocks, of the inverse of its root values; its factors are not separately required to fix the roots. The inner endpoint u e u_e u e of a current link across the boundary of the block region may move before the tree gauge fixing, while the based transport is preserved, since its endpoints are fixed.
Assume at every integrated stage the adjusted classical perturbation bound X c l ( P ‾ ) ≤ 2 σ ∗ X^{\mathrm{cl}}(\overline{P})\le2\sigma_* X cl ( P ) ≤ 2 σ ∗ and the uniform conditional-force, Hessian, and sequential margins of Theorems 19.9 and 20.14 ; the initial norm bounds and coupling ceiling of the latter theorem supply these bounds along its analytic recursion. Choose all history-produced small top representatives in the coordinate domains of part (1), retaining the native collar bounds of Theorem 21.5 (2). Define the reconstruction maps on their full spatially varying radii first, and descend through them, aligning the target at each step by the accumulated output gauge change. The unperturbed coordinates, conditional minimizers, and inverse square roots of the Hessians transform by blockwise adjoint orthogonal transformations, and the perturbing force is estimated separately. This constructs representatives satisfying the reconstruction identities at all stages, with
( T M s C t ( c , ζ ; b r e c ) ) ∣ Ω t ′ = c , \left.(T_M^s C_t(c,\zeta;b_{\mathrm{rec}}))\right|_{\Omega'_t}=c, ( T M s C t ( c , ζ ; b rec )) ∣ Ω t ′ = c ,
and for every point of every fiber over every resulting output datum, its normalized fluctuation coordinates satisfy
max i , e g i ∣ w i , e a c t ∣ ≤ δ e n = C e n , J ( r + β r e c + σ ∗ ρ c l ) . \max_{i,e} g_i\left|w^{\mathrm{act}}_{i,e}\right|
\le\delta_{\mathrm{en}}
= C_{\mathrm{en},J}\left(r+\beta_{\mathrm{rec}}+\sigma_*\rho_{\mathrm{cl}}\right). i , e max g i w i , e act ≤ δ en = C en , J ( r + β rec + σ ∗ ρ cl ) .
With R e n = 2 δ e n / λ ∗ R_{\mathrm{en}} = 2\delta_{\mathrm{en}}/\lambda_* R en = 2 δ en / λ ∗ , each such point lies in the interior ∣ w i , e a c t ∣ ≤ R e n / ( 2 t max ) |w^{\mathrm{act}}_{i,e}| \le R_{\mathrm{en}}/(2t_{\max}) ∣ w i , e act ∣ ≤ R en / ( 2 t m a x ) of the enclosing product domain K t K_t K t . Each gauge transformation of Lemma 19.5 used here is I I I at all protected vertices, including the current endpoints of the links across the boundary of the block region. No estimate on the logarithm of the accumulated gauge transformation is used. The statement concerns every supported fiber point, including points near the boundary of the coordinate domains, not one selected point or almost every point under an auxiliary measure.
At the first RG step for a stopped measure, every nonempty fiber over an output field and all input frozen coordinates that meets D r D_r D r has a representative in which all output links and input frozen coordinates are at most C a d , J r C_{\mathrm{ad},Jr} C ad , J r . The scaled fluctuations of the unperturbed objective, K 0 1 / 2 ( z a c t − z 0 ) K_0^{1/2}(z^{\mathrm{act}}-z_0) K 0 1/2 ( z act − z 0 ) , are bounded by C J ( r + C a d , J r ) C_J(r+C_{\mathrm{ad},J}r) C J ( r + C ad , J r ) .
At the first RG step under the Wilson measure, let b n e a r b^{\mathrm{near}} b near be the near values, that is, the values on the near links of (18.32), which are the links outside the block region read by a plaquette depending on an integration variable or by a path product of the block region in the initial good-cell complex. Let η d e e p \eta_{\mathrm{deep}} η deep be the deep frozen group variables, the values on all remaining links. Then we have the exact identity
S W ( U ) = F b a r e ( U i n t ; b n e a r ) + E h o l e ( b n e a r , η d e e p ) . S_W(U) = F_{\mathrm{bare}}(U_{\mathrm{int}};b^{\mathrm{near}}) + E_{\mathrm{hole}}(b^{\mathrm{near}},\eta_{\mathrm{deep}}). S W ( U ) = F bare ( U int ; b near ) + E hole ( b near , η deep ) .
where U i n t U_{\mathrm{int}} U int denotes the integration variables and F b a r e F_{\mathrm{bare}} F bare is the constrained Wilson-action objective of the first block integration, defined in Definition 18.13 and written in these variables; in the coordinates of Definition 18.6 this identity is (18.38). No smallness is imposed on η d e e p \eta_{\mathrm{deep}} η deep ; its factor e − E h o l e / g b a s e 2 e^{-E_{\mathrm{hole}}/g_{\mathrm{base}}^2} e − E hole / g base 2 remains with its original Haar integration and the original partition function. Every conditional minimizer, derivative, square root, and Haar disintegration factor of the integration variables is independent of η d e e p \eta_{\mathrm{deep}} η deep . A plaquette containing only frozen links, near or deep, belongs to E h o l e E_{\mathrm{hole}} E hole . At this step there is no inherited effective interaction, and the scaled fluctuations are bounded by C J ( r + β r e c ) C_J(r+\beta_{\mathrm{rec}}) C J ( r + β rec ) . The deep frozen group variables and exceptional values of the Borel extension in them are never logarithmized or inserted into a torus analytic action.
The constants have the dependence stated at the beginning of this section, and they are uniform in the deep frozen group variables.
Proof. We first work at a small output representative and an arbitrary supported fiber point over it (Steps 1–3). Step 4 transports the estimate to the representatives produced from above, using only the unperturbed objective, and Step 5 estimates the perturbation. Step 6 constructs the representatives satisfying the reconstruction identities, Step 7 converts the bound to the enclosing product domain, and Steps 8 and 9 treat the first step for the two initial measures. For s < J s<J s < J , the image hypothesis includes the separately imposed interior image condition of that group of steps. Only its conditional local chart and force estimates are used here; for fewer than J J J steps there is no restoration of norms and no recursion of normalization constants.
Step 1: a smallness gauge for an arbitrary fiber point. We work at a small output datum ( c , b r e c ) (c,b_{\mathrm{rec}}) ( c , b rec ) supplied either by Theorem 21.5 or, for a descended inter-step datum, by Lemma 21.8, and keep this notation until Step 4. Its retained interior output links are bounded by r / 2 r/2 r /2 , while current output-band links and newly used near and based data are bounded by β r e c \beta_{\mathrm{rec}} β rec ; its analytic old frozen coordinates obey (21.8). The root-increment argument below uses only ρ G ( c a a ′ ) ≤ r + β r e c \rho_G(c_{aa'}) \le r+\beta_{\mathrm{rec}} ρ G ( c a a ′ ) ≤ r + β rec for an internal final link, whether that link is retained or in the current output band. At the relative-admission comparison of Lemma 21.8, the input point is separately put in a smallness gauge as below; the collar estimate alone is not used as an input smallness assumption. An arbitrary analytic background on a larger spatially varying radius need not satisfy the absolute bound on c c c used in this step. Fix any configuration U U U in the support of the fiber over this representative. By Theorem 21.5, there exists a smallness gauge S S S for the input links of the block region. We restrict S S S to the input vertices of the block region and set it equal to I I I elsewhere; internal links are unchanged by this restriction. By the endpoint protection rule, its lifts fix every earlier endpoint and basepoint. A current prescribed based transport is not a protected earlier record. It may move under this intermediate operation; the combined correction below fixes its endpoints and restores its prescribed value. We write V = U S V=U^S V = U S on the block region and let u i n u_{\mathrm{in}} u in be its bound for interior links. At the initial input for a stopped measure we use u i n = r u_{\mathrm{in}}=r u in = r ; at later inputs u i n ≤ r / 2 u_{\mathrm{in}}\le r/2 u in ≤ r /2 ; at the first input under the Wilson measure u i n ≤ r u_{\mathrm{in}}\le r u in ≤ r , after the fixed choice of A 8 A_8 A 8 . By Lemma 21.1, for internal dependence neighborhoods reading only interior links, we have
max e i n t e r n a l ρ G ( ( T M i V ) e ) ≤ ( 10 M ) i u i n , 0 ≤ i ≤ s . (122) \max_{e\ \mathrm{internal}} \rho_G\left((T_M^iV)_e\right) \le(10M)^i u_{\mathrm{in}}, \qquad0\le i\le s. \tag*{(122)} e internal max ρ G ( ( T M i V ) e ) ≤ ( 10 M ) i u in , 0 ≤ i ≤ s . (122)
These dependence sets lie inside full blocks of the region. The display is not used for a path product reading a prescribed near value; such path products are bounded after the alignment below, using their frozen data as well as the interior bound.
Step 2: the root values. For each final root a a a of the region, let h a = S ( a ) − 1 h_a=S(a)^{-1} h a = S ( a ) − 1 . On an internal final link ( a , a ′ ) (a,a') ( a , a ′ ) , by the gauge covariance of the block averaging constraint, we have
c a a ′ = h a ( T M S V ) a a ′ h a ′ − 1 . c_{aa'}=h_a(T_M^SV)_{aa'}h_{a'}^{-1}. c a a ′ = h a ( T M S V ) a a ′ h a ′ − 1 .
Multiplying this identity by the appropriate inverse factor and using bi-invariance, we obtain
ρ G ( h a h a ′ − 1 ) ≤ ρ G ( c a a ′ ) + ρ G ( ( T M S V ) a a ′ ) ≤ r + β r e c + ( 10 M ) s u i n . (123) \rho_G(h_ah_{a'}^{-1})\le\rho_G(c_{aa'})+\rho_G\left((T_M^SV)_{aa'}\right)\le r+\beta_{\mathrm{rec}}+(10M)^s u_{\mathrm{in}}. \tag*{(123)} ρ G ( h a h a ′ − 1 ) ≤ ρ G ( c a a ′ ) + ρ G ( ( T M S V ) a a ′ ) ≤ r + β rec + ( 10 M ) s u in . (123)
We lift h h h as a constant on each final block of the region, with value I I I off these blocks, and denote this lift by L L L (a gauge transformation, not the side of the torus). The total correction is H = L S H=LS H = L S , since ( U S ) L = U L S (U^S)^L=U^{LS} ( U S ) L = U L S , and it equals I I I at every final root. At the vertices of the frozen boundary layer on the side of the fixed complement both factors are I I I , and the endpoint protection rule fixes the earlier frozen endpoints after lifting to their own lattices. An intermediate internal link either stays in one final block or crosses a single boundary between adjacent final blocks. In the first case L L L conjugates its value; in the second case its angle increases by at most the root increment in (21.21). Together with (21.20), this bounds all internal iterates by C J ( r + β r e c ) C_J(r+\beta_{\mathrm{rec}}) C J ( r + β rec ) . A component consisting of a single block has only the first case, and its root value causes no additional estimate.
The bound ρ G ( c a a ′ ) ≤ r + β r e c \rho_G(c_{aa'})\le r+\beta_{\mathrm{rec}} ρ G ( c a a ′ ) ≤ r + β rec used in (21.21) belongs both to a forward small representative and to the pre-freeze small comparison datum supplied by Lemma 21.8: its retained interior links are below r / 2 r/2 r /2 , its current analytic output band is below β r e c \beta_{\mathrm{rec}} β rec , and any other internal link read here obeys the same sum bound. At the first step for a stopped measure, the representative of Corollary 21.2 has output bound C a d , J r < β r e c C_{\mathrm{ad},J}r<\beta_{\mathrm{rec}} C ad , J r < β rec , by (21.6), so the displayed bound holds with u i n = r u_{\mathrm{in}}=r u in = r .
Step 3: based transports and the branch. The endpoints of a current based transport are a root of the region and the endpoint of its link in the fixed complement, and the factors of H H H at both are I I I . The transport is therefore unchanged, although the endpoint u e u_e u e in the region can have moved. We now apply the hierarchical based tree gauge fixing of Lemma 21.1. Since its paths have length at most L J L_J L J , all internal links after the tree gauge fixing satisfy the same bound with a larger C J C_J C J . A link across the boundary of the block region that is read by a path product of the region is the inverse of such a bounded path times its prescribed based transport. It is therefore bounded by C J ( r + β r e c ) C_J(r+\beta_{\mathrm{rec}}) C J ( r + β rec ) as well. Every path product of the region mixing integration variables and near values has at most L J L_J L J factors of these links and of its prescribed frozen data; its bound is C J ( r + β r e c ) C_J(r+\beta_{\mathrm{rec}}) C J ( r + β rec ) , including the contribution of the frozen data. We include the initial C a d , J C_{\mathrm{ad},J} C ad , J and the constants for these mixed path products in this fixed C J C_J C J before choosing ϵ e n , J \epsilon_{\mathrm{en},J} ϵ en , J or the final radii. At the small representative, the relative vector X i X_i X i and its free part z i z_i z i then satisfy the estimate in part (1). We choose ϵ e n , J \epsilon_{\mathrm{en},J} ϵ en , J so that this bound is less than half the common uniqueness radius and all the fixed construction margins. Since the configuration solves the block averaging constraint, its eliminated entries equal the unique eliminated entries on this branch. This reasoning applies to the arbitrary point U U U itself, not to another point of its fiber, and neither S S S nor L L L has to fix the roots separately.
Step 4: transport to representatives produced from above. We carry the estimate to an output datum produced from above, using only the unperturbed objective with torus input. Let k k k be the accumulated output gauge change. Its lift, constant on blocks, preserves the tree identities, and the coarse-field lift of the target satisfies U 0 ( c k ) = U 0 ( c ) k ~ U_0(c^{k})=U_0(c)^{\widetilde{k}} U 0 ( c k ) = U 0 ( c ) k . Each relative factor therefore transforms as
e X e ⟼ e Ad k a X e , z i ⟼ A k , i z i , (124) e^{X_e}\longmapsto e^{\operatorname{Ad}_{k_a}X_e},\qquad z_i\longmapsto A_{k,i}z_i, \tag*{(124)} e X e ⟼ e Ad k a X e , z i ⟼ A k , i z i , (124)
where a a a is the root of its block and A k , i A_{k,i} A k , i is block diagonal and orthogonal on the real color vectors. The equation for the eliminated edges transforms in the same way, as we see by conjugating the constraint and using the uniqueness of its small inverse. We repeat this over the s ≤ J s \le J s ≤ J intermediate scales. Since the changes at earlier marks are the identity, the prescribed frozen data are preserved. For a comparison of two representatives that are not related by a based change, the same formula uses their correspondingly transformed output data.
For a descended inter-step datum, we first use Lemma 21.8 on the pre-freeze output restored from its next-stage records. This produces a small datum by a relative gauge K K K equal to I I I at every earlier frozen mark. The inverse change k = K − 1 k = K^{-1} k = K − 1 carries that small datum back to the given one; its lift is endpoint preserving, so the old records and every coordinate completed by zero agree in both comparisons. The current output band is still part of c c c at this step. Its non-collar links may change under K K K , but they are not old records in b r e c b_{\mathrm{rec}} b rec in the local comparison. Only after the comparison is the actual output frozen by the raw freezing map, which records its actual values. The ordered tests and the initial measure are gauge invariant, so lifting K K K takes each supported fiber point to a supported point over the comparison datum with the same old records. We apply this change separately to each supported fiber point. Formula (124) transports the relative bound and the eliminated-edge inverse back to the given output datum, even if its output links are bounded only by a larger analytic radius; no absolute bound r / 2 r/2 r /2 on its iterates is asserted. For part (1), an output datum already related to a small representative by a change of Lemma 21.6(1)–(3) uses that stated change. All old frozen coordinates in the descended comparison satisfy (21.8), and no arbitrary endpoint change of part (4) is used for this analytic estimate.
All fields in this calculation, taken before the target-dependent gauge transformation, lie in the common larger construction domain of Theorem 20.14. In particular, the backgrounds at block boundaries, which need not be of order M − s ρ c l M^{-s}\rho_{\mathrm{cl}} M − s ρ cl , lie in that domain. The torus classical values used there agree under a compact periodic gauge transformation whenever both representatives lie in the common overlap chart, by Theorem 11.6, as recorded in (14.3). Their membership in that common domain follows from Lemma 19.7(1) and the construction-domain hypothesis of part (1). In a regional completion, the prescribed small frozen data transform at the endpoints of their paths. The earlier frozen coordinates that are completed by zero have both endpoints fixed by the endpoint protection rule, so appending them commutes with this change. For the unperturbed objective F 0 F_0 F 0 and the product color map A k A_k A k ,
F 0 ( c k , A k z ; b r e c k ) = F 0 ( c , z ; b r e c ) . (125) F_0(c^k,A_k z;b^k_{\mathrm{rec}})=F_0(c,z;b_{\mathrm{rec}}). \tag*{(125)} F 0 ( c k , A k z ; b rec k ) = F 0 ( c , z ; b rec ) . (125)
This identity is compact overlap applied to the torus classical restriction after regional completion. The endpoint-preserving change commutes with zero completion of the old coordinates, and the local constraint inverse transports the remaining coordinates by (21.22). No covariance of the regional perturbation is asserted.
By stationarity in z z z and the uniqueness of the unperturbed conditional minimizer, we have
z 0 ( c k ; b r e c k ) = A k z 0 ( c ; b r e c ) , K 0 ( c k ; b r e c k ) = A k K 0 ( c ; b r e c ) A k − 1 . (126) z_0(c^k;b_{\mathrm{rec}}^k)=A_kz_0(c;b_{\mathrm{rec}}),\qquad K_0(c^k;b_{\mathrm{rec}}^k)=A_kK_0(c;b_{\mathrm{rec}})A_k^{-1}. \tag*{(126)} z 0 ( c k ; b rec k ) = A k z 0 ( c ; b rec ) , K 0 ( c k ; b rec k ) = A k K 0 ( c ; b rec ) A k − 1 . (126)
The minimizer stays in the common uniqueness box; if boxes in scalar coordinates are used, the fixed comparison of color norms is included in the margin of the larger construction domain. An orthogonal conjugate of a positive square root is positive and squares to the conjugated matrix, so uniqueness of positive square roots gives the same identity for K 0 1 / 2 K_0^{1/2} K 0 1/2 and K 0 − 1 / 2 K_0^{-1/2} K 0 − 1/2 . Applying this argument after each conditional minimization, we obtain it for every initial segment. At a small output representative, Theorem 18.12(4) gives
∥ z 0 ∥ ∞ ≤ C J ( ∥ c ∥ ∞ + ∥ ξ ∥ ∞ ) ≤ C J ( r + β r e c ) . \lVert z_0\rVert_\infty\le C_J(\lVert c\rVert_\infty+\lVert\xi\rVert_\infty)\le C_J(r+\beta_{\mathrm{rec}}). ∥ z 0 ∥ ∞ ≤ C J (∥ c ∥ ∞ + ∥ ξ ∥ ∞ ) ≤ C J ( r + β rec ) .
Equations (21.22) and (21.24) carry the bound
∥ K 0 − 1 / 2 ( z a c t − z 0 ) ∥ ∞ ≤ C J ( r + β r e c ) (127) \lVert K_0^{-1/2}(z^{\mathrm{act}}-z_0)\rVert_\infty\le C_J(r+\beta_{\mathrm{rec}}) \tag*{(127)} ∥ K 0 − 1/2 ( z act − z 0 ) ∥ ∞ ≤ C J ( r + β rec ) (127)
to every output datum described in part (1). No bound on Log k \operatorname{Log} k Log k is needed, since blockwise adjoint transformations have norm one for every compact value at the roots.
Step 5: the perturbation. We treat the perturbation separately, at fixed real b = ( ξ , b r e c ) b=(\xi,b_{\mathrm{rec}}) b = ( ξ , b rec ) . The uniform incoming radius, the boundary-distance comparison, and the conditional force estimates are Lemma 21.7; the analytic norm bounds are those of the recursion of Theorem 20.14, constructed without any density inequality. In the notation of Theorem 19.9, the preconditioned force is at most C J σ ∗ ρ c l C_J\sigma_*\rho_{\mathrm{cl}} C J σ ∗ ρ cl , the change of the constrained Hessian is at most C J σ ∗ C_J\sigma_* C J σ ∗ , and the rows of all square roots for initial segments are bounded by C J C_J C J . That theorem estimates the perturbation on the reconstructed field in the target-dependent gauge, including the full boundary-completion difference; it asserts no covariance of the perturbation, and none is needed. We first place the higher coordinates of the arbitrary supported point in the domains where these conditional objects are defined. Put b e n = r + β r e c b_{\mathrm{en}}=r+\beta_{\mathrm{rec}} b en = r + β rec . The inverse-square-root row bound of Step 4 and (21.25) give, for the unperturbed joint center z 0 z_0 z 0 ,
∥ z a c t − z 0 ∥ ∞ ≤ G J b e n . \lVert z^{\mathrm{act}}-z_0\rVert_\infty\le G_Jb_{\mathrm{en}}. ∥ z act − z 0 ∥ ∞ ≤ G J b en .
Here G J G_J G J includes the fixed conversions between color-vector and scalar-coordinate sup norms; it is bounded on the larger construction products before the final radii are chosen. For each initial segment I I I , put v I = y I a c t − z 0 , I v_I=y_I^{\mathrm{act}}-z_{0,I} v I = y I act − z 0 , I , where y I a c t = z I a c t y_I^{\mathrm{act}}=z_I^{\mathrm{act}} y I act = z I act . At the fixed real coarse field and frozen datum, regard v I v_I v I as a constant vector in the shifted parameter chart y = z 0 , I + v y=z_{0,I}+v y = z 0 , I + v of Lemma 19.3(1). Its output-fixed degree-zero coefficient norm is exactly ∥ v I ∥ ∞ \lVert v_I\rVert_\infty ∥ v I ∥ ∞ , with no volume factor, and is at most G J b e n G_Jb_{\mathrm{en}} G J b en . Let η min > 0 \eta_{\min}>0 η m i n > 0 be the minimum of the finitely many fixed parameter margins defining T I T_I T I , with those same color-norm conversions included. In the already adjustable choice of ϵ e n , J \epsilon_{\mathrm{en},J} ϵ en , J in Step 3, require also G J ϵ e n , J < η min / 2 G_J\epsilon_{\mathrm{en},J}<\eta_{\min}/2 G J ϵ en , J < η m i n /2 . Since b e n < ϵ e n , J b_{\mathrm{en}}<\epsilon_{\mathrm{en},J} b en < ϵ en , J , every y I a c t y_I^{\mathrm{act}} y I act lies in T I T_I T I with a strict margin. This membership is established before evaluating any perturbed conditional minimizer or inverse fluctuation coordinate at that argument.
Let a i ( y i ) a_i(y_i) a i ( y i ) be the i i i th component of the unperturbed conditional minimizer a I i a_{I_i} a I i for the prefix I i = { 0 , … , i } I_i=\{0,\ldots,i\} I i = { 0 , … , i } , with the coarse field and frozen datum suppressed. This is the conditional center at scale i i i , distinct from the joint center z 0 , i z_{0,i} z 0 , i . It satisfies a i ( z 0 , > i ) = z 0 , i a_i(z_{0,>i})=z_{0,i} a i ( z 0 , > i ) = z 0 , i , and the derivative-row estimate of Lemma 19.3 (1) on the fixed larger products gives a constant D a , J D_{a,J} D a , J such that
∥ z i a c t − a i ( y i a c t ) ∥ ∞ ≤ ( 1 + D a , J ) G J b e n . \left\|z_i^{\mathrm{act}}-a_i(y_i^{\mathrm{act}})\right\|_\infty\leq(1+D_{a,J})G_Jb_{\mathrm{en}}. z i act − a i ( y i act ) ∞ ≤ ( 1 + D a , J ) G J b en .
We decrease the same ϵ e n , J \epsilon_{\mathrm{en},J} ϵ en , J , if necessary, so that this bound is below one quarter of the minimum fixed conditional solution radius. Thus the actual coordinate lies in the common conditional chart. At the already admitted higher argument, (115) and (116) give
∥ z P , i ∗ ( y i a c t ; b r e c ) − a i ( y i a c t ) ∥ ∞ ≤ 2 δ V ≤ C J σ ∗ ρ c l . \left\|z_{P,i}^*(y_i^{\mathrm{act}};b_{\mathrm{rec}})-a_i(y_i^{\mathrm{act}})\right\|_\infty\leq2\delta_V\leq C_J\sigma_*\rho_{\mathrm{cl}}. z P , i ∗ ( y i act ; b rec ) − a i ( y i act ) ∞ ≤ 2 δ V ≤ C J σ ∗ ρ cl .
The second point lies in the same conditional chart by the fixed force and solution margins. Only now do we use the exact inverse relation
g i w i a c t = K P , i ( y i a c t ; b r e c ) 1 / 2 ( z i a c t − z P , i ∗ ( y i a c t ; b r e c ) ) . g_i w_i^{\mathrm{act}}=K_{P,i}(y_i^{\mathrm{act}};b_{\mathrm{rec}})^{1/2}\left(z_i^{\mathrm{act}}-z_{P,i}^*(y_i^{\mathrm{act}};b_{\mathrm{rec}})\right). g i w i act = K P , i ( y i act ; b rec ) 1/2 ( z i act − z P , i ∗ ( y i act ; b rec ) ) .
If S J S_J S J bounds the rows of these perturbed positive square roots on the larger products, then, at every scale and for every supported point,
g i ∥ w i a c t ∥ ∞ ≤ S J [ ( 1 + D a , J ) G J b e n + 2 δ V ] . g_i\left\|w_i^{\mathrm{act}}\right\|_\infty\leq S_J\left[(1+D_{a,J})G_Jb_{\mathrm{en}}+2\delta_V\right]. g i w i act ∞ ≤ S J [ ( 1 + D a , J ) G J b en + 2 δ V ] .
The constants G J G_J G J , D a , J D_{a,J} D a , J , S J S_J S J and the two fixed positive margins precede the final cutoff and boundary widths. The at-most-J J J substitutions of higher arguments are already covered by the uniform conditional derivative rows used in D a , J D_{a,J} D a , J ; no neighborhood is inferred from a fluctuation bound that has yet to be proved. Absorbing these fixed constants and the bound on δ V \delta_V δ V yields
max i , e g i ∣ w i , e a c t ∣ ≤ C e n , J ( r + β r e c + σ ∗ ρ c l ) . (128) \max_{i,e}g_i\left|w_{i,e}^{\mathrm{act}}\right|\leq C_{\mathrm{en},J}\left(r+\beta_{\mathrm{rec}}+\sigma_*\rho_{\mathrm{cl}}\right). \tag*{(128)} i , e max g i w i , e act ≤ C en , J ( r + β rec + σ ∗ ρ cl ) . (128)
The construction conditions keep these conditional neighborhoods inside the joint branch; they are not replaced by arbitrary slices of a joint box of free coordinates. No perturbed derivative in a frozen coordinate ξ \xi ξ or in an earlier frozen coordinate is taken; these coordinates enter only as fixed real parameters of the uniform coefficient estimates.
Step 6: representatives satisfying the reconstruction identities. The construction downward is a finite induction. We start with all history-produced small top representatives of Theorem 21.5 , including both (110) and the native C B r C_{B_r} C B r collar records. At a boundary where coordinates are frozen, the hierarchical tree transports are I I I by Lemma 21.1, and the gauge transformation of Lemma 19.5 is I I I at their endpoints. Undoing the basing therefore inserts the already chosen new frozen values directly. At the next lower group of averaging steps, we lift the accumulated output change, take the free logarithmic coordinates of any configuration in its fiber, and invert the reconstruction defined beforehand. By the branch argument of Steps 1–4 and (128) , these inverse coordinates lie in its domain. The reconstruction has the same target and the same free coordinates, and uniqueness of the eliminated-edge inverse identifies the reconstructed configuration, before the gauge transformation, with the configuration aligned by the accumulated output gauge change. We then apply the gauge transformation of Lemma 19.5, which depends only on the target; by that lemma, the target is unchanged and every protected vertex is fixed. This produces the present representative and permits the next step of the descent. By Lemma 21.6(1)–(2), the subsequent lifts and the based correcting and tree gauges fix every frozen coordinate chosen at this step, so (110) and the native collar records persist throughout the family. At each inverse freezing, the ordinary collar links are restored literally from these records, independently of the newly chosen supported free coordinates, as in Lemma 21.8. Thus the relative-admission argument of Step 4 applies at every lower inter-step output, and all supported points over each fixed output datum are included. An arbitrary small top representative satisfying only (110) is not asserted to have this property. Each step uses only the representative above and the analytic maps defined beforehand on their full radii, so the induction assumes no completed family of representatives. Keeping all choices gives all fibers over all resulting output data and the identity for the target in part (2).
Step 7: the enclosing product domain. At output side m m m , we have g i = t max M ( i − s + 1 ) / 2 ≥ λ ∗ t max g_i=t_{\max}M^{(i-s+1)/2}\geq\lambda_*t_{\max} g i = t m a x M ( i − s + 1 ) /2 ≥ λ ∗ t m a x . Dividing (128) by g i g_i g i , we obtain
∣ w i , e a c t ∣ ≤ δ e n λ ∗ t max = R e n 2 t max . \left|w^{\mathrm{act}}_{i,e}\right|\leq\frac{\delta_{\mathrm{en}}}{\lambda_*t_{\max}}=\frac{R_{\mathrm{en}}}{2t_{\max}}. w i , e act ≤ λ ∗ t m a x δ en = 2 t m a x R en .
Every point of K t \mathcal{K}_t K t has g i ∣ w i , e ∣ ≤ R e n g_i|w_{i,e}|\leq R_{\mathrm{en}} g i ∣ w i , e ∣ ≤ R en . Both the classical and the quantum image conditions are imposed on the whole product K t \mathcal{K}_t K t , not only on the supported subset or on the logarithmic cutoff domain.
Step 8: the first step for a stopped measure. We fix one point of a nonempty fiber over an output field and all input frozen coordinates, and apply Corollary 21.2 to it. In the resulting representative, all output links and all based input transports are at most C a d , J r C_{\mathrm{ad},J}r C ad , J r , and the remaining input frozen links are smaller than r r r . For any other point of the same fiber, we take its own gauge transformation into B r B_r B r , cancel its root values by the lift of the inverse root values, and cancel its gauge transformation at the vertices of the fixed complement. Equations (21.20) and (21.21) hold with u i n = r u_{\mathrm{in}}=r u in = r and the output increment C a d , J r + ( 10 M ) s r C_{\mathrm{ad},J}r+(10M)^sr C ad , J r + ( 10 M ) s r . All input variables used analytically, including all input frozen coordinates, are small after the tree gauge fixing. The unperturbed estimate is C J ( r + C a d , J r ) C_J(r+C_{\mathrm{ad},J}r) C J ( r + C ad , J r ) . This is the regional form of the coordinates of Theorem 14.3 at the first step; it is uniform in the preceding depth and is not applied to frozen coordinates created at later steps.
Step 9: the first step under the Wilson measure. We use the initial good-cell complex of Theorem 21.5 (4), together with the internal coordinates of Definition 18.6 and the completion of Definition 18.13: the links across the boundary of the block region are completed by (18.33), the other links by (18.34), and P v a r \mathcal{P}_{\mathrm{var}} P var is a fixed list as in (18.35)–(18.36), namely the list of the split made in this step. The near links of (18.32) carry the near values b n e a r b^{\mathrm{near}} b near , and the deep links carry the deep frozen group variables η d e e p \eta_{\mathrm{deep}} η deep . The hypothesis of Definition 18.13(3) that the near values lie in the identity chart is the conclusion of Theorem 21.5 (4), since every near link lies in the initial good-cell complex. Indeed, the links of the path products in L Ω d e p \mathcal{L}^{\mathrm{dep}}_{\Omega} L Ω dep lie there because of the inset by the full dependence neighborhood. The plaquettes incident to L 0 \mathcal{L}_0 L 0 lie within one input step of a block of the region, while Theorem 21.5 (4) moves inward by a full layer of cells and the full dependence neighborhood. The conversion between the value on a link across the boundary and its based value is included in C c o n v , J C_{\mathrm{conv},J} C conv , J .
In the hierarchical tree coordinates, we partition the plaquettes according to whether their Wilson term depends on an integration variable. By the definition of the near links, every term in the first class uses only integration variables and near values, and the sum of these terms is F b a r e F_{\mathrm{bare}} F bare . Every remaining term is a function of frozen variables only and is assigned to E h o l e E_{\mathrm{hole}} E hole , even if its plaquette contains both near and deep links. This partitions the sum S W = 4 ∑ p q ( U p ) S_W=4\sum_p q(U_p) S W = 4 ∑ p q ( U p ) , and Lemma 18.14(1) implies the exact split (18.38), which is the identity in part (4). In particular, every objective of the integration variables and all its derivatives are independent of η d e e p \eta_{\mathrm{deep}} η deep , without any estimate of a small remainder. The Haar disintegration path products also avoid the deep links, because of the inset by the full dependence neighborhood. By (18.39), the objective agrees with the objective of Theorem 18.12 at preceding depth zero, up to a function constant in the integration variables. The external factor e − E h o l e / g b a s e 2 e^{-E_{\mathrm{hole}}/g_{\mathrm{base}}^2} e − E hole / g base 2 remains the one in (18.38), with its original Haar variables and partition function.
We apply Lemma 18.14(2)–(3) to the block region of the first step, with the prescribed near completion. Its embedding of the free coordinates is A Ω A_{\Omega} A Ω in (18.15), and its extraction matrix is (18.16). Equations (18.40)–(18.42) apply to the list P v a r \mathcal{P}_{\mathrm{var}} P var , including every case near the boundary. By (18.43), the flat joint Hessian in the free coordinates then satisfies
K Ω = A ^ Ω T d v a r T d v a r A ^ Ω ≥ λ M , J I . K_{\Omega}=\widehat{A}_{\Omega}^{\mathsf{T}}d_{\mathrm{var}}^{\mathsf{T}}d_{\mathrm{var}}\widehat{A}_{\Omega}\geq\lambda_{M,J}I. K Ω = A Ω T d var T d var A Ω ≥ λ M , J I .
It has uniformly bounded absolute rows and columns and the range (18.44), independently of the region and of the deep frozen group variables. By (18.39), this is also the case d 0 = 0 d_0=0 d 0 = 0 , κ ∗ = 1 \kappa_*=1 κ ∗ = 1 of Theorem 18.12. Every principal free matrix and every real Schur complement of an initial segment keeps a fixed-J J J gap, by (18.47). No Hessian of a field outside the region is inverted.
All Wilson and constraint derivatives at this step are derivatives of finite local path products. Their absolute row sums are bounded on a fixed smaller identity chart, independently of the deep frozen group variables. We obtain the weighted inverses without a volume estimate. First let K K K be the flat joint matrix K Ω K_{\Omega} K Ω , or one of its principal free submatrices, before the minimization over a lower scale. By Lemma 18.14(3), λ M , J I ≤ K ≤ Q M , J I \lambda_{M,J}I \le K \le Q_{M,J}I λ M , J I ≤ K ≤ Q M , J I , and K K K has range D M , J D_{M,J} D M , J . We set a = 2 Q M , J a=2Q_{M,J} a = 2 Q M , J and T = I − K / a T=I-K/a T = I − K / a . Then T T T has norm at most q = 1 − λ M , J / a < 1 q=1-\lambda_{M,J}/a<1 q = 1 − λ M , J / a < 1 , and the entries of T k T^k T k vanish beyond maximum distance k D M , J kD_{M,J} k D M , J , hence beyond graph distance 3 k D M , J 3kD_{M,J} 3 k D M , J . By the bound on the density of labels in the proof of Lemma 18.14, each row of T k T^k T k has at most C J ( k + 1 ) 3 C_J(k+1)^3 C J ( k + 1 ) 3 nonzero entries, each of absolute value at most q k q^k q k . For b > 0 b>0 b > 0 with q e 3 b D M , J < 1 qe^{3bD_{M,J}}<1 q e 3 b D M , J < 1 , the series (18.48) for K − 1 K^{-1} K − 1 and K − 1 / 2 K^{-1/2} K − 1/2 have uniformly summable rows and columns, weighted in the graph distance; the second is the scalar binomial series applied to T T T , whose coefficients are at most one. Small complex perturbations are inverted by the Neumann series about these matrices. For the conditional matrices obtained after minimization over lower scales, we use (18.47) and the bounds for Schur complements in the proof of Lemma 18.14(3), which rest on the argument of Combes and Thomas [61]. These matrices and their inverse square roots have uniform exponential row and column bounds, although the Schur complements need not have finite range. The factors M i M^i M i are included in the corresponding matrices.
The equation for the conditional minimizers is a contraction on a fixed smaller box. We first prove this at the small representative used in Steps 1–3. There, its force at zero fluctuation is C J ( r + β r e c ) C_J(r+\beta_{\mathrm{rec}}) C J ( r + β rec ) , and by the third-derivative bound in the proof of Lemma 18.14(4) its derivative changes by at most C J C_J C J times the radius of the box. At this representative, the coarse argument of (18.45) is Log c \operatorname{Log} c Log c , and ∥ Log c ∥ ∞ \lVert\operatorname{Log} c\rVert_{\infty} ∥ Log c ∥ ∞ and ∥ ξ ∥ ∞ \lVert\xi\rVert_{\infty} ∥ ξ ∥ ∞ are at most C J ( r + β r e c ) C_J(r+\beta_{\mathrm{rec}}) C J ( r + β rec ) . In the choice of ϵ e n , J \epsilon_{\mathrm{en},J} ϵ en , J , we require in addition C J ( r + β r e c ) ≤ ϵ M , J C_J(r+\beta_{\mathrm{rec}})\le\epsilon_{M,J} C J ( r + β rec ) ≤ ϵ M , J , and we require the radius of the box to be at most ϵ M , J \epsilon_{M,J} ϵ M , J , so that (18.45) applies on the whole box, including the coarse field. We reduce the radius of the box so that the inverse times the change of the derivative is at most 1 / 2 1/2 1/2 ; the solution then has size C J ( r + β r e c ) C_J(r+\beta_{\mathrm{rec}}) C J ( r + β rec ) . The finitely many conditional products and their image margins are those required in Theorems 19.9 and 20.14, and their specialization at this step has no inherited effective interaction. Applying the inverse formula for the normalized fluctuations, we obtain the bound on the scaled fluctuations in part (4) at the small representative, without the perturbative term.
For an output representative transported from above, we use the output gauge change of Step 4, with the frozen values transformed consistently. The Wilson terms are traces of plaquette variables, so the exact split in part (4) gives for F b a r F_{\mathrm{bar}} F bar the same covariance identity as (125) . Its conditional minimizers and Hessians therefore transform as in (126) , and their square roots transform by the same orthogonal conjugations. Hence the minimizers at the transported representative are the images of those at the small representative, and each normalized fluctuation w i , e w_{i,e} w i , e is rotated by an orthogonal map, so the bound on the scaled fluctuations carries over. No bound of order r + β r e c r+\beta_{\mathrm{rec}} r + β rec on the transported coarse links is needed.
All constants arose from local path products, fixed construction products, and their row-sum bounds.
Pointwise gauge-invariant densities
The densities of the stage measures will be evaluated at representatives satisfying the reconstruction identities, which form a set of Haar measure zero. We therefore construct an explicit Borel density and prove that it is gauge invariant at every point. The coverage hypothesis of Lemma 18.7 is supplied by Theorem 21.9(1), which uses no density bound.
Theorem 21.10. Fix a finite lattice, a preceding depth that is a multiple of J J J , an initial measure (a stopped measure or the Wilson measure), and a large-field decomposition X \mathcal{X} X of Theorem 21.5 with a range containing at least one integrated full RG step. At stage t t t , let X t X_t X t be the product of the link variables of the current block region and of all independent frozen links and based transports, each on the lattice where it was frozen. Let H t H_t H t be the product of the Haar probability measures on these factors, and let η t X \eta_t^{\mathcal{X}} η t X be the stage measure, a subprobability measure. Let Γ t \Gamma_t Γ t be the group of consistent gauge assignments on the finite multiscale transport graph formed by these endpoints: repeated physical vertices are identified, and each recorded object transforms at the endpoints of its path. In particular, Γ t \Gamma_t Γ t contains arbitrary values at the output roots and their lifts, constant on the blocks of the region, together with the induced changes of earlier frozen coordinates used in Theorem 21.5 . For a frozen link on the lattice of level ℓ \ell ℓ and for a based transport, the action is
b x y ⟼ k x [ ℓ ] b x y ( k y [ ℓ ] ) − 1 , b ^ e ⟼ k x e [ ℓ ] b ^ e ( k v e [ ℓ ] ) − 1 , b ^ e = U ( σ e ) b e . b_{xy} \longmapsto k_x^{[\ell]}b_{xy}\bigl(k_y^{[\ell]}\bigr)^{-1}, \qquad\widehat{b}_e \longmapsto k_{x_e}^{[\ell]}\widehat{b}_e\bigl(k_{v_e}^{[\ell]}\bigr)^{-1}, \qquad\widehat{b}_e=U(\sigma_e)b_e. b x y ⟼ k x [ ℓ ] b x y ( k y [ ℓ ] ) − 1 , b e ⟼ k x e [ ℓ ] b e ( k v e [ ℓ ] ) − 1 , b e = U ( σ e ) b e . For each fixed assignment, this is a left and right multiplication in the independent group variables, and it preserves H t H_t H t .
There exists an explicitly defined nonnegative Borel density p t p_t p t such that
η t X = p t H t , p t ( x k ) = p t ( x ) for every x ∈ X t and k ∈ Γ t . \eta_t^{\mathcal{X}}=p_tH_t,\qquad p_t(x^k)=p_t(x)\quad\text{for every }x\in X_t\text{ and }k\in\Gamma_t. η t X = p t H t , p t ( x k ) = p t ( x ) for every x ∈ X t and k ∈ Γ t . On the gauge-invariant chart domain containing the measure, p t p_t p t is given by the parameter integrals of the Haar disintegration formulas, with every indicator of the decomposition and every protected external factor retained, patched by their exact equalities on overlaps of fixed gauge translates. On the invariant complement of this domain, which has zero mass, its value is zero. In particular, the values of p t p_t p t at representatives satisfying the reconstruction identities are given by these formulas pointwise.
More explicitly, if Z c , b r e c Z_{c,b_{\mathrm{rec}}} Z c , b rec is the domain of the free coordinates of the fiber over an output datum ( c , b r e c ) (c,b_{\mathrm{rec}}) ( c , b rec ) satisfying the coordinate conditions of Theorem 21.9, then the outgoing density is
ρ t o u t ( c , b r e c ) = ∫ Z c , b r e c χ t ( c , z , b r e c ) ρ t ( U c , b r e c ( z ) , b r e c ) J C O ( c , z ; b r e c ) d z . \rho_t^{\mathrm{out}}(c,b_{\mathrm{rec}})=\int_{Z_{c,b_{\mathrm{rec}}}}\chi_t(c,z,b_{\mathrm{rec}})\rho_t(U_{c,b_{\mathrm{rec}}}(z),b_{\mathrm{rec}})J_{\mathrm{CO}}(c,z;b_{\mathrm{rec}})\,dz. ρ t out ( c , b rec ) = ∫ Z c , b rec χ t ( c , z , b rec ) ρ t ( U c , b rec ( z ) , b rec ) J CO ( c , z ; b rec ) d z .
where χ t \chi_t χ t is the product of the indicators of the decomposition, the incoming support, and the outgoing pre-freeze support, and the normalized integrals over the tree variables are kept until the covariance under based gauges has been applied. The factor at scale i i i of the Haar disintegration density J C O J_{\mathrm{CO}} J CO is exactly
c H a a r f i ∏ e o u t s i d e t h e t r e e s j H ( Γ i , e ) ∏ a det D X e a C a ( Γ i ) , j H ( X ) = sinc 2 ( ∣ X ∣ / 2 ) , c_{\mathrm{Haar}}^{f_i}\frac{\prod_{e\ \mathrm{outside\ the\ trees}}j_H(\Gamma_{i,e})}{\prod_a\det D X_{e_a}C_a(\Gamma_i)},\qquad j_H(X)=\operatorname{sinc}^{2}(|X|/2), c Haar f i ∏ a det D X e a C a ( Γ i ) ∏ e outside the trees j H ( Γ i , e ) , j H ( X ) = sinc 2 ( ∣ X ∣/2 ) ,
where the product in the denominator runs over the internal constraints, f i f_i f i is the number of free group variables at scale i i i , and the numerator includes the Haar factors of the eliminated edges. The reference measure already contains the Haar measure of the output field; no further coarse Haar density, Gram determinant, or orbit-volume factor occurs. For the stage densities and the tests of the decomposition, the covariance under based tree gauges makes each tree Haar integral equal to one; this elimination is not valid for arbitrary noncovariant Borel densities. Basing at a fixed field on the block region is the triangular Haar translation b e ↦ U ( σ e ) b e b_e\mapsto U(\sigma_e)b_e b e ↦ U ( σ e ) b e , with Jacobian one.
These statements concern the stage measures and the block averaging map with its Borel extension; they assert no independent gauge covariance or overlap identity for the auxiliary regional functions. The torus classical restriction used in the analytic comparisons does use Theorem 11.6 on the common charts, as in (125) . Deep frozen group variables, including exceptional values of the Borel extension, are allowed. All identities hold in each finite dimension and introduce no factor depending on the size, the number of steps, the shape, or the depth. The analytic chart bounds used in the construction have the dependence stated at the beginning of this section.
For a range with no complete RG step, use the original pointwise invariant initial density, with the retained decomposition and support indicators and zero extension off its support. This gives a Borel density for either initial measure directly; no mass-carrying small-chart domain is asserted. The theorem above includes stage zero of every positive-length decomposition, as well as its empty active regions and empty fluctuation products.
Proof. We construct the density and prove its pointwise invariance together. This avoids evaluating an unspecified Radon–Nikodym derivative on a set of representatives.
Step 1: the endpoint group. We identify repeated physical vertices in the finite list of endpoints. An endpoint assignment is an element of the resulting finite product of copies of S U ( 2 ) \mathrm{SU}(2) SU ( 2 ) , and any additional consistency identifications give the corresponding closed subgroup. Its action on an independent group variable is a left and a right multiplication by fixed matrices. Hence H t H_t H t is preserved by each fixed assignment, with no assumption of probabilistic independence of the frozen coordinates under η t x \eta_t^x η t x .
Step 2: the initial density. At the initial scale we use the pointwise density of the Wilson measure, or the gauge-invariant density of the stopped measure, multiplied by the indicators of the decomposition and of K 0 K_0 K 0 , and set it to zero off K 0 K_0 K 0 . At every later stage the support indicator is retained in the defining parameter integral and the version is zero off K t K_t K t . The initial density is invariant: the Wilson plaquette variables transform by conjugation, and the density of the stopped measure is the gauge-invariant density of Theorem 14.6. The indicator of the decomposition is invariant as well, since the decomposition is determined by traces of plaquette variables and by deterministic dependence lists. The block averaging map is equivariant also on the exceptional set of its Borel extension: conjugation fixes − I -I − I , and the replacement reference transport transforms at the same two endpoints. These are pointwise statements. The first freezing is a splitting into independent variables followed by the triangular basing, treated in Step 5.
Step 3: the outgoing density on a chart. Assume that the incoming density has been constructed. At an output datum satisfying the coordinate conditions, we use Lemma 18.7 with all indicators; its coverage hypothesis is part (1) of Theorem 21.9. We keep the normalized tree variables f f f at first. If J i J_i J i denotes the factor at scale i i i , we define the local outgoing value to be
∫ d H ( f ) ∫ 1 Z c , b r e c ( z , f ) χ t ( c , z , f , b r e c ) p t ( U c , b r e c ( z , f ) , b r e c ) ∏ i = 0 s − 1 J i d z . (129) \int dH(f)\int\mathbf{1}_{Z_{c,b_{\mathrm{rec}}}(z,f)}\chi_t(c,z,f,b_{\mathrm{rec}})p_t(U_{c,b_{\mathrm{rec}}}(z,f),b_{\mathrm{rec}})\prod_{i=0}^{s-1}J_i\,dz. \tag*{(129)} ∫ d H ( f ) ∫ 1 Z c , b rec ( z , f ) χ t ( c , z , f , b rec ) p t ( U c , b rec ( z , f ) , b rec ) i = 0 ∏ s − 1 J i d z . (129)
An indicator already contained in p t p_t p t is not counted again; it may equally be written again, since indicators are idempotent. This convention keeps all tests of the decomposition, including those whose information has already been integrated. By the disintegration identity, testing (21.27) against any nonnegative Borel function of the output gives the pushforward measure on this chart.
Step 4: pointwise covariance. We prove that these integrals are equal under the full group of consistent endpoint assignments. A fixed output assignment has a lift, constant on blocks, through the centered partitions of the block region; at earlier marks we use its prescribed values on the lattices of the marks. There is no conflict, since a mark in the preimage of a block of the region would have excluded that block, and repeated physical values are identified by consistency. Additional changes of the interior frames can be absorbed in the tree variables. The block averaging map and the map recording the frozen coordinates are equivariant, so the lift is a bijection of the fiber over ( c , b r e c ) (c,b_{\mathrm{rec}}) ( c , b rec ) onto the fiber over ( c k , b r e c k ) (c^k,b_{\mathrm{rec}}^k) ( c k , b rec k ) . It preserves every test, including an equality in a threshold and every condition at earlier scales.
On the hierarchical tree section, the part of the lift that is constant on blocks acts as in (124) . The adjoint action is orthogonal because it preserves − 2 tr ( X Y ) -2\operatorname{tr}(XY) − 2 tr ( X Y ) . Its determinant is 1 1 1 , since it is a continuous function on the connected group S U ( 2 ) ≃ S 3 \mathrm{SU}(2)\simeq S^3 SU ( 2 ) ≃ S 3 with values in { − 1 , 1 } \{-1,1\} { − 1 , 1 } and equals 1 1 1 at I I I . Hence the change of free coordinates preserves d z dz d z , and uniqueness of the eliminated-edge inverse transports every eliminated coordinate as well. The factors j H a a r f i j_{\mathrm{Haar}}^{f_i} j Haar f i are radial and unchanged. The derivative matrix of the eliminated edges is multiplied on its row and column frames by these adjoints of determinant one, so its determinant is unchanged, including its positive sign on the real branch. The constants c H a a r f i c_{\mathrm{Haar}}^{f_i} c Haar f i are the same. If the lift changes interior tree frames, the tree variables change by left and right Haar translations, before any tree integral is removed. Using the inductive pointwise identity for p t p_t p t , we obtain
p t o u t ( c k , b r e c k ) = ∫ Z c k , b r e c k χ t ( c k , z ′ , b r e c k ) p t ( U c k , b r e c k ( z ′ ) , b r e c k ) J c k , b r e c k ( z ′ ) d z ′ = ∫ Z c , b r e c χ t ( c , z , b r e c ) p t ( U c , b r e c ( z ) , b r e c ) J c , b r e c ( z ) d z = p t o u t ( c , b r e c ) . (130) p_t^{\mathrm{out}}(c^k,b_{\mathrm{rec}}^k)=\int_{Z_{c^k,b_{\mathrm{rec}}^k}}\chi_t(c^k,z',b_{\mathrm{rec}}^k)p_t(U_{c^k,b_{\mathrm{rec}}^k}(z'),b_{\mathrm{rec}}^k)J_{c^k,b_{\mathrm{rec}}^k}(z')\,dz'
=\int_{Z_{c,b_{\mathrm{rec}}}}\chi_t(c,z,b_{\mathrm{rec}})p_t(U_{c,b_{\mathrm{rec}}}(z),b_{\mathrm{rec}})J_{c,b_{\mathrm{rec}}}(z)\,dz=p_t^{\mathrm{out}}(c,b_{\mathrm{rec}}). \tag*{(130)} p t out ( c k , b rec k ) = ∫ Z c k , b rec k χ t ( c k , z ′ , b rec k ) p t ( U c k , b rec k ( z ′ ) , b rec k ) J c k , b rec k ( z ′ ) d z ′ = ∫ Z c , b rec χ t ( c , z , b rec ) p t ( U c , b rec ( z ) , b rec ) J c , b rec ( z ) d z = p t out ( c , b rec ) . (130)
The normalized tree integrals are understood in both lines when they have not yet been removed. This is an identity of explicit parameter integrals at each output datum; no uniqueness almost everywhere is used.
Step 5: freezing. Freezing has the required covariance too. We orient a link across the boundary of the block region from its endpoint u e u_e u e in the region to its frozen endpoint v e v_e v e , and let α e \alpha_e α e be the root of its hierarchical transport. Cancellation at the interior vertices gives
U v k ( σ e ) a e k = k α e U v ( σ e ) k u e − 1 k u e a e k v e − 1 = k α e b ^ e k v e − 1 . (131) U_{v_k}(\sigma_e)a_e^k=k_{\alpha_e}U_v(\sigma_e)k_{u_e}^{-1}k_{u_e}a_ek_{v_e}^{-1}
=k_{\alpha_e}\widehat b_ek_{v_e}^{-1}. \tag*{(131)} U v k ( σ e ) a e k = k α e U v ( σ e ) k u e − 1 k u e a e k v e − 1 = k α e b e k v e − 1 . (131)
At fixed v v v , the map from the independent variables a e a_e a e to their based transports is a product of left Haar translations. Together with the unchanged v v v , this is a triangular measure-preserving bijection; its inverse multiplies by U v ( σ e ) − 1 U_v(\sigma_e)^{-1} U v ( σ e ) − 1 . Recording any other link as a frozen link is Fubini’s splitting of product Haar measure. The endpoint protection rule makes these endpoint changes consistent at every later stage. Equations (130) and (131) prove the inductive covariance for arbitrary values at the output roots and for all induced changes of earlier frozen coordinates.
Step 6: Borel patching. Before inserting the tests, we take the open chart domains of small block-region variables and small near values, with the positive margins of their inverses; the deep frozen group variables need no logarithm. Their gauge saturation O t O_t O t is open and invariant, and the proofs of Theorems 21.5 and 21.9 put all the mass in O t O_t O t . A finite product of spheres is second countable and has a countable dense subset, for instance obtained by approximating quaternion coordinates by rational ones and normalizing; the finite endpoint subgroup has the same property after the consistency identifications. If a point is moved into an open chart by some fixed gauge assignment, the set of assignments doing so is open and meets this countable dense set. Hence countably many fixed gauge translates of the small charts cover O t O_t O t , and their coordinate charts can be chosen countably as well.
On one such chart, the integrand of (129) , extended by zero outside its domain in a fixed box of free and tree coordinates, is a nonnegative Borel function. Indeed, the Haar disintegration denominators and the reconstruction maps are continuous on their branches, and the tests, including values of the Borel extension at the deep frozen group variables, are Borel. By Tonelli’s theorem, its integral with respect to the finite product of Lebesgue and Haar measures on the box is a Borel function of ( c , b r e c ) (c,b_{\mathrm{rec}}) ( c , b rec ) .
On an overlap, the fixed gauge assignments relating two charts give the change of variables in (130) . Since both formulas include the same domain, they agree at every point of the overlap. We use the first chart in the countable cover containing a point to define p t p_t p t there, retaining the support mask 1 K t \mathbf{1}_{K_t} 1 K t , and set p t = 0 p_t=0 p t = 0 on X t ∖ ( O t ∩ K t ) X_t \setminus(O_t \cap K_t) X t ∖ ( O t ∩ K t ) . The raw outgoing version likewise retains the indicator of the support before freezing. This function is Borel, it is a density by the disintegration identities, and it does not depend on the chart used. To check invariance under an arbitrary k ∈ Γ t k \in\Gamma_t k ∈ Γ t , we take charts containing x x x and x k x^k x k and perform the same change of variables; this proves equality of the two values also when k k k is not in the countable dense set. Both O t O_t O t and K t K_t K t are invariant: invariance of the actual stage measures follows inductively from the equivariant restrictions and pushforwards, so their closed supports are invariant. Their complements have zero mass, so the stated zero extension is invariant there as well. Countability was used only to define a Borel function, not to sum densities or to introduce a multiplicity of charts.
Step 7: removal of the tree integrals. Finally, at fixed target and frozen coordinates, we apply a based tree gauge fixing. The pointwise covariance just proved makes the density and its tests independent of the tree variables in (129) , so each Haar probability integral over tree variables equals one. Removing them, we obtain the formula in the statement. Its factors at the scales i i i are exactly those of Lemma 18.7, in particular the numerator with the eliminated edges and the single inverse determinant of the eliminated-edge derivative. The Haar measure of the output is already the reference measure at the next scale, so this change of variables has no room for a further coarse Haar density, a Gram determinant, or an orbit-volume factor. For an arbitrary noncovariant Borel input density, the tree average would remain in (129) , and its removal uses the covariance of the stage density. The argument uses the Haar disintegration before the change to normalized fluctuation variables. The deep frozen group variables, including exceptional values of the Borel extension, remain compact Borel data throughout.
Freezing and substitution of frozen coordinates
When integration stops on part of the region, the corresponding variables become frozen coordinates. The next lemma shows that, at the representatives of the next stage satisfying the reconstruction identities, the inverse of the freezing map inserts the frozen values directly into the old variables. This identifies the density after freezing with the density before freezing at the same configuration, also on sets of Haar measure zero, and gives the exact substitution rules for the analytic functions.
Lemma 21.11. At every boundary of a large-field decomposition where coordinates are frozen, use the hierarchical tree transports σ e \sigma_e σ e of Lemma 21.1, with the tree conditions at coarser scales fixed first. Split the old output into the remaining link variables v v v , the variables a e a_e a e on the links across the boundary of the block region, and the other links to be frozen. The freezing map A t \mathcal{A}_t A t replaces each such variable by its based transport
b ^ e = U v ( σ e ) a e \widehat{b}_e = U_v(\sigma_e)a_e b e = U v ( σ e ) a e and records the other links as frozen links on that lattice. At fixed v v v this map preserves the product Haar measure and has Jacobian one.
Let b ♯ b^\sharp b ♯ denote the new frozen group values and ξ ♯ = Log b ♯ \xi^\sharp= \operatorname{Log} b^\sharp ξ ♯ = Log b ♯ their logarithms. For every reconstructed configuration of the next stage in the family of representatives satisfying the reconstruction identities,
v = C t + 1 ( c ′ , ζ ′ ; b r e c , b ♯ ) v = C_{t+1}(c',\zeta';b_{\mathrm{rec}},b^\sharp) v = C t + 1 ( c ′ , ζ ′ ; b rec , b ♯ ) the hierarchical tree transport has value I I I : each of its tree factors is I I I , and the gauge transformation of Lemma 19.5 is I I I at its two endpoints. All protected earlier endpoints are fixed as well. Hence
a e = U v ( σ e ) − 1 b ^ e = b ^ e , A t − 1 ( v , b r e c , b ♯ ) = ( v ⊕ b ♯ , b r e c ) , a_e = U_v(\sigma_e)^{-1}\widehat{b}_e = \widehat{b}_e,\qquad
\mathcal{A}_t^{-1}(v,b_{\mathrm{rec}},b^\sharp) = (v\mathbin{\oplus}b^\sharp,b_{\mathrm{rec}}), a e = U v ( σ e ) − 1 b e = b e , A t − 1 ( v , b rec , b ♯ ) = ( v ⊕ b ♯ , b rec ) , and the density versions of Theorem 21.10 satisfy the pointwise identity
p t + 1 ( v , b r e c , b ♯ ) = p t o u t ( v ⊕ b ♯ , b r e c ) . p_{t+1}(v,b_{\mathrm{rec}},b^\sharp)
=
p_t^{\mathrm{out}}(v\mathbin{\oplus}b^\sharp,b_{\mathrm{rec}}). p t + 1 ( v , b rec , b ♯ ) = p t out ( v ⊕ b ♯ , b rec ) . At this configuration, all omitted old variables are exactly their new frozen values, based transports or links, with the same weights and indicators. In analytic functions, ξ ♯ \xi^\sharp ξ ♯ is inserted in the corresponding analytic arguments. In particular, the freezing of the effective interaction subtracts only the value R o u t ( 0 ⊕ ξ ♯ ) R^{\mathrm{out}}(0\mathbin{\oplus}\xi^\sharp) R out ( 0 ⊕ ξ ♯ ) and transfers it with a minus sign to Λ \Lambda Λ , and the freezing of the classical action is E n e w ( v ) = E o u t ( v ⊕ ξ ♯ ) E_{\mathrm{new}}(v)=E^{\mathrm{out}}(v\mathbin{\oplus}\xi^\sharp) E new ( v ) = E out ( v ⊕ ξ ♯ ) . These are exact finite identities, uniform in the number of steps, the topology, the depth, and the coupling. They compare the same configuration and require no overlap identity between auxiliary functions at different gauges.
Proof. We fix the deterministic geometry of the freezing. The remaining variables and the independent variables on the links across the boundary are different coordinates of the old output product. At fixed v v v , left multiplication by each fixed matrix U v ( σ e ) U_v(\sigma_e) U v ( σ e ) preserves the Haar probability measure of such a variable, and the inverse multiplication is defined for every value. Integrating first in these variables, we obtain the triangular statement with Jacobian one for every nonnegative Borel integrand; earlier frozen coordinates and the other frozen links do not change.
We verify the value of the transport at a reconstructed configuration. On the based section of one scale, the tree equations set each ordered tree factor to I I I , and the fixing gauges of the coarser scales, lifted as constants on the finer blocks, conjugate these factors and leave them equal to I I I . By induction over the scales, with the coarser conditions fixed first, every factor of the prescribed hierarchical transport is I I I on the hierarchical tree section. This is the finite algorithm of Lemma 21.1 , and it uses no path through an eliminated edge. Hence the product has value I I I for every target and every free coordinate in the domain, not only at zero fluctuation.
The reconstructed configuration is obtained from that section by the gauge transformation of Lemma 19.5 . This gauge transformation multiplies a transport from α e \alpha_e α e to u e u_e u e by its values at exactly these two vertices, and both values are I I I by that lemma, with the full set of protected vertices supplied by Theorem 21.5 . The transport remains I I I . On the lattices of earlier frozen coordinates, the endpoint protection rule fixes all their endpoints. This proves, at this reconstructed configuration,
U v ( σ e ) − 1 b ^ e = b ^ e . U_v(\sigma_e)^{-1}\widehat b_e=\widehat b_e. U v ( σ e ) − 1 b e = b e .
Inverting the freezing map therefore inserts the new based transports directly into the old variables on the links across the boundary, and every other omitted variable is its new frozen link value. At this configuration, the inverse map is ( v , b r e c , b ♯ ) ↦ ( v ⊕ b ♯ , b r e c ) (v,b_{\mathrm{rec}},b^\sharp)\mapsto(v\oplus b^\sharp,b_{\mathrm{rec}}) ( v , b rec , b ♯ ) ↦ ( v ⊕ b ♯ , b rec ) .
The general formula for the density after freezing is the composition of the old outgoing density with the inverse triangular map, since its Haar Jacobian is one. We use the versions of Theorem 21.10 on both sides; that construction uses the same formulas and their pointwise overlap identities. Composing with the inverse, we obtain the pointwise identity, even when the reconstructed configuration lies in a set of representatives of Haar measure zero. All weights and indicators of the decomposition are values of the same functions at the same old configuration.
For a variable that was analytic before this freezing, its new group value lies in the small chart of frozen coordinates, and we insert ξ ♯ = Log b ♯ \xi^\sharp=\operatorname{Log}b^\sharp ξ ♯ = Log b ♯ there. The deep frozen group variables at the first step under the Wilson measure are fixed parameters instead and are not part of this logarithmic substitution. For every old analytic function F F F we have the exact identity
F ( v ⊕ ξ ♯ ) = [ F ( v ⊕ ξ ♯ ) − F ( 0 ⊕ ξ ♯ ) ] + F ( 0 ⊕ ξ ♯ ) . (132) F(v\mathbin{\oplus}\xi^\sharp)=\left[F(v\mathbin{\oplus}\xi^\sharp)-F(0\mathbin{\oplus}\xi^\sharp)\right]+F(0\mathbin{\oplus}\xi^\sharp). \tag*{(132)} F ( v ⊕ ξ ♯ ) = [ F ( v ⊕ ξ ♯ ) − F ( 0 ⊕ ξ ♯ ) ] + F ( 0 ⊕ ξ ♯ ) . (132)
Applying it to R o u t R^{\mathrm{out}} R out , the factor e Λ o u t − R o u t e^{\Lambda^{\mathrm{out}}-R^{\mathrm{out}}} e Λ out − R out becomes
e Λ o u t − R o u t ( 0 ⊕ ξ ♯ ) e − [ R o u t ( v ⊕ ξ ♯ ) − R o u t ( 0 ⊕ ξ ♯ ) ] , e^{\Lambda^{\mathrm{out}}-R^{\mathrm{out}}(0\oplus\xi^{\sharp})}e^{-[R^{\mathrm{out}}(v\oplus\xi^{\sharp})-R^{\mathrm{out}}(0\oplus\xi^{\sharp})]}, e Λ out − R out ( 0 ⊕ ξ ♯ ) e − [ R out ( v ⊕ ξ ♯ ) − R out ( 0 ⊕ ξ ♯ )] ,
which proves the minus sign of the transfer to Λ \Lambda Λ . For the classical function we keep the whole value E o u t ( v ⊕ ξ ♯ ) E^{\mathrm{out}}(v\oplus\xi^{\sharp}) E out ( v ⊕ ξ ♯ ) , including its constant. Centering the effective interaction changes no first derivative: the new first derivative in the remaining variables is the old one evaluated at ( 0 , ξ ♯ ) (0,\xi^{\sharp}) ( 0 , ξ ♯ ) , and it is kept.
Conditional integration and normalization on large-field decompositions
This section proves the density bounds used in Section 22. The induction has three stages. First, the analytic effective actions, effective interactions, normalization constants, and reconstruction maps are defined from fine to coarse scales, by the recursion of Theorem 20.14 , for every value of the frozen coordinates and without reference to any measure. Second, representatives satisfying the reconstruction identities are chosen from coarse to fine scales, using these maps. Third, the stage densities are bounded from fine to coarse scales at exactly these representatives. The construction is not circular, since each stage uses only objects defined in the preceding ones. One step of the third stage is the following comparison, which needs an incoming density bound; the bound is supplied by the induction itself.
Lemma 21.12. Assume the hypotheses of Theorem 21.9, together with the conditions (21.18) and (21.19), which give the hypotheses of Lemma 20.5 on the whole enclosing product domain. Fix a stage t t t and an output datum ( c , b r e c ) (c,b_{\mathrm{rec}}) ( c , b rec ) of the family of representatives satisfying the reconstruction identities in Theorem 21.9(2), with real frozen coordinates, and define the fluctuation cutoff domain
W t = ∏ i , e { ∣ w i , e ∣ ≤ p ( t t ) } , t t = max i g t , i , p ( t ) = A W log ( 1 / t ) . (133) \mathcal{W}_t=\prod_{i,e}\{|w_{i,e}|\le p(t_t)\},\qquad t_t=\max_i g_{t,i},\qquad p(t)=A_W\log(1/t). \tag*{(133)} W t = i , e ∏ { ∣ w i , e ∣ ≤ p ( t t )} , t t = i max g t , i , p ( t ) = A W log ( 1/ t ) . (133) Let V = V c , b r e c V=V_{c,b_{\mathrm{rec}}} V = V c , b rec be the conditional fluctuation potential of Theorem 20.14 at this datum, and let S c , b r e c ⊂ K t S_{c,b_{\mathrm{rec}}}\subset K_t S c , b rec ⊂ K t be the set of normalized fluctuation coordinates of the supported fiber points over ( c , b r e c ) (c,b_{\mathrm{rec}}) ( c , b rec ) . For an invariant Borel function 0 ≤ χ ≤ 1 0\le\chi\le1 0 ≤ χ ≤ 1 , let G t , χ ( c , b r e c ) G_{t,\chi}(c,b_{\mathrm{rec}}) G t , χ ( c , b rec ) be the outgoing density of Theorem 21.10 at ( c , b r e c ) (c,b_{\mathrm{rec}}) ( c , b rec ) with χ \chi χ inserted in the integrand. Suppose that there exist a finite prefactor P t ( c , b r e c ) ≥ 0 P_t(c,b_{\mathrm{rec}})\ge0 P t ( c , b rec ) ≥ 0 and a Borel function χ t \chi_t χ t on S c , b r e c S_{c,b_{\mathrm{rec}}} S c , b rec with values in [ 0 , 1 ] [0,1] [ 0 , 1 ] such that
G t , χ ( c , b r e c ) ≤ P t ( c , b r e c ) ∫ S c , b r e c χ t ( w ) e − V ( w ) d w . (134) G_{t,\chi}(c,b_{\mathrm{rec}})\le P_t(c,b_{\mathrm{rec}})\int_{S_{c,b_{\mathrm{rec}}}}\chi_t(w)e^{-V(w)}\,dw. \tag*{(134)} G t , χ ( c , b rec ) ≤ P t ( c , b rec ) ∫ S c , b rec χ t ( w ) e − V ( w ) d w . (134) Then S c , b r e c S_{c,b_{\mathrm{rec}}} S c , b rec is a Borel set, 0 < ∫ K t e − V d w < ∞ 0<\int_{K_t}e^{-V}\,dw<\infty 0 < ∫ K t e − V d w < ∞ , and
G t , χ ( c , b r e c ) ≤ ( 1 − δ t ) − 1 F ~ t ( c , b r e c ) , F ~ t = P t ( c , b r e c ) ∫ W t e − V ( w ) d w , G_{t,\chi}(c,b_{\mathrm{rec}})\le(1-\delta_t)^{-1}\widetilde F_t(c,b_{\mathrm{rec}}),\qquad\widetilde F_t=P_t(c,b_{\mathrm{rec}})\int_{\mathcal{W}_t}e^{-V(w)}\,dw, G t , χ ( c , b rec ) ≤ ( 1 − δ t ) − 1 F t ( c , b rec ) , F t = P t ( c , b rec ) ∫ W t e − V ( w ) d w ,
with 0 ≤ δ t < 1 0 \le\delta_t < 1 0 ≤ δ t < 1 , δ t ≤ 6 b t e − p ( t t ) 2 / 48 \delta_t \le6b_t e^{-p(t_t)^2/48} δ t ≤ 6 b t e − p ( t t ) 2 /48 , and b t ≤ C J m 3 b_t \le C_Jm^3 b t ≤ C J m 3 , where b t b_t b t is the number of three-dimensional free blocks of fluctuation variables of the step. Factors depending only on the output datum multiply both sides. The indicators are removed only in this upper bound. The constants have the dependence stated at the beginning of this section, and they are uniform in the deep frozen group variables.
At the first RG step, for either initial measure, (21.32) holds with equality. In that case P t P_t P t consists of the inverse partition function of the initial measure, the external weights, the Haar disintegration constants, the powers of the couplings, the output classical factor, and the zero-fluctuation factors, and χ t \chi_t χ t is the product of χ \chi χ and the indicators composed with the inverse coordinates. At a later stage, (21.32) is obtained by inserting an incoming bound on the density at the representatives satisfying the reconstruction identities, with the incoming multiplicative loss included in the same prefactor P t P_t P t ; this incoming bound is proved in Theorem 21.13 .
Proof. At an output datum, the set S c , b r e c S_{c,b_{\mathrm{rec}}} S c , b rec is defined with all indicators still inserted, and before any upper comparison the Haar disintegration is the exact formula of Lemma 18.7. The radii of W t \mathcal{W}_t W t belong to the logarithmic family of Theorem 20.14 and satisfy t t p ( t t ) < R e n t_t p(t_t) < R_{\mathrm{en}} t t p ( t t ) < R en by (121) ; for a complete RG step t t = t max t_t=t_{\max} t t = t m a x . The product K t \mathcal{K}_t K t is used only in this positive comparison, never as the product for the expansion defining the analytic output. At the first step, the Haar disintegration and the change to normalized fluctuation variables give the density integral in the form
G t , χ ( c , b r e c ) = P t ( c , b r e c ) ∫ S c , b r e c χ t ( w ) e − V c , b r e c ( w ) d w , F ~ t = P t ( c , b r e c ) ∫ W t e − V c , b r e c ( w ) d w (135) G_{t,\chi}(c,b_{\mathrm{rec}})=P_t(c,b_{\mathrm{rec}})\int_{S_{c,b_{\mathrm{rec}}}}\chi_t(w)e^{-V_{c,b_{\mathrm{rec}}}(w)}\,dw,\qquad\widetilde{F}_t=P_t(c,b_{\mathrm{rec}})\int_{\mathcal{W}_t}e^{-V_{c,b_{\mathrm{rec}}}(w)}\,dw \tag*{(135)} G t , χ ( c , b rec ) = P t ( c , b rec ) ∫ S c , b rec χ t ( w ) e − V c , b rec ( w ) d w , F t = P t ( c , b rec ) ∫ W t e − V c , b rec ( w ) d w (135)
with the prefactor described in the statement. At a later step the equality in the first formula is replaced by the inequality (21.32).
By Theorem 21.9(2), S c , b r e c ⊂ K t S_{c,b_{\mathrm{rec}}}\subset\mathcal{K}_t S c , b rec ⊂ K t for every such datum. We check the measure-theoretic hypotheses of Lemma 20.5 at this fixed real datum. The finite products K t \mathcal{K}_t K t and W t \mathcal{W}_t W t are compact Borel sets. The map from free logarithmic coordinates to normalized fluctuation coordinates is a homeomorphism onto its open image, with the continuous inverse already constructed, so the image of the Borel set of free coordinates of supported points is Borel. Composing the Borel tests with that inverse gives a Borel function χ t \chi_t χ t , extended by zero outside S c , b r e c S_{c,b_{\mathrm{rec}}} S c , b rec , with 0 ≤ χ t ≤ 1 0\le\chi_t\le1 0 ≤ χ t ≤ 1 . The conditional fluctuation potential is analytic on a neighbourhood of K t \mathcal{K}_t K t and real at its real points, by Theorem 20.14 . Its finite minimum V ‾ \underline{V} V and maximum V ‾ \overline{V} V on this compact product give
0 < e − V ‾ vol ( K t ) ≤ Z K t = ∫ K t e − V d w ≤ e − V ‾ vol ( K t ) < ∞ . (136) 0<e^{-\overline{V}}\operatorname{vol}(\mathcal{K}_t)\le Z_{\mathcal{K}_t}=\int_{\mathcal{K}_t}e^{-V}\,dw\le e^{-\underline{V}}\operatorname{vol}(\mathcal{K}_t)<\infty. \tag*{(136)} 0 < e − V vol ( K t ) ≤ Z K t = ∫ K t e − V d w ≤ e − V vol ( K t ) < ∞. (136)
The volume is finite and positive, since every ball has a finite positive radius. For an empty free product its zero-dimensional measure is one and V = 0 V=0 V = 0 , so Z K t = 1 Z_{\mathcal{K}_t}=1 Z K t = 1 . No uniform bound on V ‾ \underline{V} V or V ‾ \overline{V} V is needed here. The prefactor is finite at each real datum; if it vanishes, both sides vanish, and otherwise it is positive.
The hypotheses of Lemma 20.5 on the whole product give its potential estimate and, for the normalized enclosing-product comparison measure ν K t \nu_{\mathcal{K}_t} ν K t , the bound
δ t : = ν K t ( K t ∖ W t ) ≤ 6 b t e − p ( t t ) 2 / 48 , b t ≤ C J m 3 . \delta_t := \nu_{\mathcal{K}_t}(\mathcal{K}_t \setminus\mathcal{W}_t) \le6b_t e^{-p(t_t)^2/48}, \qquad b_t \le C J m^3. δ t := ν K t ( K t ∖ W t ) ≤ 6 b t e − p ( t t ) 2 /48 , b t ≤ C J m 3 .
Every ball of W t \mathcal{W}_t W t has positive radius, so its volume is positive. Since V V V is finite on K t \mathcal{K}_t K t , we have ∫ W t e − V d w > 0 \int_{\mathcal{W}_t} e^{-V}\,dw > 0 ∫ W t e − V d w > 0 , and hence ν K t ( W t ) > 0 \nu_{\mathcal{K}_t}(\mathcal{W}_t) > 0 ν K t ( W t ) > 0 and 0 ≤ δ t < 1 0 \le\delta_t < 1 0 ≤ δ t < 1 . This strict inequality concerns the actual missing mass; the displayed union-bound majorant need not be below one under the local hypotheses. For an empty free product, δ t = 0 \delta_t = 0 δ t = 0 .
Since 0 ≤ χ t ≤ 1 0 \le\chi_t \le1 0 ≤ χ t ≤ 1 , the right side of (21.32) is at most P t ∫ K t e − V P_t\int_{\mathcal{K}_t} e^{-V} P t ∫ K t e − V , and part (1) of Lemma 20.5 bounds the latter by ( 1 − δ t ) − 1 F ~ t (1-\delta_t)^{-1}\widetilde{F}_t ( 1 − δ t ) − 1 F t . No inclusion of W t \mathcal{W}_t W t in S c , b r e c \mathcal{S}_{c,b_{\mathrm{rec}}} S c , b rec is needed. The same reasoning applies if P t = 0 P_t = 0 P t = 0 or if the free list is empty. All constants arose from local path products, the fixed construction products, and their row-sum bounds.
In the theorem below, E t E_t E t , R t R_t R t , and Λ t \Lambda_t Λ t are the effective classical action, the effective interaction, and the normalization constant at stage t t t ; the last is the field-independent additive term in the log-density. Further, W t \mathcal{W}_t W t is the product of the external weights (the factors of frozen variables that are not integrated), C t C_t C t is the reconstruction map, and ℓ t \ell_t ℓ t is the level of the output of stage t t t . For the normalization comparison, Λ t T \Lambda_t^{\mathrm{T}} Λ t T denotes the normalization constant of the corresponding recursion on the whole torus. Here H t H_t H t denotes the set of scalar Gaussian coordinates of stage t t t that are integrated in the corresponding renormalization group recursion on the whole torus but absent from the regional free list (it is not the Haar measure of Theorem 21.10), and d i ( H t ) d_i(H_t) d i ( H t ) is the number of its elements at scale i i i . The numbers N m i s s i n g , t N_{\mathrm{missing},t} N missing , t , N b d , t N_{\mathrm{bd},t} N bd , t , and N f r e e z e , t N_{\mathrm{freeze},t} N freeze , t count, on the lattices where they were created, the missing labels of field-independent generated terms, the labels of the fixed interface band, and the newly frozen analytic labels, as in Theorem 20.13.
Theorem 21.13. Fix a large-field decomposition of Theorem 21.5, an initial measure ν \nu ν (a stopped measure or the Wilson measure), an initial scale and a stopping scale differing by T ≥ 1 T \ge1 T ≥ 1 full RG steps, and the cutoff domains and norm ball of Theorem 20.14. Every integrated group of averaging steps is a complete RG step, s t = J s_t = J s t = J . A remaining number 0 ≤ u p h < J 0 \le u_{\mathrm{ph}} < J 0 ≤ u ph < J of averaging steps within the given hierarchy is kept through the pullback C ↦ F ( T M u p h C ) C \mapsto F(T_{\mathrm{M}}^{u_{\mathrm{ph}}} C) C ↦ F ( T M u ph C ) of an observable F F F at the stopping scale; this requires m s t o p = M u p h m r e m m_{\mathrm{stop}} = M^{u_{\mathrm{ph}}}m_{\mathrm{rem}} m stop = M u ph m rem with m r e m ∈ 4 N m_{\mathrm{rem}} \in4\mathbb{N} m rem ∈ 4 N . For a fixed additional number r s t o p ∈ N 0 r_{\mathrm{stop}} \in\mathbb{N}_0 r stop ∈ N 0 of RG steps, the same condition is imposed with the total exponent r s t o p J + u p h r_{\mathrm{stop}}J + u_{\mathrm{ph}} r stop J + u ph , and all iterates must lie within the finite hierarchy. No recursion for normalization constants is asserted for an integrated group of fewer than J J J averaging steps. Assume the construction conditions and the two reconstruction-image conditions (21.17) of Theorem 21.9, including r + β r e c < ϵ e n , J r + \beta_{\mathrm{rec}} < \epsilon_{\mathrm{en},J} r + β rec < ϵ en , J , the uniform incoming radii obtained by freezing every label of reduced radius, the boundary-distance comparison (21.10), the conditional force conditions (21.13) and (21.14), the cutoff radii of (21.19), the initial norm bounds of Theorem 20.14, and the bounds (21.8) on frozen coordinates. Let W t \mathcal{W}_t W t
be the fluctuation cutoff domain (21.31), which is the instance of Theorem 20.14 with common radius p ( t t ) p(t_t) p ( t t ) ; the enclosing product domain K t \mathcal{K}_t K t is used only in Lemma 21.12 . Write Z ν Z_\nu Z ν for the partition function of the initial measure, without centering: for a stopped measure
Z ν = ∫ D r e − E d 0 , s a t / g b a s e 2 − R d 0 , s a t d H , Z_\nu=\int_{D_r}e^{-E_{d_0,\mathrm{sat}}/g_{\mathrm{base}}^2-R_{d_0,\mathrm{sat}}}\,\mathrm{d}H, Z ν = ∫ D r e − E d 0 , sat / g base 2 − R d 0 , sat d H ,
and Z ν = Z m b a s e , γ Z_\nu=Z_{m_{\mathrm{base},\gamma}} Z ν = Z m base , γ for the Wilson measure. Use the stage measures and density versions of Theorem 21.10 . Then the following constructions and bounds hold at every finite depth and number of steps.
(Analytic functions, from fine to coarse. ) For every real value of the frozen coordinates satisfying (21.8), the functions E t E_t E t , R t R_t R t , Λ t \Lambda_t Λ t , W t W_t W t , the classical objectives, the reconstruction maps C t C_t C t , and the auxiliary fluctuation integrals are defined on the full spatially varying radii by the recursion of Theorem 20.14 , with the substitution of frozen coordinates of Lemma 21.11 . This definition uses no density inequality and no choice of representatives. The effective interaction is centered only in its constant term, all its first derivatives are kept, and the full classical values are kept. For a stopped measure the normalization constant on the torus is Λ 0 T = − R d 0 ( I ) \Lambda^T_0=-R_{d_0}(I) Λ 0 T = − R d 0 ( I ) , and an initial freezing subtracts in addition the frozen value of the centered effective interaction. For the Wilson measure we use the exact split of Theorem 21.9(4), with its original external factor.
(Representatives satisfying the reconstruction identities, from coarse to fine. ) Take all history-produced small top representatives in the coordinate domains of Theorem 21.9(1), retaining their native collar records as well as (21.8), and construct such representatives downward with the maps of part (1). At a boundary with freezing, let D t + 1 \mathcal{D}_{t+1} D t + 1 be the set of these representatives at the next stage, and define the set of output data satisfying the freezing and reconstruction identities
P t + 1 = A t − 1 ( D t + 1 ) . \mathcal{P}_{t+1}=\mathcal{A}^{-1}_{t}(\mathcal{D}_{t+1}). P t + 1 = A t − 1 ( D t + 1 ) .
For each such datum, include every supported fiber point:
D t = { ( C t ( c , ζ ; b r e c ) , b r e c ) : ( c , b r e c ) ∈ P t + 1 , ζ ∈ Z c , b r e c } , \mathcal{D}_t=\{(C_t(c,\zeta;b_{\mathrm{rec}}),b_{\mathrm{rec}}):(c,b_{\mathrm{rec}})\in\mathcal{P}_{t+1},\ \zeta\in\mathcal{Z}_{c,b_{\mathrm{rec}}}\}, D t = {( C t ( c , ζ ; b rec ) , b rec ) : ( c , b rec ) ∈ P t + 1 , ζ ∈ Z c , b rec } ,
where Z c , b r e c \mathcal{Z}_{c,b_{\mathrm{rec}}} Z c , b rec is the set of supported fiber coordinates over ( c , b r e c ) (c,b_{\mathrm{rec}}) ( c , b rec ) . At the last RG step, define D T \mathcal{D}_T D T to be the set of these history-produced small top representatives after terminal freezing, retaining their native collar records; the terminal freezing map is the identity if no coordinates are frozen there. Put P T = A T − 1 − 1 ( D T ) \mathcal{P}_T=\mathcal{A}^{-1}_{T-1}(\mathcal{D}_T) P T = A T − 1 − 1 ( D T ) when T > 0 T>0 T > 0 . Every configuration in the support of the terminal stage measure has one of these representatives in its gauge orbit by Theorem 21.5 and the closed-support argument above. At every lower boundary, inverse freezing restores the ordinary collar links literally from the fixed new records, and Lemma 21.8 applies to every supported fiber point. Other representatives are compared only by the changes of Lemma 21.6(1)–(3) when transporting analytic estimates, or by the pointwise invariance of Theorem 21.10 when extending the final density bound. These sets contain all choices allowed by this history-produced initialization, and no probability law is assigned to a representative map.
(Density bounds, from fine to coarse. ) For every real value of the frozen coordinates, every ( c , b r e c ) ∈ P t + 1 (c,b_{\mathrm{rec}}) \in\mathcal{P}_{t+1} ( c , b rec ) ∈ P t + 1 , and every supported fiber coordinate ζ \zeta ζ over it,
p t ( C t ( c , ζ ; b r e c ) , b r e c ) ≤ A t Z ν − 1 e Λ t ( b r e c ) W t ( b r e c ) p_t(C_t(c,\zeta;b_{\mathrm{rec}}),b_{\mathrm{rec}}) \le A_t Z_\nu^{-1} e^{\Lambda_t(b_{\mathrm{rec}})} W_t(b_{\mathrm{rec}}) p t ( C t ( c , ζ ; b rec ) , b rec ) ≤ A t Z ν − 1 e Λ t ( b rec ) W t ( b rec )
× exp { − E t ( Log C t ( c , ζ ; b r e c ) ; b r e c ) / g ℓ t 2 } {}\times\exp\{-E_t(\operatorname{Log} C_t(c,\zeta;b_{\mathrm{rec}});b_{\mathrm{rec}})/g_{\ell_t}^2\} × exp { − E t ( Log C t ( c , ζ ; b rec ) ; b rec ) / g ℓ t 2 }
× exp { − R t ( Log C t ( c , ζ ; b r e c ) ; b r e c ) } . {}\times\exp\{-R_t(\operatorname{Log} C_t(c,\zeta;b_{\mathrm{rec}});b_{\mathrm{rec}})\}. × exp { − R t ( Log C t ( c , ζ ; b rec ) ; b rec )} .
At the terminal stage the bound is, directly, for every ( v , b r e c ) ∈ D T (v,b_{\mathrm{rec}}) \in\mathcal{D}_T ( v , b rec ) ∈ D T ,
p T ( v , b r e c ) ≤ A T Z ν − 1 e Λ T ( b r e c ) W T ( b r e c ) exp { − E T ( Log v ; b r e c ) / g ℓ T 2 − R T ( Log v ; b r e c ) } . p_T(v,b_{\mathrm{rec}}) \le A_T Z_\nu^{-1}e^{\Lambda_T(b_{\mathrm{rec}})}W_T(b_{\mathrm{rec}})\exp\{-E_T(\operatorname{Log}v;b_{\mathrm{rec}})/g_{\ell_T}^2-R_T(\operatorname{Log}v;b_{\mathrm{rec}})\}. p T ( v , b rec ) ≤ A T Z ν − 1 e Λ T ( b rec ) W T ( b rec ) exp { − E T ( Log v ; b rec ) / g ℓ T 2 − R T ( Log v ; b rec )} .
Here the terminal quantities include the same freezing substitution, if any. Also A 0 = 1 A_0=1 A 0 = 1 and
A t + 1 = A t ( 1 − ϵ t ) − 1 , ϵ t ≤ 6 b t e − p ( t t ) 2 / 48 , b t ≤ C J m t 3 , log A T ≤ C δ , A_{t+1}=A_t(1-\epsilon_t)^{-1}, \qquad\epsilon_t\le6b_t e^{-p(t_t)^2/48}, \qquad b_t\le C_Jm_t^3,\qquad\log A_T\le C_\delta, A t + 1 = A t ( 1 − ϵ t ) − 1 , ϵ t ≤ 6 b t e − p ( t t ) 2 /48 , b t ≤ C J m t 3 , log A T ≤ C δ ,
after the uniform choice of the minimal side in Lemma 20.5. Every step uses the exact Haar disintegration and then a single triangular change to normalized fluctuation variables, with determinant
∏ i = 0 s t − 1 g i 3 f i det K i ( c i + 1 ( w > i ) ; b r e c ) − 1 / 2 . \prod_{i=0}^{s_{t-1}}g_i^{3f_i}\det K_i(c_{i+1}(w_{>i});b_{\mathrm{rec}})^{-1/2}. i = 0 ∏ s t − 1 g i 3 f i det K i ( c i + 1 ( w > i ) ; b rec ) − 1/2 .
The resulting Borel set of fluctuation coordinates of supported points is contained in the enclosing product domain, and Lemma 21.12, with identical prefactors and the smaller cutoff domain, gives the stated loss. The next stage evaluates the same configuration. The indicators are kept until this comparison, and no domination at arbitrary points of the enclosing product domain is required.
(Normalization. ) With Δ Λ t = Λ t − Λ T \Delta\Lambda_t=\Lambda_t-\Lambda_T Δ Λ t = Λ t − Λ T in the same centered renormalization group recursion on the whole torus, Theorem 20.13 gives
Δ Λ t + 1 ≤ Δ Λ t + ∑ i i n t d i ( H t ) log ( 1 / g i ) + C J ( N m i s s i n g , t + N b d , t + N f r e e z e , t ) , \Delta\Lambda_{t+1}\le\Delta\Lambda_t+\sum_{i\ \mathrm{in}\ t}d_i(H_t)\log(1/g_i)+C_J(N_{\mathrm{missing},t}+N_{\mathrm{bd},t}+N_{\mathrm{freeze},t}), Δ Λ t + 1 ≤ Δ Λ t + i in t ∑ d i ( H t ) log ( 1/ g i ) + C J ( N missing , t + N bd , t + N freeze , t ) ,
including the initial freezings. This comparison uses the same centered recursion on the whole torus and the same partition function Z ν Z_\nu Z ν throughout.
(External classical terms and the measure of frozen coordinates. ) For the comparison of the output classical actions, the protected external classical exponents are recombined exactly once. At fixed frozen coordinates, with their original couplings included in S e x t ( b r e c ) S_{\mathrm{ext}}(b_{\mathrm{rec}}) S ext ( b rec ) ,
e − S e x t ( b r e c ) e − E i n t / g t 2 = e − E f u l l / g t 2 , E f u l l = E i n t + g t 2 S e x t ( b r e c ) , e^{-S_{\mathrm{ext}}(b_{\mathrm{rec}})}e^{-E^{\mathrm{int}}/g_t^2}=e^{-E^{\mathrm{full}}/g_t^2},\qquad E^{\mathrm{full}}=E^{\mathrm{int}}+g_t^2S_{\mathrm{ext}}(b_{\mathrm{rec}}), e − S ext ( b rec ) e − E int / g t 2 = e − E full / g t 2 , E full = E int + g t 2 S ext ( b rec ) ,
where E i n t E^{\mathrm{int}} E int is the effective classical action of the block region, and this factor is removed from the external weight. Under an RG step both the action and g t 2 g_t^2 g t 2 are multiplied by M s t M^{s_t} M s t , so each such exponent is preserved and counted once. At the top we keep the measure
ρ X ( d b r e c ) = w X ( b r e c ) H X ( d b r e c ) ≤ H X ( d b r e c ) , w X = I X a c t I X i n v , e n d , \rho_{\mathsf X}(db_{\mathrm{rec}})=w_{\mathsf X}(b_{\mathrm{rec}})H_{\mathsf X}(db_{\mathrm{rec}})\le H_{\mathsf X}(db_{\mathrm{rec}}),\qquad w_{\mathsf X}=I_{\mathsf X}^{\mathrm{act}}I_{\mathsf X}^{\mathrm{inv,end}}, ρ X ( d b rec ) = w X ( b rec ) H X ( d b rec ) ≤ H X ( d b rec ) , w X = I X act I X inv , end ,
on the frozen coordinates, where H X H_{\mathsf X} H X is their product Haar measure. The first indicator keeps every failure q f ( b r e c ) ≥ d ∗ q_f(b_{\mathrm{rec}})\ge d_* q f ( b rec ) ≥ d ∗ , which is a function of the frozen coordinates since the whole dependence set of f f f was frozen. The second keeps any chosen invariant tests depending only on the frozen coordinates at the output. There is no second copy of the frozen-coordinate terms included in the full classical action. Optional noninvariant chart restrictions may be bounded by one.
All local constants have the dependence stated at the beginning of this section, including independence of the later widths, while C δ C_\delta C δ has the dependence described in item (4) of Definition 4.14. No comparison of regional integrals from below or independent covariance or overlap identity for the auxiliary regional functions is part of this induction. Compact overlap for the torus classical restriction on the common charts is used in (125) .
For a range containing no complete RG step, this theorem asserts no small terminal representative, for either initial measure. The empty analytic recursion retains its initial values, and estimates for the initial measure are made directly. Empty active regions and empty fluctuation products within a positive-length decomposition retain the conventions of the theorem.
Proof. Every integrated step has s t = J s_t=J s t = J , so ℓ t + 1 = ℓ t + J \ell_{t+1}=\ell_t+J ℓ t + 1 = ℓ t + J and g ℓ t + 1 2 = B g ℓ t 2 g_{\ell_{t+1}}^2=B g_{\ell_t}^2 g ℓ t + 1 2 = B g ℓ t 2 . We reserve z i z_i z i for the free logarithmic coordinates and write ζ i = g i w i \zeta_i=g_iw_i ζ i = g i w i for the scaled fluctuation arguments of C t C_t C t . All analytic functions below are functions of the identity-chart variables of the block region at fixed real ( ξ , b r e c ) (\xi,b_{\mathrm{rec}}) ( ξ , b rec ) . We prove the five parts in order, followed by the statement on the remaining averaging steps.
Part (1): initialization. For a stopped measure, we set
r d 0 = R d 0 − R d 0 ( I ) , Λ 0 T = − R d 0 ( I ) . r_{d_0}=R_{d_0}-R_{d_0}(I),\qquad\Lambda_0^{\mathsf T}=-R_{d_0}(I). r d 0 = R d 0 − R d 0 ( I ) , Λ 0 T = − R d 0 ( I ) .
The partition function without centering then gives exactly Z ν − 1 e Λ 0 T − E d 0 / g b a s e 2 − r d 0 Z_\nu^{-1}e^{\Lambda_0^{\mathsf T}-E_{d_0}/g_{\mathrm{base}}^2-r_{d_0}} Z ν − 1 e Λ 0 T − E d 0 / g base 2 − r d 0 at a small representative in D r D_r D r . In particular Z ^ = e R d 0 ( I ) Z ν \widehat Z=e^{R_{d_0}(I)}Z_\nu Z = e R d 0 ( I ) Z ν for the centered partition function in Theorem 21.5 ; we keep Z ν Z_\nu Z ν throughout this proof. For an initial freezing at ξ ♯ \xi^\sharp ξ ♯ , we set
E 0 ( v ; b r e c ) = E d 0 ( v ⊕ ξ ♯ ) , R 0 ( v ; b r e c ) = r d 0 ( v ⊕ ξ ♯ ) − r d 0 ( 0 ⊕ ξ ♯ ) , Λ 0 ( b r e c ) = − R d 0 ( I ) − r d 0 ( 0 ⊕ ξ ♯ ) . (137) \begin{aligned}
E_0(v;b_{\mathrm{rec}})&=E_{d_0}(v\oplus\xi^\sharp),\\
R_0(v;b_{\mathrm{rec}})&=r_{d_0}(v\oplus\xi^\sharp)-r_{d_0}(0\oplus\xi^\sharp),\\
\Lambda_0(b_{\mathrm{rec}})&=-R_{d_0}(I)-r_{d_0}(0\oplus\xi^\sharp). \tag*{(137)}
\end{aligned} E 0 ( v ; b rec ) R 0 ( v ; b rec ) Λ 0 ( b rec ) = E d 0 ( v ⊕ ξ ♯ ) , = r d 0 ( v ⊕ ξ ♯ ) − r d 0 ( 0 ⊕ ξ ♯ ) , = − R d 0 ( I ) − r d 0 ( 0 ⊕ ξ ♯ ) . (137)
Additional frozen coordinates are held as real parameters. For the Wilson measure, we set E 0 = F b a r e E_0=F_{\mathrm{bare}} E 0 = F bare , R 0 = 0 R_0=0 R 0 = 0 , Λ 0 = 0 \Lambda_0=0 Λ 0 = 0 , and W 0 = e − E h o l e / g b a s e 2 W_0=e^{-E_{\mathrm{hole}}/g_{\mathrm{base}}^2} W 0 = e − E hole / g base 2 , together with any kept indicators at the output. The exact split in Theorem 21.9(4) implies this initialization with the partition function of the Wilson measure.
We verify the initial norm bounds. The initial analytic values obey (110) , because C c o n v , J ≥ C a d , J C_{\mathrm{conv},J}\ge C_{\mathrm{ad},J} C conv , J ≥ C ad , J , which was fixed before Theorem 21.5 , and the near values at the first step under the Wilson measure satisfy its conversion bound. For a stopped measure, the decomposition with zero completion has classical perturbation P 0 ( v ) = E d 0 ( v ⊕ ξ ♯ ) − E d 0 ( v ⊕ 0 ) P_0(v)=E_{d_0}(v\oplus\xi^\sharp)-E_{d_0}(v\oplus0) P 0 ( v ) = E d 0 ( v ⊕ ξ ♯ ) − E d 0 ( v ⊕ 0 ) , including its constant. With the matched completion ι ♯ \iota^\sharp ι ♯ of Theorem 19.9 , this is exactly its boundary-completion difference, and the adjusted perturbation is P ‾ 0 = 0 \overline{P}_0=0 P 0 = 0 . In the norm with zero completion, its nonconstant part is bounded by C b c , c l E c l β r e c / ρ c l ≤ σ ∗ / 4 C_{\mathrm{bc,cl}}E_{\mathrm{cl}}\beta_{\mathrm{rec}}/\rho_{\mathrm{cl}}\le\sigma_*/4 C bc , cl E cl β rec / ρ cl ≤ σ ∗ /4 , by (118) . For the quantum input, splitting the centered function r d 0 ( v ⊕ ξ ♯ ) r_{d_0}(v\oplus\xi^\sharp) r d 0 ( v ⊕ ξ ♯ ) into its restriction to the torus and its boundary-completion difference gives the bound C b c , q E β r e c / ρ ≤ σ ∗ / 4 C_{\mathrm{bc,q}}E\beta_{\mathrm{rec}}/\rho\le\sigma_*/4 C bc , q E β rec / ρ ≤ σ ∗ /4 . The constant in (137) is kept exactly and is not bounded by this norm.
At the first step under the Wilson measure, the plaquettes depending on integration variables use only the small near values of the initial good-cell complex. The zero-completed Wilson function evaluated at those data differs from the sum of these plaquette terms by plaquettes involving only frozen variables, which are constant in the integration variables. Hence the adjusted classical perturbation is constant at this step, and its nonconstant boundary-completion difference with zero completion obeys the same classical bound. These classical constants remain in F b a r e F_{\mathrm{bare}} F bare , and the energy of the deep frozen group variables stays in W 0 W_0 W 0 . The quantum input is zero. This proves x 0 , x 0 c l ≤ σ ∗ x_0,x_0^{\mathrm{cl}}\le\sigma_* x 0 , x 0 cl ≤ σ ∗ for either initial measure, on uniform radii of the remaining inputs. No deep frozen group variable has entered an analytic logarithm.
Part (1): the analytic recursion. At a later input, let D i n D_{\mathrm{in}} D in be the list of remaining analytic variables of the block region, with their original locations and volume factors; it does not include the frozen parameters. Since the preceding freezing of the union removed every label of reduced radius, the classical and quantum input radii are uniform by (117) . The previous fixed complement and all earlier frozen basepoints obey (112) . The initialization satisfies the same statement, with zero boundary-weighted input when there is no incoming fixed complement, and additional excluded regions created by the tests preserve this inclusion. All completion adjustments use only newly frozen small analytic variables; previously nonintegrated torus variables remain zero, and the deep frozen group variables remain fixed external parameters.
We apply Theorems 19.9 and 20.14 inductively, with the fixed amplitude A w A_w A w and the coupling ceiling g ∗ g_* g ∗ chosen for it, to define the classical objective, the conditional minimizers for initial segments, their positive square roots, and the reconstruction map on the full radii. The hypotheses of each application are checked before any set of representatives is constructed. The input norm just restored bounds the adjusted perturbation by 2 σ ∗ 2\sigma_* 2 σ ∗ . The images and every connecting segment lie strictly in the uniform input ball, by (119) and the fixed margin of the larger construction domain. The force and Hessian estimates, uniformly for y ∈ T I y \in\mathcal{T}_I y ∈ T I for each initial segment I I I , followed by (116) , give all conditional minimizers and sequential neighborhoods; these checks are Lemma 21.7 . All perturbation derivatives are taken in the integration and free variables at fixed real b b b . The fixed rates and spatial margins are those of the common tuple of Theorem 20.14 and are not changed by freezing. The torus input is the member at accumulated classical depth d t = d 0 + t J d_t=d_0+tJ d t = d 0 + t J of its uniform analytic class, with ceiling E E E ; a new nonconstant boundary term is accounted for by the two input norms rather than added to this ceiling. We include the exact boundary-completion difference when the unperturbed objective uses newly frozen input variables. By Theorem 19.9 , this objective agrees on the connected chart with E t E_t E t evaluated at its reconstructed field; only the torus term uses compact overlap there, and the perturbation is evaluated at that field from the outset. The full outgoing action is defined to be the value at the conditional minimizer multiplied by M s t M^{s_t} M s t .
We next define the update of the effective interaction by an integral; this also fixes all normalization conventions. Let J t 0 ( c , b r e c ) J_t^0(c,b_{\mathrm{rec}}) J t 0 ( c , b rec ) be the product of the Haar disintegration factors of the intermediate scales evaluated along the zero-fluctuation chain, and define
D t 0 ( c , b r e c ) = J t 0 ( c , b r e c ) ∏ i g i 3 f i det K i ( c ^ i + 1 ; b r e c ) − 1 / 2 , (138) D_t^0(c,b_{\mathrm{rec}})=J_t^0(c,b_{\mathrm{rec}})\prod_i g_i^{3f_i}\det K_i(\widehat{c}_{i+1};b_{\mathrm{rec}})^{-1/2}, \tag*{(138)} D t 0 ( c , b rec ) = J t 0 ( c , b rec ) i ∏ g i 3 f i det K i ( c i + 1 ; b rec ) − 1/2 , (138)
where c ^ i + 1 \widehat{c}_{i+1} c i + 1 denotes the corresponding higher argument at zero fluctuation. All field-independent constants of the Haar disintegration are included. Sequential minimization and normalization give the exact decomposition of the action: the output value divided by g ℓ t + 1 2 g_{\ell_{t+1}}^2 g ℓ t + 1 2 , the quadratic term ∑ i ∣ w i ∣ 2 / 2 \sum_i |w_i|^2/2 ∑ i ∣ w i ∣ 2 /2 , and the higher classical remainders add up to the input classical exponent. Indeed, at each intermediate scale we subtract the conditional minimum and its quadratic Taylor term; the latter is ∣ w i ∣ 2 / 2 |w_i|^2/2 ∣ w i ∣ 2 /2 by the definition of K i K_i K i , and the factor M M M in the minimum cancels the factor M M M in the next coupling. Repeating this over the s t s_t s t scales, we obtain the decomposition, including its constants. Adding the inherited difference of effective interactions, all logarithmic differences of the Haar disintegration factors, and all logarithmic differences of the determinants that depend on the higher fluctuation variables, we obtain the conditional fluctuation potential V t ( c , b r e c ; w ) V_t(c,b_{\mathrm{rec}};w) V t ( c , b rec ; w ) , including its quadratic part; it is the potential of Theorem 20.14 .
In this integral, we use the fluctuation cutoff domain (133) : every radius is p ( t t ) p(t_t) p ( t t ) with t t = max i g t , i t_t=\max_i g_{t,i} t t = max i g t , i . This is in the allowed logarithmic family; it is not the radius R e n / t t R_{\mathrm{en}}/t_t R en / t t of the enclosing product. The containment, positive denominator, interaction, and auxiliary half-disc conditions hold by (119) and (121) ; they are deterministic conditions on the parameters and on g ∗ g_* g ∗ , not restrictions on the support of the measure. On the fixed product W t \mathcal{W}_t W t , we define
I t ( c ; b r e c ) = D t 0 ( c , b r e c ) e − R t ( Log C t ( c , 0 ; b r e c ) ; b r e c ) ∫ W t e − V t ( c , b r e c ; w ) d w , \mathcal{I}_t(c;b_{\mathrm{rec}})=D_t^0(c,b_{\mathrm{rec}})e^{-R_t(\operatorname{Log} C_t(c,0;b_{\mathrm{rec}});b_{\mathrm{rec}})}\int_{\mathcal{W}_t}e^{-V_t(c,b_{\mathrm{rec}};w)}\,dw, I t ( c ; b rec ) = D t 0 ( c , b rec ) e − R t ( Log C t ( c , 0 ; b rec ) ; b rec ) ∫ W t e − V t ( c , b rec ; w ) d w ,
Q t ( c ; b r e c ) = − log I t ( c ; b r e c ) , Q_t(c;b_{\mathrm{rec}})=-\log\mathcal{I}_t(c;b_{\mathrm{rec}}), Q t ( c ; b rec ) = − log I t ( c ; b rec ) ,
R t o u t ( c ; b r e c ) = Q t ( c ; b r e c ) − Q t ( 0 ; b r e c ) , Λ t o u t = Λ t − Q t ( 0 ; b r e c ) . (139) R_t^{\mathrm{out}}(c;b_{\mathrm{rec}})=Q_t(c;b_{\mathrm{rec}})-Q_t(0;b_{\mathrm{rec}}),\qquad\Lambda_t^{\mathrm{out}}=\Lambda_t-Q_t(0;b_{\mathrm{rec}}). \tag*{(139)} R t out ( c ; b rec ) = Q t ( c ; b rec ) − Q t ( 0 ; b rec ) , Λ t out = Λ t − Q t ( 0 ; b rec ) . (139)
The logarithm is real on real data, where it agrees with the real logarithm of the positive integral; its zero-free analytic continuation is supplied by Theorem 20.14 , and its sign is unrestricted. The integrand and its domain are defined on the full radii before any support set is considered. We have the exact identity
Z ν − 1 Λ t W t e − E t o u t ( c ) / g t + 1 2 I t ( c ; b r e c ) = Z ν − 1 Λ t o u t W t e − E t o u t ( c ) / g t + 1 2 e − R t o u t ( c ; b r e c ) . (140) Z_{\nu}^{-1}\Lambda_t\mathcal{W}_t e^{-E_t^{\mathrm{out}}(c)/g_{t+1}^2}\mathcal{I}_t(c;b_{\mathrm{rec}})=Z_{\nu}^{-1}\Lambda_t^{\mathrm{out}}\mathcal{W}_t e^{-E_t^{\mathrm{out}}(c)/g_{t+1}^2}e^{-R_t^{\mathrm{out}}(c;b_{\mathrm{rec}})}. \tag*{(140)} Z ν − 1 Λ t W t e − E t out ( c ) / g t + 1 2 I t ( c ; b rec ) = Z ν − 1 Λ t out W t e − E t out ( c ) / g t + 1 2 e − R t out ( c ; b rec ) . (140)
Changing the distribution of constant factors between D t 0 D_t^0 D t 0 and the definition of Q t Q_t Q t leaves this identity unchanged; we use the convention of Theorem 20.13 , which assigns each field-independent factor once.
At a boundary we use the substitution of Lemma 21.11 . This defines all functions at the next level for every real value of the frozen coordinates, independently of any density bound. The simultaneously frozen list is the union of the full boundary band of reduced radius, the removed band, the new boundary layers, and the based transports; the corresponding initial variables were included at the initial scale. Their analytic values satisfy (21.8) for both radii. At a representative of the next stage satisfying the reconstruction identities, the freezing identity inserts these values, not small values from a different frame, and the total boundary-completion difference obeys (21.16). The resulting norm bounds are exactly those of Theorem 20.14 : the output ceilings before freezing are y t ∙ y_t^\bullet y t ∙ and y t c l y_t^{\mathrm{cl}} y t cl , the boundary-completion differences are at most σ ∗ / 4 \sigma_*/4 σ ∗ /4 , and the removed band makes each remaining term at most σ ∗ / 2 \sigma_*/2 σ ∗ /2 . With α q = α \alpha_{\mathrm{q}}=\alpha α q = α and the corresponding classical rate, we have for both types
x t + 1 ∙ ≤ C b c , ∙ E ∙ β r e c / ρ ∙ + 2 e − α ∙ ( s e r − 1 ) y t ∙ ≤ σ ∗ / 4 + σ ∗ / 2 , x_{t+1}^\bullet\le C_{\mathrm{bc},\bullet}E_\bullet\beta_{\mathrm{rec}}/\rho_\bullet+2e^{-\alpha_\bullet(s_{\mathrm{er}}-1)}y_t^\bullet\le\sigma_*/4+\sigma_*/2, x t + 1 ∙ ≤ C bc , ∙ E ∙ β rec / ρ ∙ + 2 e − α ∙ ( s er − 1 ) y t ∙ ≤ σ ∗ /4 + σ ∗ /2 ,
so x t + 1 , x t + 1 c l ≤ 3 σ ∗ / 4 x_{t+1},x_{t+1}^{\mathrm{cl}}\le3\sigma_*/4 x t + 1 , x t + 1 cl ≤ 3 σ ∗ /4 whenever the input is in its ball. The output ceilings y t ∙ y_t^\bullet y t ∙ are computed on the input ball before the width s e r s_{\mathrm{er}} s er is chosen. By induction from the initial ball, we obtain the analytic recursion at every finite number of steps. It keeps the terms of degree one; no first derivative has been removed by recentering, and every restoration of the bounds happens at a complete RG step.
It remains to verify the Borel regularity in the real parameters required at each perturbation step. The initial coefficients and chart path products are Borel functions of the protected real frozen coordinates; the deep frozen group variables at the first step under the Wilson measure enter only the Borel external factor. Suppose this holds at an input. The fixed-point iterations for the conditional minimizers use Borel coefficients and converge uniformly on their fixed neighborhoods, so their limits are Borel in b b b . The positive matrix square roots and inverses are Borel as uniform limits of the matrix-function constructions. The resulting values, Haar disintegration determinants, and coefficients are Borel. By Tonelli’s theorem, integration over the fixed finite product preserves Borel dependence; we apply it to the positive and negative parts and to the real and imaginary parts, which have finite integrals. The logarithm can be chosen jointly Borel without a branch choice depending on the frozen coordinates: on the star-shaped output domain,
log I t ( c ; b ) = log I t ( 0 ; b ) + ∫ 0 1 D c I t ( s c ; b ) [ c ] I t ( s c ; b ) d s \log\mathcal{I}_t(c;b)=\log\mathcal{I}_t(0;b)+\int_0^1\frac{D_c\mathcal{I}_t(sc;b)[c]}{\mathcal{I}_t(sc;b)}\,\mathrm{d}s log I t ( c ; b ) = log I t ( 0 ; b ) + ∫ 0 1 I t ( sc ; b ) D c I t ( sc ; b ) [ c ] d s
where the first term is the real logarithm. Zero-freeness and holomorphy make this the required analytic branch, and the parameter integral is Borel. Finally, the substitution of frozen coordinates evaluates these Borel families at the Borel coordinates of the frozen values. This proves the Borel regularity inductively, without differentiating in a real frozen parameter.
Part (2): representatives satisfying the reconstruction identities. We start from all history-produced small representatives of top configurations, in the coordinate domains of Theorem 21.9(1), retaining both (21.8) and the native C B r C_{Br} C B r collar records; every supported top orbit has such a representative by Theorem 21.5 . At a boundary with freezing, we pull the next set of representatives satisfying the reconstruction identities back by the freezing bijection A t A_t A t ; this defines P t + 1 \mathscr{P}_{t+1} P t + 1 , which is not enlarged to an arbitrary comparison set. For each such datum, we include every supported fiber coordinate, as in the definition of D t \mathscr{D}_t D t . Theorem 21.9, applied at this step of the descent, shows that each of them has its inverse normalized fluctuation coordinate in the interior half of the enclosing product, and that its representative has the correct target and protected frozen coordinates. This defines the whole finite family downward using only the reconstruction maps C t C_t C t of part (1). The sets integrated below are the Borel sets of free coordinates of supported fiber points at fixed output data, so no measurable representative map is needed.
The bounds (21.8) persist along this descent, by Lemma 21.6(1)–(3) and, at a new boundary, the direct insertion of Lemma 21.11 . Changes as in Lemma 21.6(4) are not used to enlarge these sets. Over each output datum in P t + 1 \mathscr{P}_{t+1} P t + 1 , the set still contains every supported fiber point.
These sets meet every supported gauge orbit. Indeed, suppose this holds at stage t + 1 t+1 t + 1 , and take a supported point at stage t t t . Map it to its raw output and then through the freezing map. An admissible endpoint gauge sends that next-stage orbit to D t + 1 \mathscr{D}_{t+1} D t + 1 . The equivariance of freezing and the fiber bijection used in (130) lift this change to the incoming point. Its output then lies in A t − 1 ( D t + 1 ) A_t^{-1}(\mathscr{D}_{t+1}) A t − 1 ( D t + 1 ) , and the resulting supported fiber point is included in D t \mathscr{D}_t D t by the all-fiber definition. This proves the assertion by downward induction from the history-produced top representatives. This orbit selection uses only the exact gauge action on the stage variables and their support; it does not assert covariance of the auxiliary regional perturbations under arbitrary endpoint gauges. Analytic comparisons of the torus classical restriction within the selected family use compact overlap after the endpoint-preserving completions proved above.
Part (3): density bounds. We use the densities of Theorem 21.10. For a stopped measure, the gauge-invariant density agrees with its analytic expression at each representative satisfying the reconstruction identities, and the indicators of the decomposition and of the support are between zero and one. Then (21.35) implies the claimed bound at the initial scale with A 0 = 1 A_0 = 1 A 0 = 1 . For the Wilson measure, the exact split implies the same bound with the kept factor W 0 W_0 W 0 . At a point outside the support, the density is zero, and no analytic identity is needed there.
Suppose that the inequality holds on D t \mathcal{D}_t D t , and fix an output datum ( c , b r e c ) ∈ P t + 1 (c,b_{\mathrm{rec}}) \in\mathcal{P}_{t+1} ( c , b rec ) ∈ P t + 1 . We apply the exact Haar disintegration to the incoming density, keeping its tree variables at first. By Theorem 21.10, the covariance under based gauges makes their normalized Haar integrals equal to one. For this fixed target, the gauge transformation of Lemma 19.5 is a fixed gauge transformation, independent of the free variables and equal to I I I at the roots of the target and at every protected vertex. The invariance of the density therefore allows us to evaluate the integrand at C t ( c , ζ ; b r e c ) C_t(c,\zeta;b_{\mathrm{rec}}) C t ( c , ζ ; b rec ) instead of at its representative before the gauge transformation, and the tests take the same values there. This changes neither the Haar disintegration nor any determinant, and it is not a covariance statement for the auxiliary functions on the right side of the inequality.
We now perform the single sequential change to normalized fluctuation variables for the classical objective. At scale i i i its derivative with respect to w i w_i w i is g i K i − 1 / 2 g_iK_i^{-1/2} g i K i − 1/2 , and its other dependence is on higher variables. Ordering the higher variables first makes the derivative block triangular, so its positive real determinant is exactly
∏ i = 0 s t − 1 g i 3 f i det K i ( c i + 1 ( w > i ) ; b r e c ) − 1 / 2 . (141) \prod_{i=0}^{s_t-1} g_i^{3f_i}\det K_i\left(c_{i+1}(w_{>i});b_{\mathrm{rec}}\right)^{-1/2}. \tag*{(141)} i = 0 ∏ s t − 1 g i 3 f i det K i ( c i + 1 ( w > i ) ; b rec ) − 1/2 . (141)
In particular, the determinant at a varying higher argument is not replaced by its value at zero. (138) and the corresponding logarithmic differences split this one determinant; they do not add a second copy. Injectivity and the inverse map on the supported domain follow from the branches for initial segments, with their positive margins, established before. As in the proof of Lemma 21.12, the image is a Borel set S c , b r e c S_{c,b_{\mathrm{rec}}} S c , b rec in the enclosing product, and the remaining indicators composed with the inverse form a Borel function χ t \chi_t χ t on K t \mathcal{K}_t K t with values in [ 0 , 1 ] [0,1] [ 0 , 1 ] . The potential has a finite nonzero integral over K t \mathcal{K}_t K t at each fixed real datum, also for an empty product.
Every supported point of that domain belongs to D t \mathscr{D}_t D t . We insert the inductive bound there, and let P t 0 ( c , b r e c ) P_t^0(c,b_{\mathrm{rec}}) P t 0 ( c , b rec ) be
Z ν − 1 e A t ( b r e c ) W t ( b r e c ) D t 0 ( c , b r e c ) e − R t o u t ( c ; b r e c ) / g t + 1 2 − R t ( Log C t ( c , 0 ; b r e c ) ; b r e c ) . Z_\nu^{-1}e^{A_t(b_{\mathrm{rec}})}W_t(b_{\mathrm{rec}})D_t^0(c,b_{\mathrm{rec}})e^{-R_t^{\mathrm{out}}(c;b_{\mathrm{rec}})/g_{t+1}^2-R_t(\operatorname{Log} C_t(c,0;b_{\mathrm{rec}});b_{\mathrm{rec}})}. Z ν − 1 e A t ( b rec ) W t ( b rec ) D t 0 ( c , b rec ) e − R t out ( c ; b rec ) / g t + 1 2 − R t ( Log C t ( c , 0 ; b rec ) ; b rec ) .
Tests depending only on the output datum are kept in this common factor. If this factor is zero, the outgoing upper bound is zero; otherwise it is finite and positive. With χ t \chi_t χ t denoting the remaining indicators, the exact disintegration followed by the inductive bound gives
p t o u t ( c , b r e c ) ≤ A t P t 0 ( c , b r e c ) ∫ S c , b r e c χ t ( w ) e − V t ( w ) d w ≤ A t P t 0 ( c , b r e c ) ∫ K t e − V t ( w ) d w ≤ A t 1 − ϵ t P t 0 ( c , b r e c ) ∫ W t e − V t ( w ) d w . (142) \begin{aligned}
p_t^{\mathrm{out}}(c,b_{\mathrm{rec}})&\le A_tP_t^0(c,b_{\mathrm{rec}})\int_{\mathscr{S}_{c,b_{\mathrm{rec}}}}\chi_t(w)e^{-V_t(w)}\,dw\\
&\le A_tP_t^0(c,b_{\mathrm{rec}})\int_{\mathcal{K}_t}e^{-V_t(w)}\,dw\\
&\le\frac{A_t}{1-\epsilon_t}P_t^0(c,b_{\mathrm{rec}})\int_{\mathcal{W}_t}e^{-V_t(w)}\,dw. \tag*{(142)}
\end{aligned} p t out ( c , b rec ) ≤ A t P t 0 ( c , b rec ) ∫ S c , b rec χ t ( w ) e − V t ( w ) d w ≤ A t P t 0 ( c , b rec ) ∫ K t e − V t ( w ) d w ≤ 1 − ϵ t A t P t 0 ( c , b rec ) ∫ W t e − V t ( w ) d w . (142)
Here ϵ t \epsilon_t ϵ t is a common upper bound for the datum-dependent δ t \delta_t δ t of Lemma 21.12 at stage t t t , over all admitted output data, after the uniform minimum-side choice in part (3) of Lemma 20.5 . One may take ϵ t = 6 b t e − p ( t t ) 2 / 48 ≤ 1 / 2 \epsilon_t=6b_te^{-p(t_t)^2/48}\le1/2 ϵ t = 6 b t e − p ( t t ) 2 /48 ≤ 1/2 . The first inequality is the hypothesis (21.32 of that lemma with χ = 1 \chi=1 χ = 1 and P t = A t P t 0 P_t=A_tP_t^0 P t = A t P t 0 , and the remaining two inequalities are its conclusion. The middle step uses the enclosure of Theorem 21.9, positivity on the whole larger product, and only 0 ≤ χ t ≤ 1 0\le\chi_t\le1 0 ≤ χ t ≤ 1 . The last step is Lemma 20.5 with identical prefactors; its radii are the common logarithmic radii, enclosed by t t p ( t t ) < R e n t_tp(t_t)<R_{\mathrm{en}} t t p ( t t ) < R en , and its Hessian and mean conditions hold by (121) , in particular with mean at most p ( t t ) / ( 2 3 ) p(t_t)/(2\sqrt{3}) p ( t t ) / ( 2 3 ) . The larger product K t \mathcal{K}_t K t is used only in this comparison. This verifies the hypothesis of Lemma 21.12 at every later stage, with the incoming loss A t A_t A t in the common prefactor.
By (140) , the last line of (142) is the analytic outgoing bound, defined independently, with the factor A t + 1 = A t / ( 1 − ϵ t ) A_{t+1}=A_t/(1-\epsilon_t) A t + 1 = A t / ( 1 − ϵ t ) . By Lemma 21.11 , the inverse under freezing of a representative of the next stage is this output configuration. That lemma substitutes the based transports and frozen links directly, transfers − R t o u t ( 0 ⊕ ξ ♯ ) -R_t^{\mathrm{out}}(0\oplus\xi^\sharp) − R t out ( 0 ⊕ ξ ♯ ) to the normalization constant, and keeps the full classical value. This proves the incoming inequality at stage t + 1 t+1 t + 1 on exactly D t + 1 \mathscr{D}_{t+1} D t + 1 , including the displayed terminal bound when t + 1 = T t+1=T t + 1 = T , and completes the induction.
There is one comparison per output datum, jointly over all components and scales of the step. The probability estimate used here is ϵ t ≤ 6 b t e − p ( t t ) 2 / 48 \epsilon_t\le6b_te^{-p(t_t)^2/48} ϵ t ≤ 6 b t e − p ( t t ) 2 /48 with b t ≤ C J m t 3 b_t\le CJm_t^3 b t ≤ C J m t 3 . Ordering the finite list of output data along the sequence of sides and taking the uniform minimal side of Lemma 20.5 , we obtain
log A T = ∑ t < T − log ( 1 − ϵ t ) ≤ 2 ∑ t < T ϵ t ≤ C δ . \log A_T=\sum_{t<T}-\log(1-\epsilon_t)\le2\sum_{t<T}\epsilon_t\le C_\delta. log A T = t < T ∑ − log ( 1 − ϵ t ) ≤ 2 t < T ∑ ϵ t ≤ C δ .
Empty free products have determinant one, logarithmic determinant zero, and fluctuation integral one, and an empty block region has zero nonconstant norm. Such an empty step contributes no loss.
Part (4): normalization. Since s t = J s_t = J s t = J and the initial and stopping scales differ by complete RG steps, every application below is within the scope of Theorem 20.13 . The recursion just constructed is the same centered recursion, with the same extraction of field-independent factors, as in that theorem. All its norm and domain hypotheses were verified in part (1), including the common logarithmic cutoffs and the freezing of the whole union. The column bound used for the analytic norm propagation in Theorem 20.10 is that of the perturbed zero-fluctuation map. Indeed, the doubled output radius and the inherited half radius are part of the two image conditions, and the differences with each index fixed in Theorem 19.9 give, in the same boundary-weighted sum over distinguished outputs,
K c o l , J a c t u a l ≤ K c o l , J 0 + C J σ 0 , J , K^{\mathrm{actual}}_{\mathrm{col},J} \le K^{0}_{\mathrm{col},J} + C_J\sigma_{0,J}, K col , J actual ≤ K col , J 0 + C J σ 0 , J ,
where K c o l , J a c t u a l K^{\mathrm{actual}}_{\mathrm{col},J} K col , J actual and K c o l , J 0 K^{0}_{\mathrm{col},J} K col , J 0 are the column ceilings of the perturbed and unperturbed maps. The fixed allocations of dependence neighborhoods and exponents are already included in C J C_J C J . More explicitly, the margin in the coefficient exponent, after the transfer of boundary distances in (21.10), leaves a positive exponent c ′ c' c ′ for summing over the incoming label at a fixed output, and this sum is bounded by v J ∑ q ≥ 0 ( q + 2 ) 3 e − c ′ q < ∞ v_J\sum_{q\geq0}(q+2)^3e^{-c'q}<\infty v J ∑ q ≥ 0 ( q + 2 ) 3 e − c ′ q < ∞ . The volume factors of the lattices where the labels were created are kept in this column sum. Deleting labels or adding excluded regions cannot increase this bound of bounded density. This proves that the enlargement has the dependence stated at the beginning of this section. We fix this enlarged ceiling on the whole larger perturbation domain before choosing s e r s_{\mathrm{er}} s er or later widths. In Theorem 20.10 , its coefficient is K J = 4 ( log 2 ) K c o l , J K_J=4(\log2)K_{\mathrm{col},J} K J = 4 ( log 2 ) K col , J for this enlarged ceiling; no column bound for the unperturbed map alone is substituted.
Applying the generated-increment comparison of Theorem 20.13 with the incoming Δ Λ t = Λ t − Λ t T \Delta\Lambda_t=\Lambda_t-\Lambda_t^{\mathsf T} Δ Λ t = Λ t − Λ t T , we obtain the increment in part (4), including the initial freezings. To check the sign of the dimension term, note that one scalar Gaussian coordinate integrated in the corresponding renormalization group recursion on the whole torus contributes a factor g i g_i g i . When that coordinate is absent from the regional free list, the logarithmic difference between the regional and torus terms is − log g i = log ( 1 / g i ) -\log g_i=\log(1/g_i) − log g i = log ( 1/ g i ) . Summing over the missing scalar coordinates, we obtain d j ( H t ) log ( 1 / g i ) d_j(H_t)\log(1/g_i) d j ( H t ) log ( 1/ g i ) , once for each scale. By that theorem, the remaining generated differences of Haar disintegration factors, minimizers, and cluster expansions, and the freezings, are bounded by C J C_J C J times the three label counts, taken on the lattices where the labels were created. This comparison is additional analytic information: neither the exact disintegration, nor (21.37), nor normalization would by itself bound these generated differences. The factor A t A_t A t remains separate and is not added again to Λ t \Lambda_t Λ t .
The remaining averaging steps. Let a = ℓ T − ℓ 0 = J T a=\ell_T-\ell_0=JT a = ℓ T − ℓ 0 = J T be the number of integrated averaging steps (here a a a is not a mesh), and let ℓ max \ell_{\max} ℓ m a x be the last level of the given finite hierarchy. A remaining number 0 ≤ u p h < J 0\leq u_{\mathrm{ph}}<J 0 ≤ u ph < J of averaging steps is chosen only when it is available: at the stopping scale we require
m s t o p = M u p h m r e m , m r e m ∈ 4 N , ℓ T + u p h ≤ ℓ max . m_{\mathrm{stop}}=M^{u_{\mathrm{ph}}}m_{\mathrm{rem}},\qquad m_{\mathrm{rem}}\in4\mathbb{N},\qquad\ell_T+u_{\mathrm{ph}}\leq\ell_{\max}. m stop = M u ph m rem , m rem ∈ 4 N , ℓ T + u ph ≤ ℓ m a x .
For a fixed additional number r s t o p ∈ N 0 r_{\mathrm{stop}}\in\mathbb{N}_0 r stop ∈ N 0 of RG steps, we set v = r s t o p J + u p h v=r_{\mathrm{stop}}J+u_{\mathrm{ph}} v = r stop J + u ph and require
m s t o p = M v m r e m , m r e m ∈ 4 N , ℓ T + v ≤ ℓ max ; m_{\mathrm{stop}}=M^v m_{\mathrm{rem}},\qquad m_{\mathrm{rem}}\in4\mathbb{N},\qquad\ell_T+v\leq\ell_{\max}; m stop = M v m rem , m rem ∈ 4 N , ℓ T + v ≤ ℓ m a x ;
the case of u p h u_{\mathrm{ph}} u ph alone is r s t o p = 0 r_{\mathrm{stop}}=0 r stop = 0 . For each 0 ≤ k ≤ v 0\leq k\leq v 0 ≤ k ≤ v , the intermediate side is M v − k m r e m ∈ 4 N M^{v-k}m_{\mathrm{rem}}\in4\mathbb{N} M v − k m rem ∈ 4 N , which is at least four. The initial side is M a m s t o p M^a m_{\mathrm{stop}} M a m stop , and the bound on the index keeps every intermediate level in the given hierarchy. Hence both T M a T_M^a T M a at the stopping scale and T M a + v T_M^{a+v} T M a + v at the initial scale exist; a lower bound on m s t o p m_{\mathrm{stop}} m stop alone would not give these integer conditions. We work on whole configurations, before regional coordinates are introduced. The block averaging map with its Borel extension is Borel, and its iterates are compositions of that map, so for every initial configuration V V V , we have
F ( T M a + v V ) = ( F ∘ T M v ) ( T M a V ) . F(T_M^{a+v}V)=(F\circ T_M^v)(T_M^aV). F ( T M a + v V ) = ( F ∘ T M v ) ( T M a V ) .
and the function at the stopping scale is exactly C ↦ F ( T M v C ) C\mapsto F(T_M^vC) C ↦ F ( T M v C ) , in particular C ↦ F ( T M u p h C ) C\mapsto F(T_M^{u_{\mathrm{ph}}}C) C ↦ F ( T M u ph C ) for u p h u_{\mathrm{ph}} u ph alone. We define the measure μ a , X = ( T M a ) # ( 1 A X ν ) \mu_{a,X}=(T_M^a)_\#(1_{A_X}\nu) μ a , X = ( T M a ) # ( 1 A X ν ) on the whole configurations at the stopping scale. For every nonnegative Borel F F F we have
∫ 1 A X ( V ) F ( T M a + v V ) d ν ( V ) = ∫ ( F ∘ T M v ) ( C ) d μ a , X ( C ) , (143) \int1_{A_X}(V)F(T_M^{a+v}V)\,d\nu(V)=\int(F\circ T_M^v)(C)\,d\mu_{a,X}(C), \tag*{(143)} ∫ 1 A X ( V ) F ( T M a + v V ) d ν ( V ) = ∫ ( F ∘ T M v ) ( C ) d μ a , X ( C ) , (143)
by the definition of the pushforward; the identity extends to integrable F F F through positive and negative parts. Each fixed additional number of steps uses the total exponent v v v whose existence was just checked, together with its finite enlarged dependence neighborhood. For the local link and plaquette observables of Theorem 21.5 , these iterates use its already enlarged final dependence neighborhood. The decomposition keeps this observable on each piece. There is no Haar disintegration, change of variables, restoration of bounds, or update of normalization for v v v . When the integrated part of the decomposition is empty, the identity is the initial pushforward itself.
Part (5): external classical factors and failure indicators. At fixed frozen coordinates, let S e x t ( b r e c ) S_{\mathrm{ext}}(b_{\mathrm{rec}}) S ext ( b rec ) be the sum of the protected external classical exponents with their original couplings; for the Wilson measure it includes E h o l e / g b a s e 2 E_{\mathrm{hole}}/g_{\mathrm{base}}^2 E hole / g base 2 . The term g t 2 S e x t g_t^2S_{\mathrm{ext}} g t 2 S ext is constant in all integration variables, so adding it to the full action of the block region changes no minimizer, Hessian, map, or nonconstant norm. We have the exact identity
e − S e x t e − E i n t / g t 2 = e − ( E i n t + g t 2 S e x t ) / g t 2 . e^{-S_{\mathrm{ext}}}e^{-E^{\mathrm{int}}/g_t^2}=e^{-(E^{\mathrm{int}}+g_t^2S_{\mathrm{ext}})/g_t^2}. e − S ext e − E int / g t 2 = e − ( E int + g t 2 S ext ) / g t 2 .
We delete this external factor from W t W_t W t when the full classical action including the frozen-coordinate terms is used. After an RG step, the corresponding constant becomes M t s t g t 2 S e x t = g t + 1 2 S e x t M_t^{st}g_t^2S_{\mathrm{ext}}=g_{t+1}^2S_{\mathrm{ext}} M t s t g t 2 S ext = g t + 1 2 S ext , so its exponent is preserved exactly. At each classical freezing, the full value was kept. Induction over the finite decomposition shows that each frozen-coordinate term of the full classical action has one copy in the output action and none in the remaining weight of the frozen coordinates. The constants from the freezing of the effective interaction remain separate exact terms of Λ \Lambda Λ .
For each failing plaquette f f f , the whole dependence set was frozen at the input of its RG step. Its value q f q_f q f is therefore a Borel function of the frozen coordinates on that lattice, computed with the same path products of the block averaging map, and later operations fix its marked data exactly. Its trace is invariant under the consistent endpoint action. We keep the product of all indicators { q f ≥ d ∗ } \{q_f\ge d_*\} { q f ≥ d ∗ } as I χ a c t I_\chi^{\mathrm{act}} I χ act , and any chosen invariant tests depending only on the frozen coordinates at the output as I χ i n v , e n d I_\chi^{\mathrm{inv,end}} I χ inv , end . Every other optional coordinate restriction is between zero and one and may be removed in an upper bound. After the classical recombination, this leaves exactly the measure on frozen coordinates in part (5), which is bounded by its product Haar measure. It contains no second copy of an energy that could be used for a lower bound on the output action or for an energy bound on failing plaquettes.
Every local bound has the dependence stated at the beginning of this section on the fixed larger domains, and only C δ C_\delta C δ has the dependence of item (4) of Definition 4.14. □ \square □
Exponential-moment bounds between renormalization scales
In this section, we prove exponential-moment bounds for sums of plaquette energies of block-averaged configurations. The bounds are uniform in the number of averaging steps between the lattice on which the measure is given and the lattice on which the energies are evaluated. The main result is Theorem 22.1. Under its hypotheses, fix a coupling ceiling Γ > 0 \Gamma> 0 Γ > 0 and take the smallest side u ≥ n min ( Γ ) u \ge n_{\min}(\Gamma) u ≥ n m i n ( Γ ) . Let b ∈ N 0 b \in\mathbb{N}_0 b ∈ N 0 with 0 ≤ b ≤ j − 1 0 \le b \le j-1 0 ≤ b ≤ j − 1 , and let ν \nu ν be either the Wilson measure on the lattice of side n b n_b n b or a normalized stopped measure there at a preceding depth in N 0 \mathbb{N}_0 N 0 . The theorem states that the exponential moments under ν \nu ν of the sum Q j T Q_j^T Q j T of the energies of the predecessor plaquettes of a chessboard tiling, (145) , are bounded by exp { 3 C v n j − 1 3 + C s n j − 1 2 log ( 1 / g n j − 1 ) + C b } \exp\{3C_v n_{j-1}^3+C_s n_{j-1}^2\log(1/g_{n_{j-1}})+C_b\} exp { 3 C v n j − 1 3 + C s n j − 1 2 log ( 1/ g n j − 1 ) + C b } . The constants are independent of γ ∈ ( 0 , Γ ] \gamma\in(0,\Gamma] γ ∈ ( 0 , Γ ] , of the lattice sides, and of j − 1 − b j-1-b j − 1 − b , and the expectation contains no small-field or regularity indicator. We also prove a chessboard estimate in which the regularity indicators stay inside the root (Lemma 22.11). Finally, we compare the Wilson measure, restricted to the first a a a conditions of the small-field event, with the stopped measures; this bounds the probability that the next condition fails, without conditioning on later scales (Lemmas 22.12 and 22.13). The inputs are the large-field decomposition and the density bounds of Section 21 (Theorems 21.5, 21.10, and 21.13); the regional integration estimates of § 2.2 (Theorems 20.13 and 20.14 and Lemma 20.5); the classical lower bound of Theorems 11.2 and 11.6; the energy transport and predecessor geometry of Theorem 7.3 and Lemma 7.4; the reference normalization of Theorem 14.8; the chessboard estimate of Theorem 6.6; and, for the comparison of restrictions, Theorems 7.7 and 14.7. The results are used in Section 23 to prove the bound on large fields, Theorem 2.7.
Section 22.1 states the main theorem and outlines its proof, Sections 22.2 to 22.6 prove the estimates for one large-field decomposition, Section 22.7 completes the proof, and Sections 22.8 and 22.9 contain the chessboard estimate and the comparison of restrictions.
Setting and the main estimate
Scales and couplings. Fix L > 0 L > 0 L > 0 , an odd integer M ≥ 3 M \ge3 M ≥ 3 , an integer J ≥ 1 J \ge1 J ≥ 1 , and B = M J B = M^J B = M J . For u ∈ 4 N u \in4\mathbb{N} u ∈ 4 N and K ∈ N 0 K \in\mathbb{N}_0 K ∈ N 0 let N = u B K N = uB^K N = u B K , n i = N / M i n_i = N/M^i n i = N / M i , and V i = T M i U V^i = T_M^i U V i = T M i U for 0 ≤ i ≤ J K 0 \le i \le JK 0 ≤ i ≤ J K , where T M T_M T M is the block averaging map of Definition 4.3, including its Borel extension at the exceptional values. Increasing i i i moves toward the infrared, and all the sides n i n_i n i are divisible by four. One application of T M T_M T M is an averaging step, and J J J consecutive averaging steps form an RG step. The levels i ∈ J N 0 i \in J\mathbb{N}_0 i ∈ J N 0 are the boundaries of RG steps. At side n n n the coupling is h n = g n 2 = γ / n h_n = g_n^2 = \gamma/n h n = g n 2 = γ / n , as in Definition 4.1, so that h n i + 1 = M h n i h_{n_{i+1}} = M h_{n_i} h n i + 1 = M h n i . All Haar measures in this section are probability measures.
Plaquette energies. For A ∈ S U ( 2 ) A \in\mathrm{SU}(2) A ∈ SU ( 2 ) let ρ G ( A ) ∈ [ 0 , 2 π ] \rho_G(A) \in[0,2\pi] ρ G ( A ) ∈ [ 0 , 2 π ] be the distance from A A A to I I I in the bi-invariant metric of Section 4, and recall that q ( A ) = 1 − ( tr A ) / 2 q(A) = 1 - (\operatorname{tr} A)/2 q ( A ) = 1 − ( tr A ) /2 . On the whole group,
ρ G ( A ) 2 2 π 2 ≤ q ( A ) ≤ ρ G ( A ) 2 8 , 0 ≤ q ( A ) ≤ 2. (144) \frac{\rho_G(A)^2}{2\pi^2} \leq q(A) \leq\frac{\rho_G(A)^2}{8}, \qquad0 \leq q(A) \leq2. \tag*{(144)} 2 π 2 ρ G ( A ) 2 ≤ q ( A ) ≤ 8 ρ G ( A ) 2 , 0 ≤ q ( A ) ≤ 2. (144)
Indeed, if t = ρ G ( A ) / 4 ∈ [ 0 , π / 2 ] t = \rho_G(A)/4 \in[0,\pi/2] t = ρ G ( A ) /4 ∈ [ 0 , π /2 ] , then q ( A ) = 1 − cos ( 2 t ) = 2 sin 2 t q(A) = 1 - \cos(2t) = 2\sin^2 t q ( A ) = 1 − cos ( 2 t ) = 2 sin 2 t by (4.2), and 2 t / π ≤ sin t ≤ t 2t/\pi\leq\sin t \leq t 2 t / π ≤ sin t ≤ t . For a set P \mathcal{P} P of positive plaquettes at level i i i and a configuration V V V on the lattice of side n i n_i n i we write
S i P ( V ) = 4 h n i ∑ p ∈ P q ( V p ) . S_i^{\mathcal{P}}(V) = \frac{4}{h_{n_i}} \sum_{p \in\mathcal{P}} q(V_p). S i P ( V ) = h n i 4 p ∈ P ∑ q ( V p ) .
If P \mathcal{P} P is the set of all positive plaquettes, then S i P ( V ) = S W ( V ) / h n i S_i^{\mathcal{P}}(V) = S_W(V)/h_{n_i} S i P ( V ) = S W ( V ) / h n i is the Wilson exponent.
The plaquette-energy sum. For a set A \mathcal{A} A of positive plaquettes at level i + 1 i+1 i + 1 , let Pa ( A ) \operatorname{Pa}(\mathcal{A}) Pa ( A ) be the union, counted once, of the plaquettes at level i i i in the complete predecessor rectangles of the elements of A \mathcal{A} A (Theorem 7.3). Let 1 ≤ j ≤ J K 1 \leq j \leq JK 1 ≤ j ≤ J K . For each of the eight tilings T \mathcal{T} T of the level-j j j lattice into cells of 2 × 2 × 2 2 \times2 \times2 2 × 2 × 2 sites in Lemma 7.4, and each cell C ∈ T C \in\mathcal{T} C ∈ T , let P j − 1 ( C ) \mathcal{P}_{j-1}(C) P j − 1 ( C ) be the union of the complete predecessor rectangles of the six internal level-j j j plaquettes of C C C , and let P j − 1 T = ⋃ C ∈ T P j − 1 ( C ) \mathcal{P}_{j-1}^{\mathcal{T}} = \bigcup_{C \in\mathcal{T}} \mathcal{P}_{j-1}(C) P j − 1 T = ⋃ C ∈ T P j − 1 ( C ) , counted once. By Lemma 7.4, ∣ P j − 1 ( C ) ∣ ≤ 24 M 3 |\mathcal{P}_{j-1}(C)| \leq24M^3 ∣ P j − 1 ( C ) ∣ ≤ 24 M 3 . For a configuration V V V on the lattice of side n j − 1 n_{j-1} n j − 1 we set
Q j T ( V ) = 4 h n j − 1 ∑ p ∈ P j − 1 T q ( V p ) = S j − 1 P j − 1 T ( V ) . (145) Q_j^{\mathcal{T}}(V) = \frac{4}{h_{n_{j-1}}} \sum_{p \in\mathcal{P}_{j-1}^{\mathcal{T}}} q(V_p) = S_{j-1}^{\mathcal{P}_{j-1}^{\mathcal{T}}}(V). \tag*{(145)} Q j T ( V ) = h n j − 1 4 p ∈ P j − 1 T ∑ q ( V p ) = S j − 1 P j − 1 T ( V ) . (145)
This is the function that enters the chessboard estimate of Section 22.8.
Initial measures. Every estimate of this section starts from a probability measure ν \nu ν on the configurations of some side n b n_b n b of the hierarchy, which we call the initial measure. It is one of the following.
The Wilson measure ν = μ n b , γ \nu= \mu_{n_b,\gamma} ν = μ n b , γ , with partition function Z ν = Z n b , γ Z_\nu= Z_{n_b,\gamma} Z ν = Z n b , γ . It is not restricted to D r D_r D r .
A normalized stopped measure at a depth d 0 ∈ J N 0 d_0 \in J\mathbb{N}_0 d 0 ∈ J N 0 ,
d ν = Z ν − 1 1 D r e − E d 0 , s a t / h n b − R d 0 , s a t d H n b , Z ν = ∫ D r e − E d 0 , s a t / h n b − R d 0 , s a t d H n b . d\nu= Z_\nu^{-1}\mathbf{1}_{D_r}e^{-E_{d_0,\mathrm{sat}}/h_{n_b}-R_{d_0,\mathrm{sat}}}\,dH_{n_b}, \qquad Z_\nu= \int_{D_r} e^{-E_{d_0,\mathrm{sat}}/h_{n_b}-R_{d_0,\mathrm{sat}}}\,dH_{n_b}. d ν = Z ν − 1 1 D r e − E d 0 , sat / h n b − R d 0 , sat d H n b , Z ν = ∫ D r e − E d 0 , sat / h n b − R d 0 , sat d H n b .
as in Definition 4.4 and Theorem 14.6. Here E d , s a t E_{d,\mathrm{sat}} E d , sat is the extension of the effective classical action E d E_d E d to D ρ D_\rho D ρ by gauge invariance (Theorem 11.6), and R d , s a t R_{d,\mathrm{sat}} R d , sat is the corresponding effective interaction. The partition function Z ν Z_\nu Z ν is not centered: the value R d 0 ( I ) R_{d_0}(I) R d 0 ( I ) is kept in it.
Large-field decompositions. We use the partition of Theorem 21.5 . Given an initial measure ν \nu ν and a stopping level, the support of ν \nu ν is partitioned into finitely many Borel sets A X \mathcal{A}_{\mathcal{X}} A X , indexed by large-field decompositions X \mathcal{X} X . A decomposition is produced by testing the plaquettes of the iterates T M ℓ V T^\ell_M V T M ℓ V in increasing order of the level ℓ \ell ℓ , and in a fixed order within one level. A plaquette f f f at level ℓ f \ell_f ℓ f is a recorded bad plaquette of X \mathcal{X} X if q f ≥ d ∗ q_f \ge d_* q f ≥ d ∗ and its full dependence set avoids the set excluded before it. The variables of that dependence set that are not yet frozen are then frozen at the beginning of the RG step containing the level ℓ f \ell_f ℓ f , together with a neighborhood, the excluded neighborhood of f f f , which is enlarged at each later RG step. We write n f = n ℓ f n_f=n_{\ell_f} n f = n ℓ f for the side on which f f f is recorded and x f = n f / γ x_f=n_f/\gamma x f = n f / γ . Away from the excluded regions, the renormalization group is run on block regions, in RG steps of J J J averaging steps each, from the initial level to the stopping level; both levels are boundaries of RG steps. At the stopping level, of side m m m , the decomposition leaves a stopping field c ∈ S U ( 2 ) F c\in\mathrm{SU}(2)^F c ∈ SU ( 2 ) F on the set F F F of links of the block region at the stopping level. We write H = E m ∖ F H=\mathcal{E}_m\setminus F H = E m ∖ F for the other links of the stopping lattice, including the links across the boundary of the block region. For z ∈ S U ( 2 ) H z\in\mathrm{SU}(2)^H z ∈ SU ( 2 ) H , we write c ⊕ z c\oplus z c ⊕ z for the configuration on the stopping lattice that equals c c c on F F F and z z z on H H H , and we call it a completion of c c c . The frozen coordinates b r e c b_{\mathrm{rec}} b rec consist of all frozen group variables, each kept on the lattice of the scale at which it was frozen. The map V ↦ O T X ( V ) = ( c , b r e c ) V\mapsto O^\mathcal{X}_T(V)=(c,b_{\mathrm{rec}}) V ↦ O T X ( V ) = ( c , b rec ) records the stopping field and the frozen coordinates. For a nonnegative function Q Q Q and s ≥ 0 s\ge0 s ≥ 0 , we write
η X = ( O T X ) # ( 1 A X ν ) , η X s = ( O T X ) # ( e s Q 1 A X ν ) . \eta_{\mathcal{X}}=(O^\mathcal{X}_T)_{\#}(1_{\mathcal{A}_{\mathcal{X}}}\nu),\qquad
\eta^s_{\mathcal{X}}=(O^\mathcal{X}_T)_{\#}(e^{sQ}1_{\mathcal{A}_{\mathcal{X}}}\nu). η X = ( O T X ) # ( 1 A X ν ) , η X s = ( O T X ) # ( e s Q 1 A X ν ) .
For a Wilson initial measure, the first RG step of Theorem 21.5 separates the frozen group variables adjacent to the block region from those deep inside the excluded regions. The latter are arbitrary elements of S U ( 2 ) \mathrm{SU}(2) SU ( 2 ) ; they are never inserted into an analytic function and enter only through an exact Wilson factor.
The analytic estimates are applied at gauge representatives of the stopping field and the frozen coordinates that satisfy the domain conditions and the reconstruction identities of Theorem 21.13 . They are obtained from the representatives with small frozen coordinates at the stopping level by applying the reconstruction maps of the successive RG steps, and they also satisfy the identities for the insertion of frozen coordinates stated there. Below, we call them representatives satisfying the reconstruction identities.
Table 22.1 lists the measures that occur in the proof of Theorem 22.1 , with their total masses and normalizations. The measures η X \eta_{\mathcal{X}} η X , η X s \eta^s_{\mathcal{X}} η X s , and ζ ~ T \widetilde{\zeta}_T ζ T are not renormalized; they keep the partition function Z ν Z_\nu Z ν of the initial measure.
Table 22.1. Measures used in the proof of Theorem 22.1 .
Dependence of constants. The constants of the local analytic estimates have the dependence described in item (3) of Definition 4.14. They depend only on M M M , J J J , the larger construction domains and their radius ratios, the spatial and boundary decay rates, the density of variables, and the norm bounds. They are independent of L L L , γ \gamma γ , the lattice sides, the depth, the number of RG steps since a frozen coordinate was created, the shape and topology of the block regions, and the real values of the frozen coordinates in their domain. Where stated, they are also independent of the widths of the boundary layers and excluded neighborhoods chosen later; these widths enter only through explicit counts. The constants in the probability estimates have the dependence described in item (4) of Definition 4.14. They may in addition depend on L L L , a fixed coupling ceiling, the fixed parameters, and the finite family of stopping offsets, but not on the coupling below the ceiling, the sides, the depth, or the geometry, age, and frozen values of a decomposition. We claim no useful bound as J → ∞ J \to\infty J → ∞ .
Thresholds. Fix d c u t > 0 d_{\mathrm{cut}}>0 d cut > 0 and set
d ∗ = d c u t / 2 , c E = 1 + C M π 2 d c u t , d_* = d_{\mathrm{cut}}/2,\qquad c_E = 1+C_M\pi\sqrt{2d_{\mathrm{cut}}}, d ∗ = d cut /2 , c E = 1 + C M π 2 d cut ,
where δ ∗ \delta_* δ ∗ and C M C_M C M are the constants of Theorem 22.4 (a). For the chessboard estimate of Lemma 22.11 we assume π 2 d c u t ≤ δ ∗ \pi\sqrt{2d_{\mathrm{cut}}} \le\delta_* π 2 d cut ≤ δ ∗ . In Theorem 22.1 , these numbers enter only through the range of s s s and the definition of the recorded bad plaquettes.
The main estimate. The constants K M , K 1 , κ ∗ , κ 0 , c A a c t , α W K_M,K_1,\kappa_*,\kappa_0,c_A^{\mathrm{act}},\alpha_W K M , K 1 , κ ∗ , κ 0 , c A act , α W in the following theorem are defined in Theorem 22.4 ; they depend only on M M M and on the fixed classical data.
Theorem 22.1. Fix L > 0 L > 0 L > 0 , an odd integer M ≥ 3 M \ge3 M ≥ 3 , an integer J ≥ 1 J \ge1 J ≥ 1 , and B = M J B = M^J B = M J , and fix parameters satisfying the hypotheses of Theorems 20.14, 21.5, and 21.13. These include the coefficient-norm bounds at initialization, both substitution bounds (22.6), the boundary-distance comparison, and the product domains verified in Lemma 22.3. The cutoff amplitude A W A_{\mathrm{W}} A W is fixed before the coupling ceiling g ∗ < 1 g_* < 1 g ∗ < 1 . Fix a finite nonempty set R ⊂ N 0 \mathcal{R} \subset\mathbb{N}_0 R ⊂ N 0 of stopping offsets, all remainders 0 ≤ u p h < J 0 \le u_{\mathrm{ph}} < J 0 ≤ u ph < J , the buffers containing their full dependence sets, and the bound F J F_J F J of Theorem 21.5 for the number of affected source plaquettes. Fix χ ∗ ≥ 1 \chi_* \ge1 χ ∗ ≥ 1 such that the ratio of every stopping side of this family to the requested side n j − 1 n_{j-1} n j − 1 is at most χ ∗ \chi_* χ ∗ , and such that n b / n j − 1 ≤ χ ∗ n_b/n_{j-1} \le\chi_* n b / n j − 1 ≤ χ ∗ whenever the stopping level lies at or below the initial level. Every iterate of T M T_M T M that is used must belong to the given finite hierarchy, and only full RG steps are integrated.
Let K M , K 1 , κ ∗ , κ 0 , c A a c t , α W K_M,K_1,\kappa_*,\kappa_0,c_A^{\mathrm{act}},\alpha_W K M , K 1 , κ ∗ , κ 0 , c A act , α W be as in Theorem 22.4, and put K W = 48 K 1 / α W K_W = 48K_1/\alpha_W K W = 48 K 1 / α W and θ = 1 / 2 \theta= 1/2 θ = 1/2 . Choose s 0 > 0 s_0 > 0 s 0 > 0 , after F J F_J F J , such that
s 0 ≤ min { 1 2 , κ ∗ 2 c E − ( J − 1 ) , 1 4 K W , θ c A a c t d ∗ 16 F J , κ 0 2 K M } . s_0 \le\min\left\{\frac{1}{2},\frac{\kappa_*}{2}c_E^{-(J-1)},\frac{1}{4K_W},\frac{\theta c_A^{\mathrm{act}}d_*}{16F_J},\frac{\kappa_0}{2K_M}\right\}. s 0 ≤ min { 2 1 , 2 κ ∗ c E − ( J − 1 ) , 4 K W 1 , 16 F J θ c A act d ∗ , 2 K M κ 0 } . Any further fixed decrease of s 0 s_0 s 0 is allowed. Then for every 0 < Γ < ∞ 0 < \Gamma< \infty 0 < Γ < ∞ there exist n min ∈ 4 N n_{\min} \in4\mathbb{N} n m i n ∈ 4 N and finite nonnegative constants C v , C s , C b C_v,C_s,C_b C v , C s , C b with C s ≥ 9 χ ∗ 2 C_s \ge9\chi_*^2 C s ≥ 9 χ ∗ 2 , chosen before the coupling, the side, and the depth, such that the following holds.
Let 0 < γ ≤ Γ 0 < \gamma\le\Gamma 0 < γ ≤ Γ , u ∈ 4 N u \in4\mathbb{N} u ∈ 4 N with u ≥ n min u \ge n_{\min} u ≥ n m i n , K ∈ N 0 K \in\mathbb{N}_0 K ∈ N 0 , N = u B K N = uB^K N = u B K , n i = N / M i n_i = N/M^i n i = N / M i , 1 ≤ j ≤ J K 1 \le j \le JK 1 ≤ j ≤ J K , and b ∈ J N 0 b \in J\mathbb{N}_0 b ∈ J N 0 with 0 ≤ b ≤ j − 1 0 \le b \le j-1 0 ≤ b ≤ j − 1 . Let ν \nu ν be either the Wilson measure μ n b , γ \mu_{n_b,\gamma} μ n b , γ or a normalized stopped measure on U n b U_{n_b} U n b at a depth a J aJ a J , a ∈ N 0 a \in\mathbb{N}_0 a ∈ N 0 . Then, for each of the eight tilings T \mathcal{T} T and every 0 ≤ s ≤ s 0 0 \le s \le s_0 0 ≤ s ≤ s 0 ,
log E ν exp { s Q j T ( T M j − 1 − b V ) } ≤ 3 C v n j − 1 3 + C s n j − 1 2 log ( 1 / g n j − 1 ) + C b . \log\mathbb{E}_{\nu}\exp\{sQ^{\mathcal{T}}_j(T_M^{j-1-b}V)\}
\le3C_v n_{j-1}^3 + C_s n_{j-1}^2\log(1/g_{n_{j-1}}) + C_b. log E ν exp { s Q j T ( T M j − 1 − b V )} ≤ 3 C v n j − 1 3 + C s n j − 1 2 log ( 1/ g n j − 1 ) + C b . This covers every remainder u p h u_{\mathrm{ph}} u ph , stopping levels at or below the initial level, an empty block region at the stopping level, and decompositions whose block region becomes empty before the stopping level and which are continued by empty RG steps.
More precisely, a recorded bad plaquette f f f created at side n f n_f n f contributes at least P 0 x f P_0x_f P 0 x f to the classical part of the exponent, where x f = n f / γ x_f = n_f/\gamma x f = n f / γ and P 0 = 4 θ c A a c t d ∗ P_0 = 4\theta c_A^{\mathrm{act}}d_* P 0 = 4 θ c A act d ∗ . The plaquette-energy sum costs at most 8 s F J x f ≤ P 0 x f / 4 8sF_Jx_f \le P_0x_f/4 8 s F J x f ≤ P 0 x f /4 per recorded bad plaquette, and all the corrections assigned to f f f , from the normalization constants, the missing coordinates, the boundary bands, the substitutions, and the integration over the omitted links, are at most
C c h , J Γ ( 1 + log x f ) 2 ≤ P 0 x f / 4. C_{\mathrm{ch},J}^{\Gamma}(1+\log x_f)^2 \le P_0x_f/4. C ch , J Γ ( 1 + log x f ) 2 ≤ P 0 x f /4. Here one may take C c h , J Γ = ( 1 + log + Γ ) C c h , J b a s e C_{\mathrm{ch},J}^{\Gamma}=(1+\log_+\Gamma)C_{\mathrm{ch},J}^{\mathrm{base}} C ch , J Γ = ( 1 + log + Γ ) C ch , J base , where C c h , J b a s e C_{\mathrm{ch},J}^{\mathrm{base}} C ch , J base depends only on the fixed parameters and the family of stopping offsets. Each decomposition retains the factor ∏ f e − a a c t n f / γ \prod_f e^{-a_{\mathrm{act}}n_f/\gamma} ∏ f e − a act n f / γ , with a a c t = P 0 / 2 a_{\mathrm{act}}=P_0/2 a act = P 0 /2 , and the sum of these factors over all decompositions is at most
exp { ∑ ℓ 3 n ℓ 3 e − a a c t n ℓ / γ } , \exp\left\{\sum_{\ell}3n_{\ell}^3e^{-a_{\mathrm{act}}n_{\ell}/\gamma}\right\}, exp { ℓ ∑ 3 n ℓ 3 e − a act n ℓ / γ } ,
which is bounded uniformly in the depth once u ≥ n min u \ge n_{\min} u ≥ n m i n . At a stopping side m = χ n j − 1 m = \chi n_{j-1} m = χ n j − 1 one has g m = g n j − 1 / χ g_m = g_{n_{j-1}}/\sqrt{\chi} g m = g n j − 1 / χ , and the only surface term is 9 χ 2 n j − 1 2 log ( 1 / g n j − 1 ) 9\chi^2 n_{j-1}^2 \log(1/g_{n_{j-1}}) 9 χ 2 n j − 1 2 log ( 1/ g n j − 1 ) .
The constants n min , C v , C s , C b n_{\min}, C_v, C_s, C_b n m i n , C v , C s , C b have the dependence described in item (4) of Definition 4.14, with the fixed ceiling Γ \Gamma Γ , the fixed parameters, and the finite family of stopping offsets and remainders. They are independent of the coupling γ ∈ ( 0 , Γ ] \gamma\in(0,\Gamma] γ ∈ ( 0 , Γ ] , the sides, the depth, the position and the length j − 1 − b j-1-b j − 1 − b of the range of scales, and the age, shape, topology, and frozen values of the decompositions. The constants of the local analytic estimates keep the dependence described in item (3) of Definition 4.14; the widths enter only through the displayed counts.
Remark 22.2. There is no restriction on j − b j-b j − b , and the expectation contains no indicator of a cell condition or of a condition of the small-field event. Taking b = 0 b=0 b = 0 and ν = μ N , γ \nu=\mu_{N,\gamma} ν = μ N , γ gives the estimate under the Wilson measure with its own partition function. The proof does not use the bounds on the Wilson measure at successive scales of Section 23, where the theorem is used in Theorems 23.2 and 23.6.
Outline of the proof. Fix j j j , b b b , and T \mathcal{T} T , and write n = n j − 1 n=n_{j-1} n = n j − 1 . If the planned stopping level lies at or below the initial level, then only a bounded number of averaging steps separates the initial lattice from level j − 1 j-1 j − 1 , and the estimate follows from the partition-function bounds on the initial lattice. Otherwise we integrate full RG steps from level b b b to a stopping level t s t o p ≤ j − 1 t_{\mathrm{stop}}\le j-1 t stop ≤ j − 1 and keep the remaining v = j − 1 − t s t o p v=j-1-t_{\mathrm{stop}} v = j − 1 − t stop averaging steps inside the function Q j T ∘ T M T Q_j^{\mathcal{T}}\circ T_M^{\mathcal{T}} Q j T ∘ T M T .
Step 1: classical lower bounds. Along a decomposition X \mathcal{X} X , the density of η X \eta_{\mathcal{X}} η X at a representative satisfying the reconstruction identities is bounded by e − E T / g m 2 e^{-E_{\mathcal{T}}/g_m^2} e − E T / g m 2 times normalization factors, where E T E_{\mathcal{T}} E T is the effective classical action of the stopping field with all external classical exponents included (Theorem 21.13 ). We reconstruct one configuration U ∗ U_* U ∗ on the initial lattice with zero fluctuations. Using an energy transport estimate that is uniform in the number of averaging steps, we show that E T / g m 2 E_{\mathcal{T}}/g_m^2 E T / g m 2 dominates both Ψ F ( c ) / g m 2 \Psi_F(c)/g_m^2 Ψ F ( c ) / g m 2 , where Ψ F ( c ) \Psi_F(c) Ψ F ( c ) is the minimum of α W S W ( c ⊕ z ) \alpha_W S_W(c\oplus z) α W S W ( c ⊕ z ) over all values z z z of the links in H H H , and A X = c A a c t ∑ f 4 q f / h n f ≥ 4 c A a c t d ∗ ∑ f x f A_{\mathcal{X}}=c_A^{\mathrm{act}}\sum_f 4q_f/h_{n_f}\ge4c_A^{\mathrm{act}}d_*\sum_f x_f A X = c A act ∑ f 4 q f / h n f ≥ 4 c A act d ∗ ∑ f x f (Theorem 22.4 ). Since both lower bounds are gauge invariant, the density bound extends to all points (Theorem 22.10).
Step 2: the plaquette-energy sum. At every point of A X A_{\mathcal{X}} A X , the plaquette-energy sum is at most K W Ψ F ( c ) / g m 2 + 8 F W ∑ f x f K_W\Psi_F(c)/g_m^2+8F_W\sum_f x_f K W Ψ F ( c ) / g m 2 + 8 F W ∑ f x f (Lemma 22.8). Multiplication by e s Q e^{sQ} e s Q therefore leaves the Wilson factor e − λ s Ψ F ( c ) / g m 2 e^{-\lambda_s\Psi_F(c)/g_m^2} e − λ s Ψ F ( c ) / g m 2 with λ s = 1 / 2 − s K W ≥ 1 / 4 \lambda_s=1/2-sK_W\ge1/4 λ s = 1/2 − s K W ≥ 1/4 , bounded above by e − Ψ F ( c ) / ( 4 g m 2 ) e^{-\Psi_F(c)/(4g_m^2)} e − Ψ F ( c ) / ( 4 g m 2 ) , together with e 8 s F W ∑ f x f e^{8sF_W\sum_f x_f} e 8 s F W ∑ f x f . Replacing the minimum by a Haar integral over the links in H H H adds 3 ∣ H ∣ log ( 1 / ( c ∗ g m ) ) + 8 λ s α W ∣ H ∣ 3|H|\log(1/(c_*g_m))+8\lambda_s\alpha_W|H| 3∣ H ∣ log ( 1/ ( c ∗ g m )) + 8 λ s α W ∣ H ∣ to the exponent (Lemma 22.7).
Step 3: summation over decompositions. Each recorded bad plaquette contributes − P 0 x f -P_0x_f − P 0 x f to the exponent. All normalization corrections, boundary terms, substitution terms, and the correction from the integration over the omitted links can be assigned to recorded bad plaquettes, and the corrections assigned to f f f are at most C ( 1 + log x f ) 2 C(1+\log x_f)^2 C ( 1 + log x f ) 2 . For a large minimum side, they are absorbed by P 0 x f / 4 P_0x_f/4 P 0 x f /4 . This leaves a factor e − P 0 x f / 2 e^{-P_0x_f/2} e − P 0 x f /2 per recorded bad plaquette, and the sum over decompositions converges uniformly in the depth. The remaining ratio of a Wilson integral to the reference partition function on the stopping lattice is at most e C m 3 + 9 m 2 log ( 1 / g m ) e^{Cm^{3}+9m^{2}\log(1/g_{m})} e C m 3 + 9 m 2 l o g ( 1/ g m ) (see Remark 22.5).
The number of ultraviolet scales causes no accumulated loss, for the following reason. A recorded bad plaquette created at side n f n_f n f has q f ≥ d ∗ q_f \ge d_* q f ≥ d ∗ , and therefore contributes at least a fixed multiple of x f = n f / γ x_f = n_f/\gamma x f = n f / γ to the classical action. At each later RG step, its excluded neighborhood contains a number of variables bounded independently of the scale, and there are at most O ( 1 + log n f ) O(1+\log n_f) O ( 1 + log n f ) later RG steps. Each missing Gaussian coordinate changes the normalization by log ( 1 / g i ) = log ( n i / γ ) / 2 ≤ ( log x f ) / 2 + O ( 1 ) \log(1/g_i) = \log(n_i/\gamma)/2 \le(\log x_f)/2 + O(1) log ( 1/ g i ) = log ( n i / γ ) /2 ≤ ( log x f ) /2 + O ( 1 ) , and each other correction is bounded per variable. The total correction is therefore O Γ ( ( 1 + log x f ) 2 ) O_\Gamma((1+\log x_f)^2) O Γ (( 1 + log x f ) 2 ) , which is small compared to x f x_f x f once x f x_f x f is large.
We distinguish two sets attached to a recorded bad plaquette f f f ; see Figure 22.1 . The first is its energy-estimate predecessor set B f B_f B f on the initial lattice, which carries the classical lower bound for q f q_f q f . These sets are disjoint for recorded bad plaquettes at different levels and overlap at most 48 times at one level. The second is the full excluded neighborhood assigned to f f f , which is much larger and may overlap with the neighborhoods of other recorded bad plaquettes. Source plaquettes whose full dependence sets meet such a neighborhood are discarded, and their number per recorded bad plaquette is bounded by F J F_J F J .
Figure 22.1. Schematic picture of the sets attached to a recorded bad plaquette f f f , drawn at one scale. The energy-estimate predecessor set B f B_f B f carries the lower bound for the action. The source plaquette p 1 p_1 p 1 is discarded because its full dependence set D ( p 1 ) D(p_1) D ( p 1 ) meets the excluded neighborhood of f f f ; the source plaquette p 2 p_2 p 2 is kept because D ( p 2 ) D(p_2) D ( p 2 ) consists only of links of the block region at the stopping level.
Product domains along a large-field decomposition
The following lemma checks that the regional integration estimates of § 2.2 , on which the density bounds of Theorem 21.13 rest, apply along an arbitrary large-field decomposition. Only the small frozen coordinates are substituted into analytic functions (Theorems 21.5 and 21.13 ).
Lemma 22.3. Fix A W ≥ 1 A_W \ge1 A W ≥ 1 first, and then fix parameters satisfying the hypotheses of Theorems 20.14 and 21.5 , including the product domains and the coefficient-norm bounds at initialization in Theorem 21.13 . Assume also the construction and reconstruction-image hypotheses of that theorem, in particular r + β r e c < ϵ e n , J r+\beta_{\mathrm{rec}}<\epsilon_{\mathrm{en},J} r + β rec < ϵ en , J . Let X \mathcal{X} X be a large-field decomposition whose initial and stopping levels are boundaries of RG steps, with T ≥ 1 T \ge1 T ≥ 1 integrated RG steps. Then every integrated step consists of J J J averaging steps, the total number of integrated averaging steps is J T JT J T , and Theorems 20.13 and 20.14 and Lemma 20.5 apply at the gauge representatives of X \mathcal{X} X satisfying the domain conditions and reconstruction identities of Theorem 21.13 , with the frozen coordinates satisfying the bounds of Theorem 21.5 . In each such application the fluctuation cutoff domain is
W t = ∏ i = 0 J − 1 ∏ e { ∣ w i , e ∣ ≤ p ( t t ) } , t t = max i g t , i , p ( t ) = A W log ( 1 / t ) , (146) \mathcal{W}_t=\prod_{i=0}^{J-1}\prod_e\left\{|w_{i,e}|\le p(t_t)\right\},\qquad t_t=\max_i g_{t,i},\qquad p(t)=A_W\log(1/t), \tag*{(146)} W t = i = 0 ∏ J − 1 e ∏ { ∣ w i , e ∣ ≤ p ( t t ) } , t t = i max g t , i , p ( t ) = A W log ( 1/ t ) , (146) where g t , i g_{t,i} g t , i is the coupling of the i i i -th averaging step of the t t t -th RG step. The coupling ceiling may depend on the fixed amplitude A W A_W A W ; this does not change the constants of the local analytic estimates or the bounds that are uniform before J J J is chosen. The conclusion is uniform in the number of RG steps since the creation of the frozen coordinates and in the real values of the frozen coordinates in their stated domain, including the first step for a Wilson initial measure. No derivative with respect to a frozen coordinate is involved.
Proof. The cutoff amplitude is fixed before the coupling ceiling of Theorem 20.14 is chosen. By Theorem 21.13 , every integrated step has J J J averaging steps, and Eq. (146) holds. If m t m_t m t is the output side of an RG step, then
t t 2 = γ M m t , g t , i = t t M ( i − J + 1 ) / 2 ( 0 ≤ i < J ) . t_t^2=\frac{\gamma}{M m_t},\qquad g_{t,i}=t_tM^{(i-J+1)/2}\qquad(0\le i<J). t t 2 = M m t γ , g t , i = t t M ( i − J + 1 ) /2 ( 0 ≤ i < J ) .
In particular, the radii p t , i = p ( t t ) p_{t,i}=p(t_t) p t , i = p ( t t ) used in the small-field event belong to the family of radii allowed in Theorem 20.14 ,
p ( t t ) ≤ p t , i ≤ p ( t t ) + c J log , c J log = 1 2 A W ( J − 1 ) log M . p(t_t)\le p_{t,i}\le p(t_t)+c_J^{\log},\qquad c_J^{\log}=\frac{1}{2}A_W(J-1)\log M. p ( t t ) ≤ p t , i ≤ p ( t t ) + c J l o g , c J l o g = 2 1 A W ( J − 1 ) log M .
The larger enclosing product domain is used only in the comparison of Lemma 20.5 ; its radii do not replace the radii p ( t t ) p(t_t) p ( t t ) .
We now check the conditions on the coupling for this logarithmic family, which contains the common radius as a special case. Let p + ( t ) = p ( t ) + c J log p^{+}(t)=p(t)+c_J^{\log} p + ( t ) = p ( t ) + c J l o g . The condition that the cutoff domain lies in the enclosing product domain is t p + ( t ) < R e n t p^{+}(t)<R_{\mathrm{en}} t p + ( t ) < R en . With the local constants of Theorem 20.14 , and with E + σ ∗ E+\sigma_{*} E + σ ∗ for the norm bound of the input effective interaction, the small-interaction condition and the condition on the auxiliary complex parameter domain required there read
C J t p + ( t ) ≤ 1 2 , C J t [ ( p + ( t ) ) 3 + p + ( t ) + ( E + σ ∗ ) p + ( t ) ] 1 − C J t p + ( t ) ≤ ϵ J , C_J t p^{+}(t)\leq\frac{1}{2},\qquad
\frac{C_J t[(p^{+}(t))^3+p^{+}(t)+(E+\sigma_{*})p^{+}(t)]}{1-C_J t p^{+}(t)}\leq\epsilon_J, C J t p + ( t ) ≤ 2 1 , 1 − C J t p + ( t ) C J t [( p + ( t ) ) 3 + p + ( t ) + ( E + σ ∗ ) p + ( t )] ≤ ϵ J ,
2 t [ ( p + ( t ) ) 3 + p + ( t ) + ( E + σ ∗ ) p + ( t ) ] ≤ c z , J . (147) 2t[(p^{+}(t))^3+p^{+}(t)+(E+\sigma_{*})p^{+}(t)]\leq c_{z,J}. \tag*{(147)} 2 t [( p + ( t ) ) 3 + p + ( t ) + ( E + σ ∗ ) p + ( t )] ≤ c z , J . (147)
These are the expansion conditions of Theorem 20.14 , and by the same choice of parameters the denominators of its rational image bounds are positive on the outer product domains. With v = log ( 1 / t ) v=\log(1/t) v = log ( 1/ t ) , every term of the form t k ( p + ( t ) ) q t^k(p^{+}(t))^q t k ( p + ( t ) ) q with k > 0 k>0 k > 0 equals e − k v ( A W v + c J log ) q e^{-kv}(A_Wv+c_J^{\log})^q e − k v ( A W v + c J l o g ) q , which tends to zero as v → ∞ v\to\infty v → ∞ . Hence one fixed ceiling on t t t gives the containment, all the inequalities (147) , and the required smallness of the perturbative term
C J t 2 ( E + σ ∗ ) [ ( p + ( t ) ) 4 + ( E + σ ∗ ) ( p + ( t ) ) 2 ] . C_Jt^2(E+\sigma_{*})[(p^{+}(t))^4+(E+\sigma_{*})(p^{+}(t))^2]. C J t 2 ( E + σ ∗ ) [( p + ( t ) ) 4 + ( E + σ ∗ ) ( p + ( t ) ) 2 ] .
This ceiling is one of the fixed parameters. Since g ∗ < e − 1 g_{*}<e^{-1} g ∗ < e − 1 and A W ≥ 1 A_W\geq1 A W ≥ 1 , the radii p ( t t ) p(t_t) p ( t t ) are at least one. At each fixed value of the frozen coordinates, the bound for the conditional mean in Lemma 20.5 stays uniformly below the cutoff radius: the ratio of the mean bound to p ( t ) p(t) p ( t ) is at most C 1 t ( L V + 2 ) / p ( t ) C_1t(\mathcal{L}_V+2)/p(t) C 1 t ( L V + 2 ) / p ( t ) , which tends to zero. Together with the containment, this verifies the hypotheses of the tail estimate of Lemma 20.5 on the chosen cutoff domain.
We also verify the measure-theoretic hypotheses of that comparison, at every fixed real value of the frozen coordinates and every fixed coarse background. The enclosing product domain K t K_t K t is a finite product of closed balls of positive radius in R 3 \mathbb{R}^3 R 3 , so it is Borel, compact, and of finite positive Lebesgue volume. The conditional fluctuation potential V V V given by Theorem 20.14 is real and of class C 2 C^2 C 2 on a neighborhood of K t K_t K t , because the image bounds keep its inverse maps, Jacobians, and positive Hessian determinants within their domains. Thus there exist finite real numbers v − ≤ v + v_{-}\leq v_{+} v − ≤ v + , which may depend on the fixed background and frozen coordinates, such that v − ≤ V ≤ v + v_{-}\leq V\leq v_{+} v − ≤ V ≤ v + on K t K_t K t . In particular,
0 < e − v + ∣ K t ∣ ≤ Z K t = ∫ K t e − V d w ≤ e − v − ∣ K t ∣ < ∞ . (148) 0<e^{-v_{+}}|K_t|\leq Z_{K_t}=\int_{K_t}e^{-V}\,dw\leq e^{-v_{-}}|K_t|<\infty. \tag*{(148)} 0 < e − v + ∣ K t ∣ ≤ Z K t = ∫ K t e − V d w ≤ e − v − ∣ K t ∣ < ∞. (148)
The set S S S of supported fluctuation coordinates is the inverse image under a Borel coordinate map of the finitely many tests of the decomposition and of the coordinate domains, and the cutoff domain W t \mathcal{W}_t W t is defined by finitely many inequalities; both are Borel. Every indicator or optional test 0 ≤ χ ≤ 1 0\leq\chi\leq1 0 ≤ χ ≤ 1 that is kept is Borel, and e − V e^{-V} e − V is positive and Borel. The Haar disintegration, Gaussian, and constant prefactors are identical on both sides of the comparison. Since these checks are made at each fixed preceding output point and fixed frozen coordinates, no uniform bound on v ± v_{\pm} v ± is needed. If there are no fluctuation blocks, the product domain is a point of mass one; its partition function is positive and finite, its determinant is one, and its logarithmic determinant is zero.
For the substitution of frozen coordinates into analytic functions, Theorem 21.5 includes the initial bound for finite products of links and all conversion constants in C c o n v , J ≥ C a d , J C_{\mathrm{conv},J} \ge C_{\mathrm{ad},J} C conv , J ≥ C ad , J . At the representatives of the lemma, we have, at each scale at which variables are frozen and for both spatially varying radii ϕ c l , ϕ q \phi_{\mathrm{cl}},\phi_{\mathrm{q}} ϕ cl , ϕ q ,
∣ c j ♯ ∣ ≤ min { β r e c , ϕ ∙ ( j ) / 4 } , C b c , ∙ E ∙ β r e c / ρ ∙ ≤ σ ∗ / 4 , ∙ ∈ { c l , q } . (149) \lvert c_j^\sharp\rvert\le\min\{\beta_{\mathrm{rec}},\phi_\bullet(j)/4\}, \qquad C_{\mathrm{bc},\bullet}E_\bullet\beta_{\mathrm{rec}}/\rho_\bullet\le\sigma_*/4, \qquad\bullet\in\{\mathrm{cl},\mathrm{q}\}. \tag*{(149)} ∣ c j ♯ ∣ ≤ min { β rec , ϕ ∙ ( j ) /4 } , C bc , ∙ E ∙ β rec / ρ ∙ ≤ σ ∗ /4 , ∙ ∈ { cl , q } . (149)
For a frozen group variable b j b_j b j kept on the lattice where it was frozen, we have c j ♯ = Log b j c_j^\sharp=\operatorname{Log} b_j c j ♯ = Log b j in the identity chart, and hence ∣ c j ♯ ∣ = ρ G ( b j ) \lvert c_j^\sharp\rvert=\rho_G(b_j) ∣ c j ♯ ∣ = ρ G ( b j ) . A longer fixed product of links uses the conversion included in C c o n v , J C_{\mathrm{conv},J} C conv , J , which is fixed before β r e c \beta_{\mathrm{rec}} β rec and r r r . At initialization, for a stopped initial measure, we use the bound by r r r coming from D r D_r D r , together with C a d , J r < β r e c C_{\mathrm{ad},J}r<\beta_{\mathrm{rec}} C ad , J r < β rec ; for a Wilson initial measure, the bound for the frozen variables adjacent to the block region is the other initialization clause of Theorem 21.5 . The bound (149) covers all frozen variables, including the boundary band, the boundary neighborhoods, the links across the boundary of the block region, and the initialization, and its constant C b c , ∙ C_{\mathrm{bc},\bullet} C bc , ∙ includes the sum over all these types. The arbitrary frozen group variables deep inside excluded regions are fixed external parameters and are not inserted into analytic functions.
All variables in the band of reduced analytic radius are frozen. Thus every variable that remains in the incoming block region D i n , t D_{\mathrm{in},t} D in , t has the classical radius ρ c l \rho_{\mathrm{cl}} ρ cl , and the incoming norm of the classical perturbation in the coefficient-norm bounds is evaluated at that radius; the spatially varying radius is used only for the output. By Theorem 21.5 , the old fixed complement and the points of creation of frozen variables remain in the fixed complement of the next block region, up to the fixed buffer of the finite link products. In the current units, the incoming and current distances to the fixed complement satisfy
d D i n , t ( i ) ≥ B d c u r r e n t ( e ) − B dist ( e , i / B ) − C b u f , J , d_{D_{\mathrm{in},t}}(i) \ge B d_{\mathrm{current}}(e) - B\operatorname{dist}(e,i/B) - C_{\mathrm{buf},J}, d D in , t ( i ) ≥ B d current ( e ) − B dist ( e , i / B ) − C buf , J ,
where i i i is the point of creation of an incoming variable, e e e is a current point, and C b u f , J C_{\mathrm{buf},J} C buf , J is the fixed buffer constant. This is the boundary-distance comparison required in Theorem 21.13 (compare Theorem 19.9 ); enlarging the fixed complement preserves the inclusion, and at initialization there is no boundary. The uniform bound on the conditional force kept in that induction is the following condition on the same product domains:
δ V , t = max I sup y ∈ T I ∥ A I − 1 V I ( c , a I ( c , y ; ξ ) , y ) ∥ c o e f f ′ \delta_{V,t}=\max_I\sup_{y\in\mathcal{T}_I}\left\|A_I^{-1}V_I\bigl(c,a_I(c,y;\xi),y\bigr)\right\|_{\mathrm{coeff}'} δ V , t = I max y ∈ T I sup A I − 1 V I ( c , a I ( c , y ; ξ ) , y ) coeff ′
C s e q , J ( 2 δ V , t + r q ) < 1 2 ε t u b e . (150) C_{\mathrm{seq},J}(2\delta_{V,t}+r_q)<\frac{1}{2}\varepsilon_{\mathrm{tube}}. \tag*{(150)} C seq , J ( 2 δ V , t + r q ) < 2 1 ε tube . (150)
Here I I I runs over the finitely many initial segments of the conditional integrations, A I A_I A I is the positive reference block of the segment I I I , a I a_I a I is its conditional constrained minimizer without perturbation, V I V_I V I is the z I z_I z I -derivative, in the original free coordinates, of the added force V V V of (114) , T I \mathcal{T}_I T I is the domain of the remaining parameters, and r q r_q r q is the fixed radius of the free coordinates. The coefficient norm includes the common product domain of the coarse fields, uniformly at fixed real frozen coordinates. The condition (150) is a hypothesis on the product domains in Theorem 21.13 (see Lemma 19.3 ), uniform on the prescribed ball of classical perturbations; it is part of the fixed parameters and is verified in Section 23, see (23.29). The supremum is over every remaining parameter used in the sequential construction, not only over the point with zero fluctuations, and all real frozen coordinates are held fixed. Hence the image bounds and the sequential containment used in the induction hold on their stated domains.
Finally, x 0 x_0 x 0 , x 0 c l ≤ σ ∗ x_0^{\mathrm{cl}} \leq\sigma_* x 0 cl ≤ σ ∗ are the coefficient-norm hypotheses of the induction at initialization, including the substitution at initialization. Its constants are those of the perturbed maps: K c o l , J K_{\mathrm{col},J} K col , J in Theorem 20.14 bounds the columns of the perturbed zero-fluctuation map uniformly on the larger perturbation domain, and we use the resulting K J K_J K J in the output bound of that theorem and in its choice of the width s e r s_{\mathrm{er}} s er of the boundary layer removed at each step. By the substitution bounds (149) , we have z t , z t c l ≤ σ ∗ / 4 z_t,z_t^{\mathrm{cl}} \leq\sigma_*/4 z t , z t cl ≤ σ ∗ /4 , and the choice of s e r s_{\mathrm{er}} s er gives 2 e − α ( s e r − 1 ) y t ≤ σ ∗ / 2 2e^{-\alpha(s_{\mathrm{er}}-1)}y_t \leq\sigma_*/2 2 e − α ( s er − 1 ) y t ≤ σ ∗ /2 , with the same inequality for the classical output bound. Thus both incoming norms at the next boundary are at most 3 σ ∗ / 4 < σ ∗ 3\sigma_*/4 < \sigma_* 3 σ ∗ /4 < σ ∗ , and by induction the coefficient-norm hypothesis holds after any number of RG steps. Since every integrated step is a full RG step, Theorem 20.13 applies at every step, including its incoming value and the initialization.
The sets of such representatives in Theorem 21.13 start from the representatives at which the bounds on the frozen coordinates just checked hold, and they contain every supported fiber point over each preceding output point. A gauge transformation acting independently at the endpoints of the frozen variables need not preserve the bound on their logarithms. We therefore use analytic estimates only on these sets, and we use the pointwise gauge covariance of Theorem 21.10 only after a gauge-invariant upper bound has been obtained. This verifies the hypotheses for both initial measures.
Energy transport, feasible configurations, and the reference normalization
The next theorem collects the classical estimates: (a) is used in Sections 22.5 and 22.8, (b) in Section 22.6, and (c) in Section 22.7.
Theorem 22.4. Fix an odd integer M ≥ 3 M \geq3 M ≥ 3 .
(a) There are δ ∗ ( M ) > 0 \delta_*(M)>0 δ ∗ ( M ) > 0 and C M < ∞ C_M<\infty C M < ∞ , depending only on M M M , such that for every level i i i , every set P ′ \mathcal{P}' P ′ of positive plaquettes at level i + 1 i+1 i + 1 , every configuration V V V at level i i i , and every δ \delta δ ,
max p ∈ Pa ( P ′ ) ρ G ( V p ) ≤ δ ≤ δ ∗ ⟹ S i + 1 P ′ ( T M V ) ≤ ( 1 + C M δ ) S i Pa ( P ′ ) ( V ) . \max_{p\in\operatorname{Pa}(\mathcal{P}')} \rho_G(V_p)\leq\delta\leq\delta_* \quad\Longrightarrow\quad\mathcal{S}_{i+1}^{\mathcal{P}'}(T_MV)\leq(1+C_M\delta)\mathcal{S}_{i}^{\operatorname{Pa}(\mathcal{P}')}(V). p ∈ Pa ( P ′ ) max ρ G ( V p ) ≤ δ ≤ δ ∗ ⟹ S i + 1 P ′ ( T M V ) ≤ ( 1 + C M δ ) S i Pa ( P ′ ) ( V ) .
Under this hypothesis no based loop holonomy in these predecessor rectangles equals − I -I − I , so all the logarithms in the definition of T M T_M T M there are the analytic ones. Define
B ∗ = min { δ ∗ 2 π 2 , [ ( M − 1 ) log 2 π C M ] 2 } , σ ( b ) = π C M b M − 1 . B_*=\min\left\{\frac{\delta_*^2}{\pi^2},\left[\frac{(\sqrt{M}-1)\log2}{\pi C_M}\right]^2\right\},\qquad\sigma(b)=\frac{\pi C_M\sqrt{b}}{\sqrt{M}-1}. B ∗ = min ⎩ ⎨ ⎧ π 2 δ ∗ 2 , [ π C M ( M − 1 ) log 2 ] 2 ⎭ ⎬ ⎫ , σ ( b ) = M − 1 π C M b .
Let ℓ ≥ 0 \ell\geq0 ℓ ≥ 0 be such that the side at level ℓ \ell ℓ is at least four, let P ℓ \mathcal{P}_\ell P ℓ be a set of positive plaquettes at level ℓ \ell ℓ , and let P i = Pa ( P i + 1 ) \mathcal{P}_i=\operatorname{Pa}(\mathcal{P}_{i+1}) P i = Pa ( P i + 1 ) for 0 ≤ i < ℓ 0\leq i<\ell 0 ≤ i < ℓ ; we call the P i \mathcal{P}_i P i the energy-estimate predecessor sets of P ℓ \mathcal{P}_\ell P ℓ . Then for every configuration U U U at level 0 0 0 and every 0 < b ≤ B ∗ 0<b\leq B_* 0 < b ≤ B ∗ ,
S ℓ P ℓ ( T M ℓ U ) ≤ max { e σ ( b ) , 8 ∣ P ℓ ∣ / b } S 0 P 0 ( U ) . (151) \mathcal{S}_{\ell}^{\mathcal{P}_\ell}(T_M^\ell U)\leq\max\{e^{\sigma(b)},8|\mathcal{P}_\ell|/b\}\mathcal{S}_{0}^{\mathcal{P}_0}(U). \tag*{(151)} S ℓ P ℓ ( T M ℓ U ) ≤ max { e σ ( b ) , 8∣ P ℓ ∣/ b } S 0 P 0 ( U ) . (151)
This holds for every holonomy and with the Borel extension of T M T_M T M . If S 0 P 0 ( U ) ≤ b / h n ℓ \mathcal{S}_{0}^{\mathcal{P}_0}(U)\leq b/h_{n_\ell} S 0 P 0 ( U ) ≤ b / h n ℓ , then for 0 ≤ i ≤ ℓ 0\leq i\leq\ell 0 ≤ i ≤ ℓ every plaquette in P i \mathcal{P}_i P i has angle at most π b M − ( ℓ − i ) / 2 \pi\sqrt{b}M^{-(\ell-i)/2} π b M − ( ℓ − i ) /2 , and the factor in (22.8) can be replaced by e σ ( b ) ≤ 2 e^{\sigma(b)}\leq2 e σ ( b ) ≤ 2 . If q < d c u t q<d_{\mathrm{cut}} q < d cut for every plaquette in P ℓ \mathcal{P}_\ell P ℓ , then 8 ∣ P ℓ ∣ / b 8|\mathcal{P}_\ell|/b 8∣ P ℓ ∣/ b may be replaced by 4 ∣ P ℓ ∣ d c u t / b 4|\mathcal{P}_\ell|d_{\mathrm{cut}}/b 4∣ P ℓ ∣ d cut / b . Let p M ≤ 24 M 3 p_M\leq24M^3 p M ≤ 24 M 3 be the largest number of plaquettes in a set P j − 1 ( C ) \mathcal{P}_{j-1}(\mathsf{C}) P j − 1 ( C ) , and put
K M = max ( 2 , 8 p M / B ∗ ) , K 1 = max ( 2 , 8 / B ∗ ) . K_M=\max(2,8p_M/B_*),\qquad K_1=\max(2,8/B_*). K M = max ( 2 , 8 p M / B ∗ ) , K 1 = max ( 2 , 8/ B ∗ ) .
Then for 0 ≤ k ≤ j − 1 0\leq k\leq j-1 0 ≤ k ≤ j − 1 and every configuration V V V at level k k k , and for every depth ℓ \ell ℓ ,
Q j T ( T M j − 1 − k V ) ≤ K M S W ( V ) h n k , S W ( T M ℓ U ) h n ℓ ≤ 48 K 1 S W ( U ) h n 0 . Q_j^{\mathsf{T}}(T_M^{j-1-k}V)\leq K_M\frac{S_W(V)}{h_{n_k}},\qquad\frac{S_W(T_M^\ell U)}{h_{n_\ell}}\leq48K_1\frac{S_W(U)}{h_{n_0}}. Q j T ( T M j − 1 − k V ) ≤ K M h n k S W ( V ) , h n ℓ S W ( T M ℓ U ) ≤ 48 K 1 h n 0 S W ( U ) .
These constants are independent of the depth, the side, J J J , L L L , γ \gamma γ , and the geometry. The factor K M K_M K M may be replaced by 2 2 2 if q < d c u t q<d_{\mathrm{cut}} q < d cut for all the relevant plaquettes at level j − 1 j-1 j − 1 and 4 p M d c u t ≤ B ∗ 4p_Md_{\mathrm{cut}}\leq B_* 4 p M d cut ≤ B ∗ .
(b) Let X \mathcal{X} X be a large-field decomposition as in Theorem 21.13 , with ℓ \ell ℓ integrated averaging steps and stopping side m m m , and let g b a s e 2 = γ / n b g_{\mathrm{base}}^2=\gamma/n_b g base 2 = γ / n b be the coupling of the initial lattice. At every representative ( c ^ , b ^ r e c ) (\widehat{c},\widehat{b}_{\mathrm{rec}}) ( c , b rec ) satisfying the domain conditions and reconstruction identities of Theorem 21.13 , the zero-fluctuation reconstructions of the successive RG steps, with the frozen coordinates inserted at the scales at which they were frozen, give one configuration U ∗ U_* U ∗ on the initial lattice such that
( T M ℓ U ∗ ) ∣ F = c ^ , E T ( c ^ ; b ^ r e c ) g m 2 = E b a s e ( U ∗ ) g b a s e 2 , g m 2 = M ℓ g b a s e 2 . (T_M^\ell U_*)|_F=\widehat{c},\qquad\frac{E_T(\widehat{c};\widehat{b}_{\mathrm{rec}})}{g_m^2}=\frac{\mathcal{E}_{\mathrm{base}}(U_*)}{g_{\mathrm{base}}^2},\qquad g_m^2=M^\ell g_{\mathrm{base}}^2. ( T M ℓ U ∗ ) ∣ F = c , g m 2 E T ( c ; b rec ) = g base 2 E base ( U ∗ ) , g m 2 = M ℓ g base 2 .
Here E T E_T E T is the effective classical action of the stopping field after all external classical exponents have been incorporated once into the classical exponent. For a stopped initial measure, U ∗ ∈ D ρ U_\ast\in D_\rho U ∗ ∈ D ρ and E b a s e = E d 0 , s a t \mathcal{E}_{\mathrm{base}} = E_{d_0,\mathrm{sat}} E base = E d 0 , sat on this larger gauge-invariant domain; for a Wilson initial measure, U ∗ U_\ast U ∗ is a configuration in U n b \mathcal{U}_{\mathrm{nb}} U nb and E b a s e = S W \mathcal{E}_{\mathrm{base}} = S_W E base = S W . Consequently
E T ( c ^ ; b ^ r e c ) ≥ M ℓ inf { E b a s e ( U ) : U ∈ S , ( T M ℓ U ) ∣ F = c ^ } , E_T(\hat{c};\hat{b}_{\mathrm{rec}}) \ge M^\ell\inf\{\mathcal{E}_{\mathrm{base}}(U): U \in S,\ (T^\ell_M U)|_F = \hat{c}\}, E T ( c ^ ; b ^ rec ) ≥ M ℓ inf { E base ( U ) : U ∈ S , ( T M ℓ U ) ∣ F = c ^ } ,
where S = D ρ S = D_\rho S = D ρ or S = U n b S = \mathcal{U}_{\mathrm{nb}} S = U nb , respectively. No minimizer in D r D_r D r is asserted. Let κ ∗ > 0 \kappa_\ast> 0 κ ∗ > 0 be such that E d , s a t ≥ κ ∗ S W E_{d,\mathrm{sat}} \ge\kappa_\ast S_W E d , sat ≥ κ ∗ S W on D ρ D_\rho D ρ for every depth d ≥ 0 d \ge0 d ≥ 0 (Theorems 11.2 and 11.6), and set
κ 0 = min ( 1 , κ ∗ ) , c A a c t = α W = κ 0 48 K 1 , V m = α W S W . \kappa_0 = \min(1,\kappa_\ast),\qquad c_A^{\mathrm{act}} = \alpha_W = \frac{\kappa_0}{48K_1},\qquad\mathcal{V}_m = \alpha_W S_W. κ 0 = min ( 1 , κ ∗ ) , c A act = α W = 48 K 1 κ 0 , V m = α W S W .
The two names c A a c t c_A^{\mathrm{act}} c A act and α W \alpha_W α W denote the same number, in its two roles below. For c ∈ S U ( 2 ) F c \in\mathrm{SU}(2)^F c ∈ SU ( 2 ) F and for frozen coordinates b r e c b_{\mathrm{rec}} b rec put
Ψ F ( c ) = V m ; F ( c ) = min z ∈ S U ( 2 ) H V m ( c ⊕ z ) , A X ( b r e c ) = c A a c t ∑ f ∈ X 4 q f ( b r e c ) n f , \Psi_F(c) = \mathcal{V}_{m;F}(c) = \min_{z\in\mathrm{SU}(2)^H}\mathcal{V}_m(c\oplus z),\qquad A_{\mathcal{X}}(b_{\mathrm{rec}}) = c_A^{\mathrm{act}}\sum_{f\in\mathcal{X}}\frac{4q_f(b_{\mathrm{rec}})}{n_f}, Ψ F ( c ) = V m ; F ( c ) = z ∈ SU ( 2 ) H min V m ( c ⊕ z ) , A X ( b rec ) = c A act f ∈ X ∑ n f 4 q f ( b rec ) ,
where q f ( b r e c ) q_f(b_{\mathrm{rec}}) q f ( b rec ) is the energy of the recorded bad plaquette f f f , a function of the frozen coordinates because the full dependence set of f f f is frozen. Then, at the same representative,
E T g m 2 ≥ Ψ F ( c ^ ) g m 2 , E T g m 2 ≥ A X ( b r e c ) , (152) \frac{E_T}{g_m^2} \ge\frac{\Psi_F(\hat{c})}{g_m^2},\qquad\frac{E_T}{g_m^2} \ge A_{\mathcal{X}}(b_{\mathrm{rec}}), \tag*{(152)} g m 2 E T ≥ g m 2 Ψ F ( c ^ ) , g m 2 E T ≥ A X ( b rec ) , (152)
and
A X ( b r e c ) ≥ 4 c A a c t d ∗ ∑ f ∈ X n f γ on I X a c t = { q f ( b r e c ) ≥ d ∗ for all f ∈ X } . A_{\mathcal{X}}(b_{\mathrm{rec}}) \ge4c_A^{\mathrm{act}}d_\ast\sum_{f\in\mathcal{X}}\frac{n_f}{\gamma}\qquad\text{on }I_{\mathcal{X}}^{\mathrm{act}}=\{q_f(b_{\mathrm{rec}})\ge d_\ast\text{ for all }f\in\mathcal{X}\}. A X ( b rec ) ≥ 4 c A act d ∗ f ∈ X ∑ γ n f on I X act = { q f ( b rec ) ≥ d ∗ for all f ∈ X } .
(c) Use the full-torus recursion of the stopped measures (Theorem 14.6) with the same RG steps as the decomposition; we call it the reference recursion. At a depth d ≥ 0 d \ge0 d ≥ 0 of the reference recursion on the stopping lattice put
Z ^ d ( g ) = ∫ D r e − E d , s a t / g 2 − [ R d , s a t − R d ( I ) ] d H m , g 2 = γ / m . \widehat{Z}_d(g)=\int_{D_r}e^{-E_{d,\mathrm{sat}}/g^2-[R_{d,\mathrm{sat}}-R_d(I)]}\,dH_m,\qquad g^2=\gamma/m. Z d ( g ) = ∫ D r e − E d , sat / g 2 − [ R d , sat − R d ( I )] d H m , g 2 = γ / m .
The restriction of the initial measure to the small-field conditions of the reference recursion and its pushforward to the stopping lattice has the form
d ξ ~ T = Z ν − 1 e Λ T T 1 D r e − E d , s a t / g 2 − [ R d , s a t − R d ( I ) ] d H m , d\widetilde{\xi}_T=Z_\nu^{-1}e^{\Lambda_T^T}\mathbf{1}_{D_r}e^{-E_{d,\mathrm{sat}}/g^2-[R_{d,\mathrm{sat}}-R_d(I)]}\,dH_m, d ξ T = Z ν − 1 e Λ T T 1 D r e − E d , sat / g 2 − [ R d , sat − R d ( I )] d H m ,
where Λ T T \Lambda_T^T Λ T T is the normalization constant of the reference recursion; for a Wilson initial measure, only this reference measure is restricted to D r D_r D r at the start. With m T T = ξ ~ T ( 1 ) ≤ 1 m_T^T=\widetilde{\xi}_T(1)\le1 m T T = ξ T ( 1 ) ≤ 1 , one has exactly
e Λ T T Z ν = m T T Z ^ d ( g ) ≤ Z ^ d ( g ) − 1 . (153) \frac{e^{\Lambda_T^T}}{Z_\nu}=\frac{m_T^T}{\widehat{Z}_d(g)}\le\widehat{Z}_d(g)^{-1}. \tag*{(153)} Z ν e Λ T T = Z d ( g ) m T T ≤ Z d ( g ) − 1 . (153)
All Haar disintegration, Gaussian, cluster expansion, and substitution constants are included in Λ T T \Lambda^{\mathsf{T}}_T Λ T T , and for a stopped initial measure Λ 0 T = − R d 0 ( I ) \Lambda^{\mathsf{T}}_0=-R_{d_0}(I) Λ 0 T = − R d 0 ( I ) . For every fixed 0 < λ 0 ≤ λ 1 < ∞ 0<\lambda_0\le\lambda_1<\infty 0 < λ 0 ≤ λ 1 < ∞ there are C , C ′ , C ′ ′ < ∞ C,C',C''<\infty C , C ′ , C ′′ < ∞ , uniform in the depth d d d of the reference recursion, in m m m , and in 0 < g ≤ g ∗ < 1 0<g\le g_*<1 0 < g ≤ g ∗ < 1 , such that
Z ^ d ( g ) ≥ e − C 3 m 3 g 3 ( 2 m 3 + 1 ) , ∫ e − λ V m / g 2 d H m ≤ e C ′ 3 m 3 g 3 ( 2 m 3 − 3 m 2 + 1 ) \widehat{Z}_d(g)\ge e^{-C3m^3}g^{3(2m^3+1)},\qquad\int e^{-\lambda V_m/g^2}\,dH_m\le e^{C'3m^3}g^{3(2m^3-3m^2+1)} Z d ( g ) ≥ e − C 3 m 3 g 3 ( 2 m 3 + 1 ) , ∫ e − λ V m / g 2 d H m ≤ e C ′ 3 m 3 g 3 ( 2 m 3 − 3 m 2 + 1 )
for λ ∈ [ λ 0 , λ 1 ] \lambda\in[\lambda_0,\lambda_1] λ ∈ [ λ 0 , λ 1 ] with 4 λ α W / g 2 ≥ 1 4\lambda\alpha_W/g^2\ge1 4 λ α W / g 2 ≥ 1 . Consequently the logarithm of the ratio of these two integrals is at most C ′ ′ 3 m 3 + 9 m 2 log ( 1 / g ) C''3m^3+9m^2\log(1/g) C ′′ 3 m 3 + 9 m 2 log ( 1/ g ) . The normalized stopped measure ν ^ \widehat{\nu} ν on the stopping lattice, with the same support D r D_r D r , satisfies
log E ν ^ e v S W / g 2 ≤ C ′ ′ 3 m 3 + 9 m 2 log ( 1 / g ) , 0 ≤ v ≤ κ ∗ / 2. \log\mathbb{E}_{\widehat{\nu}}e^{vS_W/g^2}\le C''3m^3+9m^2\log(1/g),\qquad0\le v\le\kappa_*/2. log E ν e v S W / g 2 ≤ C ′′ 3 m 3 + 9 m 2 log ( 1/ g ) , 0 ≤ v ≤ κ ∗ /2.
For fixed 0 < λ 0 ≤ 1 0<\lambda_0\le1 0 < λ 0 ≤ 1 , the quantity
Z ^ d ( ξ ) − 1 ∫ D r e − λ ( E d , s a t / g 2 + R d , s a t − R d ( I ) ) + v S W / g 2 d H m , r , λ ∈ [ λ 0 , 1 ] , 0 ≤ v ≤ λ 0 κ ∗ / 2 , \widehat{Z}_d(\xi)^{-1}\int_{D_r}e^{-\lambda(E_{d,\mathrm{sat}}/g^2+R_{d,\mathrm{sat}}-R_d(I))+vS_W/g^2}\,dH_{m,r},\qquad\lambda\in[\lambda_0,1],\quad0\le v\le\lambda_0\kappa_*/2, Z d ( ξ ) − 1 ∫ D r e − λ ( E d , sat / g 2 + R d , sat − R d ( I )) + v S W / g 2 d H m , r , λ ∈ [ λ 0 , 1 ] , 0 ≤ v ≤ λ 0 κ ∗ /2 ,
satisfies the same bound with a larger volume constant. In each of these ratios the boundary correction 9 m 2 log ( 1 / g ) 9m^2\log(1/g) 9 m 2 log ( 1/ g ) occurs once. The denominator is always the integral over exactly D r D_r D r ; the periodic boundary plaquettes are deleted only in the upper bound for the numerator.
The constants of the reconstruction and of the normalization constants have the dependence described in item (3) of Definition 4.14, excluding the widths chosen later. The constants in (c) depend on the fixed domains on the torus, r r r , the norm bounds, and the interval [ λ 0 , λ 1 ] [\lambda_0,\lambda_1] [ λ 0 , λ 1 ] of action coefficients, not on the depth, the side, or the coupling. Applications to probability estimates have the dependence described in item (4) of Definition 4.14.
Remark 22.5. The factor g − 9 m 2 g^{-9m^2} g − 9 m 2 in (c) comes from a count of independent coordinates. The periodic lattice of side m m m has 3 m 3 3m^3 3 m 3 links and m 3 m^3 m 3 sites, so the complement of a spanning tree has 2 m 3 + 1 2m^3+1 2 m 3 + 1 links; restricting each of them to a ball of radius proportional to g g g in SU(2), a group of dimension three, gives the lower bound e − C 3 m 3 g 3 ( 2 m 3 + 1 ) e^{-C3m^3}g^{3(2m^3+1)} e − C 3 m 3 g 3 ( 2 m 3 + 1 ) for Z ^ d ( g ) \widehat{Z}_d(g) Z d ( g ) , with the constant of part (c). The open cube with m m m sites per side, used for the numerator, has 3 m 2 ( m − 1 ) 3m^2(m-1) 3 m 2 ( m − 1 ) links and m 3 m^3 m 3 sites, hence 2 m 3 − 3 m 2 + 1 2m^3-3m^2+1 2 m 3 − 3 m 2 + 1 links outside a spanning tree, which give the upper bound e C ′ 3 m 3 g 3 ( 2 m 3 − 3 m 2 + 1 ) e^{C'3m^3}g^{3(2m^3-3m^2+1)} e C ′ 3 m 3 g 3 ( 2 m 3 − 3 m 2 + 1 ) , with the constant of part (c). The difference 3 [ ( 2 m 3 + 1 ) − ( 2 m 3 − 3 m 2 + 1 ) ] = 9 m 2 3[(2m^3+1)-(2m^3-3m^2+1)]=9m^2 3 [( 2 m 3 + 1 ) − ( 2 m 3 − 3 m 2 + 1 )] = 9 m 2 is a boundary correction in an inequality; it is not an identification of a physical quantity. At a side m = χ n m=\chi n m = χ n it becomes 9 χ 2 n 2 log ( 1 / g n ) + ( 9 / 2 ) χ 2 n 2 log χ 9\chi^2n^2\log(1/g_n)+(9/2)\chi^2n^2\log\chi 9 χ 2 n 2 log ( 1/ g n ) + ( 9/2 ) χ 2 n 2 log χ . Gauge fixing along a tree, deleting nonnegative plaquette terms, and factorizing one-plaquette integrals are elementary stability arguments in the tradition of Faria da Veiga and O’Carroll [64, 65]; the precise periodic and open counts used here are proved in Theorem 14.8, with the open-cube estimate of Lemma 6.10.
Proof of Theorem 22.4(a). We write h i = h n i h_i=h_{n_i} h i = h n i within this proof.
Transport over one averaging step. Multiplying the restricted conclusion of Theorem 7.3 by four and dividing by h i + 1 = M h i h_{i+1}=Mh_i h i + 1 = M h i , we obtain the displayed implication. Its hypotheses are local: the maximum is over the union of the complete predecessor rectangles, and the side at level i + 1 i+1 i + 1 is at least four. We may decrease δ ∗ ( M ) \delta_*(M) δ ∗ ( M ) so that the exceptional based loop holonomies are excluded as well. To check this, note that a contour F c , x C c − 1 F_{c,x}C_c^{-1} F c , x C c − 1 in (4.3) is filled by moving its straight segment of length M M M across the coordinate segments of the path σ v , x \sigma_{v,x} σ v , x . At most M ∑ k = 1 3 ∣ x k − v k ∣ ≤ 3 M ( M − 1 ) / 2 M\sum_{k=1}^{3}|x_k-v_k|\leq3M(M-1)/2 M ∑ k = 1 3 ∣ x k − v k ∣ ≤ 3 M ( M − 1 ) /2 elementary moves across one plaquette are needed, and each move inserts a conjugate of a plaquette transport or of its inverse. All these plaquettes belong to the complete predecessor rectangle, so by the bi-invariant triangle inequality the distance of this based loop holonomy from I I I is at most 3 M 2 δ 3M^2\delta 3 M 2 δ . If we also require 3 M 2 δ ∗ < 2 π 3M^2\delta_*<2\pi 3 M 2 δ ∗ < 2 π , then this holonomy is not − I -I − I , and its logarithm is the analytic one. This decrease of δ ∗ \delta_* δ ∗ changes only a constant depending on M M M .
Transport over many averaging steps. Let V i = T M i U V^i=T_M^iU V i = T M i U and write
e i = S i P i ( V i ) , δ i = max p ∈ P i ρ G ( V p i ) , Π i = ∏ k < i ( 1 + C M δ k ) . e_i=S_i^{\mathcal{P}_i}(V^i),\qquad\delta_i=\max_{p\in\mathcal{P}_i}\rho_G(V_p^i),\qquad\Pi_i=\prod_{k<i}(1+C_M\delta_k). e i = S i P i ( V i ) , δ i = p ∈ P i max ρ G ( V p i ) , Π i = k < i ∏ ( 1 + C M δ k ) .
For an empty set of plaquettes we set the maximum to zero; the corresponding energy is also zero. Suppose first that e 0 ≤ b / h ℓ e_0\leq b/h_\ell e 0 ≤ b / h ℓ . We prove by induction on i ≤ ℓ i\leq\ell i ≤ ℓ that the first i i i steps satisfy
e i ≤ Π i e 0 , Π i ≤ exp { π C M b ∑ k < i M − ( ℓ − k ) / 2 } ≤ 2. (154) e_i\leq\Pi_i e_0,\qquad\Pi_i\leq\exp\left\{\pi C_M\sqrt{b}\sum_{k<i}M^{-(\ell-k)/2}\right\}\leq2. \tag*{(154)} e i ≤ Π i e 0 , Π i ≤ exp { π C M b k < i ∑ M − ( ℓ − k ) /2 } ≤ 2. (154)
The case i = 0 i=0 i = 0 is immediate. Suppose that (154) holds for some i i i . Since all plaquette energies are nonnegative, (144) implies that, for every p ∈ P i p\in\mathcal{P}_i p ∈ P i ,
ρ G ( V p i ) 2 ≤ 2 π 2 q ( V p i ) ≤ π 2 h i 2 2 e i ≤ π 2 2 Π i b M − ( ℓ − i ) ≤ π 2 b M − ( ℓ − i ) . \rho_G(V_p^i)^2\leq2\pi^2q(V_p^i)\leq\frac{\pi^2h_i^2}{2}e_i\leq\frac{\pi^2}{2}\Pi_i bM^{-(\ell-i)}\leq\pi^2bM^{-(\ell-i)}. ρ G ( V p i ) 2 ≤ 2 π 2 q ( V p i ) ≤ 2 π 2 h i 2 e i ≤ 2 π 2 Π i b M − ( ℓ − i ) ≤ π 2 b M − ( ℓ − i ) .
where we used h i / h ℓ = M − ( ℓ − i ) h_i/h_\ell=M^{-(\ell-i)} h i / h ℓ = M − ( ℓ − i ) and Π i ≤ 2 \Pi_i\leq2 Π i ≤ 2 . For i < ℓ i<\ell i < ℓ this is at most π 2 b ≤ δ ∗ 2 \pi^2b\leq\delta_*^2 π 2 b ≤ δ ∗ 2 . Then the transport over one averaging step applies with Pa ( P i + 1 ) = P i \operatorname{Pa}(\mathcal{P}_{i+1})=\mathcal{P}_i Pa ( P i + 1 ) = P i ; it gives e i + 1 ≤ ( 1 + C M δ i ) e i e_{i+1}\leq(1+C_M\delta_i)e_i e i + 1 ≤ ( 1 + C M δ i ) e i and excludes all exceptional values used in that step. Further,
log Π i + 1 ≤ C M ∑ k ≤ i δ k ≤ π C M b ∑ r ≥ 1 M − r / 2 = σ ( b ) ≤ log 2. \log\Pi_{i+1}\leq C_M\sum_{k\leq i}\delta_k\leq\pi C_M\sqrt{b}\sum_{r\geq1}M^{-r/2}=\sigma(b)\leq\log2. log Π i + 1 ≤ C M k ≤ i ∑ δ k ≤ π C M b r ≥ 1 ∑ M − r /2 = σ ( b ) ≤ log 2.
This closes the induction (154) . The bound on individual plaquettes also holds at i = ℓ i=\ell i = ℓ , and no further transport step is needed there. This proves all the stated conclusions under the hypothesis e 0 ≤ b / h ℓ e_0\leq b/h_\ell e 0 ≤ b / h ℓ , including the case ℓ = 0 \ell=0 ℓ = 0 .
If instead e 0 > b / h ℓ e_0>b/h_\ell e 0 > b / h ℓ , then since q ≤ 2 q\leq2 q ≤ 2 , we have e ℓ ≤ 8 ∣ P ℓ ∣ / h ℓ < ( 8 ∣ P ℓ ∣ / b ) e 0 e_\ell\leq8|\mathcal{P}_\ell|/h_\ell<(8|\mathcal{P}_\ell|/b)e_0 e ℓ ≤ 8∣ P ℓ ∣/ h ℓ < ( 8∣ P ℓ ∣/ b ) e 0 . If every plaquette of P ℓ \mathcal{P}_\ell P ℓ has q < d c u t q<d_{\mathrm{cut}} q < d cut , the upper bound is 4 ∣ P ℓ ∣ d c u t / h ℓ 4|\mathcal{P}_\ell|d_{\mathrm{cut}}/h_\ell 4∣ P ℓ ∣ d cut / h ℓ . These two cases prove (151) for arbitrary configurations. Since no logarithm is taken in the second case, the exceptional values of the Borel extension cause no difficulty. If q < d c u t q<d_{\mathrm{cut}} q < d cut for all the relevant plaquettes and 4 p M d c u t ≤ B ∗ 4p_Md_{\mathrm{cut}} \le B_* 4 p M d cut ≤ B ∗ , we may choose b = 4 p M d c u t b = 4p_Md_{\mathrm{cut}} b = 4 p M d cut ; then for every cell the second factor in (151) is at most one and the first is at most two.
Cells and all plaquettes. We apply (151) with b = B ∗ b = B_* b = B ∗ separately to the set P j − 1 ( C ) \mathcal{P}_{j-1}(C) P j − 1 ( C ) of each cell, with level k k k in place of level 0. Each such set has at most p M p_M p M plaquettes, so the factor is at most K M K_M K M . By Lemma 7.4, the energy-estimate predecessor sets at level k k k of distinct cells of one tiling are disjoint, and the sum of their energies is therefore at most S W ( V ) / h n k S_W(V)/h_{n_k} S W ( V ) / h n k . This proves the bound with K M K_M K M . For an arbitrary set of plaquettes at level ℓ \ell ℓ , we apply (151) to each plaquette separately, with ∣ P ℓ ∣ = 1 |\mathcal{P}_\ell| = 1 ∣ P ℓ ∣ = 1 and factor at most K 1 K_1 K 1 . By the same lemma, a plaquette of the initial lattice lies in at most 48 of these predecessor sets, and summation gives the factor 48 K 1 48K_1 48 K 1 . Both arguments apply with any intermediate level as initial level.
Proof of Theorem 22.4 (b). The effective classical action along the decomposition. By Theorem 21.13 , the initial and stopping levels are boundaries of RG steps, and there are T T T integrated RG steps of J J J averaging steps each; thus ℓ = J T \ell= JT ℓ = J T . The logarithmic product domains, the bounds on the frozen coordinates that are inserted into analytic functions, the coefficient-norm bounds, and the representatives are those checked in Lemma 22.3 . For the t t t -th RG step, let C t 0 C_t^0 C t 0 be its reconstruction map with every normalized fluctuation variable set to zero, that is, its zero-fluctuation reconstruction. By the definition of the classical part of its objective, the outgoing effective classical action is
E t + 1 o u t ( c ; b r e c ( t ) ) = B E t ( C t 0 ( c ; b r e c ( t ) ) ; b r e c ( t ) ) , B = M J . (155) E^{\mathrm{out}}_{t+1}(c;b_{\mathrm{rec}}^{(t)}) = B E_t\left(C_t^0(c;b_{\mathrm{rec}}^{(t)});b_{\mathrm{rec}}^{(t)}\right), \qquad B = M^J. \tag*{(155)} E t + 1 out ( c ; b rec ( t ) ) = B E t ( C t 0 ( c ; b rec ( t ) ) ; b rec ( t ) ) , B = M J . (155)
If new coordinates c ♯ c^\sharp c ♯ are frozen at the end of the step, the next incoming action is
E t + 1 ( v ; b r e c ( t ) , c ♯ ) = E t + 1 o u t ( v ⊕ c ♯ ; b r e c ( t ) ) . E_{t+1}(v;b_{\mathrm{rec}}^{(t)},c^\sharp) = E^{\mathrm{out}}_{t+1}\left(v \oplus c^\sharp;b_{\mathrm{rec}}^{(t)}\right). E t + 1 ( v ; b rec ( t ) , c ♯ ) = E t + 1 out ( v ⊕ c ♯ ; b rec ( t ) ) .
Thus freezing is a direct substitution in this identity; in particular, the value at v = 0 v=0 v = 0 is kept together with the nonconstant part. For an external classical exponent S e x t S_{\mathrm{ext}} S ext , the convention of Theorem 21.13 is the exact identity
e − S e x t e − E t i n t / g t 2 = e − [ E t i n t + g t 2 S e x t ] / g t 2 . e^{-S_{\mathrm{ext}}}e^{-E_t^{\mathrm{int}}/g_t^2}
=
e^{-\left[E_t^{\mathrm{int}}+g_t^2S_{\mathrm{ext}}\right]/g_t^2}. e − S ext e − E t int / g t 2 = e − [ E t int + g t 2 S ext ] / g t 2 .
It adds the term once to the effective classical action and removes it from the weight of the frozen coordinates. Under (155) the added action and g t 2 g_t^2 g t 2 are both multiplied by B B B , which leaves the exponent unchanged.
One feasible configuration. We start from such a representative at which the bounds on the frozen coordinates hold. We insert the frozen coordinates at each preceding scale at which they were frozen, and apply the maps C t 0 C_t^0 C t 0 downward. The reconstruction identities and the protection of all stored endpoints in Theorems 21.5 and 21.13 ensure that each insertion gives back the coordinate of the preceding output point, that each reconstruction solves its block averaging constraint exactly, and that every older frozen dependence is preserved. The image bounds of Theorem 20.14 , applied with the hypotheses verified in Lemma 22.3 , keep this downward sequence of reconstructions in its product domains. For a stopped initial measure, the resulting representative on the initial lattice lies in the larger common chart, so its configuration U ∗ U_\ast U ∗ lies in D ρ D_\rho D ρ ; this uses the larger image domain, not membership in D r D_r D r . For a Wilson initial measure, the frozen variables deep inside excluded regions remain arbitrary group elements, and the initial classical exponent is exactly S W ( U ∗ ) S_W(U_\ast) S W ( U ∗ ) . Composing the constraint identities, we obtain ( T M ℓ U ∗ ) ∣ F = c ^ (T_M^\ell U_\ast)|_F=\widehat{c} ( T M ℓ U ∗ ) ∣ F = c . Dividing (22.12) by g t + 1 2 = B g t 2 g_{t+1}^2=Bg_t^2 g t + 1 2 = B g t 2 and iterating over the T T T RG steps gives the identity for E T / g m 2 E_T/g_m^2 E T / g m 2 , and J T = ℓ JT=\ell J T = ℓ implies g m 2 = M ℓ g b a s e 2 g_m^2=M^\ell g_{\mathrm{base}}^2 g m 2 = M ℓ g base 2 . Since U ∗ U_\ast U ∗ belongs to the constrained set, the lower bound by the infimum follows; no interchange of minima is involved.
The first lower bound. By the lower bound of Theorem 11.2 on the common radius ρ \rho ρ , transported by the fixed gauge transformation of a representative on the initial lattice (Theorem 11.6), we have
E b a s e ( U ∗ ) g b a s e 2 ≥ κ 0 S W ( U ∗ ) g b a s e 2 . (156) \frac{\mathcal{E}_{\mathrm{base}}(U_\ast)}{g_{\mathrm{base}}^2}\geq\kappa_0\frac{S_W(U_\ast)}{g_{\mathrm{base}}^2}. \tag*{(156)} g base 2 E base ( U ∗ ) ≥ κ 0 g base 2 S W ( U ∗ ) . (156)
The radius is within the domain of Theorem 11.2, and the constant κ ∗ \kappa_\ast κ ∗ there is valid for every integer depth, whether or not an effective interaction is attached to that depth, as also stated in Theorem 14.8; for a Wilson initial measure (22.13) holds with constant one. Combining this with the bound for all plaquettes in (a), we obtain
E T g m 2 ≥ κ 0 48 K 1 S W ( T M ℓ U ∗ ) g m 2 ≥ Ψ F ( c ^ ) g m 2 . \frac{E_T}{g_m^2}\geq\frac{\kappa_0}{48K_1}\frac{S_W(T_M^\ell U_\ast)}{g_m^2}\geq\frac{\Psi_F(\widehat{c})}{g_m^2}. g m 2 E T ≥ 48 K 1 κ 0 g m 2 S W ( T M ℓ U ∗ ) ≥ g m 2 Ψ F ( c ) .
In the last inequality we used that T M ℓ U ∗ = c ^ ⊕ z T_M^\ell U_\ast=\widehat{c}\oplus z T M ℓ U ∗ = c ⊕ z for some z ∈ S U ( 2 ) H z\in\mathrm{SU}(2)^H z ∈ SU ( 2 ) H , and that Ψ F \Psi_F Ψ F minimizes over all such z z z .
The second lower bound. For a recorded bad plaquette f f f , let B f \mathcal{B}_f B f be the energy-estimate predecessor set of { f } \{f\} { f } on the initial lattice. We first justify disjointness at different levels. On the input lattice of the RG step in which f f f is recorded, the endpoints of its full dependence list cover every vertex of the four centered blocks obtained by pulling back the corners of f f f to that input lattice. Indeed, each one-step formula reads the root paths σ a , x \sigma_{a,x} σ a , x for every vertex x x x of each endpoint block; their trees span those blocks, and the same assertion follows recursively from the centered nesting identity. For a plaquette at the input level, its four boundary links already cover its four corner vertices. By Lemma 7.4, every plaquette of B f \mathcal{B}_f B f lies in the centered preimages of these blocks.
All these dependence variables are frozen together with their excluded neighborhood by Definition 21.4 (2). The neighborhood includes the protected endpoint margins, so it excludes the incident dependence links as well as the marked vertices; the endpoint protection rule of Theorem 21.3 (4) preserves this exclusion at later scales. Suppose that f f f and f ′ f' f ′ are recorded at levels ℓ f < ℓ f ′ \ell_f<\ell_{f'} ℓ f < ℓ f ′ and that their initial energy-estimate predecessor sets share a plaquette. The centered partitions are nested. A shared vertex therefore places a marked vertex from a corner block of f f f in a corner block of f ′ f' f ′ , when both are read on the input lattice where f f f was recorded. If the two tests belong to the same RG step, the full dependence list of f ′ f' f ′ meets the excluded neighborhood just added for f f f . If they belong to different RG steps, the centered ancestor of that mark is excluded by the endpoint protection rule, and again the full dependence list of f ′ f' f ′ meets the earlier excluded set. Both cases contradict the rule that a recorded test avoid that set. Thus the sets B f \mathcal{B}_f B f at different levels are disjoint. This argument uses endpoint coverage and the protected neighborhoods; it does not require the minimal list of read links to contain every link of an energy-estimate patch.
At one level, a plaquette lies in at most 48 of the sets B f \mathcal{B}_f B f by Lemma 7.4. Since plaquette energies are nonnegative, we have
S W ( U ∗ ) g b a s e 2 ≥ 1 48 ∑ f S 0 B f ( U ∗ ) . \frac{S_{\mathrm{W}}(U_*)}{g_{\mathrm{base}}^2} \ge\frac{1}{48}\sum_f S_0^{\mathcal{B}_f}(U_*). g base 2 S W ( U ∗ ) ≥ 48 1 f ∑ S 0 B f ( U ∗ ) .
The comparison (22.8) for one plaquette and arbitrary configurations, proved in (a), applies at the level ℓ f \ell_f ℓ f of each f f f with constant K 1 K_1 K 1 . Since the feasible configuration U ∗ U_* U ∗ preserves every frozen dependence, the energy of f f f in it is q f ( b ^ r e c ) q_f(\widehat{b}_{\mathrm{rec}}) q f ( b rec ) . Thus
S W ( U ∗ ) g b a s e 2 ≥ 1 48 K 1 ∑ f 4 q f ( b ^ r e c ) h n f . \frac{S_{\mathrm{W}}(U_*)}{g_{\mathrm{base}}^2} \ge\frac{1}{48K_1}\sum_f \frac{4q_f(\widehat{b}_{\mathrm{rec}})}{h_{n_f}}. g base 2 S W ( U ∗ ) ≥ 48 K 1 1 f ∑ h n f 4 q f ( b rec ) .
Multiplying by κ 0 \kappa_0 κ 0 and using (22.13), we find E T / g m 2 ≥ A X ( b ^ r e c ) E_T/g_m^2 \ge A_X(\widehat{b}_{\mathrm{rec}}) E T / g m 2 ≥ A X ( b rec ) . On I X a c t I_X^{\mathrm{act}} I X act we have q f ≥ d ∗ q_f \ge d_* q f ≥ d ∗ and h n f = γ / n f h_{n_f}=\gamma/n_f h n f = γ / n f for every f f f , which gives the lower bound for A X A_X A X .
Proof of Theorem 22.4(c). The reference identity. We apply the identity at the end of Theorem 14.8 to the same reference operators and the same partition function Z ν Z_\nu Z ν as in Theorem 21.13. Every reference operator is a full RG step on the same levels. The depth d 0 d_0 d 0 of a stopped initial measure is a multiple of J J J , and the depth at the stopping level is d = d 0 + J T d=d_0+JT d = d 0 + J T , with d 0 = 0 d_0=0 d 0 = 0 for a Wilson initial measure. Thus the reference recursion lies within the scope of Theorem 14.8. The stopping side is m m m , the preceding depth is d d d , and the centered effective interaction is R d , s a t − R d ( I ) R_{d,\mathrm{sat}}-R_d(I) R d , sat − R d ( I ) . Restriction and pushforward give a subprobability measure, also for a Wilson initial measure. Its density is the displayed ξ ~ T \widetilde{\xi}_T ξ T , with mass at most one. Integrating it, we obtain (22.10). The initial value Λ 0 T = − R d 0 ( I ) \Lambda_0^T=-R_{d_0}(I) Λ 0 T = − R d 0 ( I ) for a stopped initial measure holds because Z ν Z_\nu Z ν is not centered. All later constants are those of this one recursion.
The partition-function bounds. These are the bounds of Theorem 14.8 on the lattice of side m m m , with D = 3 m 3 D=3m^3 D = 3 m 3 . The trial configurations of that theorem have 2 m 2 + 1 2m^2+1 2 m 2 + 1 links outside a spanning tree, each with logarithm of norm at most a 0 g < min ( r / 2 , 1 ) a_0g<\min(r/2,1) a 0 g < min ( r /2 , 1 ) , and the tree