Introduction
A quantum sensor record can encode both when a signal begins and how strongly it acts. Estimating these quantities connects change point detection with quantum sensing and metrology [1–4]. For an abrupt onset, moving the boundary changes a short segment of the record, while changing the amplitude affects the whole suffix. This spatial structure offers a way to allocate information between the two tasks. The question is how much precision a common readout can retain for each.
Quantum change point studies have established identification with minimum error [5], exact identification with inconclusive outcomes [6], and sequential local strategies [7]. Related settings include boundaries between unknown domain states [8], multiple anomalies [9], and sequential detection delay [10, 11]. Here, the amplitude that makes the boundary distinguishable is also an estimation target. Each unknown parameter affects the information available for estimating the other, connecting the problem to nuisance parameters and semiparametric estimation [12, 13]. Measurement compatibility determines whether local precision limits for several parameters can be attained by one readout [14–16]. Quantum Fisher information and multiparameter bounds relate local distinguishability to estimation precision [17–23]. Bayesian studies characterize incompatibility [24], derive bounds from monotone metrics [25], and optimize projective readouts [26]. Classical change point theory relates localization to jump size and transition width [27, 28]. For quantum change point records, the central question is how onset geometry and decision loss determine the common readout cost and its asymptotic scaling.
Within Bayesian decision theory [29–33], we define the common readout cost as the weighted joint optimum minus the weighted marginal optima. All three optimizations use the full record, the same joint prior and complete measurement class, with both parameters unknown. The model consists of product records on a known smooth curve of pure states, with independent uniform priors on position and a calibrated amplitude offset (Figure 1 ). Position has clipped absolute loss in resolution units, and the offset has squared loss in fixed calibrated units. A common optimum exists precisely when the positive operator valued measure (POVM) that is optimal for position commutes with the optimal amplitude observable. Quantum local asymptotic normality and Gaussian equivalence provide a framework for comparisons with coherent state models [34–36]. We use these comparisons to control the optimal joint risk and both optimal marginal risks.
Figure 1. Quantum record and onset profiles. (a) The illustrative phase family ∣ ϕ ( θ ) ⟩ = ( ∣ 0 ⟩ + e i θ ∣ 1 ⟩ ) / 2 |\phi(\theta)\rangle= (|0\rangle+ e^{i\theta}|1\rangle)/\sqrt{2} ∣ ϕ ( θ )⟩ = ( ∣0 ⟩ + e i θ ∣1 ⟩) / 2 . Sites 1 , … , k − 1 1,\ldots,k-1 1 , … , k − 1 are in ρ ( 0 ) \rho(0) ρ ( 0 ) and sites k , … , n k,\ldots,n k , … , n in ρ ( θ ) \rho(\theta) ρ ( θ ) , with first changed site k = ⌊ n u ⌋ + 1 k=\lfloor nu\rfloor+1 k = ⌊ n u ⌋ + 1 . (b) Normalized profiles for an ideal step and a smooth rise F ( ( t − u ) / w ) F((t-u)/w) F (( t − u ) / w ) , with unknown onset u u u and known width w w w .
We certify a positive common readout cost for the ideal step at a reference signal strength, including for phase records of sixteen qubits. As signal strength grows, a common readout recovers both marginal optima. Recovery assigns a retained quantum boundary region to position and the remaining suffix to amplitude. With the offset prior fixed at any finite positive width, physical recovery holds along s n → ∞ s_n \to\infty s n → ∞ with s n = o ( n ) s_n=o(\sqrt{n}) s n = o ( n ) , where n n n is the record length. Both marginal optimal risks remain bounded away from zero, as do the risk reductions relative to using the prior alone. For known smooth rise profiles, the coherent experiment also admits recovery to leading order as the signal strengthens and the rise width shrinks. For a quintic rise of known width w w w , we obtain two scaling laws for the residual cost. The residual cost is Θ ( s − 2 ) \Theta(s^{-2}) Θ ( s − 2 ) as s → ∞ s \to\infty s → ∞ , uniformly for s 2 w ∈ [ 1 / 2 , 2 ] s^2w \in[1/2,2] s 2 w ∈ [ 1/2 , 2 ] . Taking the strong signal limit at fixed w w w yields a limiting cost of Θ ( w 3 ) \Theta(w^3) Θ ( w 3 ) as w → 0 w \to0 w → 0 . For quadratic position loss in the resolved local experiment, a readout based on amplitude prediction attains both marginal optima within the complete covariant class. With clipped absolute position loss, the optimal readout requires a cubic correction whose squared contribution accounts for the cubic width law under the full position prior.
Section 2 defines the model and estimation task, and Section 3 presents the main results and their mechanisms. Section 4 presents numerical conflict certificates and comparisons with physical records. The Supplementary Material contains the proofs and certificate error budgets.
Model and estimation task
Quantum record and prior information
The quantum record contains n n n systems with the same finite dimensional Hilbert space H \mathcal{H} H . An unknown boundary u u u marks the onset of a response controlled by an unknown calibrated amplitude offset z z z . The known response curve ϕ ( θ ) \phi(\theta) ϕ ( θ ) consists of normalized pure states and is C 3 C^3 C 3 in norm on [ − r 0 , r 0 ] [-r_0,r_0] [ − r 0 , r 0 ] . Define the state and the squared projective speed at θ = 0 \theta=0 θ = 0 by
ρ ( θ ) = ∣ ϕ ( θ ) ⟩ ⟨ ϕ ( θ ) ∣ , q 0 = ∥ ( I − ρ ( 0 ) ) ϕ ′ ( 0 ) ∥ 2 > 0. (1) \rho(\theta)=\lvert\phi(\theta)\rangle\langle\phi(\theta)\rvert,
\qquad
q_0=\lVert(I-\rho(0))\phi'(0)\rVert^2>0. \tag*{(1)} ρ ( θ ) = ∣ ϕ ( θ )⟩ ⟨ ϕ ( θ )∣ , q 0 = ∥( I − ρ ( 0 )) ϕ ′ ( 0 ) ∥ 2 > 0. (1)
The positive constant q 0 q_0 q 0 sets the local scale of the curve parameter.
Let L ∈ ( 0 , ∞ ) L\in(0,\infty) L ∈ ( 0 , ∞ ) be the known half width of the offset prior. Position and amplitude offset have independent uniform priors,
( u , z ) ∼ Unif [ 0 , 1 ] ⊗ Unif [ − L , L ] . (2) (u,z)\sim\operatorname{Unif}[0,1]\otimes\operatorname{Unif}[-L,L]. \tag*{(2)} ( u , z ) ∼ Unif [ 0 , 1 ] ⊗ Unif [ − L , L ] . (2)
We fix the calibration a = 2 log 2 a=2\sqrt{\log2} a = 2 log 2 , s c = 3 a / 2 s_c=3a/2 s c = 3 a /2 and K = s c 2 K=s_c^2 K = s c 2 . All algorithms are natural. The signal center s > 0 s>0 s > 0 is known. The offset z z z determines the calibrated strength h = s + a z h=s+az h = s + a z , which is the coherent field amplitude in the comparison below. The physical curve parameter is θ = h / n q 0 \theta=h/\sqrt{nq_0} θ = h / n q 0 .
For an ideal step, define k n ( u ) = ⌊ n u ⌋ + 1 k_n(u)=\lfloor nu\rfloor+1 k n ( u ) = ⌊ n u ⌋ + 1 . The state on H ⊗ n \mathcal{H}^{\otimes n} H ⊗ n is
ρ n ( u , z ; s ) = ρ ( 0 ) ⊗ ( k n − 1 ) ⊗ ρ ( θ ) ⊗ ( n − k n + 1 ) . (3) \rho_n(u,z;s)=\rho(0)^{\otimes(k_n-1)}\otimes\rho(\theta)^{\otimes(n-k_n+1)}. \tag*{(3)} ρ n ( u , z ; s ) = ρ ( 0 ) ⊗ ( k n − 1 ) ⊗ ρ ( θ ) ⊗ ( n − k n + 1 ) . (3)
When u < 1 u<1 u < 1 , the unchanged sites 1 , … , k n − 1 1,\ldots,k_n-1 1 , … , k n − 1 precede the first unchanged site k n k_n k n . Under the position prior, k n k_n k n is uniform on { 1 , … , n } \{1,\ldots,n\} { 1 , … , n } ; the zero-mass endpoint u = 1 u=1 u = 1 represents the unchanged auxiliary state. The position target for a finite record is u n = ( k n − 1 ) / n u_n=(k_n-1)/n u n = ( k n − 1 ) / n . Reports lie in X n = { 0 , 1 / n , … , ( n − 1 ) / n } \mathcal{X}_n=\{0,1/n,\ldots,(n-1)/n\} X n = { 0 , 1/ n , … , ( n − 1 ) / n } , and x x x gives the site estimate k ^ = n x + 1 \hat{k}=nx+1 k ^ = n x + 1 .
The condition ( s + a L ) / n q 0 ≤ r 0 (s+aL)/\sqrt{nq_0}\le r_0 ( s + a L ) / n q 0 ≤ r 0 ensures that the full physical parameter interval lies within the domain of the known curve. Where a construction requires a strictly positive amplitude h h h , the additional condition s > a L s>aL s > a L is imposed.
The coherent comparison experiment is defined on bosonic Fock space over L 2 [ 0 , 1 ] L^2[0,1] L 2 [ 0 , 1 ] by
σ s ( u , z ) = ∣ f u , h ⟩ ⟨ f u , h ∣ c o h , f u , h = h 1 [ u , 1 ] . (4) \sigma_s(u,z)=\lvert f_{u,h}\rangle\langle f_{u,h}\rvert_{\mathrm{coh}},
\qquad
f_{u,h}=h\mathbf{1}_{[u,1]}. \tag*{(4)} σ s ( u , z ) = ∣ f u , h ⟩ ⟨ f u , h ∣ coh , f u , h = h 1 [ u , 1 ] . (4)
It retains the same prior on ( u , z ) (u,z) ( u , z ) , but has continuous location target u u u and report space X ∞ = [ 0 , 1 ] \mathcal{X}_\infty= [0,1] X ∞ = [ 0 , 1 ] . For normalized coherent vectors, the squared fidelity is Fid ( ∣ f ⟩ , ∣ g ⟩ ) = e − ∥ f − g ∥ 2 \operatorname{Fid}(|f\rangle,|g\rangle)=e^{-\lVert f-g\rVert^2} Fid ( ∣ f ⟩ , ∣ g ⟩) = e − ∥ f − g ∥ 2 . Distance is measured by the half trace distance D ( ρ , σ ) = ∥ ρ − σ ∥ 1 / 2 D(\rho,\sigma)=\lVert\rho-\sigma\rVert_1/2 D ( ρ , σ ) = ∥ ρ − σ ∥ 1 /2 , which for these pure states satisfies D = 1 − Fid D=\sqrt{1-\operatorname{Fid}} D = 1 − Fid .
Common measurement and calibrated losses
A complete POVM M ( d x , d y ) M(dx,dy) M ( d x , d y ) produces paired reports x ∈ X x\in\mathcal{X} x ∈ X for position and y ∈ [ − L , L ] y\in[-L,L] y ∈ [ − L , L ] for the calibrated offset z z z , where X \mathcal{X} X denotes the report space of the chosen experiment. Completeness is expressed by
M ( d x , d y ) ≥ 0 , ∫ X × [ − L , L ] M ( d x , d y ) = I . M(dx,dy)\geq0,\qquad\int_{\mathcal{X}\times[-L,L]}M(dx,dy)=I. M ( d x , d y ) ≥ 0 , ∫ X × [ − L , L ] M ( d x , d y ) = I .
The measurement design depends only on the known model, prior and task weight. The POVM description covers adaptive strategies based on observed outcomes and retained quantum systems, together with randomized and ancillary implementations. Every outcome ( x , y ) (x,y) ( x , y ) is scored for both tasks, both risks are averaged over the full outcome distribution.
The resolution in the location loss is set by the known signal center. Define b ( θ ) = 1 − ∣ ⟨ ϕ ( 0 ) , ϕ ( θ ) ⟩ ∣ 2 b(\theta)=1-\lvert\langle\phi(0),\phi(\theta)\rangle\rvert^2 b ( θ ) = 1 − ∣⟨ ϕ ( 0 ) , ϕ ( θ )⟩ ∣ 2 and
b n ( s ) = b ( s / n q 0 ) , τ n ( s ) = n b n ( s ) . (5) b_n(s)=b\left(s/\sqrt{nq_0}\right),\qquad\tau_n(s)=nb_n(s). \tag*{(5)} b n ( s ) = b ( s / n q 0 ) , τ n ( s ) = n b n ( s ) . (5)
Position error is scored by clipped absolute loss in resolution units, and amplitude offset error by squared loss in fixed units:
ℓ l o c , n ( x , u ) = 2 min { 1 , τ n ( s ) ∣ x − u n ∣ / K } , ℓ l o c , ∞ ( x , u ) = 2 min { 1 , s 2 ∣ x − u ∣ / K } , ℓ a m p ( y , z ) = ( y − z ) 2 , y ∈ [ − L , L ] . (6) \begin{aligned}
\ell_{\mathrm{loc},n}(x,u)&=2\min\{1,\tau_n(s)\lvert x-u_n\rvert/K\},\\
\ell_{\mathrm{loc},\infty}(x,u)&=2\min\{1,s^2\lvert x-u\rvert/K\},\\
\ell_{\mathrm{amp}}(y,z)&=(y-z)^2,\qquad y\in[-L,L]. \tag*{(6)}
\end{aligned} ℓ loc , n ( x , u ) ℓ loc , ∞ ( x , u ) ℓ amp ( y , z ) = 2 min { 1 , τ n ( s ) ∣ x − u n ∣ / K } , = 2 min { 1 , s 2 ∣ x − u ∣ / K } , = ( y − z ) 2 , y ∈ [ − L , L ] . (6)
Amplitude reports may be clipped to [ − L , L ] [-L,L] [ − L , L ] without increasing the loss. The bounds for location and amplitude are B l o c = 2 B_{\mathrm{loc}}=2 B loc = 2 and B a m p = 4 L 2 B_{\mathrm{amp}}=4L^2 B amp = 4 L 2 , respectively. At s = s c s=s_c s = s c , the coherent location loss becomes 2 ∣ x − u ∣ 2\lvert x-u\rvert 2 ∣ x − u ∣ . For finite records, b n ( s ) ∣ k ^ − k n ∣ b_n(s)\lvert\hat{k}-k_n\rvert b n ( s ) ∣ k ^ − k n ∣ expresses position error in the same resolution units, whereas the physical parameter estimate θ ^ = ( s + a y ) / n q 0 \hat{\theta}=(s+ay)/\sqrt{nq_0} θ ^ = ( s + a y ) / n q 0 has squared error a 2 ( y − z ) 2 / ( n q 0 ) a^2(y-z)^2/(nq_0) a 2 ( y − z ) 2 / ( n q 0 ) under a physical prior of half width a L / n q 0 aL/\sqrt{nq_0} a L / n q 0 .
We compare prior widths while keeping a a a , s s s , q 0 q_0 q 0 , K K K and the loss units fixed. Along physical recovery paths, we take s n → ∞ s_n\to\infty s n → ∞ with s n = o ( n ) s_n=o(\sqrt{n}) s n = o ( n ) and keep the state curve, offset prior and amplitude units fixed. The physical displacement θ \theta θ then vanishes at each site, while the prior spread in z z z remains fixed and position errors are scored in resolution units.
Bayesian risks and the cost of a common readout
For either experiment, write ω u , z \omega_{u,z} ω u , z for the input state and ℓ j \ell_j ℓ j for the loss associated with j ∈ { l o c , a m p } j\in\{\mathrm{loc},\mathrm{amp}\} j ∈ { loc , amp } . The Bayes risk of a paired POVM, averaged over the joint prior in (2), is
r j ( M ) = E u , z ∫ X × [ − L , L ] ℓ j Tr [ ω u , z M ( d x , d y ) ] . r_j(M)=\mathbb{E}_{u,z}\int_{\mathcal{X}\times[-L,L]}\ell_j\operatorname{Tr}[\omega_{u,z}M(dx,dy)]. r j ( M ) = E u , z ∫ X × [ − L , L ] ℓ j Tr [ ω u , z M ( d x , d y )] .
Here ℓ l o c \ell_{\mathrm{loc}} ℓ loc depends on ( x , u ) (x,u) ( x , u ) and ℓ a m p \ell_{\mathrm{amp}} ℓ amp on ( y , z ) (y,z) ( y , z ) . The minimum Bayes risk for task j j j is r j ⋆ = inf M r j ( M ) r_j^\star=\inf_M r_j(M) r j ⋆ = inf M r j ( M ) and optimizes task j j j over the full record and joint prior, averaging over unknown z z z for position and unknown u u u for amplitude. Appending a fixed report x x x for the other task embeds each single task measurement in the paired class, so all three optimizations use the same complete measurements.
With e j ( M ) = r j ( M ) − r j ⋆ e_j(M)=r_j(M)-r_j^\star e j ( M ) = r j ( M ) − r j ⋆ denoting the marginal excess risk, the joint optimum and cost of a common readout at fixed weight 0 ≤ λ ≤ 1 0\leq\lambda\leq1 0 ≤ λ ≤ 1 are
R λ ⋆ = inf M { λ r l o c ( M ) + ( 1 − λ ) r a m p ( M ) } , R_\lambda^\star=\inf_M\{\lambda r_{\mathrm{loc}}(M)+(1-\lambda)r_{\mathrm{amp}}(M)\}, R λ ⋆ = M inf { λ r loc ( M ) + ( 1 − λ ) r amp ( M )} ,
G λ = R λ ⋆ − λ r l o c ⋆ − ( 1 − λ ) r a m p ⋆ = inf M { λ e l o c ( M ) + ( 1 − λ ) e a m p ( M ) } . (7) \mathcal{G}_\lambda=R_\lambda^\star-\lambda r_{\mathrm{loc}}^\star-(1-\lambda)r_{\mathrm{amp}}^\star
=\inf_M\{\lambda e_{\mathrm{loc}}(M)+(1-\lambda)e_{\mathrm{amp}}(M)\}. \tag*{(7)} G λ = R λ ⋆ − λ r loc ⋆ − ( 1 − λ ) r amp ⋆ = M inf { λ e loc ( M ) + ( 1 − λ ) e amp ( M )} . (7)
The cost G λ \mathcal{G}_\lambda G λ is the additional Bayes risk incurred by serving both tasks from the same record. It is nonnegative and satisfies G 0 = G 1 = 0 \mathcal{G}_0=\mathcal{G}_1=0 G 0 = G 1 = 0 . For the weights α + β > 0 \alpha+\beta>0 α + β > 0 , the corresponding unnormalized cost is ( α + β ) G α / ( α + β ) (\alpha+\beta)\mathcal{G}_{\alpha/(\alpha+\beta)} ( α + β ) G α / ( α + β ) .
An exact operator characterization of the gap is obtained from Ω = E u , z ω u , z \Omega=\mathbb{E}_{u,z}\omega_{u,z} Ω = E u , z ω u , z and Ω 1 = E u , z [ z ω u , z ] \Omega_1=\mathbb{E}_{u,z}[z\omega_{u,z}] Ω 1 = E u , z [ z ω u , z ] , both averaged under the full joint prior. The amplitude observable thus incorporates uncertainty in position. All operators in the following statement act on the average state support H Ω = supp Ω \mathcal{H}_\Omega=\operatorname{supp}\Omega H Ω = supp Ω .
Proposition 1 (Exact common optimality). *Assume that the report spaces are compact and the losses are bounded and continuous in the reports. For squared amplitude loss, the spectral measurement of the Peresonic observable A A A minimizes the
Bayes risk [30]. On H Ω \mathcal{H}_{\Omega} H Ω , A A A is the unique bounded self-adjoint solution of
1 2 ( Ω A + A Ω ) = Ω 1 , ∥ A ∥ ≤ L . (8) \frac{1}{2}(\Omega A + A\Omega)=\Omega_1,\qquad\lVert A\rVert\le L. \tag*{(8)} 2 1 ( Ω A + A Ω ) = Ω 1 , ∥ A ∥ ≤ L . (8)
For every 0 < λ < 1 0<\lambda<1 0 < λ < 1 , the gap G λ G_{\lambda} G λ is zero if and only if there exists a POVM N N N that is optimal for position and satisfies
[ N ( B ) , A ] = 0 , (9) [N(B),A]=0, \tag*{(9)} [ N ( B ) , A ] = 0 , (9)
for every Borel set B ⊆ X B\subseteq\mathcal{X} B ⊆ X .
Every amplitude optimal readout has the spectral measure of A A A as its amplitude marginal. A position POVM is jointly measurable with this sharp marginal if and only if all its effects commute with A A A [37]. The gap vanishes at one interior task weight if and only if it vanishes at all. The supplementary proof establishes that the optimal risks are attained.
The prior and loss shape the optimal decisions, connecting the criterion to Bayesian measurement incompatibility [24].
In the smooth onset models of Section 3.2, the position parameter shifts a response profile whose shape and width are known. The prior and paired measurement class are unchanged, while position remains a continuous target even in finite records. Section 4 provides the physical comparison and its resolution units. Throughout, † \dagger † denotes an adjoint and ⋆ \star ⋆ an optimum. In comparisons of physical and coherent records, the subscript n n n denotes a finite physical record; an omitted record size subscript denotes the coherent experiment. For the ideal step, we write G λ ( s ; L ) = G λ ( s , 0 ; L ) G_{\lambda}(s;L)=G_{\lambda}(s,0;L) G λ ( s ; L ) = G λ ( s , 0 ; L ) . Known design arguments are suppressed within a fixed experiment.
Main results
Figures 2–4 summarize the spatial allocation of information, the two limiting procedures, and the role of decision loss.
Ideal steps: conflict and spatial recovery
An ideal step concentrates position information near the boundary and amplitude information across the suffix. This spatial separation gives strong signal recovery for fixed priors with finite strength conflict.
Theorem 1 (Conflict and recovery). Consider the prior and calibrated losses in (2) and (6).
(i) At s = s c s=s_{\mathrm{c}} s = s c , there are constants c , δ > 0 c,\delta>0 c , δ > 0 such that, for every 0 < λ < 1 0<\lambda<1 0 < λ < 1 , the coherent ideal step experiment satisfies
G λ ( s c ; L ) > c min { λ , 1 − λ } , ∣ L − 1 2 ∣ ≤ δ . G_{\lambda}(s_{\mathrm{c}};L)>c\min\{\lambda,1-\lambda\},\qquad\left|L-\frac{1}{2}\right|\le\delta. G λ ( s c ; L ) > c min { λ , 1 − λ } , L − 2 1 ≤ δ . The gap is also positive throughout this interval in all sufficiently long physical records.
(ii) For fixed 0 < L < ∞ 0<L<\infty 0 < L < ∞ and s n → ∞ s_n\to\infty s n → ∞ with s n = o ( n ) s_n=o(\sqrt{n}) s n = o ( n ) , there are complete common measurements M n M_n M n such that
max j e j , n ( M n ) ⟶ 0. \max_j e_{j,n}(M_n)\longrightarrow0. j max e j , n ( M n ) ⟶ 0. Thus G λ , n ( L ) → 0 G_{\lambda,n}(L)\to0 G λ , n ( L ) → 0 . Both marginal optimal risks and their improvements over the prior alone remain bounded away from zero.
The offset prior remains fixed in calibrated units, while the known center sets the position resolution. Conflict and recovery hold under the same prior and loss rule at each fixed prior width in the interval in part (i).
For u ≤ v u \le v u ≤ v , the coherent fields satisfy
∥ f u , h − f v , g ∥ 2 2 = h 2 ( v − u ) + ( h − g ) 2 ( 1 − v ) . (10) \lVert f_{u,h}-f_{v,g}\rVert_{2}^{2}=h^{2}(v-u)+(h-g)^{2}(1-v). \tag*{(10)} ∥ f u , h − f v , g ∥ 2 2 = h 2 ( v − u ) + ( h − g ) 2 ( 1 − v ) . (10)
The first term comes from the interval between the candidate boundaries, and the second from their shared suffix. A complete prelocalization instrument uses the observed coarse label to retain a quantum boundary block for the position decision and the remaining suffix for amplitude (Figure 2 ). The block size balances coarse localization error against the comparison error from replacing the block amplitude by the known center. The risk comparison covers all localization outcomes and the full position prior.
Supplementary Material Sections T1, O1 and D1 establish conflict and its robustness in prior width. Sections A2 and A3 give the recovery proof and the marginal risk bounds, respectively.
Shrinking rises: leading recovery and inverse square cost
For a smooth onset, the overlap between location and amplitude information is controlled by the rise width. Let the known profile F ∈ C 2 ( R ) F \in C^{2}(\mathbb{R}) F ∈ C 2 ( R ) be nondecreasing, equal to zero for x ≤ 0 x \le0 x ≤ 0 and to one for x ≥ 1 x \ge1 x ≥ 1 . The coherent field is f u , h ( w ) ( t ) = h F ( ( t − u ) / w ) f_{u,h}^{(w)}(t)=hF((t-u)/w) f u , h ( w ) ( t ) = h F (( t − u ) / w ) , evaluated at record coordinate t ∈ [ 0 , 1 ] t \in[0,1] t ∈ [ 0 , 1 ] . Let χ = s 2 w \chi=s^{2}w χ = s 2 w be the relative rise width. The coherent state and location resolution are
σ s , w ( u , z ) = ∣ f u , h ( w ) ⟩ ⟨ f u , h ( w ) ∣ c o h , δ s , w = 1 + χ s 2 . (11) \sigma_{s,w}(u,z)=\lvert f_{u,h}^{(w)}\rangle\langle f_{u,h}^{(w)}\rvert_{\mathrm{coh}},
\qquad
\delta_{s,w}=\frac{\sqrt{1+\chi}}{s^{2}}. \tag*{(11)} σ s , w ( u , z ) = ∣ f u , h ( w ) ⟩ ⟨ f u , h ( w ) ∣ coh , δ s , w = s 2 1 + χ . (11)
χ \chi χ compares the rise width with the ideal step resolution. The location scale δ s , w \delta_{s,w} δ s , w changes from s − 2 s^{-2} s − 2 to w / s \sqrt{w}/s w / s as the rise becomes resolved. The signal center s s s , rise width w w w and profile F F F are known. The location loss is defined by
ℓ l o c ( w ) ( x , u ) = 2 min { 1 , ∣ x − u ∣ / ( K δ s , w ) } . (12) \ell_{\mathrm{loc}}^{(w)}(x,u)=2\min\{1,\lvert x-u\rvert/(K\delta_{s,w})\}. \tag*{(12)} ℓ loc ( w ) ( x , u ) = 2 min { 1 , ∣ x − u ∣ / ( K δ s , w )} . (12)
The amplitude loss remains ( y − z ) 2 (y-z)^{2} ( y − z ) 2 . At w = 0 w=0 w = 0 , the convention f u , h ( 0 ) = f u , h f_{u,h}^{(0)}=f_{u,h} f u , h ( 0 ) = f u , h recovers the coherent ideal step location loss. For a smooth rise, position remains a continuous target in both the coherent experiment and finite records.
Theorem 2 (Leading recovery and inverse square cost).
(i) As s → ∞ s \to\infty s → ∞ and w → 0 w \to0 w → 0 , there are complete common readouts whose two marginal excess risks vanish. Consequently, G λ ( s , w ) → 0 \mathcal{G}_{\lambda}(s,w) \to0 G λ ( s , w ) → 0 for 0 ≤ λ ≤ 1 0 \le\lambda\le1 0 ≤ λ ≤ 1 , with arbitrary χ \chi χ . Each marginal optimal risk and its reduction from the prior only risk stay bounded away from zero.
(ii) Take F ( x ) = 10 x 3 − 15 x 4 + 6 x 5 F(x)=10x^{3}-15x^{4}+6x^{5} F ( x ) = 10 x 3 − 15 x 4 + 6 x 5 on [ 0 , 1 ] [0,1] [ 0 , 1 ] and L = 1 / 2 L=1/2 L = 1/2 . There exist constants 0 < C − ≤ C + < ∞ 0<C_{-}\le C_{+}<\infty 0 < C − ≤ C + < ∞ and s 0 > 0 s_{0}>0 s 0 > 0 such that, for all s ≥ s 0 s\ge s_{0} s ≥ s 0 and χ ∈ [ 1 / 2 , 2 ] \chi\in[1/2,2] χ ∈ [ 1/2 , 2 ] ,
C − ≤ s 2 G λ ( s , χ / s 2 ) ≤ C + . (13) C_{-}\le s^{2}\mathcal{G}_{\lambda}(s,\chi/s^{2})\le C_{+}. \tag*{(13)} C − ≤ s 2 G λ ( s , χ / s 2 ) ≤ C + . (13) Coarse localization guides the spatial allocation of the retained field. A shrinking quantum boundary region serves position estimation, and the remaining suffix serves amplitude estimation. Supplementary Material Section J compares this labeled product experiment with the original record in both directions, including late changes, and tracks each marginal optimum separately. At inverse square precision, optimal localization constrains a mode used by the amplitude decision. Its linear response gives the residual cost in the gray region χ ∈ [ 1 / 2 , 2 ] \chi\in[1/2,2] χ ∈ [ 1/2 , 2 ] of Figure 3 . Matching lower and upper bounds establish the inverse square law under the full prior. Comparisons accurate to o ( s − 2 ) o(s^{-2}) o ( s − 2 ) for the joint optimum and both marginal optima preserve the leading residual cost.
Figure 2. Spatial allocation and recovery. The observed coarse localization label u ^ \hat{u} u ^ guides the allocation: the retained quantum boundary region serves position estimation, and the remaining changed suffix serves amplitude estimation. For nondecreasing C 2 C^{2} C 2 profiles, both excess risks vanish in the coherent experiment as s → ∞ s \to\infty s → ∞ and w → 0 w \to0 w → 0 . Risks are averaged over all localization outcomes.
Figure 3. Schematic limiting paths for the coherent quintic experiment with L = 1 / 2 L = 1/2 L = 1/2 and fixed 0 < λ < 1 0 < \lambda< 1 0 < λ < 1 . In the gray region χ ∈ [ 1 / 2 , 2 ] \chi\in[1/2,2] χ ∈ [ 1/2 , 2 ] , G λ = Θ λ ( s − 2 ) \mathcal{G}_{\lambda} = \Theta_{\lambda}(s^{-2}) G λ = Θ λ ( s − 2 ) uniformly in χ \chi χ . The dashed ordered path first takes s → ∞ s \to\infty s → ∞ at fixed w w w , then w ↓ 0 w \downarrow0 w ↓ 0 , giving G ‾ λ ( w ) = Θ λ ( w 3 ) \overline{\mathcal{G}}_{\lambda}(w) = \Theta_{\lambda}(w^{3}) G λ ( w ) = Θ λ ( w 3 ) .
Resolved rises
Theorem 3 (Cubic width cost). For the quintic profile and L = 1 / 2 L = 1/2 L = 1/2 , 0 < λ < 1 0 < \lambda< 1 0 < λ < 1 . The strong signal limit exists. As w ↓ 0 w \downarrow0 w ↓ 0 ,
G ‾ λ ( w ) : = lim s → ∞ G λ ( s , w ) = Θ λ ( w 3 ) . (14) \overline{\mathcal{G}}_{\lambda}(w) := \lim_{s \to\infty} \mathcal{G}_{\lambda}(s,w) = \Theta_{\lambda}(w^{3}). \tag*{(14)} G λ ( w ) := s → ∞ lim G λ ( s , w ) = Θ λ ( w 3 ) . (14)
The loss dependence can be seen in the resolved local experiment (Figure 4 ). At u < 1 u < 1 u < 1 , express the position displacement v v v in units of w / s \sqrt{w}/s w / s . The local state is
Figure 4. Position loss selects the local quantum readout. The panels compare readouts for the same two-mode family, with complete covariant measurements and quadratic amplitude loss. (a) For quadratic position loss, measuring Q 1 Q_1 Q 1 and the optimal residual amplitude observable A r e s = A A_{\mathrm{res}}=A A res = A gives outcomes X , Y X,Y X , Y . The reports v ^ = ( X − d Y ) / κ \hat{v}=(X-dY)/\kappa v ^ = ( X − d Y ) / κ and z ^ = Y \hat{z}=Y z ^ = Y attain both marginal optima within this class. (b) For the quintic local model with L = 1 / 2 L=1/2 L = 1/2 , the position-optimal POVM N l o c ⋆ N_{\mathrm{loc}}^\star N loc ⋆ for clipped absolute loss requires a leading correction involving d 3 C 3 d^3C_3 d 3 C 3 . For sufficiently small c > 0 c>0 c > 0 , [ C 3 , A r e s ] ≠ 0 [C_3,A_{\mathrm{res}}]\ne0 [ C 3 , A res ] = 0 produces a local cost of order d 6 d^6 d 6 as d → 0 d\to0 d → 0 . For complete quintic rises, d 2 = ( 7 log 2 / 10 ) w d^2=(7\log2/10)w d 2 = ( 7 log 2/10 ) w . Averaging the transferred local bounds over the full position prior yields the cubic width law.
∣ κ ( u ) v + d ( u ) z ⟩ ⊗ ∣ c ( u ) z ⟩ , κ ( u ) , c ( u ) > 0. (15) \lvert\kappa(u)v + d(u)z \rangle\otimes\lvert c(u)z \rangle,\qquad\kappa(u),c(u) > 0. \tag*{(15)} ∣ κ ( u ) v + d ( u ) z ⟩ ⊗ ∣ c ( u ) z ⟩ , κ ( u ) , c ( u ) > 0. (15)
The first mode has position sensitivity κ \kappa κ and signed shared amplitude displacement d d d ; the second has residual amplitude sensitivity c c c .
For quadratic local position loss, the residual mode provides an amplitude estimate that can be used to correct the first mode displacement. Let A A A be the optimal amplitude observable for ∣ c z ⟩ \lvert cz \rangle ∣ cz ⟩ . Measuring the first mode quadrature and A A A gives outcomes X X X and Y Y Y , with v ^ = ( X − d Y ) / κ \hat{v} = (X - dY)/\kappa v ^ = ( X − d Y ) / κ and z ^ = Y \hat{z} = Y z ^ = Y . This readout attains both marginal optima within the complete covariant class.
For clipped absolute position loss, the optimal position readout requires a leading cubic correction to linear prediction, involving d 3 C 3 d^{3}C_{3} d 3 C 3 . For fixed sufficiently small c > 0 c > 0 c > 0 , [ C 3 , A ] ≠ 0 [C_{3}, A] \ne0 [ C 3 , A ] = 0 , and the squared correction gives a local cost of order d 6 d^{6} d 6 as d → 0 d \to0 d → 0 . For complete quintic rises,
d 2 = ( 7 log 2 / 10 ) w d^{2} = (7 \log2/10)w d 2 = ( 7 log 2/10 ) w . A complete upper readout controls truncated rises. Comparison with the original coherent experiment and averaging the local bounds over the full position prior then yield the cubic width law.
Supplementary Material Section S gives the local model, the limit at fixed width and the quadratic comparison. Sections P and B establish the cubic response and positive local lower bound; Section U proves the full prior width law.
Numerical certification and transfer to physical records
Numerical certificates of conflict
Two certificates establish finite strength conflict in the coherent ideal step experiment and a physical phase qubit record of 16 qubits. Each compares the joint optimum with both marginal optima under unknown position and amplitude.
At s = s c s = s_c s = s c and L = 1 / 2 L = 1/2 L = 1/2 , the coherent ideal step certificate for equal weights yields
1 × 10 − 3 < G 1 / 2 < 1.83 × 10 − 3 . (16) 1 \times10^{-3} < \mathcal{G}_{1/2} < 1.83 \times10^{-3}. \tag*{(16)} 1 × 1 0 − 3 < G 1/2 < 1.83 × 1 0 − 3 . (16)
Continuity in prior width extends this bound to the conflict interval in Theorem 1. For the original n = 16 n = 16 n = 16 phase qubit record at the same center, certification over the entire known interval L ∈ [ 7 / 16 , 9 / 16 ] L \in[7/16,9/16] L ∈ [ 7/16 , 9/16 ] gives
4.17 × 10 − 4 < G 1 / 2 , 16 ( L ) < 2.242 × 10 − 3 . (17) 4.17 \times10^{-4} < \mathcal{G}_{1/2,16}(L) < 2.242 \times10^{-3}. \tag*{(17)} 4.17 × 1 0 − 4 < G 1/2 , 16 ( L ) < 2.242 × 1 0 − 3 . (17)
Relative to the equally weighted average of the actual marginal optima, the gap lies between 0.58% and 2.76%. Both tasks remain informative throughout this interval, over which calibration and raw loss units are held fixed. Figure 5 displays deterministic certified enclosures over eight prior width cells and tighter bounds at nine nodes.
Figure 5. Certified conflict for an ideal step in a phase record of 16 qubits at s = s c s=s_c s = s c and λ = 1 / 2 \lambda=1/2 λ = 1/2 . The offset prior half width L L L varies at fixed calibration and raw loss units. The gray band encloses 10 3 G 1 / 2 , 16 ( L ) 10^3\mathcal{G}_{1/2,16}(L) 1 0 3 G 1/2 , 16 ( L ) over eight width cells; black capped segments give tighter bounds at nine nodes. Both are deterministic. The cost is between 0.58% and 2.76% of ( r l o c , 16 ⋆ + r a m p , 16 ⋆ ) / 2 (r^\star_{\mathrm{loc},16}+r^\star_{\mathrm{amp},16})/2 ( r loc , 16 ⋆ + r amp , 16 ⋆ ) /2 throughout L ∈ [ 7 / 16 , 9 / 16 ] L\in[7/16,9/16] L ∈ [ 7/16 , 9/16 ] . Tables and error budgets appear in Supplementary Material Section V.
Complete POVMs and feasible decision duals give upper and lower risk bounds, respectively [38]. Continuous bounds in prior width cover the intervals between nodes. Supplementary Material Sections T1, O1, D1 and V give the constructions and error budgets.
Transfer to physical records
Comparisons with physical records resolve each conclusion at its risk scale. Table 1 lists sufficient paths for recovery and inverse square cost in physical records. For smooth rises, the site parameters are θ i = ( s n + a z ) F ( ( i / n − u ) / w n ) / n q 0 \theta_i=(s_n+az)F((i/n-u)/w_n)/\sqrt{nq_0} θ i = ( s n + a z ) F (( i / n − u ) / w n ) / n q 0 . The target u u u remains continuous, and reports lie in X n \mathcal{X}_n X n . With δ n = 1 + w n τ n ( s n ) / τ n ( s n ) \delta_n=\sqrt{1+w_n\tau_n(s_n)/\tau_n(s_n)} δ n = 1 + w n τ n ( s n ) / τ n ( s n ) , the physical position loss is
ℓ l o c , n ( w n ) ( x , u ) = 2 min { 1 , ∣ x − u ∣ / ( K δ n ) } . (18) \ell_{\mathrm{loc},n}^{(w_n)}(x,u)=2\min\{1,|x-u|/(K\delta_n)\}. \tag*{(18)} ℓ loc , n ( w n ) ( x , u ) = 2 min { 1 , ∣ x − u ∣/ ( K δ n )} . (18)
The amplitude loss remains ( y − z ) 2 (y-z)^2 ( y − z ) 2 . Comparisons in both directions transfer the joint optimum and both marginal optima. Leading recovery uses a vanishing risk error, the inverse square cost uses o ( s n − 2 ) o(s_n^{-2}) o ( s n − 2 ) precision. Supplementary Material Sections A2 and C11 establish these paths.
The amplitude loss remains ( y − z ) 2 (y-z)^2 ( y − z ) 2 . Comparisons in both directions transfer the joint optimum and both marginal optima. Leading recovery uses a vanishing risk error, the inverse square cost uses o ( s n − 2 ) o(s_n^{-2}) o ( s n − 2 ) precision. Supplementary Material Sections A2 and C11 establish these paths.
Discussion
The common readout cost quantifies how onset geometry and decision loss shape simultaneous position and amplitude estimation. The joint and marginal optimizations use the full record and the same joint prior, with both parameters unknown. For ideal steps, we certify a positive cost at a reference signal strength and recover both marginal optima at strong signal under the same prior. These results show how finite signal conflict and asymptotic compatibility arise within the same estimation task.
Recovery exploits the spatial distribution of information in the record. Coarse localization preserves a quantum boundary region for the position decision and leaves the remaining suffix for amplitude estimation. Known nondecreasing C 2 C^2 C 2 profiles also admit leading recovery in the coherent experiment. This holds along every path s → ∞ s\to\infty s → ∞ , w → 0 w\to0 w → 0 , with arbitrary variation of the relative width χ = s 2 w \chi=s^2w χ = s 2 w . Both marginal optimal risks and their improvements over the prior alone remain bounded away from zero. The common readout, therefore, approaches two informative marginal optima at their respective loss scales.
The residual laws quantify the coupling that remains beyond leading recovery. For the coherent quintic experiment with L = 1 / 2 L = 1/2 L = 1/2 , two limiting procedures give distinct scaling laws. The cost is Θ λ ( s − 2 ) \Theta_\lambda(s^{-2}) Θ λ ( s − 2 ) . Taking s → ∞ s \to\infty s → ∞ at fixed w w w before letting w ↓ 0 g i v e s w \downarrow0 gives w ↓ 0 g i v es \overline{G}_\lambda(w) = \Theta_\lambda(w^3)$. These laws describe how the relation between rise width and position resolution controls the decay of the common readout cost. The resolved local experiment explains the role of the loss function. With quadratic position and amplitude losses, amplitude prediction from the residual mode supplies a linear correction for position. The resulting complete readout attains both marginal optima within the local covariant class. With clipped absolute position loss, a cubic correction produces a local cost of order d 6 d^6 d 6 at each fixed sufficiently small positive residual sensitivity. Transferring the local bounds to the full position prior yields the cubic width law. The loss thus determines how shared information enters the optimal decisions and their residual incompatibility.
Proposition 1 gives an exact criterion: compatibility holds if and only if some optimal position POVM commutes with the optimal amplitude observable. Leading recovery describes the approach to the marginal optima, and the residual laws quantify its rate. Certified bounds establish positive conflict in finite phase records (Figure 5). Comparisons at the required risk precision transfer recovery and the inverse square crossover cost to physical records along Table 1. The spatial allocation of information and the decisions selected by the loss connect quantum change point metrology with Bayesian measurement compatibility.
Future work can examine how the onset profile and position loss determine the first correction to linear amplitude prediction. Comparing these corrections across models would clarify which features set the residual exponent and which scaling laws persist across different designs. Treating rise width or profile shape as additional unknowns would reveal how uncertainty in the onset affects spatial recovery. Finite record implementations could also assess the tradeoff between estimation precision and the quantum memory required to retain the boundary region.
Author contributions
Xu Chen conceived the initial idea, developed the physical model, and carried out the formal analysis. Xue Ma performed the numerical computations and prepared the data visualizations. Both authors contributed writing and revision of the manuscript. LLM Models were used for language polishing and cross checking numerical results. Both authors reviewed the outputs and take full responsibility for the scientific content and the final manuscript.
Acknowledgements
This work was supported by Hebei University of Science and Technology under Starting Grant No. 81/1181298.
References
Supplementary Material
This supplement supplies the proofs and numerical error budgets for the main article. The following map identifies the argument supporting each main conclusion.
Table 1 (PDF p. 11).
Table 1. Sufficient paths to finite product records. The comparisons preserve the unknown parameters, calibrated losses and complete readouts to the precision required by each conclusion.
Conventions. The priors, calibrated losses, reports and state curve are those of the main text. Trace distance is half the trace norm and fidelity is squared. For a synthesis T T T , let Ω = T T † \Omega= TT^{\dagger} Ω = T T † and ∥ D ∥ Ω = ∥ D Ω 1 / 2 ∥ H S \lVert D\rVert_{\Omega} = \lVert D\Omega^{1/2}\rVert_{\mathrm{HS}} ∥ D ∥ Ω = ∥ D Ω 1/2 ∥ HS ; finite unnormalized masses Tr Ω \operatorname{Tr}\Omega Tr Ω are retained. A Naimark isometry is denoted by J J J , with commuting report observables X , Y X,Y X , Y . Strictly positive gap constants concern a fixed 0 < λ < 1 0 < \lambda< 1 0 < λ < 1 ; the endpoint weights have zero gap. The coherent field amplitude is h = s + a z h = s + az h = s + a z , the physical curve parameter is θ \theta θ , and t t t is a record or integration coordinate.
For the residual ensemble ∣ c z ⟩ \lvert cz\rangle ∣ cz ⟩ with z ∼ Unif [ − L , L ] z \sim\operatorname{Unif}[-L,L] z ∼ Unif [ − L , L ] , write its amplitude optimum as A L ( c ) A_L(c) A L ( c ) and its suffix average as A ‾ L = ∫ 0 1 A L ( a e ) d e \overline{A}_L = \int_0^1 A_L(a\sqrt{e})\,\mathrm{d}e A L = ∫ 0 1 A L ( a e ) d e . The local covariant gap is g λ ( κ , d , c ) g_{\lambda}(\kappa,d,c) g λ ( κ , d , c ) ; g λ , w ( u ) g_{\lambda,w}(u) g λ , w ( u ) uses the actual sensitivities, including truncated rises. For complete quintic rises, κ F = 10 7 \kappa_F = \sqrt{\frac{10}{7}} κ F = 7 10 and g λ ( d ; c ) = g λ ( κ F , d , c ) g_{\lambda}(d;c) = g_{\lambda}(\kappa_F,d,c) g λ ( d ; c ) = g λ ( κ F , d , c ) . The full coherent gap is G λ ( s , w ) G_{\lambda}(s,w) G λ ( s , w ) , the physical gap is G λ , n G_{\lambda,n} G λ , n , and G λ ( w ) = lim s → ∞ G λ ( s , w ) G_{\lambda}(w) = \lim_{s\to\infty}G_{\lambda}(s,w) G λ ( w ) = lim s → ∞ G λ ( s , w ) .
Proof of exact common optimality
Let ω u , z \omega_{u,z} ω u , z be trace norm measurable input states on a separable Hilbert space, with the joint prior of Proposition 1. The position report space X \mathcal{X} X is compact metric. The location loss is measurable, nonnegative, uniformly bounded and continuous in the report for almost every parameter. Amplitude reports lie in [ − L , L ] [-L,L] [ − L , L ] . Define
Ω j = E [ z j ω u , z ] , j = 0 , 1 , 2 , Ω = Ω 0 , C x = E [ ℓ l o c ( x , u ) ω u , z ] . \Omega_j = \mathbb{E}[z^j\omega_{u,z}],\quad j = 0,1,2,\qquad\Omega= \Omega_0,\qquad C_x = \mathbb{E}[\ell_{\mathrm{loc}}(x,u)\omega_{u,z}]. Ω j = E [ z j ω u , z ] , j = 0 , 1 , 2 , Ω = Ω 0 , C x = E [ ℓ loc ( x , u ) ω u , z ] .
Dominated convergence makes C x C_x C x continuous in trace norm. Compression to H Ω = supp Ω \mathcal{H}_{\Omega} = \operatorname{supp}\Omega H Ω = supp Ω preserves all risks. On this support, Ω \Omega Ω is faithful and trace class.
Ideal step conflict and recovery
Proof of Proposition 1. Amplitude optimum. Since − L Ω ⪯ Ω 1 ⪯ L Ω -L\Omega\preceq\Omega_1 \preceq L\Omega − L Ω ⪯ Ω 1 ⪯ L Ω , the operators
A R = 2 ∫ 0 R e − t Ω Ω 1 e − t Ω d t A_R = 2\int_0^R e^{-t\Omega}\Omega_1e^{-t\Omega}\,\mathrm{d}t A R = 2 ∫ 0 R e − t Ω Ω 1 e − t Ω d t
satisfy − L ( I − e − 2 R Ω ) ⪯ A R ⪯ L ( I − e − 2 R Ω ) -L(I-e^{-2R\Omega}) \preceq A_R \preceq L(I-e^{-2R\Omega}) − L ( I − e − 2 R Ω ) ⪯ A R ⪯ L ( I − e − 2 R Ω ) .
The integrals with positive source Ω 1 + L Ω \Omega_1+L\Omega Ω 1 + L Ω increase to a bounded operator. As e − 2 R Ω → 0 e^{-2R\Omega}\to0 e − 2 R Ω → 0 strongly, A R A_R A R converges strongly to a self-adjoint A A A with ∥ A ∥ ≤ L \lVert A\rVert\le L ∥ A ∥ ≤ L . Integration gives
1 2 ( Ω A R + A R Ω ) = Ω 1 − e − R Ω Ω 1 e − R Ω . \frac{1}{2}(\Omega A_R+A_R\Omega)=\Omega_1-e^{-R\Omega}\Omega_1e^{-R\Omega}. 2 1 ( Ω A R + A R Ω ) = Ω 1 − e − R Ω Ω 1 e − R Ω .
The remainder tends to zero in trace norm, proving the Personick equation. In an eigenbasis Ω e i = ω i e i \Omega e_i=\omega_i e_i Ω e i = ω i e i , ω i > 0 \omega_i>0 ω i > 0 , its solution has entries A i j = 2 ( Ω 1 ) i j / ( ω i + ω j ) A_{ij}=2(\Omega_1)_{ij}/(\omega_i+\omega_j) A ij = 2 ( Ω 1 ) ij / ( ω i + ω j ) , which proves uniqueness among bounded operators.
Dilate a paired POVM as M ( Q ) = J † E ( Q ) J M(Q)=J^\dagger E(Q)J M ( Q ) = J † E ( Q ) J and write Y = ∫ y E ( d x , d y ) Y=\int y\,E(\mathrm{d}x,\mathrm{d}y) Y = ∫ y E ( d x , d y ) . Completing the amplitude square yields
r a m p ∗ = Tr Ω 2 − Tr ( Ω A 2 ) , e a m p ( M ) = ∥ Y J − J A ∥ Ω 2 . r_{\mathrm{amp}}^*=\operatorname{Tr}\Omega_2-\operatorname{Tr}(\Omega A^2),\qquad e_{\mathrm{amp}}(M)=\lVert YJ-JA\rVert_\Omega^2. r amp ∗ = Tr Ω 2 − Tr ( Ω A 2 ) , e amp ( M ) = ∥ Y J − J A ∥ Ω 2 .
The spectral measurement of A A A makes this square zero. Its outcomes lie in [ − L , L ] [-L,L] [ − L , L ] , so it is an admissible amplitude optimum.
Attainment. For a compact metric report space K K K , complete POVMs are positive unital maps Φ : C ( K ) → B ( H Ω ) \Phi:C(K)\to\mathcal{B}(\mathcal{H}_\Omega) Φ : C ( K ) → B ( H Ω ) . The bounds ∥ Φ ( f ) ∥ ≤ ∥ f ∥ ∞ \lVert\Phi(f)\rVert\le\lVert f\rVert_\infty ∥ Φ ( f )∥ ≤ ∥ f ∥ ∞ give compactness in the pointwise ultraweak topology. Positivity and unitality pass to the limit, and scalar Riesz representation recovers a POVM.
Choose finite-rank projections P N ↑ I P_N\uparrow I P N ↑ I , put m = Tr Ω m=\operatorname{Tr}\Omega m = Tr Ω and η N = Tr [ ( I − P N ) Ω ] \eta_N=\operatorname{Tr}[(I-P_N)\Omega] η N = Tr [( I − P N ) Ω ] , and let r j ( N ) r_j^{(N)} r j ( N ) use the compressed states P N ω u , z P N P_N\omega_{u,z}P_N P N ω u , z P N . The projection trace-distance bound and Cauchy–Schwarz give, uniformly in M M M ,
∣ r j ( M ) − r j ( N ) ( M ) ∣ ≤ B j E ∥ ω u , z − P N ω u , z P N ∥ 1 ≤ 2 B j m η N ⟶ 0. \lvert r_j(M)-r_j^{(N)}(M)\rvert\le B_j\mathbb{E}\lVert\omega_{u,z}-P_N\omega_{u,z}P_N\rVert_1\le2B_j\sqrt{m\eta_N}\longrightarrow0. ∣ r j ( M ) − r j ( N ) ( M )∣ ≤ B j E ∥ ω u , z − P N ω u , z P N ∥ 1 ≤ 2 B j m η N ⟶ 0.
Each compressed cost field is a finite sum of fixed matrix units with continuous coefficients in the report. Its risk is therefore continuous in the preceding topology. The full risk is a uniform limit of these continuous functions. Compactness gives minimizers for the two marginal risks and every weighted joint risk.
Commuting position effects. For 0 < λ < 1 0<\lambda<1 0 < λ < 1 , an attained joint optimum with G λ = 0 G_\lambda=0 G λ = 0 has e l o c = e a m p = 0 e_{\mathrm{loc}}=e_{\mathrm{amp}}=0 e loc = e amp = 0 . The amplitude square and the dense range of Ω 1 / 2 \Omega^{1/2} Ω 1/2 imply Y J = J A YJ=JA Y J = J A . With E l o c ( B ) = E ( B × [ − L , L ] ) E_{\mathrm{loc}}(B)=E(B\times[-L,L]) E loc ( B ) = E ( B × [ − L , L ]) and N ( B ) = J † E l o c ( B ) J N(B)=J^\dagger E_{\mathrm{loc}}(B)J N ( B ) = J † E loc ( B ) J ,
[ N ( B ) , A ] = J † [ E l o c ( B ) , Y ] J = 0. [N(B),A]=J^\dagger[E_{\mathrm{loc}}(B),Y]J=0. [ N ( B ) , A ] = J † [ E loc ( B ) , Y ] J = 0.
The marginal N N N is optimal for position. Functional calculus also gives E Y ( D ) J = J E A ( D ) E_Y(D)J=JE_A(D) E Y ( D ) J = J E A ( D ) , so the amplitude marginal is the spectral measure E A E_A E A of A A A .
Conversely, a position optimum N N N commuting with A A A commutes with its spectral measure E A E_A E A . On finite report partitions, N ( B i ) E A ( D j ) N(B_i)E_A(D_j) N ( B i ) E A ( D j ) are positive complete effects. Refining the partitions and applying the preceding compactness argument gives a paired POVM with marginals N , E A N,E_A N , E A . Its rectangle effects are M ( B × D ) = N ( B ) E A ( D ) M(B\times D)=N(B)E_A(D) M ( B × D ) = N ( B ) E A ( D ) , since the second marginal is sharp [37, Proposition 8]. Both marginal risks are therefore optimal at every weight. A fixed paired report on H Ω \mathcal{H}_\Omega H Ω extends the rule to the original Hilbert space.
Ideal step conflict, recovery and physical comparison
Risk transfer and finite records
Lemma A1.1 (Comparison of the three risks). Let E E E , F F F share the same prior, targets, actions and losses 0 ≤ ℓ j ≤ B j 0 \le\ell_j \le B_j 0 ≤ ℓ j ≤ B j . Let T : E → F T:E \to F T : E → F and S : F → E S:F \to E S : F → E be channels independent of the parameters with trace distance errors, uniform or averaged under the prior, ε T , ε S \varepsilon_T,\varepsilon_S ε T , ε S . For B λ = λ B l o c + ( 1 − λ ) B a m p B_\lambda=\lambda B_{\mathrm{loc}}+(1-\lambda)B_{\mathrm{amp}} B λ = λ B loc + ( 1 − λ ) B amp , the marginal risks, weighted joint risk and gap satisfy
− B j ε S ≤ r j ⋆ ( E ) − r j ⋆ ( F ) ≤ B j ε T , -B_j\varepsilon_S \le r_j^\star(E)-r_j^\star(F) \le B_j\varepsilon_T, − B j ε S ≤ r j ⋆ ( E ) − r j ⋆ ( F ) ≤ B j ε T , − B λ ε S ≤ R λ ⋆ ( E ) − R λ ⋆ ( F ) ≤ B λ ε T , -B_\lambda\varepsilon_S \le R_\lambda^\star(E)-R_\lambda^\star(F) \le B_\lambda\varepsilon_T, − B λ ε S ≤ R λ ⋆ ( E ) − R λ ⋆ ( F ) ≤ B λ ε T , ∣ G λ ( E ) − G λ ( F ) ∣ ≤ B λ ( ε T + ε S ) . \lvert G_\lambda(E)-G_\lambda(F)\rvert\le B_\lambda(\varepsilon_T+\varepsilon_S). ∣ G λ ( E ) − G λ ( F )∣ ≤ B λ ( ε T + ε S ) . A rule of excess δ j \delta_j δ j pulls back through either channel with excess at most δ j + B j ( ε T + ε S ) \delta_j+B_j(\varepsilon_T+\varepsilon_S) δ j + B j ( ε T + ε S ) . Loss changes bounded by d j d_j d j change the gap by at most 2 [ λ d l o c + ( 1 − λ ) d a m p ] 2[\lambda d_{\mathrm{loc}}+(1-\lambda)d_{\mathrm{amp}}] 2 [ λ d loc + ( 1 − λ ) d amp ] . An ordered change to the position loss alone changes it by at most λ d l o c \lambda d_{\mathrm{loc}} λ d loc .
Proof. Pullback through T T T gives r j ⋆ ( E ) ≤ r j ⋆ ( F ) + B j ε T r_j^\star(E)\le r_j^\star(F)+B_j\varepsilon_T r j ⋆ ( E ) ≤ r j ⋆ ( F ) + B j ε T . The reverse channel gives the other inequality. The joint and weighted marginal increments lie in one interval of length B λ ( ε T + ε S ) B_\lambda(\varepsilon_T+\varepsilon_S) B λ ( ε T + ε S ) , which bounds their difference. Comparison with the corresponding infimum gives the excess bound. A loss change bounds each infimum by its uniform norm. For an ordered location change, both increments lie in [ 0 , λ d l o c ] [0,\lambda d_{\mathrm{loc}}] [ 0 , λ d loc ] .
Theorem A1.2 (Compact coherent comparison). Let M j = sup ∣ θ ∣ ≤ r 0 ∥ ϕ ( j ) ( θ ) ∥ M_j=\sup_{\lvert\theta\rvert\le r_0}\lVert\phi^{(j)}(\theta)\rVert M j = sup ∣ θ ∣ ≤ r 0 ∥ ϕ ( j ) ( θ )∥ , γ 0 = − i ⟨ ϕ ( 0 ) , ϕ ′ ( 0 ) ⟩ \gamma_0=-i\langle\phi(0),\phi'(0)\rangle γ 0 = − i ⟨ ϕ ( 0 ) , ϕ ′ ( 0 )⟩ , B 2 = M 2 + 2 ∣ γ 0 ∣ M 1 + γ 0 2 B_2=M_2+2\lvert\gamma_0\rvert M_1+\gamma_0^2 B 2 = M 2 + 2 ∣ γ 0 ∣ M 1 + γ 0 2 and C ϕ = ( B 2 / q 0 + 3 ) / 2 C_\phi=(B_2/q_0+\sqrt{3})/2 C ϕ = ( B 2 / q 0 + 3 ) /2 . For finite H ≥ 0 H\ge0 H ≥ 0 , put
C H = H + C ϕ H 2 , ε n , H = min { 1 , C H / n } . C_H=H+C_\phi H^2,\qquad\varepsilon_{n,H}=\min\{1,C_H/\sqrt{n}\}. C H = H + C ϕ H 2 , ε n , H = min { 1 , C H / n } . For H / n q 0 ≤ r 0 H/\sqrt{nq_0}\le r_0 H / n q 0 ≤ r 0 , CPTP channels compare the finite and coherent records in both directions with error at most ε n , H \varepsilon_{n,H} ε n , H , uniformly in u ∈ [ 0 , 1 ] u\in[0,1] u ∈ [ 0 , 1 ] , ∣ h ∣ ≤ H \lvert h\rvert\le H ∣ h ∣ ≤ H . The channels depend only on n n n and the known curve. For ∣ s ∣ + a L ≤ H \lvert s\rvert+aL\le H ∣ s ∣ + a L ≤ H , set d l o c , n = 2 ( ∣ τ n − s ∣ + τ n / n ) / K d_{\mathrm{loc},n}=2(\lvert\tau_n-s\rvert+\tau_n/n)/K d loc , n = 2 (∣ τ n − s ∣ + τ n / n ) / K . The corresponding marginal risks, joint risk and gap obey
∣ r l o c , n ⋆ − r l o c , ∞ ⋆ ∣ ≤ 2 ε n , H + d l o c , n , ∣ r a m p , n ⋆ − r a m p , ∞ ⋆ ∣ ≤ 4 L 2 ε n , H , \lvert r_{\mathrm{loc},n}^\star-r_{\mathrm{loc},\infty}^\star\rvert\le2\varepsilon_{n,H}+d_{\mathrm{loc},n},
\qquad
\lvert r_{\mathrm{amp},n}^\star-r_{\mathrm{amp},\infty}^\star\rvert\le4L^2\varepsilon_{n,H}, ∣ r loc , n ⋆ − r loc , ∞ ⋆ ∣ ≤ 2 ε n , H + d loc , n , ∣ r amp , n ⋆ − r amp , ∞ ⋆ ∣ ≤ 4 L 2 ε n , H , ∣ R λ , n ⋆ − R λ , ∞ ⋆ ∣ ≤ B λ ε n , H + λ d l o c , n , ∣ G λ , n − G λ ∣ ≤ 2 B λ ε n , H + 2 λ d l o c , n , \lvert R_{\lambda,n}^\star-R_{\lambda,\infty}^\star\rvert\le B_\lambda\varepsilon_{n,H}+\lambda d_{\mathrm{loc},n},
\qquad
\lvert G_{\lambda,n}-G_\lambda\rvert\le2B_\lambda\varepsilon_{n,H}+2\lambda d_{\mathrm{loc},n}, ∣ R λ , n ⋆ − R λ , ∞ ⋆ ∣ ≤ B λ ε n , H + λ d loc , n , ∣ G λ , n − G λ ∣ ≤ 2 B λ ε n , H + 2 λ d loc , n , where B λ = 2 λ + 4 L 2 ( 1 − λ ) B_\lambda=2\lambda+4L^2(1-\lambda) B λ = 2 λ + 4 L 2 ( 1 − λ ) .
Proof. Use the gauge ψ ( θ ) = e − i γ 0 θ ϕ ( θ ) \psi(\theta)=e^{-i\gamma_0\theta}\phi(\theta) ψ ( θ ) = e − i γ 0 θ ϕ ( θ ) and a basis starting with e 0 = ϕ ( 0 ) e_0=\phi(0) e 0 = ϕ ( 0 ) , e 1 = ψ ′ ( 0 ) / q 0 e_1=\psi'(0)/\sqrt{q_0} e 1 = ψ ′ ( 0 ) / q 0 . Mapping e 0 e_0 e 0 to vacuum and the transverse vectors to single photons of distinct colors defines a product isometry V n V_n V n . The cell coherent vector c θ = ∣ θ q 0 e 1 ⟩ c_\theta=\lvert\theta\sqrt{q_0}e_1\rangle c θ = ∣ θ q 0 e 1 ⟩ agrees with the mapped curve to first order and satisfies ∥ V c e l l ψ ( θ ) − c θ ∥ ≤ C ϕ q 0 θ 2 \lVert V_{\mathrm{cell}}\psi(\theta)-c_\theta\rVert\le C_\phi q_0\theta^2 ∥ V cell ψ ( θ ) − c θ ∥ ≤ C ϕ q 0 θ 2 .
Product fidelity accumulates the cell error with a factor n \sqrt{n} n ; replacing u n u_n u n by u u u costs H / n H/\sqrt{n} H / n . With Π n = V n V n † \Pi_n=V_nV_n^\dagger Π n = V n V n † and any fixed density operator ρ f i x \rho_{\mathrm{fix}} ρ fix on the finite record Hilbert space, define
C n ( ω ) = V n † ω V n + Tr [ ( I − Π n ) ω ] ρ f i x , T n ( ρ ) = Tr o t h e r ( V n ρ V n † ) , S n ( σ ) = C n ( σ ⊗ ∣ 0 ⟩ ⟨ 0 ∣ o t h e r ) . C_n(\omega)=V_n^\dagger\omega V_n+\operatorname{Tr}[(I-\Pi_n)\omega]\rho_{\mathrm{fix}},\qquad
T_n(\rho)=\operatorname{Tr}_{\mathrm{other}}(V_n\rho V_n^\dagger),\qquad
S_n(\sigma)=C_n(\sigma\otimes\lvert0\rangle\langle0\rvert_{\mathrm{other}}). C n ( ω ) = V n † ω V n + Tr [( I − Π n ) ω ] ρ fix , T n ( ρ ) = Tr other ( V n ρ V n † ) , S n ( σ ) = C n ( σ ⊗ ∣ 0 ⟩ ⟨ 0 ∣ other ) .
These maps are completely positive and trace preserving, with C n ( V n ρ V n † ) = ρ C_n(V_n\rho V_n^\dagger)=\rho C n ( V n ρ V n † ) = ρ . Contractivity gives both comparisons.
Round t t t to min { n − 1 , ⌊ n x ⌋ } / n \min\{n-1,\lfloor nx\rfloor\}/n min { n − 1 , ⌊ n x ⌋} / n . The report rounding and u − u n u-u_n u − u n both lie in [ 0 , 1 / n ] [0,1/n] [ 0 , 1/ n ] , so the displacement changes by at most 1 / n 1/n 1/ n . The Lipschitz bound for the loss gives d l o c , n d_{\mathrm{loc},n} d loc , n ; Lemma A1.1 then transfers the optima.
Recovery by retaining the boundary region
Theorem A2.1 (Recovery with aligned blocks). Use the prior and losses of (2), (6), with B l o c = 2 B_{\mathrm{loc}}=2 B loc = 2 , B a m p = 4 L 2 B_{\mathrm{amp}}=4L^2 B amp = 4 L 2 . Fix 0 < L < ∞ 0<L<\infty 0 < L < ∞ so that the physical interval lies in the curve domain. Partition the record into N b = ⌈ n / m ⌉ N_b=\lceil n/m\rceil N b = ⌈ n / m ⌉ blocks of length at most m m m , including the short last block, and assume p − = inf ∣ z ∣ ≤ L b ( ( s + a z ) / n q 0 ) > 0 p_-=\inf_{|z|\le L}b((s+az)/\sqrt{nq_0})>0 p − = inf ∣ z ∣ ≤ L b (( s + a z ) / n q 0 ) > 0 . For any δ j > 0 \delta_j>0 δ j > 0 , one complete paired rule gives
e j ≤ B j E n , m + δ j , E n , m = min { 1 , 5 N b 2 n p − + M 1 a L 2 q 0 m + 1 n } . (19) e_j\le B_jE_{n,m}+\delta_j,\qquad E_{n,m}=\min\left\{1,\frac{5N_b}{2np_-}+\frac{M_1aL}{\sqrt{2q_0}}\sqrt{\frac{m+1}{n}}\right\}. \tag*{(19)} e j ≤ B j E n , m + δ j , E n , m = min { 1 , 2 n p − 5 N b + 2 q 0 M 1 a L n m + 1 } . (19) The bounds hold for both tasks, so G λ ≤ B λ E n , m G_\lambda\le B_\lambda E_{n,m} G λ ≤ B λ E n , m at every weight.
Proof. Test successive blocks against their initial state projections, stopping at the first complementary outcome and retaining its label and output record, with no click as a separate outcome. Only blocks at or after the boundary can click. For l l l changed sites in the boundary block, the missed mass is a l = ( 1 − p ) l a_l=(1-p)^l a l = ( 1 − p ) l . The distance from the untouched record with the correct label is ( 4 a l − 3 a l 2 + a l ) / 2 ≤ a l + a l / 2 (\sqrt{4a_l-3a_l^2}+a_l)/2\le\sqrt{a_l}+a_l/2 ( 4 a l − 3 a l 2 + a l ) /2 ≤ a l + a l /2 : the lost vacuum amplitude and all other labels contribute these two terms. Summing over l l l and blocks under the uniform position prior gives e ≤ 5 N b / ( 2 n p − ) e\le5N_b/(2np_-) e ≤ 5 N b / ( 2 n p − ) .
Given the correct label, replace the boundary block amplitude by s / n q 0 s/\sqrt{nq_0} s / n q 0 . The conditional prior factors into position in the block and amplitude in the suffix. Pair rules approaching the two infima, using the prior alone for an empty suffix. The replacement error is at most M 1 a ∣ z ∣ l / ( n q 0 ) M_1a|z|\sqrt{l/(nq_0)} M 1 a ∣ z ∣ l / ( n q 0 ) . Averaging with E ∣ z ∣ = L / 2 \mathbb{E}|z|=L/2 E ∣ z ∣ = L /2 and E l ≤ ( m + 1 ) / 2 \mathbb{E}l\le(m+1)/2 E l ≤ ( m + 1 ) /2 gives f ≤ M 1 a L ( m + 1 ) / ( n q 0 ) / ( 2 2 ) f\le M_1aL\sqrt{(m+1)/(nq_0)}/(2\sqrt{2}) f ≤ M 1 a L ( m + 1 ) / ( n q 0 ) / ( 2 2 ) . The forward error is e + f e+f e + f ; discarding the label defines the reverse channel with error f f f . Lemma A1.1 gives B j ( e + 2 f ) + δ j B_j(e+2f)+\delta_j B j ( e + 2 f ) + δ j , capped by B j B_j B j .
The curve bound gives p − ≥ b n − M 1 a L / n q 0 \sqrt{p_-}\ge\sqrt{b_n}-M_1aL/\sqrt{nq_0} p − ≥ b n − M 1 a L / n q 0 . For s n → ∞ s_n\to\infty s n → ∞ , s n = o ( n ) s_n=o(\sqrt n) s n = o ( n ) , b n ∼ s n 2 / n b_n\sim s_n^2/n b n ∼ s n 2 / n and eventually p − ≥ b n / 4 p_-\ge b_n/4 p − ≥ b n /4 . Taking m = ⌈ n τ n − 2 / 3 ⌉ m=\lceil n\tau_n^{-2/3}\rceil m = ⌈ n τ n − 2/3 ⌉ makes 5 N b / ( 2 n p − ) ≤ 10 τ n − 1 / 3 + 10 / τ n 5N_b/(2np_-)\le10\tau_n^{-1/3}+10/\tau_n 5 N b / ( 2 n p − ) ≤ 10 τ n − 1/3 + 10/ τ n , while the replacement term is O L ( τ n − 1 / 3 + n − 1 / 2 ) O_L(\tau_n^{-1/3}+n^{-1/2}) O L ( τ n − 1/3 + n − 1/2 ) . Letting δ j ↓ 0 \delta_j\downarrow0 δ j ↓ 0 proves Theorem 1(ii).
Lemma A3.1 (Information bound for a uniform amplitude prior). Let z ∼ Unif [ − L , L ] z\sim\operatorname{Unif}[-L,L] z ∼ Unif [ − L , L ] , L > 0 L>0 L > 0 , and let the trace norm differentiable state model have quantum Fisher information at most I \mathcal{I} I uniformly in z z z . Every measurement and amplitude report Y Y Y satisfy
E ( Y − z ) 2 ≥ { 3 / L 2 + ( 6 / 5 ) I } − 1 . \mathbb{E}(Y-z)^2\ge\left\{3/L^2+(6/5)\mathcal{I}\right\}^{-1}. E ( Y − z ) 2 ≥ { 3/ L 2 + ( 6/5 ) I } − 1 . Proof. After the report is clipped to [ − L , L ] [-L,L] [ − L , L ] , integration by parts with g = L 2 − z 2 g=L^2-z^2 g = L 2 − z 2 gives E [ ( Y − z ) ( g ′ + g ∂ z log p z ) ] = 2 L 2 / 3 \mathbb{E}[(Y-z)(g'+g\partial_z\log p_z)]=2L^2/3 E [( Y − z ) ( g ′ + g ∂ z log p z )] = 2 L 2 /3 ; the endpoint term vanishes. Cauchy–Schwarz then proves the bound using E ( g ′ ) 2 = 4 L 2 / 3 \mathbb{E}(g')^2=4L^2/3 E ( g ′ ) 2 = 4 L 2 /3 , E g 2 = 8 L 4 / 15 \mathbb{E}g^2=8L^4/15 E g 2 = 8 L 4 /15 , the zero conditional mean of the score and the classical Fisher information bound ≤ I \le\mathcal{I} ≤ I .
For zero likelihoods, choose a faithful reference law and use ( 1 − η ) p z + η (1-\eta)p_z+\eta ( 1 − η ) p z + η . Trace norm differentiability and the information bound justify the score identity; bounded reports give risk convergence as $\eta\downarrow0.
Informative marginals in Theorem 1. Let θ 0 , n = s n / n q 0 \theta_{0,n}=s_n/\sqrt{nq_0} θ 0 , n = s n / n q 0 denote the known physical center. Set κ n = M 1 a L / q 0 τ n \kappa_n=M_1aL/\sqrt{q_0\tau_n} κ n = M 1 a L / q 0 τ n , p ± = b n ( 1 ± κ n ) 2 p_\pm=b_n(1\pm\kappa_n)^2 p ± = b n ( 1 ± κ n ) 2 , d n = ⌈ b n − 1 ⌉ d_n=\lceil b_n^{-1}\rceil d n = ⌈ b n − 1 ⌉ . Eventually 0 < p − ≤ b ( 0 ) ≤ p + < 1 0<p_-\le b(0)\le p_+<1 0 < p − ≤ b ( 0 ) ≤ p + < 1 , d n < n d_n<n d n < n , p ± / b n → 1 p_\pm/b_n\to1 p ± / b n → 1 , b n d n → 1 b_nd_n\to1 b n d n → 1 and d n / n → 0 d_n/n\to0 d n / n → 0 . With the amplitude revealed, pair the positions k , k + d n k,k+d_n k , k + d n . The corresponding states have squared fidelity at least ( 1 − p + ) d n (1-p_+)^{d_n} ( 1 − p + ) d n . Applying Helstrom’s bound [40 ] and the loss incurred by a wrong classification gives
r l o c , n ⋆ ≥ 1 − d n / n 2 min { 2 , b n d n / K } [ 1 − 1 − ( 1 − p + ) d n ] . r^\star_{\mathrm{loc},n} \ge\frac{1-d_n/n}{2}\min\{2,b_nd_n/K\}\left[1-\sqrt{1-(1-p_+)^{d_n}}\right]. r loc , n ⋆ ≥ 2 1 − d n / n min { 2 , b n d n / K } [ 1 − 1 − ( 1 − p + ) d n ] .
The factor 1 / 2 1/2 1/2 counts each pair once. Testing sites until the first click gives risk at most 2 b n ( 1 − p − ) / ( K p − ) → 2 / K 2b_n(1-p_-)/(Kp_-)\to2/K 2 b n ( 1 − p − ) / ( K p − ) → 2/ K , whereas a fixed report has loss two outside at most 2 K / b n + 1 2K/b_n+1 2 K / b n + 1 sites, which gives the prior only limit.
For the amplitude lower bound, reveal k k k . For q ( θ ) = ∥ ( I − ρ ( θ ) ) ϕ ′ ( θ ) ∥ 2 q(\theta)=\|(I-\rho(\theta))\phi'(\theta)\|^2 q ( θ ) = ∥ ( I − ρ ( θ )) ϕ ′ ( θ ) ∥ 2 , the derivative bound ∣ q ′ ∣ ≤ C q = 4 M 1 M 2 + 2 M 1 3 |q'|\le C_q=4M_1M_2+2M_1^3 ∣ q ′ ∣ ≤ C q = 4 M 1 M 2 + 2 M 1 3 bounds the conditional QFI by 4 a 2 ℓ k ( 1 + δ q , n ) 4a^2\ell_k(1+\delta_{q,n}) 4 a 2 ℓ k ( 1 + δ q , n ) , where ℓ k = ( n − k + 1 ) / n \ell_k=(n-k+1)/n ℓ k = ( n − k + 1 ) / n and δ q , n = C q ( s n + a L ) / ( q 0 n q 0 ) = o ( 1 ) \delta_{q,n}=C_q(s_n+aL)/(q_0\sqrt{nq_0})=o(1) δ q , n = C q ( s n + a L ) / ( q 0 n q 0 ) = o ( 1 ) . The preceding lemma gives
r a m p , n ⋆ ≥ 1 n ∑ k = 1 n { 3 / L 2 + ( 24 a 2 / 5 ) ℓ k ( 1 + δ q , n ) } − 1 . r^\star_{\mathrm{amp},n}\ge\frac{1}{n}\sum_{k=1}^{n}\left\{3/L^2+(24a^2/5)\ell_k(1+\delta_{q,n})\right\}^{-1}. r amp , n ⋆ ≥ n 1 k = 1 ∑ n { 3/ L 2 + ( 24 a 2 /5 ) ℓ k ( 1 + δ q , n ) } − 1 .
Its Riemann limit is ( 24 a 2 / 5 ) − 1 log ( 1 + 8 a 2 L 2 / 5 ) > 0 (24a^2/5)^{-1}\log(1+8a^2L^2/5)>0 ( 24 a 2 /5 ) − 1 log ( 1 + 8 a 2 L 2 /5 ) > 0 .
For the upper bound, test m = o ( n ) m=o(n) m = o ( n ) sites ending at j 0 = ⌊ n / 2 ⌋ j_0=\lfloor n/2\rfloor j 0 = ⌊ n /2 ⌋ , with m p − → ∞ mp_-\to\infty m p − → ∞ . A click places the last N = ⌊ n / 4 ⌋ N=\lfloor n/4\rfloor N = ⌊ n /4 ⌋ sites after the change. Use O θ = ∣ ϕ ( θ ) ⟩ ⟨ χ θ ∣ + ∣ χ θ ⟩ ⟨ ϕ ( θ ) ∣ O_\theta=|\phi(\theta)\rangle\langle\chi_\theta|+|\chi_\theta\rangle\langle\phi(\theta)| O θ = ∣ ϕ ( θ )⟩ ⟨ χ θ ∣ + ∣ χ θ ⟩ ⟨ ϕ ( θ ) ∣ , χ θ = ( I − ρ ( θ ) ) ϕ ′ ( θ ) / q ( θ ) \chi_\theta=(I-\rho(\theta))\phi'(\theta)/\sqrt{q(\theta)} χ θ = ( I − ρ ( θ )) ϕ ′ ( θ ) / q ( θ ) . The known observable O θ 0 , n O_{\theta_{0,n}} O θ 0 , n is measured on those sites. Outcomes X ∈ { − 1 , 0 , 1 } X\in\{-1,0,1\} X ∈ { − 1 , 0 , 1 } have variance at most one, mean derivative 2 q ( θ 0 , n ) 2\sqrt{q(\theta_{0,n})} 2 q ( θ 0 , n ) at the known center and second derivative bounded by 2 ( M 2 + M 1 2 ) 2(M_2+M_1^2) 2 ( M 2 + M 1 2 ) . Thus Z = n − 1 / 2 ∑ i = 1 N X i Z=n^{-1/2}\sum_{i=1}^{N}X_i Z = n − 1/2 ∑ i = 1 N X i obeys E z Z = a z / 2 + o ( 1 ) \mathbb{E}_zZ=a_z/2+o(1) E z Z = a z /2 + o ( 1 ) , Var z Z ≤ N / n \operatorname{Var}_z Z\le N/n Var z Z ≤ N / n , uniformly. The clipped report 2 a Z / ( 3 / L 2 + a 2 ) 2aZ/(3/L^2+a^2) 2 a Z / ( 3/ L 2 + a 2 ) has risk at most 1 / ( 3 / L 2 + a 2 ) + o ( 1 ) 1/(3/L^2+a^2)+o(1) 1/ ( 3/ L 2 + a 2 ) + o ( 1 ) , by expansion and E z z 2 = L 2 / 3 \mathbb{E}_z z^2=L^2/3 E z z 2 = L 2 /3 . Report zero when the pilot has no click.
The ideal branch k ≤ j 0 k\le j_0 k ≤ j 0 preserves the amplitude prior, and the pilot disagrees with that branch with probability at most m / n + e − m p − = o ( 1 ) m/n+e^{-mp_-}=o(1) m / n + e − m p − = o ( 1 ) . The pilot and tail act on disjoint sites, giving total risk 1 / [ 2 ( 3 / L 2 + a 2 ) ] + L 2 / 6 + o ( 1 ) < L 2 / 3 1/[2(3/L^2+a^2)]+L^2/6+o(1)<L^2/3 1/ [ 2 ( 3/ L 2 + a 2 )] + L 2 /6 + o ( 1 ) < L 2 /3 . Both marginal risks and their gains over the prior alone are therefore positive in the limit; the prior risks are 2 + o ( 1 ) 2+o(1) 2 + o ( 1 ) and L 2 / 3 L^2/3 L 2 /3 .
Certificate construction for continuous ideal step risks
At L = 1 / 2 L=1/2 L = 1/2 and s = s c s=s_c s = s c , partition the position prior into N N N cells. The following measurements and duals give the bounds in Section O [39 ] [29].
Representation and complete upper bounds
Write u = J / N + r u=J/N+r u = J / N + r , with independent uniform J ∈ { 0 , … , N − 1 } J\in\{0,\ldots,N-1\} J ∈ { 0 , … , N − 1 } and r ∈ [ 0 , 1 / N ] r\in[0,1/N] r ∈ [ 0 , 1/ N ] . Delete the last r r r units of the field at the left grid endpoint, translate right and add vacuum to obtain the shifted field. Subtracting r r r from a clipped position report reduces radial loss and preserves amplitude risk, so the grid optima are lower bounds.
For eight Gauss amplitude nodes, put h i = a ( 3 / 2 + z i ) h_i=a(3/2+z_i) h i = a ( 3/2 + z i ) , G ( v ) i j = e − ( h i − h j ) 2 ( 1 − v ) / 2 G(v)_{ij}=e^{-(h_i-h_j)^2(1-v)/2} G ( v ) ij = e − ( h i − h j ) 2 ( 1 − v ) /2 and D i i = e − h i 2 / ( 2 N ) D_{ii}=e^{-h_i^2/(2N)} D ii = e − h i 2 / ( 2 N ) . Endpoint projections use D j j D_{jj} D j j , and the innovation Gram matrix is G ( v ) − D G ( v + 1 / N ) D G(v)-DG(v+1/N)D G ( v ) − D G ( v + 1/ N ) D . Directed Cholesky factorization gives dimension 1 + 8 N 1+8N 1 + 8 N , including vacuum. For u = l + t ∈ [ l , l + 1 / N ] u=l+t\in[l,l+1/N] u = l + t ∈ [ l , l + 1/ N ] , the innovation coordinates are
C c e l l − 1 ( G ( l ) i j e ( h j 2 / 2 − h i h j ) t − ( D G ( l + 1 / N ) D ) i j e h j 2 t / 2 ) i , j = 1 8 . (20) C_{\mathrm{cell}}^{-1}\left(G(l)_{ij}e^{(h_j^2/2-h_ih_j)t}-(DG(l+1/N)D)_{ij}e^{h_j^2t/2}\right)_{i,j=1}^{8}. \tag*{(20)} C cell − 1 ( G ( l ) ij e ( h j 2 /2 − h i h j ) t − ( D G ( l + 1/ N ) D ) ij e h j 2 t /2 ) i , j = 1 8 . (20)
Endpoint residuals in distinct cells occupy orthogonal spaces Q c Q_c Q c . For dyadic frames V , F V,F V , F , certify V T V ⪰ n V I V^\mathsf{T}V\succeq n_VI V T V ⪰ n V I and F T F ⪰ n F I F^\mathsf{T}F\succeq n_FI F T F ⪰ n F I , so V F / n V n F VF/\sqrt{n_Vn_F} V F / n V n F is contractive. Add the Q c Q_c Q c effects and the positive complement I − ∑ i E i − ∑ c Q c I-\sum_i E_i-\sum_c Q_c I − ∑ i E i − ∑ c Q c , with report ( 1 / 2 , 0 ) (1/2,0) ( 1/2 , 0 ) . Starting from fallback risks 1 / 2 1/2 1/2 and 1 / 12 1/12 1/12 , integrate each effect’s risk change across every loss kink. Integrals of e b t e^{bt} e b t and t e b t te^{bt} t e b t use entire series remainders at small exponents.
Prior integration and feasible lower bounds
Complex amplitude continuation has trace one and trace norm e 2 a 2 ( 1 − u ) ( Im ζ ) 2 e^{2a^2(1-u)(\operatorname{Im}\zeta)^2} e 2 a 2 ( 1 − u ) ( Im ζ ) 2 . Completeness gives ∥ ∫ y j M ( d x , d y ) ∥ ≤ 2 − j \left\|\int y^jM(dx,dy)\right\|\le2^{-j} ∫ y j M ( d x , d y ) ≤ 2 − j for j = 1 , 2 j=1,2 j = 1 , 2 . Integrate the quadratic term exactly. The remaining Cauchy bounds at radius R R R are ( R + 3 / 4 ) e 2 a 2 R 2 (R+3/4)e^{2a^2R^2} ( R + 3/4 ) e 2 a 2 R 2 for amplitude and 2 e 2 a 2 R 2 2e^{2a^2R^2} 2 e 2 a 2 R 2 for position, giving the Gauss allowances
Q 8 ( R ) = ( 8 ! ) 4 17 ( 16 ! ) 2 e 2 a 2 R 2 R 16 , ε l o c = 2 Q 8 ( R ) , ε a m p = ( R + 3 / 4 ) Q 8 ( R ) . (21) Q_8(R)=\frac{(8!)^4}{17(16!)^2}\frac{e^{2a^2R^2}}{R^{16}},\qquad\varepsilon_{\mathrm{loc}}=2Q_8(R),\qquad\varepsilon_{\mathrm{amp}}=(R+3/4)Q_8(R). \tag*{(21)} Q 8 ( R ) = 17 ( 16 ! ) 2 ( 8 ! ) 4 R 16 e 2 a 2 R 2 , ε loc = 2 Q 8 ( R ) , ε amp = ( R + 3/4 ) Q 8 ( R ) . (21)
At R = 6 / 5 R=6/5 R = 6/5 , both task allowances and their weighted joint allowance are below 1.2 × 10 − 7 1.2\times10^{-7} 1.2 × 1 0 − 7 , charged once per endpoint.
For K c m p = B † / c K_{\mathrm{cmp}}=B^\dagger/\sqrt{c} K cmp = B † / c , c ≥ ∥ B ∥ 2 c\ge\|B\|^2 c ≥ ∥ B ∥ 2 , retain the compression failure flag. Return fidelity is at least ⟨ ψ , K c m p † K c m p ψ ⟩ \langle\psi,K_{\mathrm{cmp}}^\dagger K_{\mathrm{cmp}}\psi\rangle ⟨ ψ , K cmp † K cmp ψ ⟩ ; omitted mass δ \delta δ gives average distance at most 2 δ \sqrt{2\delta} 2 δ . If ∥ W − W c ∥ F ≤ e \|W-W_c\|_F\le e ∥ W − W c ∥ F ≤ e , ∥ W ∥ F ≤ 1 \|W\|_F\le1 ∥ W ∥ F ≤ 1 and node error is e z e_z e z , then
κ l o c = 2 e ( 2 + e ) , κ a m p = e ( 2 + e ) + 2 e z ( 1 + e ) 2 , κ R = ( κ l o c + κ a m p ) / 2. (22) \kappa_{\mathrm{loc}}=2e(2+e),\qquad\kappa_{\mathrm{amp}}=e(2+e)+2e_z(1+e)^2,\qquad\kappa_R=(\kappa_{\mathrm{loc}}+\kappa_{\mathrm{amp}})/2. \tag*{(22)} κ loc = 2 e ( 2 + e ) , κ amp = e ( 2 + e ) + 2 e z ( 1 + e ) 2 , κ R = ( κ loc + κ amp ) /2. (22)
Indeed, ∥ W D W † − W c D W c † ∥ ≤ ∥ D ∥ e ( 2 + e ) \|W D W^\dagger-W_c D W_c^\dagger\|\le\|D\|e(2+e) ∥ W D W † − W c D W c † ∥ ≤ ∥ D ∥ e ( 2 + e ) . Shift the d c d_c d c -dimensional dual by κ j I \kappa_jI κ j I , at cost d c κ j d_c\kappa_j d c κ j , and use the zero dual on failure, charging B j 2 δ B_j\sqrt{2\delta} B j 2 δ . An entrywise residual bound ∣ Z − F F † ∣ ≤ e Z |Z-FF^\dagger|\le e_Z ∣ Z − F F † ∣ ≤ e Z gives Z ⪰ − d c e Z I Z\succeq-d_ce_ZI Z ⪰ − d c e Z I and certifies the corrected slacks.
Let S 0 , S 1 , S 2 S_0,S_1,S_2 S 0 , S 1 , S 2 be the unnormalized amplitude moments. For any self-adjoint A 0 A_0 A 0 , set E A = S 0 A 0 + A 0 S 0 − 2 S 1 E_A=S_0A_0+A_0S_0-2S_1 E A = S 0 A 0 + A 0 S 0 − 2 S 1 . The dual
Y a m p = S 2 − A 0 S 0 A 0 − ∥ E A ∥ I / 2 Y_{\mathrm{amp}}=S_2-A_0S_0A_0-\|E_A\|I/2 Y amp = S 2 − A 0 S 0 A 0 − ∥ E A ∥ I /2
is feasible: for ∣ y ∣ ≤ 1 / 2 |y|\le1/2 ∣ y ∣ ≤ 1/2 , its slack is ( y − A 0 ) S 0 ( y − A 0 ) + y E A + ∥ E A ∥ I / 2 ⪰ 0 (y-A_0)S_0(y-A_0)+yE_A+\|E_A\|I/2\succeq0 ( y − A 0 ) S 0 ( y − A 0 ) + y E A + ∥ E A ∥ I /2 ⪰ 0 . The entrywise estimate ∥ E A ∥ ≤ d c max i j ∣ ( E A ) i j ∣ \|E_A\|\le d_c\max_{ij}|(E_A)_{ij}| ∥ E A ∥ ≤ d c max ij ∣ ( E A ) ij ∣ completes the bound.
Positive gap certificate at the reference center
Set s = s c s=s_c s = s c and L = λ = 1 / 2 L=\lambda=1/2 L = λ = 1/2 . Section T supplies the measurement and integration interfaces.
Purified comparison and feasible duals
For N = 256 N=256 N = 256 , d = 1 / N d=1/N d = 1/ N and u = j / N + r u=j/N+r u = j / N + r , purify r r r in L 2 ( [ 0 , d ] , d r / d ; F ) L^2([0,d],dr/d;\mathcal{F}) L 2 ([ 0 , d ] , d r / d ; F ) by Ψ j , z ( r ) = ∣ ( s c + a z ) 1 [ j / N + r , 1 ] ⟩ \Psi_{j,z}(r)=|(s_c+az)\mathbf{1}_{[j/N+r,1]}\rangle Ψ j , z ( r ) = ∣ ( s c + a z ) 1 [ j / N + r , 1 ] ⟩ . A decomposable POVM with fiber report clip [ 0 , U ] ( x − r ) \operatorname{clip}_{[0,U]}(x-r) clip [ 0 , U ] ( x − r ) , U = 1 − d U=1-d U = 1 − d , reduces location loss 2 ∣ x − u ∣ 2|x-u| 2∣ x − u ∣ for target j / N j/N j / N and preserves amplitude risk. Thus
R p u r ⋆ ≤ R o r i g ⋆ , r j , p u r ⋆ ≤ r j , o r i g ⋆ , j ∈ { l o c , a m p } . (23) R_{\mathrm{pur}}^\star\le R_{\mathrm{orig}}^\star,\qquad r_{j,\mathrm{pur}}^\star\le r_{j,\mathrm{orig}}^\star,\qquad j\in\{\mathrm{loc},\mathrm{amp}\}. \tag*{(23)} R pur ⋆ ≤ R orig ⋆ , r j , pur ⋆ ≤ r j , orig ⋆ , j ∈ { loc , amp } . (23)
For m a b = ( h a − h b ) 2 / 2 m_{ab}=(h_a-h_b)^2/2 m ab = ( h a − h b ) 2 /2 and exprel ( t ) = ( e t − 1 ) / t \operatorname{exprel}(t)=(e^t-1)/t exprel ( t ) = ( e t − 1 ) / t , with exprel ( 0 ) = 1 \operatorname{exprel}(0)=1 exprel ( 0 ) = 1 , integration over r r r gives
K ‾ ( v ) a b = e − m a b ( 1 − v ) exprel ( m a b d ) , (24) \overline{K}(v)_{ab}=e^{-m_{ab}(1-v)}\operatorname{exprel}(m_{ab}d), \tag*{(24)} K ( v ) ab = e − m ab ( 1 − v ) exprel ( m ab d ) , (24)
⟨ Ψ j , a , Ψ k , b ⟩ = e − h a 2 ( k − j ) / ( 2 N ) K ‾ ( k / N ) a b , j ≤ k . (25) \langle\Psi_{j,a},\Psi_{k,b}\rangle=e^{-h_a^2(k-j)/(2N)}\overline{K}(k/N)_{ab},\qquad j\le k. \tag*{(25)} ⟨ Ψ j , a , Ψ k , b ⟩ = e − h a 2 ( k − j ) / ( 2 N ) K ( k / N ) ab , j ≤ k . (25)
The backward innovation is K ‾ ( v ) − D K ‾ ( v + d ) D \overline{K}(v)-D\overline{K}(v+d)D K ( v ) − D K ( v + d ) D , D a a = e − h a 2 d / 2 D_{aa}=e^{-h_a^2d/2} D aa = e − h a 2 d /2 , and the terminal block is exprel ( − m a b d ) \operatorname{exprel}(-m_{ab}d) exprel ( − m ab d ) . Directed Cholesky factorization with eight nodes gives dimension 8 N 8N 8 N . The holomorphic density has trace one and norm at most e 2 a 2 ( Im ζ ) 2 e^{2a^2(\operatorname{Im}\zeta)^2} e 2 a 2 ( Im ζ ) 2 , so Section T gives
ε p r i o r = 3 / 25000000. (26) \varepsilon_{\mathrm{prior}}=3/25000000. \tag*{(26)} ε prior = 3/25000000. (26)
For dyadic weighted coordinates W W W of rank d r = 1675 d_r=1675 d r = 1675 , let C x , y = W diag ( ∣ x − u j ∣ + ( y − z a ) 2 / 2 ) W † C_{x,y}=W\operatorname{diag}(|x-u_j|+(y-z_a)^2/2)W^\dagger C x , y = W diag ( ∣ x − u j ∣ + ( y − z a ) 2 /2 ) W † . The stored dual Y 0 Y_0 Y 0 , positive correction D f i n a l D_{\mathrm{final}} D final , p = 2 − 36 p=2^{-36} p = 2 − 36 and row repairs d i d_i d i give
Y g r i d = Y 0 − D f i n a l − p I − diag ( d i ) ⪯ C x , y . (27) Y_{\mathrm{grid}}=Y_0-D_{\mathrm{final}}-pI-\operatorname{diag}(d_i)\preceq C_{x,y}. \tag*{(27)} Y grid = Y 0 − D final − p I − diag ( d i ) ⪯ C x , y . (27)
For each grid action, the residual E k = C x , y − Y 0 + D k + p I − F k F k † E_k=C_{x,y}-Y_0+D_k+pI-F_kF_k^\dagger E k = C x , y − Y 0 + D k + p I − F k F k † has absolute row sums at most d i d_i d i . The bound 2 ∣ v i v j ∣ ≤ ∣ v i ∣ 2 + ∣ v j ∣ 2 2|v_i v_j|\leq|v_i|^2+|v_j|^2 2∣ v i v j ∣ ≤ ∣ v i ∣ 2 + ∣ v j ∣ 2 gives E k + diag d i ⪰ 0 E_k+\operatorname{diag}d_i\succeq0 E k + diag d i ⪰ 0 ; subsequent positive corrections preserve earlier constraints. This covers 256 ⋅ 129 = 33024 256\cdot129=33024 256 ⋅ 129 = 33024 actions.
The arithmetic uses binary64 products and sums, with FMA allowed, without overflow, underflow or products at reduced precision. For u f = 2 − 53 u_f=2^{-53} u f = 2 − 53 and γ m = m u f / ( 1 − m u f ) \gamma_m=mu_f/(1-mu_f) γ m = m u f / ( 1 − m u f ) , a length-k k k dot product has error at most γ 2 k ∑ ∣ a i b i ∣ \gamma_{2k}\sum|a_i b_i| γ 2 k ∑ ∣ a i b i ∣ . Row norm bounds in powers of two w i w_i w i , f i f_i f i , b i b_i b i give entry errors 2 − 38 w i w j 2^{-38}w_iw_j 2 − 38 w i w j , 2 − 39 f i f j 2^{-39}f_if_j 2 − 39 f i f j and 2 − 35 b i b j 2^{-35}b_ib_j 2 − 35 b i b j for costs, Gram factors and positive corrections, with remaining allowance 2 − 44 2^{-44} 2 − 44 . Use γ 4096 + 2 ⋅ 33024 < 2 − 35 \gamma_{4096+2\cdot33024}<2^{-35} γ 4096 + 2 ⋅ 33024 < 2 − 35 and double the positive norm sums. If A i A_i A i is the exact absolute computed residual sum and B i B_i B i its error allowance, then s ^ i ≥ ( 1 − γ d r + 16 ) ( A i + B i ) \widehat{s}_i\geq(1-\gamma_{d_r+16})(A_i+B_i) s i ≥ ( 1 − γ d r + 16 ) ( A i + B i ) with γ d r + 16 < 1 / 4 \gamma_{d_r+16}<1/4 γ d r + 16 < 1/4 . Maximizing d i r a w = ⌈ 4 s ^ i 2 64 ⌉ d_i^{\mathrm{raw}}=\lceil4\widehat{s}_i2^{64}\rceil d i raw = ⌈ 4 s i 2 64 ⌉ over actions gives
d i = 2 − 64 ⌈ d i r a w / 3 ⌉ ≥ A i + B i . (28) d_i=2^{-64}\left\lceil d_i^{\mathrm{raw}}/3\right\rceil\geq A_i+B_i. \tag*{(28)} d i = 2 − 64 ⌈ d i raw /3 ⌉ ≥ A i + B i . (28)
Position costs are affine between nodes. Amplitude chords of spacing Δ = 1 / 128 \Delta=1/128 Δ = 1/128 exceed the quadratic cost by at most Δ 2 W W † / 8 \Delta^2WW^\dagger/8 Δ 2 W W † /8 . Subtract this matrix and ε c o o r d I \varepsilon_{\mathrm{coord}}I ε coord I to extend the dual to every continuous report.
Return to the original risk
For the compression V V V , the exact bound ∥ V † V − I ∥ ≤ ∥ ( V † V − I ) 2 ∥ ∞ < 1 \|V^\dagger V-I\|\leq\sqrt{\|(V^\dagger V-I)^2\|_\infty}<1 ∥ V † V − I ∥ ≤ ∥ ( V † V − I ) 2 ∥ ∞ < 1 certifies full column rank and contraction after normalization; ∥ ⋅ ∥ ∞ \|\cdot\|_\infty ∥ ⋅ ∥ ∞ is the maximum absolute row sum. Put t = c o l d / c n e w ≥ 1 t=c_{\mathrm{old}}/c_{\mathrm{new}}\geq1 t = c old / c new ≥ 1 . The omitted mass and compressed lower objective are δ n e w = 1 − t ( 1 − δ o l d ) \delta_{\mathrm{new}}=1-t(1-\delta_{\mathrm{old}}) δ new = 1 − t ( 1 − δ old ) and q = t [ Tr Y g r i d − Δ 2 Tr W W † / 8 − d r ε c o o r d ] q=t[\operatorname{Tr}Y_{\mathrm{grid}}-\Delta^2\operatorname{Tr}WW^\dagger/8-d_r\varepsilon_{\mathrm{coord}}] q = t [ Tr Y grid − Δ 2 Tr W W † /8 − d r ε coord ] .
Lemma O1.1 (Square root lower return). For a contraction of full column rank B = U ∣ B ∣ B=U|B| B = U ∣ B ∣ from the retained space to the original space, unit vectors ψ θ \psi_\theta ψ θ , probability prior π \pi π and 0 ≤ ℓ ( a , θ ) ≤ L θ 0\leq\ell(a,\theta)\leq L_\theta 0 ≤ ℓ ( a , θ ) ≤ L θ , a compressed lower bound q q q yields
r o r i g ∗ ≥ ( q + − D ) + 2 , D ≥ ∫ L θ ( 1 − ∥ B † ψ θ ∥ 2 ) d π . (29) r_{\mathrm{orig}}^*\geq\left(\sqrt{q_+}-\sqrt{D}\right)_+^2,\qquad D\geq\int L_\theta(1-\|B^\dagger\psi_\theta\|^2)d\pi. \tag*{(29)} r orig ∗ ≥ ( q + − D ) + 2 , D ≥ ∫ L θ ( 1 − ∥ B † ψ θ ∥ 2 ) d π . (29)
Proof. The polar isometry gives the positive contraction S = U B † S=UB^\dagger S = U B † . Pulling an original POVM through U U U , the triangle inequality in the direct sum gives q ( M ) ≤ r ( M ) + [ ∫ L θ ∥ ( I − S ) ψ θ ∥ 2 d π ] 1 / 2 \sqrt{q(M)}\leq\sqrt{r(M)}+\left[\int L_\theta\|(I-S)\psi_\theta\|^2d\pi\right]^{1/2} q ( M ) ≤ r ( M ) + [ ∫ L θ ∥ ( I − S ) ψ θ ∥ 2 d π ] 1/2 . Since ( I − S ) 2 ⪯ I − S 2 (I-S)^2\preceq I-S^2 ( I − S ) 2 ⪯ I − S 2 , the last term is at most D \sqrt{D} D . Take near minimizers of the original risk. □
The joint loss envelope is 383 / 256 383/256 383/256 . With D = ( 383 / 256 ) δ ‾ n e w D=(383/256)\overline{\delta}_{\mathrm{new}} D = ( 383/256 ) δ new , the return gives
R o r i g ∗ ≥ ( q + − D ) + 2 − ε p r i o r ≥ 0.094140512957477936. (30) R_{\mathrm{orig}}^*\geq\left(\sqrt{q_+}-\sqrt{D}\right)_+^2-\varepsilon_{\mathrm{prior}}\geq0.094140512957477936. \tag*{(30)} R orig ∗ ≥ ( q + − D ) + 2 − ε prior ≥ 0.094140512957477936. (30)
For q > D q>D q > D , enclose q D \sqrt{qD} q D above by a rational number and evaluate q + D − 2 q D q+D-2\sqrt{qD} q + D − 2 q D ; for q ≤ D q\leq D q ≤ D , use zero.
Numerical bounds
The saved coordinates, dual factors and upper measurements give the following allowances. Displayed intervals are rounded outward; calculations use the exact values in the numerical archive.
Table 1 (PDF p. 18).
Subtracting the marginal upper bounds from the returned joint lower bound gives Table 1.
Table 1 (PDF p. 18, second). Bounds for the continuous prior (2) at L = 1 / 2 L = 1/2 L = 1/2 , h ∈ [ a , 2 a ] h \in[a, 2a] h ∈ [ a , 2 a ] and position loss 2 ∣ x − u ∣ 2|x-u| 2∣ x − u ∣ . Each interval is rounded outward from the exact risk certificate. Section T1 gives the measurement interfaces.
Robustness in prior width at the reference center
Corollary D1.1 (A width neighborhood in fixed units). Fix s = s c s = s_{\mathrm{c}} s = s c , a a a , q 0 q_0 q 0 , K K K and both raw losses. Put I L = { L : ∣ L − 1 / 2 ∣ ≤ 3 × 10 − 4 } I_L = \{L : |L - 1/2| \le3 \times10^{-4}\} I L = { L : ∣ L − 1/2∣ ≤ 3 × 1 0 − 4 } and g I L = 2.525 × 10 − 4 g_{I_L} = 2.525 \times10^{-4} g I L = 2.525 × 1 0 − 4 . For each known L ∈ I L L \in I_L L ∈ I L , amplitudes are positive and
G 1 / 2 ( s c ; L ) > g I L , G λ ( s c ; L ) > 2 min { λ , 1 − λ } g I L ( 0 < λ < 1 ) . (31) G_{1/2}(s_{\mathrm{c}};L) > g_{I_L}, \qquad G_{\lambda}(s_{\mathrm{c}};L) > 2\min\{\lambda,1-\lambda\}g_{I_L} \quad(0 < \lambda< 1). \tag*{(31)} G 1/2 ( s c ; L ) > g I L , G λ ( s c ; L ) > 2 min { λ , 1 − λ } g I L ( 0 < λ < 1 ) . (31)
With H = s c + a ( 0.5003 ) H = s_{\mathrm{c}} + a(0.5003) H = s c + a ( 0.5003 ) , C I L = 1 + 2 ( 0.5003 ) 2 C_{I_L} = 1 + 2(0.5003)^2 C I L = 1 + 2 ( 0.5003 ) 2 and H / n q 0 ≤ r 0 H/\sqrt{nq_0} \le r_0 H / n q 0 ≤ r 0 , uniformly on I L I_L I L ,
G 1 / 2 , n ( s c ; L ) > g I L − 2 C I L ε n , H − d l o c , n , (32) G_{1/2,n}(s_{\mathrm{c}};L) > g_{I_L} - 2C_{I_L}\varepsilon_{n,H} - d_{\mathrm{loc},n}, \tag*{(32)} G 1/2 , n ( s c ; L ) > g I L − 2 C I L ε n , H − d loc , n , (32)
where d l o c , n d_{\mathrm{loc},n} d loc , n is evaluated at s c s_{\mathrm{c}} s c in Theorem A1.2. For a fixed known curve, one sufficiently large n n n gives positivity throughout I L I_L I L and at every fixed interior weight.
Proof. Couple z = L r z = Lr z = L r and z 0 = L 0 r z_0 = L_0r z 0 = L 0 r , where L 0 = 1 / 2 L_0 = 1/2 L 0 = 1/2 and ( u , r ) ∼ Unif [ 0 , 1 ] ⊗ Unif [ − 1 , 1 ] (u,r) \sim\operatorname{Unif}[0,1] \otimes\operatorname{Unif}[-1,1] ( u , r ) ∼ Unif [ 0 , 1 ] ⊗ Unif [ − 1 , 1 ] . Set Δ L = ∣ L − L 0 ∣ \Delta_L = |L - L_0| Δ L = ∣ L − L 0 ∣ , L m = max ( L , L 0 ) L_m = \max(L,L_0) L m = max ( L , L 0 ) . The average state distance is at most ε = a Δ L / 3 \varepsilon= a\Delta_L/3 ε = a Δ L /3 , since E ∣ r ∣ = 1 / 2 \mathbb{E}|r| = 1/2 E ∣ r ∣ = 1/2 and E 1 − u = 2 / 3 \mathbb{E}\sqrt{1-u} = 2/3 E 1 − u = 2/3 . On reports [ − L m , L m ] [-L_m,L_m] [ − L m , L m ] , the amplitude loss changes by at most Δ L [ 2 L m ∣ r ∣ + ( L + L 0 ) r 2 ] \Delta_L[2L_m|r| + (L + L_0)r^2] Δ L [ 2 L m ∣ r ∣ + ( L + L 0 ) r 2 ] , with mean ε ℓ = Δ L [ L m + ( L + L 0 ) / 3 ] \varepsilon_\ell= \Delta_L[L_m + (L + L_0)/3] ε ℓ = Δ L [ L m + ( L + L 0 ) /3 ] . Comparing states at fixed targets, then changing the amplitude target, Lemma A1.1 gives, with B λ ( L m ) = 2 λ + 4 L m 2 ( 1 − λ ) B_\lambda(L_m) = 2\lambda+ 4L_m^2(1-\lambda) B λ ( L m ) = 2 λ + 4 L m 2 ( 1 − λ ) ,
∣ Δ r l o c ⋆ ∣ ≤ 2 ε , ∣ Δ r a m p ⋆ ∣ ≤ 4 L m 2 ε + ε ℓ , ∣ Δ R ⋆ ∣ ≤ B λ ( L m ) ε + ( 1 − λ ) ε ℓ , ∣ Δ G λ ∣ ≤ 2 B λ ( L m ) ε + 2 ( 1 − λ ) ε ℓ . (33) \begin{aligned}
|\Delta r_{\mathrm{loc}}^{\star}| &\le2\varepsilon, &
|\Delta r_{\mathrm{amp}}^{\star}| &\le4L_m^2\varepsilon+ \varepsilon_\ell, \\
|\Delta R^{\star}| &\le B_\lambda(L_m)\varepsilon+ (1-\lambda)\varepsilon_\ell, &
|\Delta G_\lambda| &\le2B_\lambda(L_m)\varepsilon+ 2(1-\lambda)\varepsilon_\ell. \tag*{(33)}
\end{aligned} ∣Δ r loc ⋆ ∣ ∣Δ R ⋆ ∣ ≤ 2 ε , ≤ B λ ( L m ) ε + ( 1 − λ ) ε ℓ , ∣Δ r amp ⋆ ∣ ∣Δ G λ ∣ ≤ 4 L m 2 ε + ε ℓ , ≤ 2 B λ ( L m ) ε + 2 ( 1 − λ ) ε ℓ . (33)
At equal weights, Δ L ≤ 3 × 10 − 4 \Delta_L \le3 \times10^{-4} Δ L ≤ 3 × 1 0 − 4 , L m ≤ 0.5003 L_m \le0.5003 L m ≤ 0.5003 , L + L 0 ≤ 1.0003 L + L_0 \le1.0003 L + L 0 ≤ 1.0003 and a < 1.6655 a < 1.6655 a < 1.6655 give ∣ Δ G 1 / 2 ∣ < 7.5 × 10 − 4 |\Delta G_{1/2}| < 7.5 \times10^{-4} ∣Δ G 1/2 ∣ < 7.5 × 1 0 − 4 . The bound on a a a follows from eight terms of 2 atanh ( 1 / 3 ) = log 2 2\operatorname{atanh}(1/3) = \log2 2 atanh ( 1/3 ) = log 2 and the geometric tail. Since 0.001002640149993936 > 1.0025 × 10 − 3 0.001002640149993936 > 1.0025 \times10^{-3} 0.001002640149993936 > 1.0025 × 1 0 − 3 , the central certificate gives G 1 / 2 > g I L G_{1/2} > g_{I_L} G 1/2 > g I L . Nonnegative regrets give the factor 2 min { λ , 1 − λ } 2\min\{\lambda,1-\lambda\} 2 min { λ , 1 − λ } ; (A1.3) gives the finite record bound with common H H H . ∎
Leading recovery for shrinking rises
Leading recovery for smooth rises
Consider (11)–(12) for fixed F F F , L L L and 0 ≤ λ ≤ 1 0 \le\lambda\le1 0 ≤ λ ≤ 1 . Put c χ = 1 + χ c_{\chi}=\sqrt{1+\chi} c χ = 1 + χ , δ = c χ / s 2 \delta=c_{\chi}/s^{2} δ = c χ / s 2 , and B l o c = 2 B_{\mathrm{loc}}=2 B loc = 2 , B a m p = 4 L 2 B_{\mathrm{amp}}=4L^{2} B amp = 4 L 2 , B λ = λ B l o c + ( 1 − λ ) B a m p B_{\lambda}=\lambda B_{\mathrm{loc}}+(1-\lambda)B_{\mathrm{amp}} B λ = λ B loc + ( 1 − λ ) B amp . Define J F ( r ) = ∫ R [ F ( x − r ) − F ( x ) ] 2 d x J_{F}(r)=\int_{\mathbb{R}}[F(x-r)-F(x)]^{2}\,\mathrm{d}x J F ( r ) = ∫ R [ F ( x − r ) − F ( x ) ] 2 d x and κ F 2 = ∫ 0 1 F ′ ( x ) 2 d x \kappa_{F}^{2}=\int_{0}^{1}F'(x)^{2}\,\mathrm{d}x κ F 2 = ∫ 0 1 F ′ ( x ) 2 d x .
The local location kernel is
k χ ( v ) = exp [ − χ J F ( c χ ∣ v ∣ / χ ) / 2 ] ( 0 < χ < ∞ ) , k 0 ( v ) = e − ∣ v ∣ / 2 , k ∞ ( v ) = e − κ F 2 v 2 / 2 . (34) k_{\chi}(v)=\exp[-\chi J_{F}(c_{\chi}|v|/\chi)/2]\quad(0<\chi<\infty),\qquad k_{0}(v)=e^{-|v|/2},\qquad k_{\infty}(v)=e^{-\kappa_{F}^{2}v^{2}/2}. \tag*{(34)} k χ ( v ) = exp [ − χ J F ( c χ ∣ v ∣/ χ ) /2 ] ( 0 < χ < ∞ ) , k 0 ( v ) = e − ∣ v ∣/2 , k ∞ ( v ) = e − κ F 2 v 2 /2 . (34)
The experiment P χ P_{\chi} P χ has this overlap. Its spectral density and canonical outcome density are
m χ ( p ) = 1 2 π ∫ e − i p v k χ ( v ) d v , p χ ( x ) = ∣ F − 1 m χ ( x ) ∣ 2 , (35) m_{\chi}(p)=\frac{1}{2\pi}\int e^{-ipv}k_{\chi}(v)\,\mathrm{d}v,\qquad p_{\chi}(x)=|\mathcal{F}^{-1}\sqrt{m_{\chi}(x)}|^{2}, \tag*{(35)} m χ ( p ) = 2 π 1 ∫ e − i p v k χ ( v ) d v , p χ ( x ) = ∣ F − 1 m χ ( x ) ∣ 2 , (35)
where F \mathcal{F} F is unitary on L 2 ( R ) L^{2}(\mathbb{R}) L 2 ( R ) and the inverse is in L 2 L^{2} L 2 . Write
r F ( χ ) = 2 ∫ min { 1 , ∣ x ∣ / K } p χ ( x ) d x , A ‾ L = ∫ 0 1 A L ( a e ) d e , (36) r_{F}(\chi)=2\int\min\{1,|x|/K\}p_{\chi}(x)\,\mathrm{d}x,\qquad\overline{A}_{L}=\int_{0}^{1}A_{L}(a\sqrt{e})\,\mathrm{d}e, \tag*{(36)} r F ( χ ) = 2 ∫ min { 1 , ∣ x ∣/ K } p χ ( x ) d x , A L = ∫ 0 1 A L ( a e ) d e , (36)
where A L ( c ) A_{L}(c) A L ( c ) is the amplitude optimum for ∣ c z ⟩ |cz\rangle ∣ cz ⟩ with z ∼ Unif [ − L , L ] z\sim\operatorname{Unif}[-L,L] z ∼ Unif [ − L , L ] . On e ∈ [ 0 , 1 ] e\in[0,1] e ∈ [ 0 , 1 ] , coherent state comparison gives the modulus 4 L 2 a L ∣ e − e ′ ∣ 4L^{2}a_{L}\sqrt{|e-e'|} 4 L 2 a L ∣ e − e ′ ∣ for A L ( a e ) A_{L}(a\sqrt{e}) A L ( a e ) . The three optimal risks satisfy
r l o c ⋆ = r F ( χ ) + o ( 1 ) , r a m p ⋆ = A ‾ L + o ( 1 ) , R λ ⋆ = λ r F ( χ ) + ( 1 − λ ) A ‾ L + o ( 1 ) . (37) r_{\mathrm{loc}}^{\star}=r_{F}(\chi)+o(1),\qquad r_{\mathrm{amp}}^{\star}=\overline{A}_{L}+o(1),\qquad R_{\lambda}^{\star}=\lambda r_{F}(\chi)+(1-\lambda)\overline{A}_{L}+o(1). \tag*{(37)} r loc ⋆ = r F ( χ ) + o ( 1 ) , r amp ⋆ = A L + o ( 1 ) , R λ ⋆ = λ r F ( χ ) + ( 1 − λ ) A L + o ( 1 ) . (37)
as s → ∞ s\to\infty s → ∞ , w → 0 w\to0 w → 0 , with arbitrary χ = s 2 w \chi=s^{2}w χ = s 2 w .
Geometry and spectral continuity
Lemma J1.1 (Geometry and uniform spectral comparison). For r ≥ 0 r\ge0 r ≥ 0 , r 2 / ( 1 + r ) ≤ J F ( r ) ≤ r r^{2}/(1+r)\le J_{F}(r)\le r r 2 / ( 1 + r ) ≤ J F ( r ) ≤ r ; for r ≥ 1 r\ge1 r ≥ 1 , J F ( r ) = r − C F J_{F}(r)=r-C_{F} J F ( r ) = r − C F , C F = 2 ∫ 0 1 F ( 1 − F ) C_{F}=2\int_{0}^{1}F(1-F) C F = 2 ∫ 0 1 F ( 1 − F ) ; and J F ( r ) / r 2 → κ F 2 J_{F}(r)/r^{2}\to\kappa_{F}^{2} J F ( r ) / r 2 → κ F 2 as r ↓ 0 r\downarrow0 r ↓ 0 . The spectral densities m χ m_{\chi} m χ are normalized. If χ n → χ 0 \chi_{n}\to\chi_{0} χ n → χ 0 in [ 0 , ∞ ] [0,\infty] [ 0 , ∞ ] , then ∥ m χ n − m χ 0 ∥ 1 → 0 \|m_{\chi_{n}}-m_{\chi_{0}}\|_{1}\to0 ∥ m χ n − m χ 0 ∥ 1 → 0 and ∥ m χ n − m χ 0 ∥ 2 → 0 \|\sqrt{m_{\chi_{n}}}-\sqrt{m_{\chi_{0}}}\|_{2}\to0 ∥ m χ n − m χ 0 ∥ 2 → 0 . There are parameter independent channels in both directions whose errors are bounded by this square-root distance, uniformly over the translation v ∈ R v\in\mathbb{R} v ∈ R .
Proof. The function F ( x ) − F ( x − r ) ∈ [ 0 , 1 ] F(x)-F(x-r)\in[0,1] F ( x ) − F ( x − r ) ∈ [ 0 , 1 ] has integral r r r and is supported on an interval of length at most 1 + r 1+r 1 + r . The lower bound follows from Cauchy–Schwarz, and the upper bound from squaring this function. Splitting the integral at 0 , 1 , r , r + 1 0,1,r,r+1 0 , 1 , r , r + 1 yields the large-shift identity; differentiability of translation in L 2 L^{2} L 2 gives the small-shift limit. For the quintic, κ F 2 = 10 / 7 \kappa_{F}^{2}=10/7 κ F 2 = 10/7 , C F = 50 / 231 C_{F}=50/231 C F = 50/231 . The exact coherent fields
g χ , v ( x ) = F ( ( x − c χ v ) / χ ) − F ( x / χ ) in L 2 ( R ) (38) g_{\chi,v}(x)=F((x-c_{\chi}v)/\chi)-F(x/\chi)\ \text{in }L^{2}(\mathbb{R}) \tag*{(38)} g χ , v ( x ) = F (( x − c χ v ) / χ ) − F ( x / χ ) in L 2 ( R ) (38)
give (J1.1), with indicator-difference and one-mode ∣ κ F v ⟩ |\kappa_{F}v\rangle ∣ κ F v ⟩ endpoints. Geometry gives pointwise kernel convergence and the integrable majorant
∣ k χ ( v ) ∣ ≤ e − v 2 / ( 2 ( 1 + ∣ v ∣ ) ) , χ J F ( c χ ∣ v ∣ / χ ) ≥ ( 1 + χ ) v 2 χ + c χ ∣ v ∣ ≥ v 2 1 + ∣ v ∣ . |k_{\chi}(v)|\le e^{-v^{2}/(2(1+|v|))},\qquad\chi J_{F}(c_{\chi}|v|/\chi)\ge\frac{(1+\chi)v^{2}}{\chi+c_{\chi}|v|}\ge\frac{v^{2}}{1+|v|}. ∣ k χ ( v ) ∣ ≤ e − v 2 / ( 2 ( 1 + ∣ v ∣ )) , χ J F ( c χ ∣ v ∣/ χ ) ≥ χ + c χ ∣ v ∣ ( 1 + χ ) v 2 ≥ 1 + ∣ v ∣ v 2 .
Normalization follows from positive definiteness and Fourier inversion, and the asserted norm convergence follows from dominated convergence and Scheffé’s theorem. Equality of the Gram kernels identifies the cyclic states ∣ ψ χ , v ⟩ |\psi_{\chi,v}\rangle ∣ ψ χ , v ⟩ with e − i p v m χ ( p ) e^{-ipv}\sqrt{m_{\chi}(p)} e − i p v m χ ( p ) . Their vector distance is independent of v v v ; the resulting isometries and completed inverses give both channels. □
Product comparison under the original prior
Partition [ 0 , 1 ] [0,1] [ 0 , 1 ] into N c N_c N c cells of length b = 1 / N c b=1/N_c b = 1/ N c , centered at u j u_j u j . A cell is called good if ( j + 1 ) b + w ≤ 1 (j+1)b+w\le1 ( j + 1 ) b + w ≤ 1 . Its suffix length is ℓ j = 1 − ( j + 1 ) b − w \ell_j=1-(j+1)b-w ℓ j = 1 − ( j + 1 ) b − w , and its local coordinate is v = ( u − u j ) / δ v=(u-u_j)/\delta v = ( u − u j ) / δ . The mass of the remaining cells is β = 1 − b ⌊ ( 1 − w ) / b ⌋ ≤ w + b \beta=1-b\lfloor(1-w)/b\rfloor\le w+b β = 1 − b ⌊( 1 − w ) / b ⌋ ≤ w + b . The reference experiment Q s , w Q_{s,w} Q s , w emits a good label together with
∣ ψ χ , v ⟩ ⊗ ∣ a ℓ j z ⟩ (39) |\psi_{\chi,v}\rangle\otimes|a\sqrt{\ell_j}\,z\rangle \tag*{(39)} ∣ ψ χ , v ⟩ ⊗ ∣ a ℓ j z ⟩ (39)
or one bad label with vacuum, preserving the original targets and losses. For 0 < t < 1 0<t<1 0 < t < 1 , put
P = min { 1 , 2 ( N c + 1 ) ( a L s + 2 / π 2 s t ) } , P=\min\left\{1,2(N_c+1)\left(\frac{aL}{s}+\frac{\sqrt{2/\pi}}{2s\sqrt{t}}\right)\right\}, P = min { 1 , 2 ( N c + 1 ) ( s a L + 2 s t 2/ π ) } ,
H = t ( s b / 2 + a L ) , D w = a L b + w . (40) H=t\left(s\sqrt{b/2}+aL\right),\qquad D_w=aL\sqrt{b+w}. \tag*{(40)} H = t ( s b /2 + a L ) , D w = a L b + w . (40)
Two-way channels have prior averaged errors
ε T ≤ β + P + H + D w , ε S ≤ β + D w , (41) \varepsilon_T\le\beta+P+H+D_w,\qquad\varepsilon_S\le\beta+D_w, \tag*{(41)} ε T ≤ β + P + H + D w , ε S ≤ β + D w , (41)
capped at one. The reference factors at each label. Lemma A1.1 gives 0 ≤ G λ ≤ B λ ( ε T + ε S ) 0\le\mathcal{G}_\lambda\le B_\lambda(\varepsilon_T+\varepsilon_S) 0 ≤ G λ ≤ B λ ( ε T + ε S ) and a complete rule with e j ≤ B j ( ε T + ε S ) + η e_j\le B_j(\varepsilon_T+\varepsilon_S)+\eta e j ≤ B j ( ε T + ε S ) + η for both tasks and any η > 0 \eta>0 η > 0 .
Product comparison. A fraction t t t of the intensity is assigned to the pilot. On a good cell, A w ( u ) = ∫ 0 1 F ( ( r − u ) / w ) d r = 1 − u − w ( 1 − μ F ) A_w(u)=\int_0^1 F((r-u)/w)\,dr=1-u-w(1-\mu_F) A w ( u ) = ∫ 0 1 F (( r − u ) / w ) d r = 1 − u − w ( 1 − μ F ) , μ F = ∫ 0 1 F \mu_F=\int_0^1F μ F = ∫ 0 1 F . Constant-mode homodyne gives X = t ( s + a z ) A w ( u ) + Z / 2 X=\sqrt{t}(s+az)A_w(u)+Z/2 X = t ( s + a z ) A w ( u ) + Z /2 , Z ∼ N ( 0 , 1 ) Z\sim N(0,1) Z ∼ N ( 0 , 1 ) ; the clipped report 1 − w ( 1 − μ F ) − X / ( s t ) 1-w(1-\mu_F)-X/(s\sqrt{t}) 1 − w ( 1 − μ F ) − X / ( s t ) has error at most a L / s + ∣ Z ∣ / ( 2 s t ) aL/s+|Z|/(2s\sqrt{t}) a L / s + ∣ Z ∣/ ( 2 s t ) . Integration over the boundary strips, with the pilot’s selection probabilities, gives P P P .
For an observed good label, displace the retained field by − 1 − t s F ( ( r − u j ) / w ) -\sqrt{1-t}sF((r-u_j)/w) − 1 − t s F (( r − u j ) / w ) . Before attenuation, its residual is s [ F ( ( r − u ) / w ) − F ( ( r − u j ) / w ) ] + a z F ( ( r − u ) / w ) s[F((r-u)/w)-F((r-u_j)/w)]+azF((r-u)/w) s [ F (( r − u ) / w ) − F (( r − u j ) / w )] + a z F (( r − u ) / w ) . The first term is supported on W j = [ j b , ( j + 1 ) b + w ] W_j=[jb,(j+1)b+w] W j = [ j b , ( j + 1 ) b + w ] and has norm at most s b / 2 s\sqrt{b/2} s b /2 ; the offset has norm at most a L aL a L . Thus attenuation costs H H H , and omitting the offset on W j W_j W j costs D w D_w D w . The displacement is independent of u , z u,z u , z . Spatial factorization, followed by the change of variable x = s 2 ( r − u j ) x=s^2(r-u_j) x = s 2 ( r − u j ) , gives (J1.5) and the constant suffix mode, which is vacuum when ℓ j = 0 \ell_j=0 ℓ j = 0 .
Given u , z u,z u , z , the pilot and retained field are independent. Wrong labels and terminal cells contribute P + β P+\beta P + β to the forward error. The reverse map embeds the location field in W j W_j W j , the amplitude mode in the suffix, and vacuum elsewhere, then restores the known reference. Its errors are D w + β D_w+\beta D w + β . Both maps are complete channels.
At each good label, the factored prior lets one pair rules approaching both marginal infima. Averaging over the other factor gives the matching lower bound. The vacuum label also has zero gap.
Choose N c = ⌈ s 1 / 3 ⌉ N_c=\lceil s^{1/3}\rceil N c = ⌈ s 1/3 ⌉ , b = 1 / N c b=1/N_c b = 1/ N c , t = 1 / s t=1/s t = 1/ s . Then P , H = O L ( s − 1 / 6 ) P,H=O_L(s^{-1/6}) P , H = O L ( s − 1/6 ) , D w = O L ( s − 1 / 6 + w ) D_w=O_L(s^{-1/6}+\sqrt{w}) D w = O L ( s − 1/6 + w ) , β ≤ w + b \beta\le w+b β ≤ w + b and e j = O F , L ( s − 1 / 6 + w ) e_j=O_{F,L}(s^{-1/6}+\sqrt{w}) e j = O F , L ( s − 1/6 + w ) , for every w → 0 w\to0 w → 0 and arbitrary χ \chi χ .
Covariant risk and global risk tracking
Let r F ( χ ; V ) r_F(\chi;V) r F ( χ ; V ) denote the optimum for v ∼ Unif [ − V , V ] v\sim\operatorname{Unif}[-V,V] v ∼ Unif [ − V , V ] , clipped reports and loss 2 min ( 1 , ∣ x − v ∣ / K ) 2\min(1,|x-v|/K) 2 min ( 1 , ∣ x − v ∣/ K ) .
Lemma J1.2 (Complete covariant risk). For P χ P_\chi P χ , complete covariant measurements, with failure assigned cost two, have optimal risk r F ( χ ) r_F(\chi) r F ( χ ) in (J1.3). The finite-prior optima satisfy r F ( χ ; V ) → r F ( χ ) r_F(\chi;V)\to r_F(\chi) r F ( χ ; V ) → r F ( χ ) uniformly on [ 0 , ∞ ) [0,\infty) [ 0 , ∞ ) as V → ∞ V\to\infty V → ∞ .
Proof. On the cyclic momentum space, the shift representation has Lebesgue spectral measure. Its normalized covariant POVMs therefore have measurable isometric kernel dilations [42 Theorem 2.4.3, Corollary 2.4.5 and Section 2.5.1] . If the finite-report part has total effect Q Q Q , then Q Q Q commutes with shifts. Normalize on supp Q \operatorname{supp} Q supp Q and multiply its dilation by Q 1 / 2 Q^{1/2} Q 1/2 to obtain kernels ⟨ η ( p ) , η ( q ) ⟩ \langle\eta(p),\eta(q)\rangle ⟨ η ( p ) , η ( q )⟩ with ∥ η ( p ) ∥ ≤ 1 \|\eta(p)\| \leq1 ∥ η ( p ) ∥ ≤ 1 . Assign the commuting deficit I − Q I-Q I − Q to the failure report of cost two. The triangular benefit has nonnegative transform b ^ K ( t ) = K sinc 2 ( K t / 2 ) \widehat b_K(t)=K\operatorname{sinc}^{2}(Kt/2) b K ( t ) = K sinc 2 ( K t /2 ) , where sinc ( x ) = sin ( x ) / x \operatorname{sinc}(x)=\sin(x)/x sinc ( x ) = sin ( x ) / x and sinc ( 0 ) = 1 \operatorname{sinc}(0)=1 sinc ( 0 ) = 1 . Its expectation is
( 2 π ) − 1 ∬ m χ ( p ) m χ ( q ) b ^ K ( p − q ) Re ⟨ η ( p ) , η ( q ) ⟩ d p d q . (2\pi)^{-1}\iint\sqrt{m_\chi(p)m_\chi(q)}\widehat b_K(p-q)\operatorname{Re}\langle\eta(p),\eta(q)\rangle\,\mathrm{d}p\mathrm{d}q. ( 2 π ) − 1 ∬ m χ ( p ) m χ ( q ) b K ( p − q ) Re ⟨ η ( p ) , η ( q )⟩ d p d q .
A constant unit vector maximizes this expectation and gives p χ p_\chi p χ [41 ] . Young’s inequality justifies the integral for m ∈ L 2 \sqrt{m}\in L^2 m ∈ L 2 , b ^ K ∈ L 1 \widehat b_K\in L^1 b K ∈ L 1 .
Restricting the prior and clipping reports gives r F ( χ ; V ) ≤ r F ( χ ) r_F(\chi;V)\leq r_F(\chi) r F ( χ ; V ) ≤ r F ( χ ) . For the reverse inequality, extend finite-window rules to the one-point compactification R ∪ { ∞ } \mathbb{R}\cup\{\infty\} R ∪ { ∞ } and average their translates over [ − V , V ] [-V,V] [ − V , V ] . The risk at zero is their finite-prior risk, while translation by b b b changes any continuous test of norm at most one by at most ∣ b ∣ / V |b|/V ∣ b ∣/ V . The compactness argument in the proof of Proposition 1, applied to this compact report space, gives a complete limit along a subnet realizing the risk liminf. The covariance defect vanishes. The loss at zero is continuous, with value two at infinity, so the limit risk is at least the preceding covariant minimum, including failure mass. This proves pointwise convergence. The spectral channels give equicontinuity independently of V V V ; a finite net on [ 0 , ∞ ] [0,\infty] [ 0 , ∞ ] then gives uniformity.
Theorem 2(i): global tracking and informative marginals. On good cells, V = b / ( 2 δ ) = b s 2 / ( 2 1 + s 2 w ) → ∞ V=b/(2\delta)=bs^2/(2\sqrt{1+s^2w})\to\infty V = b / ( 2 δ ) = b s 2 / ( 2 1 + s 2 w ) → ∞ uniformly for w ≤ 1 w\leq1 w ≤ 1 . The canonical lemma, the bad-cell cost 2 β 2\beta 2 β and the channels in both directions prove the location formula in (J1.4).
The amplitude optima A L ( a ℓ j ) A_L(a\sqrt{\ell_j}) A L ( a ℓ j ) converge after averaging to A ‾ L \overline A_L A L , by their continuity modulus, 0 ≤ 1 − u − ℓ j ≤ b + w 0\leq1-u-\ell_j\leq b+w 0 ≤ 1 − u − ℓ j ≤ b + w and the terminal mass bound. Transferring the amplitude and joint infima proves (J1.4).
Continuity of r F r_F r F follows from Plancherel’s theorem. Since the outcome distribution has no point masses, r F > 0 r_F>0 r F > 0 . The benefit is also positive because m χ m_\chi m χ is positive on a set of positive measure, and b ^ K \widehat b_K b K is positive near zero. Compactness makes 0 < r F < 2 0<r_F<2 0 < r F < 2 uniform.
Lemma A3.1, applied with QFI 4 a 2 e 4a^2e 4 a 2 e , gives the amplitude lower bound. For the upper bound, use homodyne X = a e z + N X=a\sqrt e\,z+N X = a e z + N , Var N = 1 / 4 \operatorname{Var}N=1/4 Var N = 1/4 , followed by its optimal linear predictor and clipping. Integration yields
5 24 a 2 log ( 1 + 8 a 2 L 2 / 5 ) ≤ A ‾ L ≤ 1 4 a 2 log ( 1 + 4 a 2 L 2 / 3 ) < L 2 / 3. (42) \frac{5}{24a^2}\log(1+8a^2L^2/5)\leq\overline A_L\leq\frac{1}{4a^2}\log(1+4a^2L^2/3)<L^2/3. \tag*{(42)} 24 a 2 5 log ( 1 + 8 a 2 L 2 /5 ) ≤ A L ≤ 4 a 2 1 log ( 1 + 4 a 2 L 2 /3 ) < L 2 /3. (42)
The prior-only location benefit is confined to length 2 K δ → 0 2K\delta\to0 2 K δ → 0 .
Inverse square cost and physical transfer
The inverse square residual cost and its physical transfer
Fixed assumptions for C1–C11. C1–C10 establish Theorem 2(ii) for the full coherent record; C11 transfers the result to the finite product record and compares the leading risks. Throughout these sections, the quintic profile is used, with a = 2 log 2 a=2\sqrt{\log2} a = 2 log 2 , K = 9 log 2 K=9\log2 K = 9 log 2 , L = 1 / 2 L=1/2 L = 1/2 , the original independent uniform u , z u,z u , z prior and a fixed 0 < λ < 1 0<\lambda<1 0 < λ < 1 . The scaling parameters are ϵ = s − 1 \epsilon=s^{-1} ϵ = s − 1 , χ = s 2 w \chi=s^2w χ = s 2 w , and δ = 1 + χ ϵ 2 \delta=\sqrt{1+\chi\epsilon^2} δ = 1 + χ ϵ 2 . All optima use the losses in main-text (12). The compact design interval I = [ 1 / 2 , 2 ] I=[1/2,2] I = [ 1/2 , 2 ] is known; rules may depend on χ \chi χ but are independent of u , z u,z u , z .
Proof route. C1 supplies the relative operator bounds and C2 the quintic packet estimates. C3 removes endpoint contributions at o ( ϵ 2 ) o(\epsilon^{2}) o ( ϵ 2 ) precision. C4 centers comparisons at each model’s own marginal optimum; C5 and C6 give the scalar upper and lower bounds. C7–C9 return that scalar decision problem to the original record. C10 verifies uniformity on I I I , and C11 gives the physical transfer.
Amplitude optima, response and source lifts
For a Hilbert–Schmidt synthesis V V V and Z = Z † Z=Z^{\dagger} Z = Z † , ∥ Z ∥ ≤ L \lVert Z\rVert\le L ∥ Z ∥ ≤ L , the common optimality proof applies to Ω = V V † \Omega=VV^{\dagger} Ω = V V † , Ω 1 = V Z V † \Omega_{1}=VZV^{\dagger} Ω 1 = V Z V † on supp Ω \operatorname{supp}\Omega supp Ω , retaining the finite source mass:
1 2 { Ω , A } = V Z V † , ∥ A ∥ ≤ L , (43) \frac{1}{2}\{\Omega,A\}=VZV^{\dagger},\qquad\lVert A\rVert\le L, \tag*{(43)} 2 1 { Ω , A } = V Z V † , ∥ A ∥ ≤ L , (43)
e a m p ( J ; V ) = ∥ Y J − J A ∥ Ω 2 . (44) e_{\mathrm{amp}}(J;V)=\lVert YJ-JA\rVert_{\Omega}^{2}. \tag*{(44)} e amp ( J ; V ) = ∥ Y J − J A ∥ Ω 2 . (44)
The spectral measurement of A A A is amplitude optimal. The Schur norm is the maximum absolute row or column sum; it is submultiplicative and dominates the operator norm.
Lemma C1.1 (A source defect controls the amplitude optimum). Let A ~ = A ~ † \widetilde{A}=\widetilde{A}^{\dagger} A = A † be bounded on the record space and let K † = − K K^{\dagger}=-K K † = − K be bounded on the source. The source defect E = A ~ V − V Z − V K E=\widetilde{A}V-VZ-VK E = A V − V Z − V K satisfies
∥ A ~ − A ∥ Ω ≤ ∥ E ∥ H S . \lVert\widetilde{A}-A\rVert_{\Omega}\le\lVert E\rVert_{\mathrm{HS}}. ∥ A − A ∥ Ω ≤ ∥ E ∥ HS . Proof. Skew-adjointness gives { Ω , H } = E V † + V E † \{\Omega,H\}=EV^{\dagger}+VE^{\dagger} { Ω , H } = E V † + V E † for H = A ~ − A H=\widetilde{A}-A H = A − A . Trace-class products and Cauchy–Schwarz yield ∥ H ∥ Ω 2 = Re Tr ( V † H E ) ≤ ∥ H ∥ Ω ∥ E ∥ H S \lVert H\rVert_{\Omega}^{2}=\operatorname{Re}\operatorname{Tr}(V^{\dagger}HE)\le\lVert H\rVert_{\Omega}\lVert E\rVert_{\mathrm{HS}} ∥ H ∥ Ω 2 = Re Tr ( V † H E ) ≤ ∥ H ∥ Ω ∥ E ∥ HS .
Use V c f = ∫ − L L ∣ c z ⟩ f ( z ) ν ( d z ) V_cf=\int_{-L}^{L}|cz\rangle f(z)\nu(dz) V c f = ∫ − L L ∣ cz ⟩ f ( z ) ν ( d z ) , ν = Unif [ − L , L ] \nu=\operatorname{Unif}[-L,L] ν = Unif [ − L , L ] , L = 1 / 2 L=1/2 L = 1/2 , 0 ≤ c ≤ a 0\le c\le a 0 ≤ c ≤ a . Set Ω j , c = V c Z j V c † \Omega_{j,c}=V_cZ^{j}V_c^{\dagger} Ω j , c = V c Z j V c † , Ω c = Ω 0 , c \Omega_c=\Omega_{0,c} Ω c = Ω 0 , c , and suppress c c c in the polynomial representation. For monic orthogonal polynomials P n P_n P n of μ c = e − c 2 z 2 ν \mu_c=e^{-c^{2}z^{2}}\nu μ c = e − c 2 z 2 ν with squared norms h n h_n h n , the unitary f ↦ e c 2 z 2 / 2 f f\mapsto e^{c^{2}z^{2}/2}f f ↦ e c 2 z 2 /2 f gives
T m n = c m m ! ∫ z m P n ( z ) h n μ c ( d z ) , Ω j = T J j T † . (45) T_{mn}=\frac{c^{m}}{\sqrt{m!}}\int z^{m}\frac{P_n(z)}{\sqrt{h_n}}\,\mu_c(dz),\qquad\Omega_j=TJ^{j}T^{\dagger}. \tag*{(45)} T mn = m ! c m ∫ z m h n P n ( z ) μ c ( d z ) , Ω j = T J j T † . (45)
Here J J J is multiplication by z z z in this polynomial basis.
Lemma C1.2 (Uniform prior recurrence bound). For the monic orthogonal polynomials of e − t x 2 d x e^{-tx^{2}}dx e − t x 2 d x on [ − 1 , 1 ] [-1,1] [ − 1 , 1 ] , write x P n = P n + 1 + β n ( t ) P n − 1 xP_n=P_{n+1}+\beta_n(t)P_{n-1} x P n = P n + 1 + β n ( t ) P n − 1 , β 0 = 0 \beta_0=0 β 0 = 0 . Their recurrence coefficients satisfy, for 0 ≤ t ≤ log 2 0\le t\le\log2 0 ≤ t ≤ log 2 ,
e − t 4 ≤ β n ( t ) ≤ 1 3 ( n ≥ 1 ) . \frac{e^{-t}}{4}\le\beta_n(t)\le\frac{1}{3}\qquad(n\ge1). 4 e − t ≤ β n ( t ) ≤ 3 1 ( n ≥ 1 ) . Proof. Norm differentiation and orthogonality give β n ′ = β n ( β n − 1 − β n + 1 ) \beta_n'=\beta_n(\beta_{n-1}-\beta_{n+1}) β n ′ = β n ( β n − 1 − β n + 1 ) , h n ′ / h n = − ( β n + β n + 1 ) h_n'/h_n=-(\beta_n+\beta_{n+1}) h n ′ / h n = − ( β n + β n + 1 ) and β n ( 0 ) = n 2 / ( 4 n 2 − 1 ) \beta_n(0)=n^2/(4n^2-1) β n ( 0 ) = n 2 / ( 4 n 2 − 1 ) . Thus 0 < β n ≤ 1 0<\beta_n\le1 0 < β n ≤ 1 implies β n ≥ e − t / 4 \beta_n\ge e^{-t}/4 β n ≥ e − t /4 , while β 1 ′ = − β 1 β 2 < 0 \beta_1'=-\beta_1\beta_2<0 β 1 ′ = − β 1 β 2 < 0 . Before a first exit above 1 / 3 1/3 1/3 , each n ≥ 2 n\ge2 n ≥ 2 satisfies β n ( t ) ≤ ( 4 / 15 ) exp [ t / 3 − ( 1 − e − t ) / 4 ] < 1 / 3 \beta_n(t)\le(4/15)\exp[t/3-(1-e^{-t})/4]<1/3 β n ( t ) ≤ ( 4/15 ) exp [ t /3 − ( 1 − e − t ) /4 ] < 1/3 on [ 0 , log 2 ] [0,\log2] [ 0 , log 2 ] . The uniform derivative bound ∣ β n ′ ∣ ≤ 1 |\beta_n'|\le1 ∣ β n ′ ∣ ≤ 1 permits continuation of the infinite family.
After z = L x z=Lx z = Lx , the recurrence coefficient is at most L 2 / 3 L^{2}/3 L 2 /3 . Put
q max = log 2 3 , U ∗ = 1 + 8 q max 1 − q max , S j = U ∗ 2 ( 2 L / 3 ) j , R A = 2 S 1 , R B = 2 ( S 2 + R A S 1 ) . (46) q_{\max}=\frac{\log2}{3},\qquad U_{*}=1+\frac{8q_{\max}}{1-q_{\max}},\qquad S_j=U_{*}^{2}(2L/\sqrt{3})^{j},\qquad R_A=2S_1,\qquad R_B=2(S_2+R_AS_1). \tag*{(46)} q m a x = 3 log 2 , U ∗ = 1 + 1 − q m a x 8 q m a x , S j = U ∗ 2 ( 2 L / 3 ) j , R A = 2 S 1 , R B = 2 ( S 2 + R A S 1 ) . (46)
Proposition C1.3 (Relative amplitude and response bounds). Let A c A_c A c be the bounded amplitude optimum in (C1.1), and let the response B c B_c B c solve
1 2 { Ω c , B c } = Ω 2 , c − 1 2 { A c , Ω 1 , c } . (47) \frac{1}{2}\{\Omega_c,B_c\}=\Omega_{2,c}-\frac{1}{2}\{A_c,\Omega_{1,c}\}. \tag*{(47)} 2 1 { Ω c , B c } = Ω 2 , c − 2 1 { A c , Ω 1 , c } . (47) On the average state support, uniformly for 0 ≤ c ≤ a 0\le c\le a 0 ≤ c ≤ a , the bounds are
∥ Ω c − 1 / 2 A c Ω c 1 / 2 ∥ ≤ R A , ∥ Ω c − 1 / 2 B c Ω c 1 / 2 ∥ ≤ R B , ∥ B c ∥ ≤ L 2 + L R A . (48) \lVert\Omega_c^{-1/2}A_c\Omega_c^{1/2}\rVert\le R_A,\qquad\lVert\Omega_c^{-1/2}B_c\Omega_c^{1/2}\rVert\le R_B,\qquad\lVert B_c\rVert\le L^2+LR_A. \tag*{(48)} ∥ Ω c − 1/2 A c Ω c 1/2 ∥ ≤ R A , ∥ Ω c − 1/2 B c Ω c 1/2 ∥ ≤ R B , ∥ B c ∥ ≤ L 2 + L R A . (48) The same bounds apply to all unnormalized finite Fock sections.
Proof. For c > 0 c>0 c > 0 , write T = C D T=CD T = C D , d n = c n h n / n ! d_n=c^n\sqrt{h_n}/\sqrt{n!} d n = c n h n / n ! , q = c 2 L 2 / 3 ≤ q max q=c^2L^2/3\le q_{\max} q = c 2 L 2 /3 ≤ q m a x . Counting downward recurrence steps and factorial log-convexity give
∣ C n + 2 j , n ∣ ≤ ( 2 q ) j / j ! , ∣ ( C − 1 ) n + 2 j , n ∣ ≤ q j / j ! . (49) |C_{n+2j,n}|\le(2q)^j/j!,\qquad|(C^{-1})_{n+2j,n}|\le q^j/j!. \tag*{(49)} ∣ C n + 2 j , n ∣ ≤ ( 2 q ) j / j ! , ∣ ( C − 1 ) n + 2 j , n ∣ ≤ q j / j ! . (49)
The inverse entries have the representation
( C − 1 ) n + 2 j , n = c 2 j n ! ( n + 2 j ) ! [ z n ] P n + 2 j ( z ) , ∣ [ z n ] P n + 2 j ∣ ≤ ( n + j j ) ( L 2 / 3 ) j . (50) (C^{-1})_{n+2j,n}=c^{2j}\sqrt{\frac{n!}{(n+2j)!}}[z^n]P_{n+2j}(z),\qquad|[z^n]P_{n+2j}|\le\binom{n+j}{j}(L^2/3)^j. \tag*{(50)} ( C − 1 ) n + 2 j , n = c 2 j ( n + 2 j )! n ! [ z n ] P n + 2 j ( z ) , ∣ [ z n ] P n + 2 j ∣ ≤ ( j n + j ) ( L 2 /3 ) j . (50)
Thus ∥ C ∥ S c h ≤ e 2 q \lVert C\rVert_{\mathrm{Sch}}\le e^{2q} ∥ C ∥ Sch ≤ e 2 q , ∥ C − 1 ∥ S c h ≤ e q \lVert C^{-1}\rVert_{\mathrm{Sch}}\le e^q ∥ C − 1 ∥ Sch ≤ e q , e − 2 q I ≤ C † C ≤ e 4 q I e^{-2q}I\le C^\dagger C\le e^{4q}I e − 2 q I ≤ C † C ≤ e 4 q I , and d n + 1 / d n ≤ q / ( n + 1 ) d_{n+1}/d_n\le\sqrt{q/(n+1)} d n + 1 / d n ≤ q / ( n + 1 ) .
For K = D H D K=DHD K = D H D , m I ≤ H ≤ M I mI\le H\le MI m I ≤ H ≤ M I , the min–max principle gives m d n 2 ≤ λ n ≤ M d n 2 md_n^2\le\lambda_n\le Md_n^2 m d n 2 ≤ λ n ≤ M d n 2 . Its eigenvectors satisfy
∣ ( v n ) i ∣ ≤ min { M d i m λ n , M λ n m d i } ≤ M m min { d i / d n , d n / d i } . (51) |(v_n)_i|\le\min\left\{\frac{Md_i}{\sqrt{m\lambda_n}},\frac{\sqrt{M\lambda_n}}{md_i}\right\}\le\frac{M}{m}\min\{d_i/d_n,d_n/d_i\}. \tag*{(51)} ∣ ( v n ) i ∣ ≤ min { m λ n M d i , m d i M λ n } ≤ m M min { d i / d n , d n / d i } . (51)
The inverse weight estimate follows from D − 1 v n = λ n − 1 H D v n ∈ ℓ 2 D^{-1}v_n=\lambda_n^{-1}HDv_n\in\ell^2 D − 1 v n = λ n − 1 H D v n ∈ ℓ 2 and ⟨ D − 1 v n , H − 1 D − 1 v n ⟩ = λ n − 1 \langle D^{-1}v_n,H^{-1}D^{-1}v_n\rangle=\lambda_n^{-1} ⟨ D − 1 v n , H − 1 D − 1 v n ⟩ = λ n − 1 . Since M / m = e 6 q M/m=e^{6q} M / m = e 6 q and q e 6 q ≤ 4 log 2 / 3 < 1 qe^{6q}\le4\log2/3<1 q e 6 q ≤ 4 log 2/3 < 1 , the min–max bands are disjoint. Parity and continuation from c = 0 + c=0^+ c = 0 + therefore determine the parity of each eigenvector. On indices of the same parity, ∣ U i n ∣ ≤ e 6 q q ∣ i − n ∣ / 2 |U_{in}|\le e^{6q}q^{|i-n|/2} ∣ U in ∣ ≤ e 6 q q ∣ i − n ∣/2 , giving ∥ U ∥ S c h ≤ U ( q ) : = 1 + 2 e 6 q q / ( 1 − q ) ≤ U ∗ \lVert U\rVert_{\mathrm{Sch}}\le U(q):=1+2e^{6q}q/(1-q)\le U_* ∥ U ∥ Sch ≤ U ( q ) := 1 + 2 e 6 q q / ( 1 − q ) ≤ U ∗ . Finite sections and the factorial tails give this continuation.
The polar factor of the injective, dense range T T T is unitary. Normalized moments are Z j = U † J j U Z_j=U^\dagger J^jU Z j = U † J j U , with
∥ Z j ∥ S c h ≤ U ( q ) 2 ( 2 L / 3 ) j ≤ S j , q = c 2 L 2 / 3 ≤ log 2 / 3. (52) \lVert Z_j\rVert_{\mathrm{Sch}}\le U(q)^2(2L/\sqrt{3})^j\le S_j,\qquad q=c^2L^2/3\le\log2/3. \tag*{(52)} ∥ Z j ∥ Sch ≤ U ( q ) 2 ( 2 L / 3 ) j ≤ S j , q = c 2 L 2 /3 ≤ log 2/3. (52)
For B = Ω − 1 / 2 A Ω 1 / 2 B=\Omega^{-1/2}A\Omega^{1/2} B = Ω − 1/2 A Ω 1/2 , the entries 2 λ j ( Z 1 ) i j / ( λ i + λ j ) 2\lambda_j(Z_1)_{ij}/(\lambda_i+\lambda_j) 2 λ j ( Z 1 ) ij / ( λ i + λ j ) give ∥ B ∥ S c h ≤ 2 S 1 \lVert B\rVert_{\mathrm{Sch}}\le2S_1 ∥ B ∥ Sch ≤ 2 S 1 . Applying this multiplier to the response source Z 2 − ( B Z 1 + Z 1 B † ) / 2 Z_2-(BZ_1+Z_1B^\dagger)/2 Z 2 − ( B Z 1 + Z 1 B † ) /2 gives R B R_B R B . The positive Sylvester integral gives the ordinary operator bound. For the projection P N P_N P N onto the first N N N Fock coordinates, lower triangularity gives T N = P N T = C N D N P N T_N=P_NT=C_ND_NP_N T N = P N T = C N D N P N , where C N = P N C P N C_N=P_NCP_N C N = P N C P N and D N = P N D P N D_N=P_NDP_N D N = P N D P N act on the retained coordinates. Hence
Ω j , N = C N D N ( P N J j P N ) D N C N † . \Omega_{j,N}=C_ND_N(P_NJ^jP_N)D_NC_N^\dagger. Ω j , N = C N D N ( P N J j P N ) D N C N † .
The bounds for C N C_N C N , C N − 1 C_N^{-1} C N − 1 and the separated min–max bands give the same Schur estimates after diagonalizing D N C N † C N D N D_NC_N^\dagger C_ND_N D N C N † C N D N . The moments use compressions P N J j P N P_NJ^jP_N P N J j P N of the full powers and retain the original mass. At c = 0 c=0 c = 0 , A 0 = 0 A_0=0 A 0 = 0 , B 0 = ( L 2 / 3 ) I B_0=(L^2/3)I B 0 = ( L 2 /3 ) I on vacuum.
Source lifts and continuity
Lemma C1.4 (Skew-adjoint source lift). The relative amplitude bound R A R_A R A gives a bounded skew-adjoint source operator K V K_V K V satisfying
A V − V Z = V K V , ∥ K V ∥ ≤ κ s r c : = 3 ( R A + L ) . (53) AV - VZ = VK_V,\qquad\lVert K_V\rVert\le\kappa_{\mathrm{src}} := 3(R_A+L). \tag*{(53)} A V − V Z = V K V , ∥ K V ∥ ≤ κ src := 3 ( R A + L ) . (53) For the coherent residual response, a bounded skew-adjoint source operator K 1 , V K_{1,V} K 1 , V satisfies
V K 1 , V = 2 B V + V [ K V , Z ] , ∥ K 1 , V ∥ ≤ 6 ( R B + L κ s r c ) . (54) VK_{1,V}=2BV+V[K_V,Z],\qquad\lVert K_{1,V}\rVert\le6(R_B+L\kappa_{\mathrm{src}}). \tag*{(54)} V K 1 , V = 2 B V + V [ K V , Z ] , ∥ K 1 , V ∥ ≤ 6 ( R B + L κ src ) . (54) Proof. Write V = Ω 1 / 2 U V=\Omega^{1/2}U V = Ω 1/2 U , U U † = I UU^\dagger=I U U † = I , P = U † U P=U^\dagger U P = U † U and D = Ω − 1 / 2 ( A V − V Z ) D=\Omega^{-1/2}(AV-VZ) D = Ω − 1/2 ( A V − V Z ) . Since ∥ D ∥ ≤ R A + L \lVert D\rVert\le R_A+L ∥ D ∥ ≤ R A + L and D U † + U D † = 0 DU^\dagger+UD^\dagger=0 D U † + U D † = 0 , the skew lift K V = U † D − D † U − U † D P K_V=U^\dagger D-D^\dagger U-U^\dagger DP K V = U † D − D † U − U † D P has U K V = D UK_V=D U K V = D and norm at most 3 ∥ D ∥ 3\lVert D\rVert 3 ∥ D ∥ . The response equation gives W V † + V W † = 0 WV^\dagger+VW^\dagger=0 W V † + V W † = 0 for W = 2 B V + V [ K V , Z ] W=2BV+V[K_V,Z] W = 2 B V + V [ K V , Z ] ; the same construction with ∥ Ω − 1 / 2 W ∥ ≤ 2 R B + 2 L κ s r c \lVert\Omega^{-1/2}W\rVert\le2R_B+2L\kappa_{\mathrm{src}} ∥ Ω − 1/2 W ∥ ≤ 2 R B + 2 L κ src proves the first lift.
Write K c = K V c K_c=K_{V_c} K c = K V c and K 1 , c = K 1 , V c K_{1,c}=K_{1,V_c} K 1 , c = K 1 , V c . For two syntheses with the same target, put Ω V = V V † \Omega_V=VV^\dagger Ω V = V V † . Applying the defect lemma to A W ( V − W ) − ( V − W ) ( Z + K W ) A_W(V-W)-(V-W)(Z+K_W) A W ( V − W ) − ( V − W ) ( Z + K W ) and subtracting the lift identities gives
∥ A V − A W ∥ Ω V ≤ ( 2 L + κ s r c ) ∥ V − W ∥ H S , (55) \lVert A_V-A_W\rVert_{\Omega_V}\le(2L+\kappa_{\mathrm{src}})\lVert V-W\rVert_{\mathrm{HS}}, \tag*{(55)} ∥ A V − A W ∥ Ω V ≤ ( 2 L + κ src ) ∥ V − W ∥ HS , (55)
∥ V ( K V − K W ) ∥ H S ≤ 2 ( 2 L + κ s r c ) ∥ V − W ∥ H S . (56) \lVert V(K_V-K_W)\rVert_{\mathrm{HS}}\le2(2L+\kappa_{\mathrm{src}})\lVert V-W\rVert_{\mathrm{HS}}. \tag*{(56)} ∥ V ( K V − K W ) ∥ HS ≤ 2 ( 2 L + κ src ) ∥ V − W ∥ HS . (56)
For the coherent family, the synthesis difference satisfies
∥ V c − V c ′ ∥ H S 2 = ∫ 2 ( 1 − e − ( c − c ′ ) 2 z 2 / 2 ) ν ( d z ) ≤ L 2 3 ∣ c − c ′ ∣ 2 . (57) \lVert V_c-V_{c'}\rVert_{\mathrm{HS}}^2=\int2(1-e^{-(c-c')^2z^2/2})\,\nu(dz)\le\frac{L^2}{3}|c-c'|^2. \tag*{(57)} ∥ V c − V c ′ ∥ HS 2 = ∫ 2 ( 1 − e − ( c − c ′ ) 2 z 2 /2 ) ν ( d z ) ≤ 3 L 2 ∣ c − c ′ ∣ 2 . (57)
On compact subsets of c > 0 c>0 c > 0 , uniqueness, weighted convergence and the bounded lift norms give strong convergence on the dense range of V c † V_c^\dagger V c † , for both K c K_c K c and A c A_c A c . Trace norm continuity of the moments identifies each weak limit of B c B_c B c through its Sylvester equation. For its source C c C_c C c , ∥ B c V c ∥ H S 2 = Tr ( B c C c ) \lVert B_cV_c\rVert_{\mathrm{HS}}^2=\operatorname{Tr}(B_cC_c) ∥ B c V c ∥ HS 2 = Tr ( B c C c ) gives Hilbert–Schmidt continuity. The first lift identity gives weighted continuity, with endpoint strips bounded by their mass.
Quintic packets and spectral bounds
Integrating the translated polynomial pieces gives
J F ( r ) = { 10 7 r 2 − 10 7 r 4 + 2 r 6 − 10 7 r 7 + 5 21 r 9 − 2 77 r 11 , 0 ≤ r ≤ 1 , r − 50 231 , r ≥ 1. (58) J_F(r)=
\begin{cases}
\frac{10}{7}r^2-\frac{10}{7}r^4+2r^6-\frac{10}{7}r^7+\frac{5}{21}r^9-\frac{2}{77}r^{11}, & 0\le r\le1,\\
r-\frac{50}{231}, & r\ge1.
\end{cases} \tag*{(58)} J F ( r ) = { 7 10 r 2 − 7 10 r 4 + 2 r 6 − 7 10 r 7 + 21 5 r 9 − 77 2 r 11 , r − 231 50 , 0 ≤ r ≤ 1 , r ≥ 1. (58)
Also ν F = ∫ 0 1 ( 1 − F 2 ) = 281 / 462 \nu_F=\int_0^1(1-F^2)=281/462 ν F = ∫ 0 1 ( 1 − F 2 ) = 281/462 and ∫ ( F ′ ) 2 = 10 / 7 \int(F')^2=10/7 ∫ ( F ′ ) 2 = 10/7 . With c χ = 1 + χ c_\chi=\sqrt{1+\chi} c χ = 1 + χ , ψ χ ( v ) = χ J F ( c χ ∣ v ∣ / χ ) \psi_\chi(v)=\chi J_F(c_\chi|v|/\chi) ψ χ ( v ) = χ J F ( c χ ∣ v ∣/ χ ) , put
m χ , τ ( p ) = 1 2 π ∫ e − i p v e − τ ψ χ ( v ) / 2 d v . (59) m_{\chi,\tau}(p)=\frac{1}{2\pi}\int e^{-ipv}e^{-\tau\psi_\chi(v)/2}\,dv. \tag*{(59)} m χ , τ ( p ) = 2 π 1 ∫ e − i p v e − τ ψ χ ( v ) /2 d v . (59)
For φ = F ′ \varphi=F' φ = F ′ , Plancherel’s theorem gives ψ χ ( v ) = ( c χ / π ) ∫ ( 1 − cos p v ) ∣ φ ^ ( χ p / c χ ) ∣ 2 p − 2 d p \psi_\chi(v)=(c_\chi/\pi)\int(1-\cos pv)|\widehat{\varphi}(\chi p/c_\chi)|^2p^{-2}\,dp ψ χ ( v ) = ( c χ / π ) ∫ ( 1 − cos p v ) ∣ φ ( χ p / c χ ) ∣ 2 p − 2 d p . The symmetric Lévy density is therefore τ c χ ∣ φ ^ ( χ p / c χ ) ∣ 2 / ( 2 π p 2 ) \tau c_\chi|\widehat{\varphi}(\chi p/c_\chi)|^2/(2\pi p^2) τ c χ ∣ φ ( χ p / c χ ) ∣ 2 / ( 2 π p 2 ) . Extract a finite uniform jump component on a symmetric interval where φ ^ ≠ 0 \widehat{\varphi}\ne0 φ = 0 . Its compound Poisson mixture has a positive density component everywhere, preserved by convolution with the remaining Lévy law. Integrability of the characteristic function then gives a continuous, strictly positive density.
The first six derivatives match at 0 , ± χ / c χ 0,\pm\chi/c_{\chi} 0 , ± χ / c χ . The polynomial gives seventh derivative jumps 7200 τ c χ 7 / χ 6 7200\tau c_{\chi}^{7}/\chi^{6} 7200 τ c χ 7 / χ 6 at zero and − 3600 τ c χ 7 χ − 6 e − τ D χ -3600\tau c_{\chi}^{7}\chi^{-6}e^{-\tau D_{\chi}} − 3600 τ c χ 7 χ − 6 e − τ D χ at each join. The regular eighth derivative has bounded variation and exponential tails. Eight distributional integrations by parts give the expansion below, including any fixed finite number of momentum and intensity derivatives, uniformly on I = [ 1 / 2 , 2 ] I=[1/2,2] I = [ 1/2 , 2 ] , T = [ 1 / 2 , 3 / 2 ] \mathcal{T}=[1/2,3/2] T = [ 1/2 , 3/2 ] :
m χ , τ ( p ) = 3600 τ c χ 7 π χ 6 p 8 { 1 − e − τ D χ cos ( d χ p ) } + O I ( ∣ p ∣ − 9 ) , D χ = 181 χ 462 , d χ = χ 1 + χ . (60) m_{\chi,\tau}(p)=\frac{3600\tau c_{\chi}^{7}}{\pi\chi^{6}p^{8}}\left\{1-e^{-\tau D_{\chi}}\cos(d_{\chi}p)\right\}+O_I\left(|p|^{-9}\right),\qquad D_{\chi}=\frac{181\chi}{462},\qquad d_{\chi}=\frac{\chi}{\sqrt{1+\chi}}. \tag*{(60)} m χ , τ ( p ) = π χ 6 p 8 3600 τ c χ 7 { 1 − e − τ D χ cos ( d χ p ) } + O I ( ∣ p ∣ − 9 ) , D χ = 462 181 χ , d χ = 1 + χ χ . (60)
The floor 1 − e − τ D χ > 0 1-e^{-\tau D_{\chi}}>0 1 − e − τ D χ > 0 , compact positivity and bounded polynomial variation give
c 1 ( 1 + ∣ p ∣ ) − 8 ≤ m χ , τ ( p ) ≤ c 2 ( 1 + ∣ p ∣ ) − 8 , (61) c_1(1+|p|)^{-8}\le m_{\chi,\tau}(p)\le c_2(1+|p|)^{-8}, \tag*{(61)} c 1 ( 1 + ∣ p ∣ ) − 8 ≤ m χ , τ ( p ) ≤ c 2 ( 1 + ∣ p ∣ ) − 8 , (61)
sup χ ∈ I , τ ∈ T , p ∣ ∂ τ j ∂ p k m χ , τ ( p ) m χ , τ ( p ) ∣ < ∞ ( j , k fixed finite ) . (62) \sup_{\chi\in I,\tau\in\mathcal{T},p}\left|\frac{\partial_{\tau}^{j}\partial_{p}^{k}m_{\chi,\tau}(p)}{m_{\chi,\tau}(p)}\right|<\infty\quad(j,k\text{ fixed finite}). \tag*{(62)} χ ∈ I , τ ∈ T , p sup m χ , τ ( p ) ∂ τ j ∂ p k m χ , τ ( p ) < ∞ ( j , k fixed finite ) . (62)
All intensities ( 1 + a ϵ z ) 2 (1+a\epsilon z)^2 ( 1 + a ϵz ) 2 lie in T \mathcal{T} T when a L ϵ ≤ 1 / 8 aL\epsilon\le1/8 a L ϵ ≤ 1/8 .
The first arrival representation
Sorting every photon sector by its first coordinate gives the exact isometry
∣ f ⟩ = e − ∥ f ∥ 2 / 2 ∣ v a c ⟩ ⊕ ∫ 0 1 f ( t ) e − 1 2 ∫ 0 t f ( r ) 2 d r ∣ t ⟩ ⊗ ∣ f ( t , 1 ] ⟩ d t . (63) |f\rangle=e^{-\lVert f\rVert^{2}/2}|\mathrm{vac}\rangle\oplus\int_{0}^{1}f(t)e^{-\frac{1}{2}\int_{0}^{t}f(r)^{2}\,\mathrm{d}r}|t\rangle\otimes|f_{(t,1]}\rangle\,\mathrm{d}t. \tag*{(63)} ∣ f ⟩ = e − ∥ f ∥ 2 /2 ∣ vac ⟩ ⊕ ∫ 0 1 f ( t ) e − 2 1 ∫ 0 t f ( r ) 2 d r ∣ t ⟩ ⊗ ∣ f ( t , 1 ] ⟩ d t . (63)
The sector factor n \sqrt{n} n gives 1 / ( n − 1 ) ! 1/\sqrt{(n-1)!} 1/ ( n − 1 )! ; symmetrization inverts the isometry and retains vacuum coherence. At the quantum coordinate t ≤ 1 − w t\le1-w t ≤ 1 − w , the known displacement − s -s − s is applied, followed by dilation of the window of width w w w and compression of the suffix to one residual mode. With F χ ( x ) = F ( x / χ ) F_{\chi}(x)=F(x/\chi) F χ ( x ) = F ( x / χ ) , A χ ( x ) = ∫ 0 x F χ 2 A_{\chi}(x)=\int_{0}^{x}F_{\chi}^{2} A χ ( x ) = ∫ 0 x F χ 2 , ζ = a ϵ z \zeta=a\epsilon z ζ = a ϵz , the normalized packet is
B ζ ( x ) = 1 x ≥ 0 ( 1 + ζ ) F χ ( x ) e − ( 1 + ζ ) 2 A χ ( x ) / 2 ∣ ( 1 + ζ ) F χ ( x + ⋅ ) − 1 ⟩ [ 0 , x ] . (64) B_{\zeta}(x)=\mathbf{1}_{x\ge0}(1+\zeta)F_{\chi}(x)e^{-(1+\zeta)^{2}A_{\chi}(x)/2}\left|(1+\zeta)F_{\chi}(x+\cdot)-1\right\rangle_{[0,x]}. \tag*{(64)} B ζ ( x ) = 1 x ≥ 0 ( 1 + ζ ) F χ ( x ) e − ( 1 + ζ ) 2 A χ ( x ) /2 ∣ ( 1 + ζ ) F χ ( x + ⋅ ) − 1 ⟩ [ 0 , x ] . (64)
The survival probability e − ( 1 + ζ ) 2 A χ ( x ) e^{-(1+\zeta)^{2}A_{\chi}(x)} e − ( 1 + ζ ) 2 A χ ( x ) normalizes the packet, giving the arrival component
ϵ − 1 B a ϵ z ( ( t − u ) / ϵ 2 ) ⊗ ∣ a z 1 − t − w ⟩ . (65) \epsilon^{-1}B_{a\epsilon z}\left((t-u)/\epsilon^{2}\right)\otimes\left|az\sqrt{1-t-w}\right\rangle. \tag*{(65)} ϵ − 1 B a ϵz ( ( t − u ) / ϵ 2 ) ⊗ a z 1 − t − w ⟩ . (65)
Vacuum and late arrivals are retained in C3.
Cross overlaps and all vector directions
Use B ^ ζ ( p ) = ( 2 π c χ ) − 1 / 2 ∫ 0 ∞ e − i p x / c χ B ζ ( x ) d x \widehat{B}_{\zeta}(p)=(2\pi c_{\chi})^{-1/2}\int_{0}^{\infty}e^{-ipx/c_{\chi}}B_{\zeta}(x)\,\mathrm{d}x B ζ ( p ) = ( 2 π c χ ) − 1/2 ∫ 0 ∞ e − i p x / c χ B ζ ( x ) d x . At separation d ≥ 0 d\ge0 d ≥ 0 , differentiate the exact first arrival packets at equal amplitudes. With η \eta η perturbing the ket amplitude at the later position, their overlap has first-order expansion
κ 1 + η , 1 P ( d ) = e − χ J F ( d / χ ) / 2 [ 1 + η 2 { d − χ J F ( d / χ ) } ] + O ( η 2 ) . \kappa^{P}_{1+\eta,1}(d)=e^{-\chi J_{F}(d/\chi)/2}\left[1+\frac{\eta}{2}\{d-\chi J_{F}(d/\chi)\}\right]+O(\eta^{2}). κ 1 + η , 1 P ( d ) = e − χ J F ( d / χ ) /2 [ 1 + 2 η { d − χ J F ( d / χ )} ] + O ( η 2 ) .
Indeed, the common coherent overlap contributes − η χ J F / 2 -\eta\chi J_{F}/2 − η χ J F /2 , the relative waiting normalization contributes η d / 2 \eta d/2 η d /2 , and the terms involving the squared amplitude difference start at order η 2 \eta^{2} η 2 . The exact packet derivative bounds below justify differentiation and Fourier inversion. At equal amplitudes,
∥ B ^ ζ ( p ) ∥ 2 = m χ , ( 1 + ζ ) 2 ( p ) . (66) \lVert\widehat{B}_{\zeta}(p)\rVert^{2}=m_{\chi,(1+\zeta)^{2}}(p). \tag*{(66)} ∥ B ζ ( p ) ∥ 2 = m χ , ( 1 + ζ ) 2 ( p ) . (66)
With m = m χ , 1 m=m_{\chi,1} m = m χ , 1 , the first ket-amplitude derivative is [ c χ ( v − v ′ ) − ψ χ ( v − v ′ ) ] k χ ( v − v ′ ) / 2 [c_{\chi}(v-v')-\psi_{\chi}(v-v')]k_{\chi}(v-v')/2 [ c χ ( v − v ′ ) − ψ χ ( v − v ′ )] k χ ( v − v ′ ) /2 ; Fourier inversion gives
w 0 = B ^ 0 / m , η ( p ) = ∂ τ log m χ , τ ( p ) ∣ τ = 1 − i c χ 2 ∂ p log m ( p ) , w 1 = ∂ ζ B ^ ζ ∣ 0 m − η w 0 , w 1 ⊥ w 0 . (67) w_0=\widehat{B}_0/\sqrt{m},\qquad\eta(p)=\left.\partial_{\tau}\log m_{\chi,\tau}(p)\right|_{\tau=1}-\frac{ic_{\chi}}{2}\partial_p\log m(p),\qquad w_1=\frac{\left.\partial_{\zeta}\widehat{B}_{\zeta}\right|_0}{\sqrt{m}}-\eta w_0,\qquad w_1\perp w_0. \tag*{(67)} w 0 = B 0 / m , η ( p ) = ∂ τ log m χ , τ ( p ) ∣ τ = 1 − 2 i c χ ∂ p log m ( p ) , w 1 = m ∂ ζ B ζ 0 − η w 0 , w 1 ⊥ w 0 . (67)
Thus q = a Re η q=a\operatorname{Re}\eta q = a Re η and r = a Im η r=a\operatorname{Im}\eta r = a Im η are the scalar intensity and phase components; w 1 w_1 w 1 is the orthogonal tangent. For every fixed finite j , k j,k j , k ,
∥ ∂ ζ j ∂ p k B ^ ζ ( p ) ∥ ≤ C j , k ( 1 + ∣ p ∣ ) − 4 , ∣ ζ ∣ ≤ ζ 0 < 1. (68) \left\|\partial_{\zeta}^{j}\partial_p^{k}\widehat{B}_{\zeta}(p)\right\|\le C_{j,k}(1+|p|)^{-4},\qquad|\zeta|\le\zeta_0<1. \tag*{(68)} ∂ ζ j ∂ p k B ζ ( p ) ≤ C j , k ( 1 + ∣ p ∣ ) − 4 , ∣ ζ ∣ ≤ ζ 0 < 1. (68)
Put N ζ ( x ) = ∥ ( 1 + ζ ) F χ ( x + ⋅ ) − 1 ∥ L 2 [ 0 , χ ] N_{\zeta}(x)=\|(1+\zeta)F_{\chi}(x+\cdot)-1\|_{L^2[0,\chi]} N ζ ( x ) = ∥ ( 1 + ζ ) F χ ( x + ⋅ ) − 1 ∥ L 2 [ 0 , χ ] . In sector n n n , the integrand is the waiting amplitude times e − N ζ ( x ) / 2 ∏ l = 1 n [ ( 1 + ζ ) F χ ( x + y l ) − 1 ] / n ! e^{-N_{\zeta}(x)/2}\prod_{l=1}^{n}[(1+\zeta)F_{\chi}(x+y_l)-1]/\sqrt{n!} e − N ζ ( x ) /2 ∏ l = 1 n [( 1 + ζ ) F χ ( x + y l ) − 1 ] / n ! . After j j j derivatives and multiplication by x k x^k x k , this integrand is C 2 C^2 C 2 with third derivative jumps at 0 , χ , χ − y l 0,\chi,\chi-y_l 0 , χ , χ − y l . Their absolute sum and the regular fourth derivative L 1 L^1 L 1 norm are bounded by C j , k ( n + 1 ) N j , k R 1 n / n ! C_{j,k}(n+1)N_{j,k}R_1^n/\sqrt{n!} C j , k ( n + 1 ) N j , k R 1 n / n ! . Four integrations by parts give ∣ p ∣ − 4 |p|^{-4} ∣ p ∣ − 4 decay, and the L 1 L^1 L 1 bound controls small p p p . Since ∑ ( n + 1 ) 2 N ( R 2 n ) / n ! < ∞ \sum(n+1)^{2N}(R_2^n)/n!<\infty ∑ ( n + 1 ) 2 N ( R 2 n ) / n ! < ∞ , the squared sector bounds sum. The relative spectral estimates then bound derivatives of the normalized packet.
On the enlarged window space, use
U 0 ( p ) = ( 0 w 0 ( p ) † w 0 ( p ) I − ∣ w 0 ( p ) ⟩ ⟨ w 0 ( p ) ∣ ) . (69) U_0(p)=\begin{pmatrix}0&w_0(p)^\dagger\\w_0(p)&I-|w_0(p)\rangle\langle w_0(p)|\end{pmatrix}. \tag*{(69)} U 0 ( p ) = ( 0 w 0 ( p ) w 0 ( p ) † I − ∣ w 0 ( p )⟩ ⟨ w 0 ( p ) ∣ ) . (69)
This self-adjoint unitary maps ( 0 , w 0 ) (0,w_0) ( 0 , w 0 ) to e 0 = ( 1 , 0 ) e_0=(1,0) e 0 = ( 1 , 0 ) . Set w ~ 1 = U 0 ( 0 , w 1 ) \widetilde{w}_1=U_0(0,w_1) w 1 = U 0 ( 0 , w 1 ) ; the derivative estimates give C7’s weighted expansion. Rescaling the window to [ 0 , 1 ] [0,1] [ 0 , 1 ] makes the sector bounds uniform on I I I .
Endpoint estimates at quadratic precision
Fix L < ∞ L<\infty L < ∞ , w = χ ϵ 2 w=\chi\epsilon^2 w = χ ϵ 2 , s = ϵ − 1 ≥ 2 a L s=\epsilon^{-1}\ge2aL s = ϵ − 1 ≥ 2 a L , w < d < 1 / 4 w<d<1/4 w < d < 1/4 , and B λ = 2 λ + 4 L 2 ( 1 − λ ) B_{\lambda}=2\lambda+4L^2(1-\lambda) B λ = 2 λ + 4 L 2 ( 1 − λ ) . For real filters ∑ j f j ( X ) 2 = I \sum_j f_j(X)^2=I ∑ j f j ( X ) 2 = I , put E j = f j ( X ) − f j ( u ) E_j=f_j(X)-f_j(u) E j = f j ( X ) − f j ( u ) , Q = ∑ j E j 2 Q=\sum_j E_j^2 Q = ∑ j E j 2 . Completeness gives ∑ j f j ( u ) E j = − Q / 2 \sum_j f_j(u)E_j=-Q/2 ∑ j f j ( u ) E j = − Q /2 , hence
∑ j f j ( X ) ρ f j ( X ) − ρ = ∑ j E j ρ E j − 1 2 ( Q ρ + ρ Q ) . (70) \sum_j f_j(X)\rho f_j(X)-\rho=\sum_j E_j\rho E_j-\frac{1}{2}(Q\rho+\rho Q). \tag*{(70)} j ∑ f j ( X ) ρ f j ( X ) − ρ = j ∑ E j ρ E j − 2 1 ( Qρ + ρQ ) . (70)
For a pure input its trace distance disturbance is at most ( ⟨ Q ⟩ + ⟨ Q 2 ⟩ ) / 2 (\langle Q\rangle+\sqrt{\langle Q^2\rangle})/2 (⟨ Q ⟩ + ⟨ Q 2 ⟩ ) /2 .
Choose filters ( f L , f M , f R ) (f_L,f_M,f_R) ( f L , f M , f R ) equal to the left, middle and right unit vectors outside the transition intervals [ d , 2 d ] [d,2d] [ d , 2 d ] , [ 1 − 2 d , 1 − d ] [1-2d,1-d] [ 1 − 2 d , 1 − d ] . Rotate linearly from zero to π / 2 \pi/2 π /2 on each transition, giving speed ℓ / d \ell/d ℓ / d , ℓ = π / 2 \ell=\pi/2 ℓ = π /2 , and ∑ j TV ( f j 2 ) = 4 \sum_j\operatorname{TV}(f_j^2)=4 ∑ j TV ( f j 2 ) = 4 . Apply the filters to the quantum arrival coordinate, assigning vacuum value one and retaining labels and states. The delay is min ( W z , 1 − u ) \min(W_z,1-u) min ( W z , 1 − u ) , where
W z ≤ s t ϵ 2 Y χ , Y χ = χ + Exp ( 1 / 4 ) . (71) W_z\le_{\mathrm{st}}\epsilon^2Y_{\chi},\qquad Y_{\chi}=\chi+\operatorname{Exp}(1/4). \tag*{(71)} W z ≤ st ϵ 2 Y χ , Y χ = χ + Exp ( 1/4 ) . (71)
because the post rise intensity is at least s 2 / 4 s^2/4 s 2 /4 . All moments of Y χ Y_{\chi} Y χ are uniformly finite on I I I . Hence
δ ‾ ϵ , d ≤ C I ( ϵ 4 d + ϵ 6 d 2 ) . (72) \overline{\delta}_{\epsilon,d}\le C_I\left(\frac{\epsilon^4}{d}+\frac{\epsilon^6}{d^2}\right). \tag*{(72)} δ ϵ , d ≤ C I ( d ϵ 4 + d 2 ϵ 6 ) . (72)
A delay W W W crosses transitions on length at most 2 d + 2 W 2d+2W 2 d + 2 W . Integrating Q ≤ ℓ 2 W 2 / d 2 Q\le\ell^2W^2/d^2 Q ≤ ℓ 2 W 2 / d 2 gives C I ( ϵ 4 / d + ϵ 6 / d 2 ) C_I(\epsilon^4/d+\epsilon^6/d^2) C I ( ϵ 4 / d + ϵ 6 / d 2 ) . For the square root term, the transition interiors have the same bound, while the exterior strips contribute at most ϵ 6 d − 2 ∫ 0 ∞ [ E ( Y χ 4 1 Y χ ≥ r ) ] 1 / 2 d r \epsilon^6d^{-2}\int_0^\infty[\mathbb{E}(Y_{\chi}^4\mathbf{1}_{Y_{\chi}\ge r})]^{1/2}\,dr ϵ 6 d − 2 ∫ 0 ∞ [ E ( Y χ 4 1 Y χ ≥ r ) ] 1/2 d r . The exponential envelope bounds this integral uniformly, including vacuum.
Posterior weights and endpoint costs
Keep the normalized conditional states and compare their label weights p j ( u , z ) p_j(u,z) p j ( u , z ) with p j ( u , 0 ) p_j(u,0) p j ( u , 0 ) . Monotone coupling of the waiting survival probabilities e − τ z A χ ( x ) e^{-\tau_z A_\chi(x)} e − τ z A χ ( x ) , τ z = ( 1 + a ε z ) 2 \tau_z=(1+a\varepsilon z)^2 τ z = ( 1 + a ε z ) 2 , gives E ∣ W z − W 0 ∣ ≤ ε 2 ∣ τ z − 1 ∣ C χ \mathbb{E}|W_z-W_0|\le\varepsilon^2|\tau_z-1|C_\chi E ∣ W z − W 0 ∣ ≤ ε 2 ∣ τ z − 1∣ C χ , C χ = χ 2 + 16 C_\chi=\chi^2+16 C χ = χ 2 + 16 . Estimate the initial segment of width χ \chi χ directly, then use A χ ( x ) = x − ν F χ A_\chi(x)=x-\nu_F\chi A χ ( x ) = x − ν F χ for τ ≥ 1 / 4 \tau\ge1/4 τ ≥ 1/4 . The bounded variation translation inequality gives
∫ 0 1 1 2 ∑ j ∣ p j ( u , z ) − p j ( u , 0 ) ∣ d u ≤ η ˉ ε : = 2 C χ ( 2 a L ε 3 + a 2 L 2 ε 4 ) . (73) \int_0^1 \frac{1}{2}\sum_j |p_j(u,z)-p_j(u,0)|\,du\le\bar\eta_\varepsilon:=2C_\chi(2aL\varepsilon^3+a^2L^2\varepsilon^4). \tag*{(73)} ∫ 0 1 2 1 j ∑ ∣ p j ( u , z ) − p j ( u , 0 ) ∣ d u ≤ η ˉ ε := 2 C χ ( 2 a L ε 3 + a 2 L 2 ε 4 ) . (73)
For π j = ∫ p j ( u , 0 ) d u \pi_j=\int p_j(u,0)\,du π j = ∫ p j ( u , 0 ) d u , let F j F_j F j have position density p j ( u , 0 ) / π j p_j(u,0)/\pi_j p j ( u , 0 ) / π j , the original amplitude prior and normalized states. The forward labelled comparison and reverse label removal have errors η ˉ \bar\eta η ˉ and δ ˉ + η ˉ \bar\delta+\bar\eta δ ˉ + η ˉ . Lemma A1.1 gives
∣ G λ ( E ) − ∑ j π j G λ ( F j ) ∣ ≤ B λ ( δ ˉ + 2 η ˉ ) . (74) \left|\mathcal{G}_\lambda(E)-\sum_j\pi_j\mathcal{G}_\lambda(F_j)\right|\le B_\lambda(\bar\delta+2\bar\eta). \tag*{(74)} G λ ( E ) − j ∑ π j G λ ( F j ) ≤ B λ ( δ ˉ + 2 η ˉ ) . (74)
On the left branch, u < 2 d u<2d u < 2 d . Splitting the field at 2 d + w 2d+w 2 d + w and freezing the amplitude in the window leaves a constant suffix mode. The contracted vectors satisfy
∥ ∥ v 0 ∥ v z ∥ v z ∥ − v 0 ∥ v 0 ∥ ∥ ≤ 2 ∥ v z − v 0 ∥ . (75) \left\|\|v_0\|\frac{v_z}{\|v_z\|}-\frac{v_0}{\|v_0\|}\right\|\le2\|v_z-v_0\|. \tag*{(75)} ∥ v 0 ∥ ∥ v z ∥ v z − ∥ v 0 ∥ v 0 ≤ 2∥ v z − v 0 ∥. (75)
The error weighted by the reference probability is at most 4 a L d 2 d + w 4aLd\sqrt{2d+w} 4 a L d 2 d + w ; a zero reference probability contributes zero mass. The product experiment has zero gap, so two comparisons through the identity channel give
π L G λ ( F L ) ≤ 8 B λ a L d 2 d + w . (76) \pi_L\mathcal{G}_\lambda(F_L)\le8B_\lambda aLd\sqrt{2d+w}. \tag*{(76)} π L G λ ( F L ) ≤ 8 B λ a L d 2 d + w . (76)
On the right branch, a near optimal location rule is paired with an amplitude report of zero. Passing to the actual weights costs at most 10 L 2 η ˉ 10L^2\bar\eta 10 L 2 η ˉ in the risk of the zero report and in the optimum. Reveal u u u and complete the rule by reporting zero on the complement of the branch. Lemma A3.1 with QFI 4 a 2 ( 1 − u ) 4a^2(1-u) 4 a 2 ( 1 − u ) bounds the full record amplitude improvement by ( 8 / 15 ) a 2 L 4 ( 1 − u ) (8/15)a^2L^4(1-u) ( 8/15 ) a 2 L 4 ( 1 − u ) . Integrate over [ 1 − 4 d , 1 ] [1-4d,1] [ 1 − 4 d , 1 ] ; earlier right labels have probability at most 4 ε 2 e − ( 2 d − w ) / ( 4 ε 2 ) 4\varepsilon^2e^{-(2d-w)/(4\varepsilon^2)} 4 ε 2 e − ( 2 d − w ) / ( 4 ε 2 ) . Thus
π R G λ ( F R ) ≤ ( 1 − λ ) { 64 15 a 2 L 4 d 2 + 4 L 2 ε 2 e − ( 2 d − w ) / ( 4 ε 2 ) + 10 L 2 η ˉ ε } . (77) \pi_R\mathcal{G}_\lambda(F_R)\le(1-\lambda)\left\{\frac{64}{15}a^2L^4d^2+4L^2\varepsilon^2e^{-(2d-w)/(4\varepsilon^2)}+10L^2\bar\eta_\varepsilon\right\}. \tag*{(77)} π R G λ ( F R ) ≤ ( 1 − λ ) { 15 64 a 2 L 4 d 2 + 4 L 2 ε 2 e − ( 2 d − w ) / ( 4 ε 2 ) + 10 L 2 η ˉ ε } . (77)
Return to the middle synthesis
For d = ε 3 / 2 d=\varepsilon^{3/2} d = ε 3/2 , the disturbance is O ( ε 5 / 2 ) O(\varepsilon^{5/2}) O ( ε 5/2 ) and the weight change is O ( ε 3 ) O(\varepsilon^3) O ( ε 3 ) . The weighted left and right gaps are O ( ε 9 / 4 ) O(\varepsilon^{9/4}) O ( ε 9/4 ) , O ( ε 3 ) O(\varepsilon^3) O ( ε 3 ) . The conditional gaps are nonnegative, so restoring the middle weights gives
G λ ( E ) = G λ ( T p h ) + o ( ε 2 ) . (78) \mathcal{G}_\lambda(E)=\mathcal{G}_\lambda(T_{\rm ph})+o(\varepsilon^2). \tag*{(78)} G λ ( E ) = G λ ( T ph ) + o ( ε 2 ) . (78)
Here T p h T_{\rm ph} T ph is C7’s exact arrival synthesis with k = f M k=f_M k = f M ; source extension adds the waiting tail. The endpoint rules give o ( ε 2 ) o(\varepsilon^2) o ( ε 2 ) excess for each task at fixed interior weights and combine with the middle rule in C9, uniformly on compact positive χ \chi χ intervals.
Comparison of marginal excess risks
Use F δ v = ( 2 π δ ) − 1 / 2 ∫ e − i p u / δ v ( u ) d u \mathcal{F}_\delta v=(2\pi\delta)^{-1/2}\int e^{-ipu/\delta}v(u)\,du F δ v = ( 2 π δ ) − 1/2 ∫ e − i p u / δ v ( u ) d u , P δ = − i δ ∂ u P_\delta=-i\delta\partial_u P δ = − i δ ∂ u , m 1 ( p ) = m χ , 1 ( p ) > 0 m_1(p)=m_{\chi,1}(p)>0 m 1 ( p ) = m χ , 1 ( p ) > 0 from C2, f = m 1 f=\sqrt{m_1} f = m 1 , and
γ ( p , q ) = f ( p ) f ( q ) ( 1 − cos K ( p − q ) ) [ π K ( p − q ) ] 2 , g ( p ) = ∫ γ ( p , q ) d q . \gamma(p,q)=\frac{f(p)f(q)(1-\cos K(p-q))}{[\pi K(p-q)]^2},\qquad g(p)=\int\gamma(p,q)\,dq. γ ( p , q ) = [ π K ( p − q ) ] 2 f ( p ) f ( q ) ( 1 − cos K ( p − q )) , g ( p ) = ∫ γ ( p , q ) d q .
Write ⟨ P δ ⟩ = ( I + P δ 2 ) 1 / 2 \langle P_\delta\rangle=(I+P_\delta^2)^{1/2} ⟨ P δ ⟩ = ( I + P δ 2 ) 1/2 , and let M E M_E M E denote multiplication by the coefficient field E ( t ) E(t) E ( t ) . For f ≍ ( 1 + ∣ p ∣ ) − 4 f\asymp(1+|p|)^{-4} f ≍ ( 1 + ∣ p ∣ ) − 4 ,
Λ δ = 2 ( 1 − g / m 1 ) ( P δ ) ⊗ I , g ( p ) m 1 ( p ) ≤ C ( 1 + p 2 ) . (79) \Lambda_\delta=2(1-g/m_1)(P_\delta)\otimes I,\qquad\frac{g(p)}{m_1(p)}\le C(1+p^2). \tag*{(79)} Λ δ = 2 ( 1 − g / m 1 ) ( P δ ) ⊗ I , m 1 ( p ) g ( p ) ≤ C ( 1 + p 2 ) . (79)
The corresponding positive source form is
L x , δ = 2 min { 1 , ∣ x − u ∣ / ( K δ ) } − Λ δ ≥ 0 (80) L_{x,\delta}=2\min\{1,|x-u|/(K\delta)\}-\Lambda_\delta\ge0 \tag*{(80)} L x , δ = 2 min { 1 , ∣ x − u ∣/ ( K δ )} − Λ δ ≥ 0 (80)
Conjugating by f ( P δ ) f(P_\delta) f ( P δ ) and the report phase gives twice the nonnegative graph Laplacian. The bound on g / m 1 g/m_1 g / m 1 follows from convolution of ( 1 + ∣ p ∣ ) − 4 (1+|p|)^{-4} ( 1 + ∣ p ∣ ) − 4 with C ( 1 + ∣ p − q ∣ ) − 2 C(1+|p-q|)^{-2} C ( 1 + ∣ p − q ∣ ) − 2 . For T ⟨ P δ ⟩ T\langle P_\delta\rangle T ⟨ P δ ⟩ Hilbert–Schmidt, define
p J ( T ) = r l o c ( J ; T ) − Tr ( T Λ δ T † ) , d δ ( R ) 2 = 2 ∥ R ∥ H S 2 + ∥ R ( − Λ δ ) 1 / 2 ∥ H S 2 . (81) \mathfrak{p}_J(T)=r_{\mathrm{loc}}(J;T)-\operatorname{Tr}(T\Lambda_\delta T^\dagger),\qquad d_\delta(R)^2=2\lVert R\rVert_{\mathrm{HS}}^2+\lVert R(-\Lambda_\delta)^{1/2}\rVert_{\mathrm{HS}}^2. \tag*{(81)} p J ( T ) = r loc ( J ; T ) − Tr ( T Λ δ T † ) , d δ ( R ) 2 = 2 ∥ R ∥ HS 2 + ∥ R ( − Λ δ ) 1/2 ∥ HS 2 . (81)
The positive source form writes p J ( T ) \mathfrak{p}_J(T) p J ( T ) as a squared direct integral row norm. Outcome partitions and domain approximation extend it to every complete POVM. The triangle inequality and 0 ≤ ℓ ≤ 2 0\le\ell\le2 0 ≤ ℓ ≤ 2 give
0 ≤ p J ( R ) ≤ d δ ( R ) 2 , ∣ p J ( T ) − p J ( S ) ∣ ≤ d δ ( T − S ) ≤ C ∥ ( T − S ) ⟨ P δ ⟩ ∥ H S . (82) 0\le\mathfrak{p}_J(R)\le d_\delta(R)^2,\qquad\left|\sqrt{\mathfrak{p}_J(T)}-\sqrt{\mathfrak{p}_J(S)}\right|\le d_\delta(T-S)\le C\lVert(T-S)\langle P_\delta\rangle\rVert_{\mathrm{HS}}. \tag*{(82)} 0 ≤ p J ( R ) ≤ d δ ( R ) 2 , p J ( T ) − p J ( S ) ≤ d δ ( T − S ) ≤ C ∥( T − S ) ⟨ P δ ⟩ ∥ HS . (82)
Taking near minimizers in both directions gives, with p ⋆ ( T ) = inf J p J ( T ) \mathfrak{p}^\star(T)=\inf_J\mathfrak{p}_J(T) p ⋆ ( T ) = inf J p J ( T ) ,
∣ p ⋆ ( T ) − p ⋆ ( S ) ∣ ≤ d δ ( T − S ) , e l o c ( J ; T ) = p J ( T ) − p ⋆ ( T ) . (83) \left|\sqrt{\mathfrak{p}^\star(T)}-\sqrt{\mathfrak{p}^\star(S)}\right|\le d_\delta(T-S),\qquad e_{\mathrm{loc}}(J;T)=\mathfrak{p}_J(T)-\mathfrak{p}^\star(T). \tag*{(83)} p ⋆ ( T ) − p ⋆ ( S ) ≤ d δ ( T − S ) , e loc ( J ; T ) = p J ( T ) − p ⋆ ( T ) . (83)
Thus each excess is centered at its own marginal optimum.
Lemma C4.1 (Centered quadratic comparison). Fix 0 < λ < 1 0<\lambda<1 0 < λ < 1 . Let T ϵ , S ϵ T_\epsilon,S_\epsilon T ϵ , S ϵ have the same source targets, losses and class of complete paired measurements, with their finite unnormalized masses retained. Assume:
d δ ( T ϵ − S ϵ ) = o ( ϵ ) d_\delta(T_\epsilon-S_\epsilon)=o(\epsilon) d δ ( T ϵ − S ϵ ) = o ( ϵ ) ;
for a common reference T 0 T_0 T 0 and a complete rule J 0 J_0 J 0 , p J 0 ( T 0 ) = 0 \mathfrak{p}_{J_0}(T_0)=0 p J 0 ( T 0 ) = 0 and d δ ( T ϵ − T 0 ) + d δ ( S ϵ − T 0 ) = O ( ϵ ) d_\delta(T_\epsilon-T_0)+d_\delta(S_\epsilon-T_0)=O(\epsilon) d δ ( T ϵ − T 0 ) + d δ ( S ϵ − T 0 ) = O ( ϵ ) ;
uniformly over all paired rules with ∣ Y ∣ ≤ L |Y|\le L ∣ Y ∣ ≤ L , ∣ e a m p ( J ; T ϵ ) − e a m p ( J ; S ϵ ) ∣ = o ( ϵ ) \left|\sqrt{e_{\mathrm{amp}}(J;T_\epsilon)}-\sqrt{e_{\mathrm{amp}}(J;S_\epsilon)}\right|=o(\epsilon) e amp ( J ; T ϵ ) − e amp ( J ; S ϵ ) = o ( ϵ ) ;
at least one model admits a complete common rule with weighted excess O ( ϵ 2 ) O(\epsilon^2) O ( ϵ 2 ) .
Then
G λ ( T ϵ ) − G λ ( S ϵ ) = o ( ϵ 2 ) . (84) G_\lambda(T_\epsilon)-G_\lambda(S_\epsilon)=o(\epsilon^2). \tag*{(84)} G λ ( T ϵ ) − G λ ( S ϵ ) = o ( ϵ 2 ) . (84) Uniform hypotheses and moduli on a known design set give the uniform conclusion there.
Proof. The reference gives p ⋆ ( T ) , p ⋆ ( S ) = O ( ϵ 2 ) \mathfrak{p}^\star(T),\mathfrak{p}^\star(S)=O(\epsilon^2) p ⋆ ( T ) , p ⋆ ( S ) = O ( ϵ 2 ) . A feasible O ( ϵ 2 ) O(\epsilon^2) O ( ϵ 2 ) gap at an interior weight bounds both excesses. Root stability and hypothesis 3 transfer these bounds between the models. For a gap optimizer within o ( ϵ 2 ) o(\epsilon^2) o ( ϵ 2 ) of the infimum, d δ ( T − S ) ( p J ( T ) + p J ( S ) ) = o ( ϵ 2 ) d_\delta(T-S)(\sqrt{\mathfrak{p}_J(T)}+\sqrt{\mathfrak{p}_J(S)})=o(\epsilon^2) d δ ( T − S ) ( p J ( T ) + p J ( S ) ) = o ( ϵ 2 ) controls the location forms and, by the same estimate at marginal minimizers, their infima. Squaring the amplitude root bound and interchanging T , S T,S T , S completes the comparison. □
Hypothesis 3 follows from a common bounded self-adjoint predictor. With Ω T = T T † \Omega_T=TT^\dagger Ω T = T T † , Ω S = S S † \Omega_S=SS^\dagger Ω S = S S † , assume
∥ A T − A ~ ϵ ∥ Ω T + ∥ A S − A ~ ϵ ∥ Ω S = o ( ϵ ) , ∥ T − S ∥ H S = o ( ϵ ) , sup ϵ ∥ A ~ ϵ ∥ < ∞ . (85) \lVert A_T-\widetilde{A}_\epsilon\rVert_{\Omega_T}+\lVert A_S-\widetilde{A}_\epsilon\rVert_{\Omega_S}=o(\epsilon),\qquad\lVert T-S\rVert_{\mathrm{HS}}=o(\epsilon),\qquad\sup_\epsilon\lVert\widetilde{A}_\epsilon\rVert<\infty. \tag*{(85)} ∥ A T − A ϵ ∥ Ω T + ∥ A S − A ϵ ∥ Ω S = o ( ϵ ) , ∥ T − S ∥ HS = o ( ϵ ) , ϵ sup ∥ A ϵ ∥ < ∞. (85)
The square identity and triangle inequality give
∣ e a m p ( J ; T ) − e a m p ( J ; S ) ∣ ≤ o ( ϵ ) + ( L + ∥ A ~ ϵ ∥ ) ∥ T − S ∥ H S = o ( ϵ ) . (86) \left|\sqrt{e_{\mathrm{amp}}(J;T)}-\sqrt{e_{\mathrm{amp}}(J;S)}\right|\le o(\epsilon)+(L+\lVert\widetilde{A}_\epsilon\rVert)\lVert T-S\rVert_{\mathrm{HS}}=o(\epsilon). \tag*{(86)} e amp ( J ; T ) − e amp ( J ; S ) ≤ o ( ϵ ) + ( L + ∥ A ϵ ∥) ∥ T − S ∥ HS = o ( ϵ ) . (86)
C1’s source defect bound verifies these estimates in C5 and C7.
Scalar experiment and a common readout
Use C3’s middle filter k k k with d = ϵ 3 / 2 d=\epsilon^{3/2} d = ϵ 3/2 . Then ∥ k ∥ 2 2 = 1 − 3 d \lVert k\rVert_2^2=1-3d ∥ k ∥ 2 2 = 1 − 3 d , ∥ k ′ ∥ 2 2 = π 2 / ( 4 d ) \lVert k'\rVert_2^2=\pi^2/(4d) ∥ k ′ ∥ 2 2 = π 2 / ( 4 d ) , and
c ( t ) = a 1 − clip [ d , 1 − d ] ( t ) − w ˉ , ∥ c ′ ∥ 2 2 = a 2 4 log 1 − d − w ˉ d − w ˉ . (87) c(t)=a\sqrt{1-\operatorname{clip}_{[d,1-d]}(t)-\bar{w}},\qquad\lVert c'\rVert_2^2=\frac{a^2}{4}\log\frac{1-d-\bar{w}}{d-\bar{w}}. \tag*{(87)} c ( t ) = a 1 − clip [ d , 1 − d ] ( t ) − w ˉ , ∥ c ′ ∥ 2 2 = 4 a 2 log d − w ˉ 1 − d − w ˉ . (87)
The source is L 2 ( d u ) ⊗ L 2 ( ν ) L^2(du)\otimes L^2(\nu) L 2 ( d u ) ⊗ L 2 ( ν ) , ν = Unif [ − 1 / 2 , 1 / 2 ] \nu=\operatorname{Unif}[-1/2,1/2] ν = Unif [ − 1/2 , 1/2 ] ; Z Z Z multiplies by z z z and D ( t ) = k ( t ) V c ( t ) D(t)=k(t)V_{c(t)} D ( t ) = k ( t ) V c ( t ) . Using the positive spectrum in (C2.2), define
m τ ( p ) = m χ , τ ( p ) , m_\tau(p)=m_{\chi,\tau}(p), m τ ( p ) = m χ , τ ( p ) ,
f = m 1 , q = a ∂ τ log m τ ∣ τ = 1 , f z = m ( 1 + a ϵ z ) 2 , f=\sqrt{m_1},\qquad q=a\,\partial_\tau\log m_\tau\big|_{\tau=1},\qquad f_z=\sqrt{m(1+a\epsilon z)^2}, f = m 1 , q = a ∂ τ log m τ τ = 1 , f z = m ( 1 + a ϵz ) 2 ,
Φ ϵ ( z , p ) = f z ( p ) f ( p ) > 0. (88) \Phi_\epsilon(z,p)=\frac{f_z(p)}{f(p)}>0. \tag*{(88)} Φ ϵ ( z , p ) = f ( p ) f z ( p ) > 0. (88)
The source multiplier Φ ϵ ( Z , P δ ) \Phi_\epsilon(Z,P_\delta) Φ ϵ ( Z , P δ ) defines the real synthesis and reference
T ϵ r = T 0 Φ ϵ ( Z , P δ ) , T 0 = 2 π δ M D f ( P δ ) . (89) T_\epsilon^{\mathrm r}=T_0\Phi_\epsilon(Z,P_\delta),\qquad T_0=\sqrt{2\pi\delta}\,M_Df(P_\delta). \tag*{(89)} T ϵ r = T 0 Φ ϵ ( Z , P δ ) , T 0 = 2 π δ M D f ( P δ ) . (89)
Put Ω r = T ϵ r ( T ϵ r ) † \Omega_r=T_\epsilon^{\mathrm r}(T_\epsilon^{\mathrm r})^\dagger Ω r = T ϵ r ( T ϵ r ) † . Both syntheses have the finite mass m = ∥ k ∥ 2 2 m=\lVert k\rVert_2^2 m = ∥ k ∥ 2 2 because ∫ f z 2 = 1 \int f_z^2=1 ∫ f z 2 = 1 . Their losses are 2 min ( 1 , ∣ x − u ∣ / ( K δ ) ) 2\min(1,|x-u|/(K\delta)) 2 min ( 1 , ∣ x − u ∣/ ( K δ )) and ( y − z ) 2 (y-z)^2 ( y − z ) 2 .
Location optimality and amplitude prediction
Coordinate readout followed by A c ( t ) A_{c(t)} A c ( t ) is complete. Multiplying the positive graph slack by k ( t ) ∣ c ( t ) z ⟩ k(t)|c(t)z\rangle k ( t ) ∣ c ( t ) z ⟩ and integrating over z z z gives
p J 0 ( T 0 ) = 0 , r l o c ⋆ ( T 0 ) = m r F ( χ ) . (90) pJ_0(T_0)=0,\qquad r^\star_{\mathrm{loc}}(T_0)=m\,r_F(\chi). \tag*{(90)} p J 0 ( T 0 ) = 0 , r loc ⋆ ( T 0 ) = m r F ( χ ) . (90)
Applied to each positive f z f_z f z , the same graph argument establishes exact optimality of coordinate readout over all complete measurements. For A ˉ = M A c \bar{A}=M_{A_c} A ˉ = M A c , K ˉ = M K c \bar{K}=M_{K_c} K ˉ = M K c , the reference defect equals 2 π δ M D [ K ˉ , f ( P δ ) ] \sqrt{2\pi\delta}M_D[\bar{K},f(P_\delta)] 2 π δ M D [ K ˉ , f ( P δ )] and satisfies
∥ A ˉ T 0 − T 0 Z − T 0 K ˉ ∥ H S ≤ 2 ( 2 L + κ s r c ) L 3 δ ∥ k ∥ ∞ ∥ c ′ ∥ 2 ( ∫ ∣ f ′ ( p ) ∣ 2 d p ) 1 / 2 = o ( ϵ ) . (91) \lVert\bar{A}T_0-T_0Z-T_0\bar{K}\rVert_{\mathrm{HS}}\le2(2L+\kappa_{\mathrm{src}})\frac{L}{\sqrt{3}}\delta\lVert k\rVert_\infty\lVert c'\rVert_2\left(\int|f'(p)|^2\,dp\right)^{1/2}=o(\epsilon). \tag*{(91)} ∥ A ˉ T 0 − T 0 Z − T 0 K ˉ ∥ HS ≤ 2 ( 2 L + κ src ) 3 L δ ∥ k ∥ ∞ ∥ c ′ ∥ 2 ( ∫ ∣ f ′ ( p ) ∣ 2 d p ) 1/2 = o ( ϵ ) . (91)
Indeed, the squared kernel norm integrates k ( t ) 2 ∣ α ( x ) ∣ 2 ∥ V c ( t ) ( K c ( t ) − K c ( t − δ x ) ) ∥ H S 2 k(t)^2|\alpha(x)|^2\lVert V_{c(t)}(K_{c(t)}-K_{c(t-\delta x)})\rVert_{\mathrm{HS}}^2 k ( t ) 2 ∣ α ( x ) ∣ 2 ∥ V c ( t ) ( K c ( t ) − K c ( t − δ x ) ) ∥ HS 2 , α = F − 1 f \alpha=\mathcal{F}^{-1}f α = F − 1 f . Apply C1’s weighted lift bound and ∥ c ( ⋅ ) − c ( ⋅ − δ x ) ∥ 2 ≤ δ ∣ x ∣ ∥ c ′ ∥ 2 \lVert c(\cdot)-c(\cdot-\delta x)\rVert_2\le\delta|x|\lVert c'\rVert_2 ∥ c ( ⋅ ) − c ( ⋅ − δ x ) ∥ 2 ≤ δ ∣ x ∣ ∥ c ′ ∥ 2 .
For b ( p ) = ∫ e i y p μ ( d y ) b(p)=\int e^{iyp}\,\mu(dy) b ( p ) = ∫ e i y p μ ( d y ) , the translation representation and Minkowski’s inequality give
2 π δ ∥ [ b ( P δ ) , M E ] f ( P δ ) ∥ H S ≤ ∥ f ∥ 2 ∫ ∥ E ( ⋅ + δ y ) − E ( ⋅ ) ∥ L 2 ( H S ) ∣ μ ∣ ( d y ) . (92) \sqrt{2\pi\delta}\lVert[b(P_\delta),M_E]f(P_\delta)\rVert_{\mathrm{HS}}\le\lVert f\rVert_2\int\lVert E(\cdot+\delta y)-E(\cdot)\rVert_{L^2(\mathrm{HS})}\,|\mu|(dy). \tag*{(92)} 2 π δ ∥[ b ( P δ ) , M E ] f ( P δ ) ∥ HS ≤ ∥ f ∥ 2 ∫ ∥ E ( ⋅ + δ y ) − E ( ⋅ ) ∥ L 2 ( HS ) ∣ μ ∣ ( d y ) . (92)
For E = D E=D E = D , δ ∥ D ′ ∥ 2 = O ( ϵ 5 / 4 ) \delta\lVert D'\rVert_2=O(\epsilon^{5/4}) δ ∥ D ′ ∥ 2 = O ( ϵ 5/4 ) . For residual coefficients, remove an endpoint strip and truncate the Fourier measure, then apply C1’s weighted and strong continuity. C10 makes these moduli uniform. Intensity Taylor expansion gives the source multiplier Φ ϵ ( Z , P δ ) = I + ϵ Z q ( P δ ) + O ( ϵ 2 ) \Phi_\epsilon(Z,P_\delta)=I+\epsilon Zq(P_\delta)+O(\epsilon^2) Φ ϵ ( Z , P δ ) = I + ϵ Z q ( P δ ) + O ( ϵ 2 ) and hence
T ϵ r = T 0 + ϵ T 1 + O H S ( ϵ 2 ) , T 1 = 2 π δ M D Z q ( P δ ) f ( P δ ) . (93) T_\epsilon^{\mathrm r}=T_0+\epsilon T_1+O_{\mathrm{HS}}(\epsilon^2),\qquad T_1=\sqrt{2\pi\delta}\,M_DZq(P_\delta)f(P_\delta). \tag*{(93)} T ϵ r = T 0 + ϵ T 1 + O HS ( ϵ 2 ) , T 1 = 2 π δ M D Z q ( P δ ) f ( P δ ) . (93)
Define the amplitude and source predictors by
A δ = M A c + ϵ 2 { q ( P δ ) , M B c } , K δ = M K c + ϵ 2 { q ( P δ ) , M K 1 , c } . (94) A_\delta=M_{A_c}+\frac{\epsilon}{2}\{q(P_\delta),M_{B_c}\},\qquad K_\delta=M_{K_c}+\frac{\epsilon}{2}\{q(P_\delta),M_{K_{1,c}}\}. \tag*{(94)} A δ = M A c + 2 ϵ { q ( P δ ) , M B c } , K δ = M K c + 2 ϵ { q ( P δ ) , M K 1 , c } . (94)
Here B c B_c B c is C1’s response. The bounded predictors are self-adjoint and skew-adjoint, respectively. The zeroth order defect is (C5.5); reordering the first order terms gives 2 B c V c + V c [ K c , Z ] − V c K 1 , c = 0 2B_cV_c+V_c[K_c,Z]-V_cK_{1,c}=0 2 B c V c + V c [ K c , Z ] − V c K 1 , c = 0 . The remaining ϵ o ( 1 ) + O ( ϵ 2 ) \epsilon o(1)+O(\epsilon^2) ϵo ( 1 ) + O ( ϵ 2 ) error and the source defect lemma imply
∥ A δ T ϵ r − T ϵ r Z − T ϵ r K δ ∥ H S = o ( ϵ ) , ∥ A δ − A ϵ ∥ Ω r = o ( ϵ ) . (95) \lVert A_\delta T_\epsilon^{\mathrm r}-T_\epsilon^{\mathrm r}Z-T_\epsilon^{\mathrm r}K_\delta\rVert_{\mathrm{HS}}=o(\epsilon),\qquad\lVert A_\delta-A_\epsilon\rVert_{\Omega_r}=o(\epsilon). \tag*{(95)} ∥ A δ T ϵ r − T ϵ r Z − T ϵ r K δ ∥ HS = o ( ϵ ) , ∥ A δ − A ϵ ∥ Ω r = o ( ϵ ) . (95)
A complete upper readout
Coordinate readout followed by A c A_c A c defines a location optimal rule M 0 M_0 M 0 satisfying Y J = J A ‾ YJ=J\overline{A} Y J = J A , A ‾ = M A c \overline{A}=M A_c A = M A c . Its amplitude calibration is
r a m p , r e f ( ϵ ) = ∫ k ( t ) 2 A L ( c ( t ) ) d t , A L ( c ) = L 2 / 3 − Tr ( Ω c A c 2 ) . (96) r_{\mathrm{amp,ref}}(\epsilon)=\int k(t)^2\mathcal{A}_L(c(t))\,dt,\qquad\mathcal{A}_L(c)=L^2/3-\operatorname{Tr}(\Omega_c A_c^2). \tag*{(96)} r amp , ref ( ϵ ) = ∫ k ( t ) 2 A L ( c ( t )) d t , A L ( c ) = L 2 /3 − Tr ( Ω c A c 2 ) . (96)
Boundedness of B c , q B_c,q B c , q gives ∥ A ‾ − A δ ∥ Ω r = O ( ϵ ) \lVert\overline{A}-A_\delta\rVert_{\Omega_r}=O(\epsilon) ∥ A − A δ ∥ Ω r = O ( ϵ ) . Equation (C5.9) and the square identity then give, uniformly on I I I ,
e a m p ( M 0 ; T ϵ r ) = ∥ A ‾ − A ϵ ∥ Ω r 2 ≤ C a m p ϵ 2 + o ( ϵ 2 ) . (97) e_{\mathrm{amp}}(M_0;T_\epsilon^r)=\lVert\overline{A}-A_\epsilon\rVert_{\Omega_r}^2\leq C_{\mathrm{amp}}\epsilon^2+o(\epsilon^2). \tag*{(97)} e amp ( M 0 ; T ϵ r ) = ∥ A − A ϵ ∥ Ω r 2 ≤ C amp ϵ 2 + o ( ϵ 2 ) . (97)
The centering uses the finite-ϵ \epsilon ϵ amplitude optimum. The identity Tr ( Ω c B c ) = A L ( c ) \operatorname{Tr}(\Omega_cB_c)=\mathcal{A}_L(c) Tr ( Ω c B c ) = A L ( c ) gives C6’s response.
Finite positions and conditional weights
Use F − 1 b ( x ) = ( 2 π ) − 1 / 2 ∫ R e i p x b ( p ) d p \mathcal{F}^{-1}b(x)=(2\pi)^{-1/2}\int_{\mathbb{R}}e^{ipx}b(p)\,dp F − 1 b ( x ) = ( 2 π ) − 1/2 ∫ R e i p x b ( p ) d p and α 0 = F − 1 f \alpha_0=\mathcal{F}^{-1}f α 0 = F − 1 f . For the real synthesis, put α ϵ , z = F − 1 f z \alpha_{\epsilon,z}=\mathcal{F}^{-1}f_z α ϵ , z = F − 1 f z ; for the phase inclusive synthesis in (C7.10), use α ϵ , z = F − 1 ( f z e i ϵ z r ) \alpha_{\epsilon,z}=\mathcal{F}^{-1}(f_z e^{i\epsilon zr}) α ϵ , z = F − 1 ( f z e i ϵz r ) instead. First use u ∈ R u\in\mathbb{R} u ∈ R with measure d u ν ( d z ) du\,\nu(dz) d u ν ( d z ) and extend k k k by zero. Let p ϵ ( u , z ) = ∫ R k ( t ) δ − 1 ∣ α ϵ , z ( ( t − u ) / δ ) ∣ 2 d t p_\epsilon(u,z)=\int_{\mathbb{R}} k(t)\delta^{-1}\lvert\alpha_{\epsilon,z}((t-u)/\delta)\rvert^2\,dt p ϵ ( u , z ) = ∫ R k ( t ) δ − 1 ∣ α ϵ , z (( t − u ) / δ ) ∣ 2 d t , and define p 0 ( u ) p_0(u) p 0 ( u ) by the same expression with α 0 \alpha_0 α 0 . These column masses weight the normalized vectors on [ 0 , 1 ] [0,1] [ 0 , 1 ] . Uniformly square integrable momentum derivatives give packet moments of every fixed order. Since supp k ⊂ [ d , 1 − d ] \operatorname{supp} k\subset[d,1-d] supp k ⊂ [ d , 1 − d ] , the added mass is at most C r m ( δ / d ) 2 r C_r m(\delta/d)^{2r} C r m ( δ / d ) 2 r . With r = 3 r=3 r = 3 , its contribution to each bounded risk is O ( ϵ 3 ) O(\epsilon^3) O ( ϵ 3 ) . Clip reports to [ 0 , 1 ] [0,1] [ 0 , 1 ] . With W 1 W_1 W 1 denoting the first Wasserstein distance, changing conditional weights while retaining normalized vectors costs
∫ 0 1 ∣ p ϵ ( u , z ) − p 0 ( u ) ∣ d u ≤ TV ( k 2 ) δ W 1 ( ∣ α ϵ , z ∣ 2 , ∣ α 0 ∣ 2 ) = O ( ϵ δ ) . (98) \int_0^1\lvert p_\epsilon(u,z)-p_0(u)\rvert\,du\leq\operatorname{TV}(k^2)\,\delta W_1(\lvert\alpha_{\epsilon,z}\rvert^2,\lvert\alpha_0\rvert^2)=O(\epsilon\delta). \tag*{(98)} ∫ 0 1 ∣ p ϵ ( u , z ) − p 0 ( u )∣ d u ≤ TV ( k 2 ) δ W 1 (∣ α ϵ , z ∣ 2 , ∣ α 0 ∣ 2 ) = O ( ϵδ ) . (98)
Equal packet masses and the weighted first moment derivative bound give the W 1 W_1 W 1 estimate by coupling unmatched mass through zero; TV ( k 2 ) = 2 \operatorname{TV}(k^2)=2 TV ( k 2 ) = 2 . Retain the reference mass π 0 = ∫ 0 1 p 0 \pi_0=\int_0^1p_0 π 0 = ∫ 0 1 p 0 at quadratic precision. These are the scalar column weights; C3 uses the physical arrival posterior.
Lower bound from rigidity and response
Lemma C6.1 (Rigidity–response for the scalar synthesis). Let T = T ϵ T=T_\epsilon T = T ϵ , Ω = T T † \Omega=TT^\dagger Ω = T T † and m = Tr Ω > 0 m=\operatorname{Tr}\Omega>0 m = Tr Ω > 0 , with actual amplitude optimum A A A . Fix a known unitary U U U on the record space and a known measurable phase g g g , ∣ g ( x ) ∣ = 1 \lvert g(x)\rvert=1 ∣ g ( x )∣ = 1 . Assume
U Ω U † ≤ C Ω , ∥ ( g ( X ) J − J U ) T ∥ H S ≤ κ r i g e l o c ( J ) (99) U\Omega U^\dagger\leq C\Omega,\qquad\lVert(g(X)J-JU)T\rVert_{\mathrm{HS}}\leq\kappa_{\mathrm{rig}}\sqrt{e_{\mathrm{loc}}(J)} \tag*{(99)} U Ω U † ≤ C Ω , ∥( g ( X ) J − J U ) T ∥ HS ≤ κ rig e loc ( J ) (99) for every complete paired rule J J J . For any bounded Hermitian predictor with ∥ A ~ − A ∥ Ω ≤ η a m p \lVert\widetilde{A}-A\rVert_\Omega\leq\eta_{\mathrm{amp}} ∥ A − A ∥ Ω ≤ η amp , define Δ ~ = Tr Ω ( A ~ − U † A ~ U ) \widetilde{\Delta}=\operatorname{Tr}\Omega(\widetilde{A}-U^\dagger\widetilde{A}U) Δ = Tr Ω ( A − U † A U ) . For fixed 0 < λ < 1 0<\lambda<1 0 < λ < 1 , the optimized gap satisfies
α = 2 L κ r i g , β = 1 + C , ρ = [ ∣ Δ ~ ∣ m − β η a m p ] + , \alpha=2L\kappa_{\mathrm{rig}},\qquad\beta=1+\sqrt{C},\qquad\rho=\left[\frac{\lvert\widetilde{\Delta}\rvert}{\sqrt{m}}-\beta\eta_{\mathrm{amp}}\right]_+, α = 2 L κ rig , β = 1 + C , ρ = [ m ∣ Δ ∣ − β η amp ] + , G λ ≥ ρ 2 α 2 / λ + β 2 / ( 1 − λ ) . (100) \mathcal{G}_\lambda\geq\frac{\rho^2}{\alpha^2/\lambda+\beta^2/(1-\lambda)}. \tag*{(100)} G λ ≥ α 2 / λ + β 2 / ( 1 − λ ) ρ 2 . (100)
Proof. Set Δ = Tr Ω ( A − U † A U ) \Delta=\operatorname{Tr}\Omega(A-U^\dagger A U) Δ = Tr Ω ( A − U † A U ) and O = J † Y J O=J^\dagger YJ O = J † Y J . The amplitude square bounds the error of replacing A A A by O O O in the unshifted and shifted states by m e a m p \sqrt{m e_{\mathrm{amp}}} m e amp and m C e a m p \sqrt{m C e_{\mathrm{amp}}} m C e amp . Since ∥ g ( X ) J T ∥ H S = \lVert g(X)JT\rVert_{\mathrm{HS}}= ∥ g ( X ) J T ∥ HS = ∥ J U T ∥ H S = m \lVert JUT\rVert_{\mathrm{HS}}=\sqrt{m} ∥ J U T ∥ HS = m and g ( X ) † Y g ( X ) = Y g(X)^{\dagger}Yg(X)=Y g ( X ) † Y g ( X ) = Y , the reported response is at most 2 L m ∥ ( g ( X ) J − J U ) T ∥ H S 2L\sqrt{m}\lVert(g(X)J-JU)T\rVert_{\mathrm{HS}} 2 L m ∥( g ( X ) J − J U ) T ∥ HS . Location rigidity therefore gives
∣ Δ ∣ m ≤ α e l o c + β e a m p . \frac{|\Delta|}{\sqrt{m}}\leq\alpha\sqrt{e_{\mathrm{loc}}}+\beta\sqrt{e_{\mathrm{amp}}}. m ∣Δ∣ ≤ α e loc + β e amp .
The predictor error gives ρ ≤ α e l o c + β e a m p \rho\leq\alpha\sqrt{e_{\mathrm{loc}}}+\beta\sqrt{e_{\mathrm{amp}}} ρ ≤ α e loc + β e amp . Weighted Cauchy–Schwarz bounds its square by [ α 2 / λ + β 2 / ( 1 − λ ) ] [ λ e l o c + ( 1 − λ ) e a m p ] [\alpha^2/\lambda+\beta^2/(1-\lambda)][\lambda e_{\mathrm{loc}}+(1-\lambda)e_{\mathrm{amp}}] [ α 2 / λ + β 2 / ( 1 − λ )] [ λ e loc + ( 1 − λ ) e amp ] ; take the infimum over paired rules.
For D γ ( v ) = ∬ γ ( p , q ) ∥ v ( p ) − v ( q ) ∥ 2 d p d q D_{\gamma}(v)=\iint\gamma(p,q)\lVert v(p)-v(q)\rVert^2dpdq D γ ( v ) = ∬ γ ( p , q ) ∥ v ( p ) − v ( q ) ∥ 2 d p d q , use the normalized scalar spectrum m 1 m_1 m 1 from C5 and assume
D γ ( v ) ≥ c D ∫ m 1 ( p ) ∥ v ( p ) − v ˉ ∥ 2 d p , v ˉ = ∫ m 1 ( p ) v ( p ) d p , (101) D_{\gamma}(v)\geq c_D\int m_1(p)\lVert v(p)-\bar{v}\rVert^2dp,\qquad\bar{v}=\int m_1(p)v(p)\,dp, \tag*{(101)} D γ ( v ) ≥ c D ∫ m 1 ( p ) ∥ v ( p ) − v ˉ ∥ 2 d p , v ˉ = ∫ m 1 ( p ) v ( p ) d p , (101)
and m 1 ( p − ζ ) ≤ C ζ m 1 ( p ) m_1(p-\zeta)\leq C_{\zeta}m_1(p) m 1 ( p − ζ ) ≤ C ζ m 1 ( p ) . Subtracting v ˉ \bar{v} v ˉ , squaring the triangle inequality and changing variable give
∫ m 1 ( p ) ∥ v ( p + ζ ) − v ( p ) ∥ 2 d p ≤ 2 ( 1 + C ζ ) c D − 1 D γ ( v ) . (102) \int m_1(p)\lVert v(p+\zeta)-v(p)\rVert^2dp\leq2(1+C_{\zeta})c_D^{-1}D_{\gamma}(v). \tag*{(102)} ∫ m 1 ( p ) ∥ v ( p + ζ ) − v ( p ) ∥ 2 d p ≤ 2 ( 1 + C ζ ) c D − 1 D γ ( v ) . (102)
For the real scalar synthesis, the location excess is the graph form of the output rows after removing the report phase, at each latent z z z . The momentum shift U ζ U_{\zeta} U ζ multiplies record functions by e i ζ t / δ e^{i\zeta t/\delta} e i ζ t / δ , with report phase g ( x ) = e i ζ x / δ g(x)=e^{i\zeta x/\delta} g ( x ) = e i ζ x / δ . Summing over rows and z z z gives
κ r i g 2 = 2 ( 1 + C ζ ) / c D . (103) \kappa_{\mathrm{rig}}^2=2(1+C_{\zeta})/c_D. \tag*{(103)} κ rig 2 = 2 ( 1 + C ζ ) / c D . (103)
Coordinate optimality centers the excess at the marginal optimum for all complete rules. Apply the graph and shift bounds to m χ , ( 1 + a ϵ z ) 2 m_{\chi,(1+a\epsilon z)}^2 m χ , ( 1 + a ϵz ) 2 with C2/C10’s common constants, then multiply by M k ∣ c z ⟩ M_k|c_z\rangle M k ∣ c z ⟩ and integrate to obtain average state domination.
Use C5’s predictor A δ A_{\delta} A δ , with weighted error o ( ϵ ) o(\epsilon) o ( ϵ ) . The zeroth order means of A c A_c A c cancel because both shifted spectra integrate to one. Evaluating the first order term at the reference spectrum and reordering gives
Δ ~ ϵ , ζ = 2 ϵ ∫ k ( t ) 2 Tr ( Ω c ( t ) B c ( t ) ) d t ∫ [ m 1 ( p ) − m 1 ( p − ζ ) ] q ( p ) d p + o ( ϵ ) = − 2 ϵ Q ζ ( χ ) A ‾ L + o ( ϵ ) , Q ζ ( χ ) = ∫ m 1 ( p ) q ( p + ζ ) d p . (104) \begin{aligned}
\widetilde{\Delta}_{\epsilon,\zeta}=2\epsilon\int k(t)^2\operatorname{Tr}(\Omega_c(t)B_c(t))\,dt\int[m_1(p)-m_1(p-\zeta)]q(p)\,dp+o(\epsilon) \\
&=-2\epsilon Q_{\zeta}(\chi)\overline{A}_L+o(\epsilon),\qquad Q_{\zeta}(\chi)=\int m_1(p)q(p+\zeta)\,dp. \tag*{(104)}
\end{aligned} Δ ϵ , ζ = 2 ϵ ∫ k ( t ) 2 Tr ( Ω c ( t ) B c ( t )) d t ∫ [ m 1 ( p ) − m 1 ( p − ζ )] q ( p ) d p + o ( ϵ ) = − 2 ϵ Q ζ ( χ ) A L + o ( ϵ ) , Q ζ ( χ ) = ∫ m 1 ( p ) q ( p + ζ ) d p . (104)
Normalization gives ∫ m 1 q = 0 \int m_1q=0 ∫ m 1 q = 0 , and Tr ( Ω c B c ) = A L ( c ) \operatorname{Tr}(\Omega_cB_c)=A_L(c) Tr ( Ω c B c ) = A L ( c ) identifies Note J’s positive suffix average A ‾ L \overline{A}_L A L . C10 gives one finite shift with inf I Q ζ > 0 \inf_I Q_{\zeta}>0 inf I Q ζ > 0 , so the rigidity bound yields
lim inf ϵ ↓ 0 ϵ − 2 G λ ( T ϵ r ) ≥ 2 Q ζ ( χ ) 2 A ‾ L 2 ( 1 + C ζ ) { 4 L 2 / ( λ c D ) + 1 / ( 1 − λ ) } > 0 , (105) \liminf_{\epsilon\downarrow0}\epsilon^{-2}\mathcal{G}_{\lambda}(T_{\epsilon}^{r})\geq\frac{2Q_{\zeta}(\chi)^2\overline{A}_L^2}{(1+C_{\zeta})\{4L^2/(\lambda c_D)+1/(1-\lambda)\}}>0, \tag*{(105)} ϵ ↓ 0 lim inf ϵ − 2 G λ ( T ϵ r ) ≥ ( 1 + C ζ ) { 4 L 2 / ( λ c D ) + 1/ ( 1 − λ )} 2 Q ζ ( χ ) 2 A L 2 > 0 , (105)
lim sup ϵ ↓ 0 ϵ − 2 G λ ( T ϵ r ) ≤ ( 1 − λ ) C a m p < ∞ . (106) \limsup_{\epsilon\downarrow0}\epsilon^{-2}\mathcal{G}_{\lambda}(T_{\epsilon}^{r})\leq(1-\lambda)C_{\mathrm{amp}}<\infty. \tag*{(106)} ϵ ↓ 0 lim sup ϵ − 2 G λ ( T ϵ r ) ≤ ( 1 − λ ) C amp < ∞. (106)
Use ( 1 + C ) 2 ≤ 2 ( 1 + C ) (1+\sqrt{C})^2\leq2(1+C) ( 1 + C ) 2 ≤ 2 ( 1 + C ) for the lower bound and C5’s rule for the upper bound, retaining the finite mass and marginal infima.
Reduction of the retained packet
Keep the weights d u ν ( d z ) du\,\nu(dz) d u ν ( d z ) of C3’s middle branch and write its exact packet as
v ϵ ; u , z ( t ) = k ( t ) δ − 1 / 2 β ϵ , z ( ( t − u ) / δ ) ⊗ ∣ c ( t ) z ⟩ , β ϵ , z ( y ) = c χ B a ϵ z ( c χ y ) , c χ = 1 + χ . (107) v_{\epsilon;u,z}(t)=k(t)\delta^{-1/2}\beta_{\epsilon,z}((t-u)/\delta)\otimes|c(t)z\rangle,\qquad\beta_{\epsilon,z}(y)=\sqrt{c_{\chi}}B_{a_{\epsilon z}}(c_{\chi}y),\qquad c_{\chi}=\sqrt{1+\chi}. \tag*{(107)} v ϵ ; u , z ( t ) = k ( t ) δ − 1/2 β ϵ , z (( t − u ) / δ ) ⊗ ∣ c ( t ) z ⟩ , β ϵ , z ( y ) = c χ B a ϵz ( c χ y ) , c χ = 1 + χ . (107)
Extension to the full source line gives mass m = ∥ k ∥ 2 2 m=\lVert k\rVert_2^2 m = ∥ k ∥ 2 2 ; negative positions require delay d − u d-u d − u and have exponentially small mass. For C2’s frame U 0 U_0 U 0 , the squared commutator norm is
∫ d ξ d p ∥ E ^ ( ξ ) ∥ H S 2 m 1 ( p ) ∥ [ U 0 ( p + δ ξ ) − U 0 ( p ) ] ( 0 , w 0 ( p ) ) ∥ 2 . (108) \int d\xi\,dp\,\lVert\widehat{E}(\xi)\rVert_{\mathrm{HS}}^2m_1(p)\left\lVert[U_0(p+\delta\xi)-U_0(p)](0,w_0(p))\right\rVert^2. \tag*{(108)} ∫ d ξ d p ∥ E ( ξ ) ∥ HS 2 m 1 ( p ) ∥ [ U 0 ( p + δ ξ ) − U 0 ( p )] ( 0 , w 0 ( p )) ∥ 2 . (108)
Ordinary derivatives bound it by C U 2 δ 2 ∥ E ′ ∥ 2 2 C_U^2\delta^2\lVert E'\rVert_2^2 C U 2 δ 2 ∥ E ′ ∥ 2 2 . For E = D E=D E = D ,
∥ D ′ ∥ 2 2 = ∥ k ′ ∥ 2 2 + L 2 3 ∫ k 2 ∣ c ′ ∣ 2 = O ( d − 1 + log ( 1 / d ) ) . (109) \lVert D'\rVert_2^2=\lVert k'\rVert_2^2+\frac{L^2}{3}\int k^2|c'|^2=O\left(d^{-1}+\log(1/d)\right). \tag*{(109)} ∥ D ′ ∥ 2 2 = ∥ k ′ ∥ 2 2 + 3 L 2 ∫ k 2 ∣ c ′ ∣ 2 = O ( d − 1 + log ( 1/ d ) ) . (109)
For real normalized coherent states, the cross derivative vanishes. The same estimate holds for D Z DZ D Z and the window multipliers.
Removal of the orthogonal tangent
The packet has the weighted expansion
U 0 ( p ) ( 0 , β ^ ϵ , z ( p ) ) = f ( p ) [ e 0 + ϵ z { e 0 ( q ( p ) + i r ( p ) ) + a w 1 ( p ) } ] + O ( ϵ 2 ) , w 1 ( p ) ⊥ e 0 . (110) U_0(p)(0,\widehat{\beta}_{\epsilon,z}(p))=f(p)\left[e_0+\epsilon z\{e_0(q(p)+ir(p))+a w_1(p)\}\right]+O(\epsilon^2),\qquad w_1(p)\perp e_0. \tag*{(110)} U 0 ( p ) ( 0 , β ϵ , z ( p )) = f ( p ) [ e 0 + ϵz { e 0 ( q ( p ) + i r ( p )) + a w 1 ( p )} ] + O ( ϵ 2 ) , w 1 ( p ) ⊥ e 0 . (110)
Here w 1 w_1 w 1 and its required derivatives are bounded, and r = − ( a c χ / 2 ) ∂ p log m 1 r=-(ac\chi/2)\partial_p\log m_1 r = − ( a cχ /2 ) ∂ p log m 1 . Let e ( t ) = 1 − clip [ d , 1 − d ] ( t ) − w e(t)=1-\operatorname{clip}_{[d,1-d]}(t)-w e ( t ) = 1 − clip [ d , 1 − d ] ( t ) − w , D 1 ( p ) = ∣ w 1 ( p ) ⟩ ⟨ e 0 ∣ D_1(p)=|w_1(p)\rangle\langle e_0| D 1 ( p ) = ∣ w 1 ( p )⟩ ⟨ e 0 ∣ , and let b b b denote annihilation in the residual mode. Define the known skew-adjoint generator
G ϵ = D 1 ( P δ ) M e − 1 / 2 ⊗ b − M e − 1 / 2 D 1 ( P δ ) † ⊗ b † . (111) G_\epsilon=D_1(P_\delta)M_{e^{-1/2}}\otimes b-M_{e^{-1/2}}D_1(P_\delta)^\dagger\otimes b^\dagger. \tag*{(111)} G ϵ = D 1 ( P δ ) M e − 1/2 ⊗ b − M e − 1/2 D 1 ( P δ ) † ⊗ b † . (111)
The generator preserves b † b + Π ⊥ b^\dagger b+\Pi_\perp b † b + Π ⊥ , Π ⊥ = I − ∣ e 0 ⟩ ⟨ e 0 ∣ \Pi_\perp=I-|e_0\rangle\langle e_0| Π ⊥ = I − ∣ e 0 ⟩ ⟨ e 0 ∣ , and has sector norm at most C N / d − w C\sqrt{N}/\sqrt{d-w} C N / d − w . Sectorwise exponentiation defines a unitary. The suffix cancels the orthogonal tangent because
b ∣ c ( t ) z ⟩ = a e ( t ) z ∣ c ( t ) z ⟩ . (112) b|c(t)z\rangle=a\sqrt{e(t)}\,z|c(t)z\rangle. \tag*{(112)} b ∣ c ( t ) z ⟩ = a e ( t ) z ∣ c ( t ) z ⟩ . (112)
For T 0 o u t = 2 π δ e 0 ⊗ M D f ( P δ ) T_0^{\mathrm{out}}=\sqrt{2\pi\delta}\,e_0\otimes M_{D f}(P_\delta) T 0 out = 2 π δ e 0 ⊗ M D f ( P δ ) , commutator control and the full vector expansion give
U 0 ( P δ ) T ϵ p h = T 0 o u t + ϵ T 1 + R ϵ , ∥ R ϵ ∥ H S ≤ C ( ϵ 2 + δ E d ) , E d = d − 1 + log ( 1 / d ) . (113) U_0(P_\delta)T_\epsilon^{\mathrm{ph}}=T_0^{\mathrm{out}}+\epsilon T_1+R_\epsilon,\qquad\lVert R_\epsilon\rVert_{\mathrm{HS}}\le C(\epsilon^2+\delta\sqrt{E_d}),\qquad E_d=d^{-1}+\log(1/d). \tag*{(113)} U 0 ( P δ ) T ϵ ph = T 0 out + ϵ T 1 + R ϵ , ∥ R ϵ ∥ HS ≤ C ( ϵ 2 + δ E d ) , E d = d − 1 + log ( 1/ d ) . (113)
The normal first order part differs from G ϵ T 0 o u t G_\epsilon T_0^{\mathrm{out}} G ϵ T 0 out by at most C δ E d C\delta\sqrt{E_d} C δ E d , while the parallel part is z ( q + i r ) z(q+ir) z ( q + i r ) . To control the exponential remainder on both sources, use W δ = ⟨ P δ ⟩ W_\delta=\langle P_\delta\rangle W δ = ⟨ P δ ⟩ and
∥ G ϵ 2 T 0 o u t W ∥ H S + ∥ G ϵ T 1 W ∥ H S ≤ C ( d − w ) − 1 / 2 , W ∈ { I , W δ } . (114) \lVert G_\epsilon^2T_0^{\mathrm{out}}W\rVert_{\mathrm{HS}}+\lVert G_\epsilon T_1W\rVert_{\mathrm{HS}}\le C(d-w)^{-1/2},\qquad W\in\{I,W_\delta\}. \tag*{(114)} ∥ G ϵ 2 T 0 out W ∥ HS + ∥ G ϵ T 1 W ∥ HS ≤ C ( d − w ) − 1/2 , W ∈ { I , W δ } . (114)
Annihilation cancels the inverse strength on e 0 e_0 e 0 ; return through the normal block leaves one inverse square root, with ∥ b † ∣ c z ⟩ ∥ 2 = 1 + c 2 z 2 \lVert b^\dagger|c z\rangle\rVert^2=1+c^2z^2 ∥ b † ∣ cz ⟩ ∥ 2 = 1 + c 2 z 2 . For W = W δ W=W_\delta W = W δ , each scalar profile gains ⟨ p ⟩ \langle p\rangle ⟨ p ⟩ . C2’s p − 8 p^{-8} p − 8 tail and vector derivative bounds preserve square integrability. The commutator estimate with weight ( 1 + p 2 ) m 1 ( p ) (1+p^2)m_1(p) ( 1 + p 2 ) m 1 ( p ) therefore gives ∥ R ϵ W δ ∥ H S ≤ C ( ϵ 2 + δ E d ) \lVert R_\epsilon W_\delta\rVert_{\mathrm{HS}}\le C(\epsilon^2+\delta\sqrt{E_d}) ∥ R ϵ W δ ∥ HS ≤ C ( ϵ 2 + δ E d ) . On each conserved sector, Taylor’s integral formula gives
[ ( e − ϵ G ϵ − I + ϵ G ϵ ) T 0 o u t + ϵ ( e − ϵ G ϵ − I ) T 1 ] W δ = ∫ 0 ϵ e − t G ϵ [ ( ϵ − t ) G ϵ 2 T 0 o u t − ϵ G ϵ T 1 ] W δ d t . \left[(e^{-\epsilon G_\epsilon}-I+\epsilon G_\epsilon)T_0^{\mathrm{out}}+\epsilon(e^{-\epsilon G_\epsilon}-I)T_1\right]W_\delta
=\int_0^\epsilon e^{-tG_\epsilon}\left[(\epsilon-t)G_\epsilon^2T_0^{\mathrm{out}}-\epsilon G_\epsilon T_1\right]W_\delta\,dt. [ ( e − ϵ G ϵ − I + ϵ G ϵ ) T 0 out + ϵ ( e − ϵ G ϵ − I ) T 1 ] W δ = ∫ 0 ϵ e − t G ϵ [ ( ϵ − t ) G ϵ 2 T 0 out − ϵ G ϵ T 1 ] W δ d t .
Summing the sector Hilbert–Schmidt bounds and setting U ϵ = e − ϵ G ϵ U 0 ( P δ ) U_\epsilon=e^{-\epsilon G_\epsilon}U_0(P_\delta) U ϵ = e − ϵ G ϵ U 0 ( P δ ) gives
∥ ( U ϵ T ϵ p h − T ϵ p ) ⟨ P δ ⟩ ∥ H S ≤ C ( ϵ 2 + ϵ 2 d − w + δ E d ) = O ( ϵ 5 / 4 ) = o ( ϵ ) . (115) \left\lVert(U_\epsilon T_\epsilon^{\mathrm{ph}}-T_\epsilon^{\mathrm{p}})\langle P_\delta\rangle\right\rVert_{\mathrm{HS}}
\le C\left(\epsilon^2+\frac{\epsilon^2}{\sqrt{d-w}}+\delta\sqrt{E_d}\right)
=O(\epsilon^{5/4})=o(\epsilon). \tag*{(115)} ( U ϵ T ϵ ph − T ϵ p ) ⟨ P δ ⟩ HS ≤ C ( ϵ 2 + d − w ϵ 2 + δ E d ) = O ( ϵ 5/4 ) = o ( ϵ ) . (115)
Here the synthesis retaining the scalar phase is
T ϵ p = e 0 ⊗ T ϵ r R ϵ s r c , R ϵ s r c = exp ( i ϵ Z r ( P δ ) ) . (116) T_\epsilon^{\mathrm{p}}=e_0\otimes T_\epsilon^{\mathrm{r}}R_\epsilon^{\mathrm{src}},\qquad R_\epsilon^{\mathrm{src}}=\exp(i\epsilon Zr(P_\delta)). \tag*{(116)} T ϵ p = e 0 ⊗ T ϵ r R ϵ src , R ϵ src = exp ( i ϵ Z r ( P δ )) . (116)
Amplitude prediction and the gap return
The source phase commutes with Z Z Z , so
T ϵ p Z J ( T ϵ p ) † = e 0 e 0 † ⊗ T ϵ † Z J ( T ϵ † ) † . (117) T_{\epsilon}^{\mathrm{p}} ZJ(T_{\epsilon}^{\mathrm{p}})^{\dagger}=e_{0}e_{0}^{\dagger}\otimes T_{\epsilon}^{\dagger}ZJ(T_{\epsilon}^{\dagger})^{\dagger}. \tag*{(117)} T ϵ p Z J ( T ϵ p ) † = e 0 e 0 † ⊗ T ϵ † Z J ( T ϵ † ) † . (117)
Put T = U ϵ T p h T=U_{\epsilon}T_{\mathrm{ph}} T = U ϵ T ph , S = T ϵ p S=T_{\epsilon}^{\mathrm{p}} S = T ϵ p , Ω T = T T † \Omega_{T}=TT^{\dagger} Ω T = T T † , Ω S = S S † \Omega_{S}=SS^{\dagger} Ω S = S S † , and let A T A_{T} A T , A S A_{S} A S be their amplitude optima. Extend A δ A_{\delta} A δ boundedly and self-adjointly to the unused directions and conjugate its source lift by the source phase. The o ( ϵ ) o(\epsilon) o ( ϵ ) synthesis error and C1’s defect bound give
∥ A T − A δ ∥ Ω T + ∥ A S − A δ ∥ Ω S = o ( ϵ ) . (118) \lVert A_{T}-A_{\delta}\rVert_{\Omega_{T}}+\lVert A_{S}-A_{\delta}\rVert_{\Omega_{S}}=o(\epsilon). \tag*{(118)} ∥ A T − A δ ∥ Ω T + ∥ A S − A δ ∥ Ω S = o ( ϵ ) . (118)
Both syntheses are within weighted O ( ϵ ) O(\epsilon) O ( ϵ ) of T 0 o u t T_{0}^{\mathrm{out}} T 0 out . Its coordinate rule has zero centered location form and o ( ϵ ) o(\epsilon) o ( ϵ ) amplitude defect; root stability gives both excesses O ( ϵ 2 ) O(\epsilon^{2}) O ( ϵ 2 ) . Equations (C4.7)–(C4.8) and (C7.9) verify Lemma C4.1, yielding
G λ ( T p h ) − G λ ( T ϵ p ) = o ( ϵ 2 ) . (119) \mathcal{G}_{\lambda}(T_{\mathrm{ph}})-\mathcal{G}_{\lambda}(T_{\epsilon}^{\mathrm{p}})=o(\epsilon^{2}). \tag*{(119)} G λ ( T ph ) − G λ ( T ϵ p ) = o ( ϵ 2 ) . (119)
The same comparison transfers individual excesses. C9 removes the phase and combines this bound with C3’s endpoint return.
Location bounds for Hermitian multipliers
Let D ( t ) D(t) D ( t ) be bounded Hilbert–Schmidt, m = ∫ ∥ D ( t ) ∥ H S 2 d t < ∞ m=\int\lVert D(t)\rVert_{\mathrm{HS}}^{2}dt<\infty m = ∫ ∥ D ( t ) ∥ HS 2 d t < ∞ , g 0 = ∫ g < ∞ g_{0}=\int g<\infty g 0 = ∫ g < ∞ , and C ( p ) = C ( p ) † C(p)=C(p)^{\dagger} C ( p ) = C ( p ) † , ∥ C ∥ ≤ B \lVert C\rVert\leq B ∥ C ∥ ≤ B . Put Q ϵ = I + ϵ C Q_{\epsilon}=I+\epsilon C Q ϵ = I + ϵ C , T ϵ = 2 π δ M D f ( P δ ) Q ϵ ( P δ ) T_{\epsilon}=\sqrt{2\pi\delta}M_{D}f(P_{\delta})Q_{\epsilon}(P_{\delta}) T ϵ = 2 π δ M D f ( P δ ) Q ϵ ( P δ ) . Write m f = f 2 m_{f}=f^{2} m f = f 2 ; m m m is the base mass. Let r ω ( p ) r_{\omega}(p) r ω ( p ) be the row of J M D F δ † JM_{D}F_{\delta}^{\dagger} J M D F δ † times e − i p x ( ω ) / δ e^{-ipx(\omega)/\delta} e − i p x ( ω ) / δ , including multiplicities. Put Δ r ω ( p , q ) = r ω ( p ) − r ω ( q ) \Delta r_{\omega}(p,q)=r_{\omega}(p)-r_{\omega}(q) Δ r ω ( p , q ) = r ω ( p ) − r ω ( q ) and
D J = 2 π δ ∬ γ ( p , q ) ∫ ∥ Δ r ω ( p , q ) ∥ H S 2 d ω d p d q . (120) \mathcal{D}_{J}=2\pi\delta\iint\gamma(p,q)\int\lVert\Delta r_{\omega}(p,q)\rVert_{\mathrm{HS}}^{2}\,d\omega\,dp\,dq. \tag*{(120)} D J = 2 π δ ∬ γ ( p , q ) ∫ ∥ Δ r ω ( p , q ) ∥ HS 2 d ω d p d q . (120)
Completeness gives
2 π δ ∫ g ( p ) ∫ ∥ r ω ( p ) ∥ H S 2 d ω d p = m g 0 . (121) 2\pi\delta\int g(p)\int\lVert r_{\omega}(p)\rVert_{\mathrm{HS}}^{2}\,d\omega\,dp=mg_{0}. \tag*{(121)} 2 π δ ∫ g ( p ) ∫ ∥ r ω ( p ) ∥ HS 2 d ω d p = m g 0 . (121)
Lemma C8.1 (Fourth order reference error). For ∣ ϵ ∣ B ≤ 1 / 8 |\epsilon|B\leq1/8 ∣ ϵ ∣ B ≤ 1/8 , let r l o c , r e f r_{\mathrm{loc,ref}} r loc , ref be the coordinate readout location risk on T ϵ T_{\epsilon} T ϵ . Every complete paired rule J J J satisfies
∣ r l o c ( J ; T ϵ ) − r l o c , r e f − D J ∣ ≤ ( 2 ∣ ϵ ∣ B + ϵ 2 B 2 ) D J + 2 ϵ 2 B 2 m g 0 D J , (122) \left|r_{\mathrm{loc}}(J;T_{\epsilon})-r_{\mathrm{loc,ref}}-\mathcal{D}_{J}\right|
\leq(2|\epsilon|B+\epsilon^{2}B^{2})\mathcal{D}_{J}+2\epsilon^{2}B^{2}\sqrt{mg_{0}\mathcal{D}_{J}}, \tag*{(122)} ∣ r loc ( J ; T ϵ ) − r loc , ref − D J ∣ ≤ ( 2∣ ϵ ∣ B + ϵ 2 B 2 ) D J + 2 ϵ 2 B 2 m g 0 D J , (122) 0 ≤ r l o c , r e f − r l o c ⋆ ( T ϵ ) ≤ 2 B 4 m g 0 ϵ 4 , (123) 0\leq r_{\mathrm{loc,ref}}-r_{\mathrm{loc}}^{\star}(T_{\epsilon})\leq2B^{4}mg_{0}\epsilon^{4}, \tag*{(123)} 0 ≤ r loc , ref − r loc ⋆ ( T ϵ ) ≤ 2 B 4 m g 0 ϵ 4 , (123) D J ≤ 4 e l o c ( J ; T ϵ ) + 16 B 4 m g 0 ϵ 4 . (124) \mathcal{D}_{J}\leq4e_{\mathrm{loc}}(J;T_{\epsilon})+16B^{4}mg_{0}\epsilon^{4}. \tag*{(124)} D J ≤ 4 e loc ( J ; T ϵ ) + 16 B 4 m g 0 ϵ 4 . (124) For direct sums of labeled branches with a common B B B , these bounds use the sum of the base masses in place of m m m .
Proof. Let S ϵ = { Q ϵ ( p ) , Q ϵ ( q ) } / 2 S_{\epsilon}=\{Q_{\epsilon}(p),Q_{\epsilon}(q)\}/2 S ϵ = { Q ϵ ( p ) , Q ϵ ( q )} /2 , K ϵ = [ Q ϵ ( p ) , Q ϵ ( q ) ] / 2 K_{\epsilon}=[Q_{\epsilon}(p),Q_{\epsilon}(q)]/2 K ϵ = [ Q ϵ ( p ) , Q ϵ ( q )] /2 . Then S ϵ ≥ I / 2 S_{\epsilon}\geq I/2 S ϵ ≥ I /2 , ∥ S ϵ − I ∥ ≤ 2 ∣ ϵ ∣ B + ϵ 2 B 2 \lVert S_{\epsilon}-I\rVert\leq2|\epsilon|B+\epsilon^{2}B^{2} ∥ S ϵ − I ∥ ≤ 2∣ ϵ ∣ B + ϵ 2 B 2 and ∥ K ϵ ∥ ≤ ϵ 2 B 2 \lVert K_{\epsilon}\rVert\leq\epsilon^{2}B^{2} ∥ K ϵ ∥ ≤ ϵ 2 B 2 . Expanding the triangular Fourier benefit and symmetrizing gives
r l o c ( J ; T ϵ ) − r l o c , r e f = 2 π δ ∫ γ ( p , q ) ∫ ( Δ r ω ( p , q ) ) S ϵ ( p , q ) ( Δ r ω ( p , q ) ) † d ω d p d q − 4 π δ Re ∫ γ ( p , q ) ∫ ( Δ r ω ( p , q ) ) K ϵ ( p , q ) r ω ( q ) † d ω d p d q . (125) \begin{aligned}
r_{\mathrm{loc}}(J;T_{\epsilon})-r_{\mathrm{loc,ref}}
={}&2\pi\delta\int\gamma(p,q)\int(\Delta r_{\omega}(p,q))S_{\epsilon}(p,q)(\Delta r_{\omega}(p,q))^{\dagger}\,d\omega\,dp\,dq\\
&-4\pi\delta\operatorname{Re}\int\gamma(p,q)\int(\Delta r_{\omega}(p,q))K_{\epsilon}(p,q)r_{\omega}(q)^{\dagger}\,d\omega\,dp\,dq. \tag*{(125)}
\end{aligned} r loc ( J ; T ϵ ) − r loc , ref = 2 π δ ∫ γ ( p , q ) ∫ ( Δ r ω ( p , q )) S ϵ ( p , q ) ( Δ r ω ( p , q ) ) † d ω d p d q − 4 π δ Re ∫ γ ( p , q ) ∫ ( Δ r ω ( p , q )) K ϵ ( p , q ) r ω ( q ) † d ω d p d q . (125)
The coordinate reference trace is 4 π δ M D [ m f ( p ) Q ϵ 2 − ∫ γ S ϵ d q ] ( P δ ) M D † 4\pi\delta M_D[mf(p)Q_\epsilon^2-\int\gamma S\epsilon\,dq](P_\delta)M_D^\dagger 4 π δ M D [ m f ( p ) Q ϵ 2 − ∫ γ S ϵ d q ] ( P δ ) M D † . Subtract r ( q ) K ϵ r ( q ) † r(q)K_\epsilon r(q)^\dagger r ( q ) K ϵ r ( q ) † , whose real part vanishes, then apply Cauchy–Schwarz and (C8.2).
In particular, r l o c ( J ) − r l o c , r e f ≥ D J / 2 − 2 ϵ 2 B 2 m g 0 D J ≥ − 2 B 4 m g 0 ϵ 4 r_{\mathrm{loc}}(J)-r_{\mathrm{loc,ref}}\geq\mathcal{D}_J/2-2\epsilon^2B^2\sqrt{m g_0\mathcal{D}_J}\geq-2B^4mg_0\epsilon^4 r loc ( J ) − r loc , ref ≥ D J /2 − 2 ϵ 2 B 2 m g 0 D J ≥ − 2 B 4 m g 0 ϵ 4 . Take near minimizers for the second bound. Young’s inequality gives e l o c ( J ) ≥ D J / 4 − 4 B 4 m g 0 ϵ 4 e_{\mathrm{loc}}(J)\geq\mathcal{D}_J/4-4B^4mg_0\epsilon^4 e loc ( J ) ≥ D J /4 − 4 B 4 m g 0 ϵ 4 . Summing branches before Cauchy–Schwarz gives the direct sum version.
For a positive diagonal multiplier F ϵ ( Z , p ) = I + O ( ϵ ) F_\epsilon(Z,p)=I+O(\epsilon) F ϵ ( Z , p ) = I + O ( ϵ ) , the graph kernel is positive at each latent amplitude. Coordinate readout is therefore optimal, with
e l o c r e a l ( J ) = ( 1 + O ( ϵ ) ) D J . (126) e_{\mathrm{loc}}^{\mathrm{real}}(J)=(1+O(\epsilon))\mathcal{D}_J. \tag*{(126)} e loc real ( J ) = ( 1 + O ( ϵ )) D J . (126)
On either model’s near optimal set, e l o c e_{\mathrm{loc}} e loc and D J \mathcal{D}_J D J are O ( ϵ 2 ) O(\epsilon^2) O ( ϵ 2 ) , so
e l o c H e r m i t i a n ( J ) − e l o c r e a l ( J ) = o ( ϵ 2 ) (127) e_{\mathrm{loc}}^{\mathrm{Hermitian}}(J)-e_{\mathrm{loc}}^{\mathrm{real}}(J)=o(\epsilon^2) \tag*{(127)} e loc Hermitian ( J ) − e loc real ( J ) = o ( ϵ 2 ) (127)
for every rule in these sets.
Removal of the scalar phase
The phase can be removed at quadratic precision:
G λ ( T ϵ p ) − G λ ( T ϵ r ) = o ( ϵ 2 ) . (128) \mathcal{G}_\lambda(T_\epsilon^{\mathrm{p}})-\mathcal{G}_\lambda(T_\epsilon^{\mathrm{r}})=o(\epsilon^2). \tag*{(128)} G λ ( T ϵ p ) − G λ ( T ϵ r ) = o ( ϵ 2 ) . (128)
Take a real sine and cosine partition ∑ j f j 2 = 1 \sum_j f_j^2=1 ∑ j f j 2 = 1 of support length at most 2 b 2b 2 b , b = ϵ 1 / 2 b=\epsilon^{1/2} b = ϵ 1/2 , with ∑ j ∣ f j ′ ∣ 2 ≤ C / b 2 \sum_j\lvert f'_j\rvert^2\leq C/b^2 ∑ j ∣ f j ′ ∣ 2 ≤ C / b 2 . Apply the corresponding complete instrument to the output coordinate, retaining the labels. The disturbance is bounded using the instrument identity from C3 and uniform second and fourth packet moments:
O ( m δ 2 / b 2 ) = O ( ϵ 3 ) (129) O(m\delta^2/b^2)=O(\epsilon^3) \tag*{(129)} O ( m δ 2 / b 2 ) = O ( ϵ 3 ) (129)
for both models. Applying or discarding the instrument compares the three risks. The direct sum fields D j = k f j V c D_j=kf_jV_c D j = k f j V c retain their weights and obey
∑ j ∥ D j ∥ L 2 ( H S ) 2 = m , ∑ j ∥ D j ′ ∥ L 2 ( H S ) 2 ≤ C ( d − 1 + b − 2 + log ( 1 / d ) ) . (130) \sum_j\lVert D_j\rVert_{L^2(\mathrm{HS})}^2=m,\qquad\sum_j\lVert D'_j\rVert_{L^2(\mathrm{HS})}^2\leq C(d^{-1}+b^{-2}+\log(1/d)). \tag*{(130)} j ∑ ∥ D j ∥ L 2 ( HS ) 2 = m , j ∑ ∥ D j ′ ∥ L 2 ( HS ) 2 ≤ C ( d − 1 + b − 2 + log ( 1/ d )) . (130)
The cross terms vanish since ∑ j f j f j ′ = 0 \sum_j f_jf'_j=0 ∑ j f j f j ′ = 0 . Reorder the summed field before applying Cauchy–Schwarz.
Output gauge and source lift
Use the known bounded output gauge
H δ = 1 2 { r ( P δ ) , M A c } , U ϵ = e − i ϵ H δ . (131) H_\delta=\tfrac{1}{2}\{r(P_\delta),MA_c\},\qquad U_\epsilon=e^{-i\epsilon H_\delta}. \tag*{(131)} H δ = 2 1 { r ( P δ ) , M A c } , U ϵ = e − i ϵ H δ . (131)
The Taylor expansion, A c V c = V c Z + V c K c A_cV_c=V_cZ+V_cK_c A c V c = V c Z + V c K c and Fourier reordering give
U ϵ T ϵ , j p = T 0 , j + ϵ 2 π δ M D j Z q ( P δ ) f ( P δ ) − i ϵ 2 π δ M D j K c r ( P δ ) f ( P δ ) + o H S , ⟨ P δ ⟩ ( ϵ ) . (132) U_\epsilon T_{\epsilon,j}^{\mathrm{p}}=T_{0,j}+\epsilon\sqrt{2\pi\delta}\,M_{D_j}Zq(P_\delta)f(P_\delta)
-i\epsilon\sqrt{2\pi\delta}\,M_{D_j}K_cr(P_\delta)f(P_\delta)+o_{\mathrm{HS},\langle P_\delta\rangle}(\epsilon). \tag*{(132)} U ϵ T ϵ , j p = T 0 , j + ϵ 2 π δ M D j Z q ( P δ ) f ( P δ ) − i ϵ 2 π δ M D j K c r ( P δ ) f ( P δ ) + o HS , ⟨ P δ ⟩ ( ϵ ) . (132)
The identity i r D j Z − i r A c D j = − i r D j K c irD_jZ-irA_cD_j=-irD_jK_c i r D j Z − i r A c D j = − i r D j K c cancels the parallel phase. The summed derivatives and tight Fourier measures bound the remainder by weighted o ( ϵ ) o(\epsilon) o ( ϵ ) , including C4’s factor ⟨ P δ ⟩ \langle P_\delta\rangle ⟨ P δ ⟩ . The exponential remainder is O ( ϵ 2 ) O(\epsilon^2) O ( ϵ 2 ) . For one known point t j t_j t j in each support, K j = K c ( t j ) K_j=K_c(t_j) K j = K c ( t j ) satisfies
∑ j ∫ ∥ D j ( t ) ( K c ( t ) − K j ) ∥ H S 2 d t ≤ C b m . (133) \sum_j\int\left\|D_j(t)(K_c(t)-K_j)\right\|_{\mathrm{HS}}^2\,\mathrm{d}t \le Cb m. \tag*{(133)} j ∑ ∫ ∥ D j ( t ) ( K c ( t ) − K j ) ∥ HS 2 d t ≤ C bm . (133)
C1’s weighted lift bound and the square root modulus of c c c give this estimate, so freezing costs O ( ϵ b ) = O ( ϵ 5 / 4 ) O(\epsilon\sqrt{b})=O(\epsilon^{5/4}) O ( ϵ b ) = O ( ϵ 5/4 ) . Set
C j ( p ) = q ( p ) Z − i r ( p ) K j = C j ( p ) † , T ^ j = T 0 , j [ I + ϵ C j ( P δ ) ] . (134) C_j(p)=q(p)Z-ir(p)K_j=C_j(p)^\dagger,\qquad\widehat{T}_j=T_{0,j}[I+\epsilon C_j(P_\delta)]. \tag*{(134)} C j ( p ) = q ( p ) Z − i r ( p ) K j = C j ( p ) † , T j = T 0 , j [ I + ϵ C j ( P δ )] . (134)
These multipliers have the common bound B = L ∥ q ∥ ∞ + κ s r c ∥ r ∥ ∞ B=L\|q\|_\infty+\kappa_{\mathrm{src}}\|r\|_\infty B = L ∥ q ∥ ∞ + κ src ∥ r ∥ ∞ . Suffixes, observables and masses still vary with the record coordinate.
Comparing amplitude and location excesses
The gauge conjugates the amplitude optimum. To verify C4’s predictor criterion for A δ A_\delta A δ , it remains to bound
[ H δ , M A c ] = 1 2 [ r ( P δ ) , M A c 2 ] . (135) [H_\delta,M_{A_c}]=\frac{1}{2}[r(P_\delta),M_{A_c^2}]. \tag*{(135)} [ H δ , M A c ] = 2 1 [ r ( P δ ) , M A c 2 ] . (135)
Expanding the difference of squares, with A c V c = V c ( Z + K c ) A_cV_c=V_c(Z+K_c) A c V c = V c ( Z + K c ) , gives
∥ ( A c ′ 2 − A c 2 ) V c ∥ H S ≤ 2 L ∥ ( A c ′ − A c ) V c ∥ H S + C ∥ V c ′ − V c ∥ H S ≤ C ′ ∣ c − c ′ ∣ . (136) \left\|(A_{c'}^2-A_c^2)V_c\right\|_{\mathrm{HS}}\le2L\left\|(A_{c'}-A_c)V_c\right\|_{\mathrm{HS}}+C\left\|V_{c'}-V_c\right\|_{\mathrm{HS}}\le C'|c-c'|. \tag*{(136)} ( A c ′ 2 − A c 2 ) V c HS ≤ 2 L ∥ ( A c ′ − A c ) V c ∥ HS + C ∥ V c ′ − V c ∥ HS ≤ C ′ ∣ c − c ′ ∣. (136)
Fourier reordering makes this commutator o ( 1 ) o(1) o ( 1 ) on the syntheses, hence o ( ϵ ) o(\epsilon) o ( ϵ ) with its prefactor. The first order predictor adds O ( ϵ 2 ) O(\epsilon^2) O ( ϵ 2 ) . Transfer the defect through freezing and apply C1. The coordinate optimal reference and C7’s quadratic rule verify the remaining hypotheses of Lemma C4.1, giving
G λ ( T p , l a b ) − G λ ( T ^ ) = o ( ϵ 2 ) . (137) \mathcal{G}_\lambda(T^{p,\mathrm{lab}})-\mathcal{G}_\lambda(\widehat{T})=o(\epsilon^2). \tag*{(137)} G λ ( T p , lab ) − G λ ( T ) = o ( ϵ 2 ) . (137)
The real labelled synthesis is T 0 , j Φ ϵ ( Z , P δ ) T_{0,j}\Phi_\epsilon(Z,P_\delta) T 0 , j Φ ϵ ( Z , P δ ) , using C5’s positive source multiplier. Let T 0 = ⨁ j T 0 , j T_0=\bigoplus_jT_{0,j} T 0 = ⨁ j T 0 , j , Q = ⨁ j ( I + ϵ C j ) Q=\bigoplus_j(I+\epsilon C_j) Q = ⨁ j ( I + ϵ C j ) and Φ = ⨁ j Φ ϵ ( Z , P δ ) \Phi=\bigoplus_j\Phi_\epsilon(Z,P_\delta) Φ = ⨁ j Φ ϵ ( Z , P δ ) . Both right factors have uniformly bounded inverses and ∥ Q − Φ ∥ = O ( ϵ ) \|Q-\Phi\|=O(\epsilon) ∥ Q − Φ∥ = O ( ϵ ) . For a rule J J J within o ( ϵ 2 ) o(\epsilon^2) o ( ϵ 2 ) of either gap optimum, define
E J = ( Y J − J A δ ) T 0 . E_J=(YJ-JA_\delta)T_0. E J = ( Y J − J A δ ) T 0 .
The amplitude root excess is O ( ϵ ) O(\epsilon) O ( ϵ ) . Replacing its optimum by A δ A_\delta A δ costs o ( ϵ ) o(\epsilon) o ( ϵ ) ; removing the invertible right factor therefore gives ∥ E J ∥ H S = O ( ϵ ) \|E_J\|_{\mathrm{HS}}=O(\epsilon) ∥ E J ∥ HS = O ( ϵ ) and
∥ E J Q − E J Φ ∥ H S ≤ ∥ E J ∥ H S ∥ Q − Φ ∥ = O ( ϵ 2 ) . \|E_JQ-E_J\Phi\|_{\mathrm{HS}}\le\|E_J\|_{\mathrm{HS}}\|Q-\Phi\|=O(\epsilon^2). ∥ E J Q − E J Φ ∥ HS ≤ ∥ E J ∥ HS ∥ Q − Φ∥ = O ( ϵ 2 ) .
Restoring each model’s own amplitude optimum yields
∣ e a m p ( J ; T ^ ) − e a m p ( J ; T r , l a b ) ∣ = o ( ϵ ) . \left|\sqrt{e_{\mathrm{amp}}(J;\widehat{T})}-\sqrt{e_{\mathrm{amp}}(J;T^{r,\mathrm{lab}})}\right|=o(\epsilon). e amp ( J ; T ) − e amp ( J ; T r , lab ) = o ( ϵ ) .
Both roots are O ( ϵ ) O(\epsilon) O ( ϵ ) , so the squared excesses differ by o ( ϵ 2 ) o(\epsilon^2) o ( ϵ 2 ) . C8 gives the same precision for location, since the summed Dirichlet form is O ( ϵ 2 ) O(\epsilon^2) O ( ϵ 2 ) . Comparing near optimizers in both directions yields
G λ ( T ^ ) − G λ ( T r , l a b ) = o ( ϵ 2 ) . (138) \mathcal{G}_\lambda(\widehat{T})-\mathcal{G}_\lambda(T^{r,\mathrm{lab}})=o(\epsilon^2). \tag*{(138)} G λ ( T ) − G λ ( T r , lab ) = o ( ϵ 2 ) . (138)
Combine the O ( ϵ 3 ) O(\epsilon^{3}) O ( ϵ 3 ) instrument disturbance with C3’s endpoint and posterior bounds and C7’s packet comparison to obtain
G λ ( E o r i g i n a l ) = G λ ( T p h ) + o ( ϵ 2 ) = G λ ( T ϵ P ) + o ( ϵ 2 ) = G λ ( T ϵ E ) + o ( ϵ 2 ) . (139) \mathcal{G}_{\lambda}(E_{\mathrm{original}})=\mathcal{G}_{\lambda}(T_{\mathrm{ph}})+o(\epsilon^{2})
=\mathcal{G}_{\lambda}(T_{\epsilon}^{P})+o(\epsilon^{2})
=\mathcal{G}_{\lambda}(T_{\epsilon}^{E})+o(\epsilon^{2}). \tag*{(139)} G λ ( E original ) = G λ ( T ph ) + o ( ϵ 2 ) = G λ ( T ϵ P ) + o ( ϵ 2 ) = G λ ( T ϵ E ) + o ( ϵ 2 ) . (139)
Pulling C5’s rule through the gauges, labels and endpoint branches gives excesses over the original marginal optima:
e l o c = o ( ϵ 2 ) , e a m p = O ( ϵ 2 ) . (140) e_{\mathrm{loc}}=o(\epsilon^{2}),\qquad e_{\mathrm{amp}}=O(\epsilon^{2}). \tag*{(140)} e loc = o ( ϵ 2 ) , e amp = O ( ϵ 2 ) . (140)
C2 gives Q = sup χ , p ∣ q χ ( p ) ∣ < ∞ Q=\sup_{\chi,p}|q_{\chi}(p)|<\infty Q = sup χ , p ∣ q χ ( p ) ∣ < ∞ and R = sup χ , p ∣ r χ ( p ) ∣ < ∞ R=\sup_{\chi,p}|r_{\chi}(p)|<\infty R = sup χ , p ∣ r χ ( p ) ∣ < ∞ under the crossover assumptions.
For R 0 = π / K R_{0}=\pi/K R 0 = π / K , P K = 2 π / K P_{K}=2\pi/K P K = 2 π / K , use
c 0 = c 1 / c 2 2 π K ( 1 + R 0 2 ) ( 1 + R 0 ) 4 , G 0 = K c 2 3 π , c D = c 0 2 P K 4 G 0 > 0. (141) c_{0}=\frac{\sqrt{c_{1}/c_{2}}}{2\pi K(1+R_{0}^{2})(1+R_{0})^{4}},\qquad G_{0}=\frac{Kc_{2}}{3\pi},\qquad c_{D}=\frac{c_{0}^{2}P_{K}}{4G_{0}}>0. \tag*{(141)} c 0 = 2 π K ( 1 + R 0 2 ) ( 1 + R 0 ) 4 c 1 / c 2 , G 0 = 3 π K c 2 , c D = 4 G 0 c 0 2 P K > 0. (141)
For ∣ r ∣ ≤ R 0 |r|\le R_{0} ∣ r ∣ ≤ R 0 , γ ( p , r ) ≥ c 0 m ( p ) ( 1 − cos K ( p − r ) ) \gamma(p,r)\ge c_{0}m(p)(1-\cos K(p-r)) γ ( p , r ) ≥ c 0 m ( p ) ( 1 − cos K ( p − r )) , g ( r ) ≤ G 0 g(r)\le G_{0} g ( r ) ≤ G 0 . The kernel ∫ − R 0 R 0 γ ( p , r ) γ ( r , q ) / g ( r ) d r \int_{-R_{0}}^{R_{0}}\gamma(p,r)\gamma(r,q)/g(r)\,\mathrm{d}r ∫ − R 0 R 0 γ ( p , r ) γ ( r , q ) / g ( r ) d r is at least c 0 2 P K m ( p ) m ( q ) / ( 2 G 0 ) c_{0}^{2}P_{K}m(p)m(q)/(2G_{0}) c 0 2 P K m ( p ) m ( q ) / ( 2 G 0 ) : the period integral of the two cosine factors is at least P K / 2 P_{K}/2 P K /2 . Its Dirichlet form is at most four times the original, whereas the m ⊗ m m\otimes m m ⊗ m form is twice the variance. This proves C6’s Hilbert valued coercivity with c D c_{D} c D . The tail comparison also gives
m χ , τ ( p − ζ ) ≤ C ζ m χ , τ ( p ) , C ζ = ( c 2 / c 1 ) ( 1 + ∣ ζ ∣ ) 8 . (142) m_{\chi,\tau}(p-\zeta)\le C_{\zeta}m_{\chi,\tau}(p),\qquad C_{\zeta}=(c_{2}/c_{1})(1+|\zeta|)^{8}. \tag*{(142)} m χ , τ ( p − ζ ) ≤ C ζ m χ , τ ( p ) , C ζ = ( c 2 / c 1 ) ( 1 + ∣ ζ ∣ ) 8 . (142)
It also gives the source domain bound g / m ≤ C I ( 1 + p 2 ) g/m\le C_{I}(1+p^{2}) g / m ≤ C I ( 1 + p 2 ) . Differentiating the spectral tail yields
q χ ( p ) = a { 1 + D χ e − D χ cos ( d χ p ) 1 − e − D χ cos ( d χ p ) } + v χ ( p ) , v χ ( p ) = O I ( ∣ p ∣ − 1 ) . (143) q_{\chi}(p)=a\left\{1+\frac{D_{\chi}e^{-D_{\chi}}\cos(d_{\chi}p)}{1-e^{-D_{\chi}}\cos(d_{\chi}p)}\right\}+v_{\chi}(p),\qquad v_{\chi}(p)=O_{I}(|p|^{-1}). \tag*{(143)} q χ ( p ) = a { 1 + 1 − e − D χ cos ( d χ p ) D χ e − D χ cos ( d χ p ) } + v χ ( p ) , v χ ( p ) = O I ( ∣ p ∣ − 1 ) . (143)
Since 181 / 924 ≤ D χ ≤ 181 / 231 181/924\le D_{\chi}\le181/231 181/924 ≤ D χ ≤ 181/231 , its periodic part is at least a [ 1 − D χ / ( e D χ + 1 ) ] ≥ 2 a / ( 2 + D χ ) ≥ q = 462 a / 643 > 0 a[1-D_{\chi}/(e^{D_{\chi}}+1)]\ge2a/(2+D_{\chi})\ge q=462a/643>0 a [ 1 − D χ / ( e D χ + 1 )] ≥ 2 a / ( 2 + D χ ) ≥ q = 462 a /643 > 0 . Choose a common P 0 P_{0} P 0 with ∣ v χ ( p ) ∣ ≤ q / 2 |v_{\chi}(p)|\le q/2 ∣ v χ ( p ) ∣ ≤ q /2 for ∣ p ∣ ≥ P 0 |p|\ge P_{0} ∣ p ∣ ≥ P 0 . The common polynomial tail permits a known finite ζ > P 0 \zeta>P_{0} ζ > P 0 such that
2 P 0 c 2 ( 1 + ζ − P 0 ) − 8 ≤ q 4 ( q / 2 + Q ) . (144) 2P_{0}c_{2}(1+\zeta-P_{0})^{-8}\le\frac{q}{4(q/2+Q)}. \tag*{(144)} 2 P 0 c 2 ( 1 + ζ − P 0 ) − 8 ≤ 4 ( q /2 + Q ) q . (144)
This bounds the mass of ∣ p + ζ ∣ < P 0 |p+\zeta|<P_{0} ∣ p + ζ ∣ < P 0 . The score is at least q / 2 q/2 q /2 outside this region and at least − Q -Q − Q everywhere, so
Q ζ ( χ ) = ∫ m χ ( p ) q χ ( p + ζ ) d p ≥ q / 4 = : Q ∗ > 0 ( χ ∈ I ) . (145) Q_{\zeta}(\chi)=\int m_{\chi}(p)q_{\chi}(p+\zeta)\,\mathrm{d}p\ge q/4=:Q_{*}>0\qquad(\chi\in I). \tag*{(145)} Q ζ ( χ ) = ∫ m χ ( p ) q χ ( p + ζ ) d p ≥ q /4 =: Q ∗ > 0 ( χ ∈ I ) . (145)
The phase score splits as
r χ ( p ) = − a c χ e − D χ d χ sin ( d χ p ) 2 { 1 − e − D χ cos ( d χ p ) } + w χ ( p ) . (146) r_{\chi}(p)=-\frac{ac_{\chi}e^{-D_{\chi}}d_{\chi}\sin(d_{\chi}p)}{2\{1-e^{-D_{\chi}}\cos(d_{\chi}p)\}}+w_{\chi}(p). \tag*{(146)} r χ ( p ) = − 2 { 1 − e − D χ cos ( d χ p )} a c χ e − D χ d χ sin ( d χ p ) + w χ ( p ) . (146)
The differentiated tails and positivity on compact sets give common H 1 H^1 H 1 bounds for both remainders. Expanding ( 1 − E cos ) − 1 (1-E\cos)^{-1} ( 1 − E cos ) − 1 , E ≤ e − 181 / 924 < 1 E \le e^{-181/924} < 1 E ≤ e − 181/924 < 1 , gives periodic Fourier measures with bounded variation and exponential tails. For an H 1 H^1 H 1 remainder, Cauchy–Schwarz bounds its inverse transform L 1 L^1 L 1 norm by its H 1 H^1 H 1 norm and bounds its tail by C Y − 1 / 2 ∥ v ′ ∥ 2 CY^{-1/2}\lVert v'\rVert_2 C Y − 1/2 ∥ v ′ ∥ 2 . Hence
sup χ ∈ I ∥ μ χ ∥ T V < ∞ , sup χ ∈ I ∣ μ χ ∣ ( { ∣ y ∣ > Y } ) ⟶ 0. (147) \sup_{\chi\in I}\lVert\mu_\chi\rVert_{\mathrm{TV}}<\infty,\qquad\sup_{\chi\in I}|\mu_\chi|(\{|y|>Y\})\longrightarrow0. \tag*{(147)} χ ∈ I sup ∥ μ χ ∥ TV < ∞ , χ ∈ I sup ∣ μ χ ∣ ({ ∣ y ∣ > Y }) ⟶ 0. (147)
The shifted score and moving atomic frequencies have the same bounds. Remove an endpoint strip, choose a common Fourier cutoff, then let ϵ → 0 \epsilon\to0 ϵ → 0 . C1’s weighted and strong continuity gives common o ( 1 ) o(1) o ( 1 ) moduli in this order.
After rescaling the window to [ 0 , 1 ] [0,1] [ 0 , 1 ] , C2’s sector majorants and C3’s waiting bounds are uniform on I I I . The labelled fields satisfy δ ( ∑ j ∥ D j ′ ′ ∥ 2 ) 1 / 2 = O I ( ϵ 5 / 4 ) \delta\left(\sum_j\lVert D_j''\rVert^2\right)^{1/2}=O_I(\epsilon^{5/4}) δ ( ∑ j ∥ D j ′′ ∥ 2 ) 1/2 = O I ( ϵ 5/4 ) and eventually d − w ≥ d / 2 d-w\ge d/2 d − w ≥ d /2 , giving alignment and freezing errors O I ( ϵ 5 / 4 ) O_I(\epsilon^{5/4}) O I ( ϵ 5/4 ) . The phase and Hermitian generators are bounded by L R LR L R and L Q + κ s r c R LQ+\kappa_{\mathrm{src}}R L Q + κ src R . With the common 2 + Exp ( 1 / 4 ) 2+\operatorname{Exp}(1/4) 2 + Exp ( 1/4 ) moment envelope, Sections C3, C5 and C9 give uniform o I ( ϵ 2 ) o_I(\epsilon^2) o I ( ϵ 2 ) returns for endpoints, posterior weights, source tails and soft labels. Hence
sup χ ∈ I ∣ G λ ( E o r i g i n a l ) − G λ ( T ϵ r ) ∣ = o ( ϵ 2 ) . (148) \sup_{\chi\in I}\left|\mathcal{G}_\lambda(E_{\mathrm{original}})-\mathcal{G}_\lambda(T_\epsilon^{r})\right|=o(\epsilon^2). \tag*{(148)} χ ∈ I sup ∣ G λ ( E original ) − G λ ( T ϵ r ) ∣ = o ( ϵ 2 ) . (148)
The response expansion and dominated residual integrals are uniform. Define
a − = 2 Q ∗ 2 A ‾ L 2 ( 1 + C ζ ) { 4 L 2 / ( λ c D ) + 1 / ( 1 − λ ) } , a + = ( 1 − λ ) C a m p . (149) a_-=\frac{2Q_*^2\overline{A}_L^2}{(1+C_\zeta)\{4L^2/(\lambda c_D)+1/(1-\lambda)\}},\qquad a_+=(1-\lambda)C_{\mathrm{amp}}. \tag*{(149)} a − = ( 1 + C ζ ) { 4 L 2 / ( λ c D ) + 1/ ( 1 − λ )} 2 Q ∗ 2 A L 2 , a + = ( 1 − λ ) C amp . (149)
C5/C6 and the uniform return give a − − r ( s ) ≤ s 2 G λ ( s , χ / s 2 ) ≤ a + + r ( s ) a_- - r(s)\le s^2\mathcal{G}_\lambda(s,\chi/s^2)\le a_+ + r(s) a − − r ( s ) ≤ s 2 G λ ( s , χ / s 2 ) ≤ a + + r ( s ) , where the common r ( s ) → 0 r(s)\to0 r ( s ) → 0 bounds predictor, response, centered and residual integral moduli together with ϵ 1 / 4 + ϵ \epsilon^{1/4}+\epsilon ϵ 1/4 + ϵ . Choose s 0 s_0 s 0 with r ≤ a − / 2 r\le a_-/2 r ≤ a − /2 ; then c − = a − / 2 c_-=a_-/2 c − = a − /2 , c + = a + + a − / 2 c_+=a_++a_-/2 c + = a + + a − /2 prove Theorem 2(ii). These comparisons transfer C5’s rule and both excesses uniformly on I I I . Continuity and compactness separate the leading risks from zero and the prior risks. C11 transfers the bounds to finite records.
Transfer to finite records
Let ϕ \phi ϕ be a known normalized C 3 C^3 C 3 curve in fixed finite dimension, q 0 = ∥ ( I − ρ ( 0 ) ) ϕ ′ ( 0 ) ∥ 2 > 0 q_0=\lVert(I-\rho(0))\phi'(0)\rVert^2>0 q 0 = ∥( I − ρ ( 0 )) ϕ ′ ( 0 ) ∥ 2 > 0 . For a known monotone profile and fixed L > 0 L>0 L > 0 , use
θ i = s n + a z n q 0 F ( ( i / n − u ) / w n ) , u ∼ Unif [ 0 , 1 ] , z ∼ Unif [ − L , L ] , (150) \theta_i=\frac{s_n+az}{\sqrt{nq_0}}F((i/n-u)/w_n),\qquad u\sim\operatorname{Unif}[0,1],\qquad z\sim\operatorname{Unif}[-L,L], \tag*{(150)} θ i = n q 0 s n + a z F (( i / n − u ) / w n ) , u ∼ Unif [ 0 , 1 ] , z ∼ Unif [ − L , L ] , (150)
Put H n = s n + a L H_n=s_n+aL H n = s n + a L , b n = 1 − ∣ ⟨ ϕ ( 0 ) , ϕ ( s n / n q 0 ) ⟩ ∣ 2 b_n=1-|\langle\phi(0),\phi(s_n/\sqrt{nq_0})\rangle|^2 b n = 1 − ∣ ⟨ ϕ ( 0 ) , ϕ ( s n / n q 0 )⟩ ∣ 2 , τ n = n b n \tau_n=nb_n τ n = n b n , k w ( t ) = t / 1 + w t k_w(t)=t/\sqrt{1+wt} k w ( t ) = t / 1 + w t . The physical and coherent position coefficients are k w n ( τ n ) / K k_{w_n}(\tau_n)/K k w n ( τ n ) / K , k w n ( s n ) / K k_{w_n}(s_n)/K k w n ( s n ) / K . The target is u u u , with reports in X n \mathcal{X}_n X n and physical loss (18).
Apply Theorem A1.2’s cell isometry and completed inverse to each θ i \theta_i θ i . The cell error C ϕ q 0 θ i 2 C_\phi q_0\theta_i^2 C ϕ q 0 θ i 2 gives product fidelity error at most C ϕ H n 2 / n C_\phi H_n^2/\sqrt{n} C ϕ H n 2 / n . For any monotone f : [ 0 , 1 ] → [ 0 , 1 ] f:[0,1]\to[0,1] f : [ 0 , 1 ] → [ 0 , 1 ] , its right step approximation obeys ∥ f n − f ∥ 2 2 ≤ ∥ f n − f ∥ 1 ≤ [ f ( 1 ) − f ( 0 ) ] / n ≤ 1 / n \lVert f_n-f\rVert_2^2\le\lVert f_n-f\rVert_1\le[f(1)-f(0)]/n\le1/n ∥ f n − f ∥ 2 2 ≤ ∥ f n − f ∥ 1 ≤ [ f ( 1 ) − f ( 0 )] / n ≤ 1/ n by telescoping. With H = H n H=H_n H = H n , these curve and grid channels give
ϵ n , H ≤ H n + C ϕ H n 2 n . (151) \epsilon_{n,H}\le\frac{H_n+C_\phi H_n^2}{\sqrt{n}}. \tag*{(151)} ϵ n , H ≤ n H n + C ϕ H n 2 . (151)
The bound is uniform in w n > 0 w_n > 0 w n > 0 . The curve expansion gives τ n / s n 2 = 1 + O ( s n / n ) \tau_n/s_n^2 = 1 + O(s_n/\sqrt{n}) τ n / s n 2 = 1 + O ( s n / n ) , and the logarithmic derivative of k w k_w k w lies in [ 1 / 2 , 1 ] [1/2,1] [ 1/2 , 1 ] . Clipping and grid rounding therefore give
d l o c , n = 2 ∣ τ n − s n 2 ∣ max { τ n , s n 2 } + 2 k w n ( τ n ) K n = O ( s n n + s n 2 n ) , d a m p , n = 0. (152) d_{\mathrm{loc},n}=2\frac{|\tau_n-s_n^2|}{\max\{\tau_n,s_n^2\}}+\frac{2k_{w_n}(\tau_n)}{K n}=O\left(\frac{s_n}{\sqrt{n}}+\frac{s_n^2}{n}\right),\qquad d_{\mathrm{amp},n}=0. \tag*{(152)} d loc , n = 2 max { τ n , s n 2 } ∣ τ n − s n 2 ∣ + K n 2 k w n ( τ n ) = O ( n s n + n s n 2 ) , d amp , n = 0. (152)
Apply Lemma A1.1 with B l o c = 2 B_{\mathrm{loc}}=2 B loc = 2 , B a m p = 4 L 2 B_{\mathrm{amp}}=4L^2 B amp = 4 L 2 , B λ = 2 λ + 4 L 2 ( 1 − λ ) B_\lambda=2\lambda+4L^2(1-\lambda) B λ = 2 λ + 4 L 2 ( 1 − λ ) and E n = B λ ε n , H + λ d l o c , n E_n=B_\lambda\varepsilon_{n,H}+\lambda d_{\mathrm{loc},n} E n = B λ ε n , H + λ d loc , n :
∣ r j , n ⋆ − r j , ∞ ⋆ ∣ ≤ B j ε n , H + d j , n , (153) |r^\star_{j,n}-r^\star_{j,\infty}|\le B_j\varepsilon_{n,H}+d_{j,n}, \tag*{(153)} ∣ r j , n ⋆ − r j , ∞ ⋆ ∣ ≤ B j ε n , H + d j , n , (153)
∣ R λ , n ⋆ − R λ , ∞ ⋆ ∣ ≤ E n , (154) |R^\star_{\lambda,n}-R^\star_{\lambda,\infty}|\le E_n, \tag*{(154)} ∣ R λ , n ⋆ − R λ , ∞ ⋆ ∣ ≤ E n , (154)
∣ G λ , n − G λ ∣ ≤ 2 E n . (155) |G_{\lambda,n}-G_\lambda|\le2E_n. \tag*{(155)} ∣ G λ , n − G λ ∣ ≤ 2 E n . (155)
Pulling back a common rule adds at most 2 B j ε n , H + 2 d j , n 2B_j\varepsilon_{n,H}+2d_{j,n} 2 B j ε n , H + 2 d j , n to each excess. For s n → ∞ s_n\to\infty s n → ∞ , the channel, calibration and report errors are O ( s n 2 / n ) O(s_n^2/\sqrt{n}) O ( s n 2 / n ) , O ( s n / n ) O(s_n/\sqrt{n}) O ( s n / n ) and O ( s n 2 / n ) O(s_n^2/n) O ( s n 2 / n ) , uniformly in w n > 0 w_n>0 w n > 0 . They vanish for s n = o ( n 1 / 4 ) s_n=o(n^{1/4}) s n = o ( n 1/4 ) , transferring J’s leading limits whenever w n → 0 w_n\to0 w n → 0 .
Multiplication by s n 2 s_n^2 s n 2 gives O ( s n 4 / n ) O(s_n^4/\sqrt{n}) O ( s n 4 / n ) , O ( s n 3 / n ) O(s_n^3/\sqrt{n}) O ( s n 3 / n ) and O ( s n 4 / n ) O(s_n^4/n) O ( s n 4 / n ) , which vanish for s n = o ( n 1 / 8 ) s_n=o(n^{1/8}) s n = o ( n 1/8 ) . Thus the joint and marginal optima, the gap and both common rule excesses transfer at o ( s n − 2 ) o(s_n^{-2}) o ( s n − 2 ) along every known path χ n ∈ [ 1 / 2 , 2 ] \chi_n\in[1/2,2] χ n ∈ [ 1/2 , 2 ] , w n = χ n / s n 2 w_n=\chi_n/s_n^2 w n = χ n / s n 2 .
Fixed width limit and cubic width cost
The fixed width limit, loss dependence and cubic width cost
Fix 0 < w < 1 0<w<1 0 < w < 1 , L = 1 / 2 L=1/2 L = 1/2 and the quintic profile from (11); in particular, F ( r ) > 0 F(r)>0 F ( r ) > 0 for 0 < r ≤ 1 0<r\le1 0 < r ≤ 1 . Use the main priors, reports and calibration, with ν = Unif [ − L , L ] \nu=\operatorname{Unif}[-L,L] ν = Unif [ − L , L ] and losses
δ s , w = 1 + s 2 w s 2 , ℓ l o c ( w ) ( x , u ) = 2 min { 1 , ∣ x − u ∣ K δ s , w } , ℓ a m p ( y , z ) = ( y − z ) 2 . (156) \delta_{s,w}=\frac{\sqrt{1+s^2w}}{s^2},\qquad\ell_{\mathrm{loc}}^{(w)}(x,u)=2\min\left\{1,\frac{|x-u|}{K\delta_{s,w}}\right\},\qquad\ell_{\mathrm{amp}}(y,z)=(y-z)^2. \tag*{(156)} δ s , w = s 2 1 + s 2 w , ℓ loc ( w ) ( x , u ) = 2 min { 1 , K δ s , w ∣ x − u ∣ } , ℓ amp ( y , z ) = ( y − z ) 2 . (156)
The loss bounds are B l o c = 2 B_{\mathrm{loc}}=2 B loc = 2 , B a m p = 4 L 2 B_{\mathrm{amp}}=4L^2 B amp = 4 L 2 and B λ = 2 λ + 4 L 2 ( 1 − λ ) B_\lambda=2\lambda+4L^2(1-\lambda) B λ = 2 λ + 4 L 2 ( 1 − λ ) . Write G \mathcal{G} G for the full record gap and g g g for the covariant fiber gap. We use D ( ∣ f ⟩ , ∣ g ⟩ ) ≤ ∥ f − g ∥ 2 D(|f\rangle,|g\rangle)\le\lVert f-g\rVert_2 D ( ∣ f ⟩ , ∣ g ⟩) ≤ ∥ f − g ∥ 2 .
Limits at fixed width and local quadratic compatibility
Truncated tangents and global comparison
For u < 1 u<1 u < 1 , set e = 1 − u e=1-u e = 1 − u , r = min { e / w , 1 } r=\min\{e/w,1\} r = min { e / w , 1 } , I ( r ) = ∫ 0 r F 2 I(r)=\int_0^r F^2 I ( r ) = ∫ 0 r F 2 , J ( r ) = ∫ 0 r F 2 J(r)=\int_0^r F^2 J ( r ) = ∫ 0 r F 2 . Then
G 11 ( u ) = I ( r ) , G 12 ( u ) = − w 2 F ( r ) 2 , G 22 ( u ) = ( e − w ) + + w J ( r ) , (157) G_{11}(u)=I(r),\qquad G_{12}(u)=-\frac{\sqrt{w}}{2}F(r)^2,\qquad G_{22}(u)=(e-w)_++wJ(r), \tag*{(157)} G 11 ( u ) = I ( r ) , G 12 ( u ) = − 2 w F ( r ) 2 , G 22 ( u ) = ( e − w ) + + w J ( r ) , (157)
κ = G 11 , d = a G 12 κ , c = a 2 G 22 − d 2 . (158) \kappa=\sqrt{G_{11}},\qquad d=\frac{aG_{12}}{\kappa},\qquad c=\sqrt{a^2G_{22}-d^2}. \tag*{(158)} κ = G 11 , d = κ a G 12 , c = a 2 G 22 − d 2 . (158)
These are the Gram entries of g l o c , u = − w − 1 / 2 F ′ ( ( t − u ) / w ) g_{\mathrm{loc},u}=-w^{-1/2}F'((t-u)/w) g loc , u = − w − 1/2 F ′ (( t − u ) / w ) and g a m p , u = F ( ( t − u ) / w ) g_{\mathrm{amp},u}=F((t-u)/w) g amp , u = F (( t − u ) / w ) . Cauchy–Schwarz gives I J > F ( r ) 4 / 4 IJ>F(r)^4/4 I J > F ( r ) 4 /4 : equality would make F ′ F' F ′ proportional to F F F , contrary to F ( 0 ) = 0 < F ( r ) F(0)=0<F(r) F ( 0 ) = 0 < F ( r ) . Thus G 11 > 0 G_{11}>0 G 11 > 0 and det G > 0 \det G>0 det G > 0 , including truncated rises. Gram–Schmidt and Fock factorization give
ρ v , z u = ∣ κ ( u ) v + d ( u ) z ⟩ ⟨ κ ( u ) v + d ( u ) z ∣ ⊗ ∣ c ( u ) z ⟩ ⟨ c ( u ) z ∣ . (159) \rho^u_{v,z}=|\kappa(u)v+d(u)z\rangle\langle\kappa(u)v+d(u)z|\otimes|c(u)z\rangle\langle c(u)z|. \tag*{(159)} ρ v , z u = ∣ κ ( u ) v + d ( u ) z ⟩ ⟨ κ ( u ) v + d ( u ) z ∣ ⊗ ∣ c ( u ) z ⟩ ⟨ c ( u ) z ∣. (159)
The continuous coefficients satisfy κ , c > 0 \kappa,c>0 κ , c > 0 and d < 0 d<0 d < 0 on [ 0 , 1 ) [0,1) [ 0 , 1 ) . The complete covariant gap is g λ ( κ , d , c ) g_\lambda(\kappa,d,c) g λ ( κ , d , c ) , defined as in (7); set g λ , w ( u ) = g λ ( κ ( u ) , d ( u ) , c ( u ) ) g_{\lambda,w}(u)=g_\lambda(\kappa(u),d(u),c(u)) g λ , w ( u ) = g λ ( κ ( u ) , d ( u ) , c ( u )) .
Partition [ 0 , 1 ] [0,1] [ 0 , 1 ] into cells of length b s = 1 / N s b_s=1/N_s b s = 1/ N s , centered at u j u_j u j . Good cells end at or before 1 − ϵ 1-\epsilon 1 − ϵ ; the others have mass at most ϵ + b s \epsilon+b_s ϵ + b s . The comparison F ϵ F_\epsilon F ϵ emits the true good label and ∣ v g l o c , u j + a z g a m p , u j ⟩ |v g_{\mathrm{loc},u_j}+a z g_{\mathrm{amp},u_j}\rangle ∣ v g loc , u j + a z g amp , u j ⟩ , with v = s ( u − u j ) / w v=s(u-u_j)/\sqrt{w} v = s ( u − u j ) / w , and emits vacuum with a bad label elsewhere. Taylor and pilot attenuation errors are T s = O w ( s b s 2 + b s ) T_s=O_w(sb_s^2+b_s) T s = O w ( s b s 2 + b s ) and H s = O w ( t s ( s b s + 1 ) ) H_s=O_w(t_s(sb_s+1)) H s = O w ( t s ( s b s + 1 )) . For q ϵ = F ( ϵ / ( 2 w ) ) > 0 q_\epsilon=F(\epsilon/(2w))>0 q ϵ = F ( ϵ / ( 2 w )) > 0 , the label error is
P s ≤ min { 1 , C w , ϵ ( N s + 1 ) ( s − 1 + ( s t s ) − 1 ) } . (160) P_s\leq\min\{1,C_{w,\epsilon}(N_s+1)(s^{-1}+(s\sqrt{t_s})^{-1})\}. \tag*{(160)} P s ≤ min { 1 , C w , ϵ ( N s + 1 ) ( s − 1 + ( s t s ) − 1 )} . (160)
The bounds are uniform over good cells and ∣ z ∣ ≤ L |z|\leq L ∣ z ∣ ≤ L . For the joint risk, take r j ⋆ = R λ ⋆ r_j^\star=R_\lambda^\star r j ⋆ = R λ ⋆ and B j = B λ B_j=B_\lambda B j = B λ .
Theorem 1 (Global comparison). Channels independent of u , z u,z u , z in both directions have errors averaged over the prior satisfying
ϵ → ≤ ϵ + b s + P s + T s + H s , ϵ ← ≤ ϵ + b s + T s , (161) \epsilon_{\to}\leq\epsilon+b_s+P_s+T_s+H_s,\qquad\epsilon_{\leftarrow}\leq\epsilon+b_s+T_s, \tag*{(161)} ϵ → ≤ ϵ + b s + P s + T s + H s , ϵ ← ≤ ϵ + b s + T s , (161) with both errors capped at one. For each marginal and the joint risk, − B j ϵ ← ≤ r j ⋆ ( E s ) − r j ⋆ ( F s ϵ ) ≤ B j ϵ → -B_j\epsilon_{\leftarrow}\leq r_j^\star(E_s)-r_j^\star(F_s^\epsilon)\leq B_j\epsilon_{\to} − B j ϵ ← ≤ r j ⋆ ( E s ) − r j ⋆ ( F s ϵ ) ≤ B j ϵ → ; hence ∣ G λ ( E s ) − G λ ( F s ϵ ) ∣ ≤ B λ ( ϵ → + ϵ ← ) |\mathcal{G}_\lambda(E_s)-\mathcal{G}_\lambda(F_s^\epsilon)|\leq B_\lambda(\epsilon_{\to}+\epsilon_{\leftarrow}) ∣ G λ ( E s ) − G λ ( F s ϵ ) ∣ ≤ B λ ( ϵ → + ϵ ← ) . At fixed ϵ \epsilon ϵ , the choices t s = s − 1 / 2 t_s=s^{-1/2} t s = s − 1/2 and N s = ⌈ s 2 / 3 ⌉ N_s=\lceil s^{2/3}\rceil N s = ⌈ s 2/3 ⌉ give P s = O ϵ ( s − 1 / 12 ) P_s=O_\epsilon(s^{-1/12}) P s = O ϵ ( s − 1/12 ) , T s = O ( s − 1 / 3 ) T_s=O(s^{-1/3}) T s = O ( s − 1/3 ) , H s = O ( s − 1 / 6 ) H_s=O(s^{-1/6}) H s = O ( s − 1/6 ) , while
V s = s b s / ( 2 w ) ⟶ ∞ . (162) V_s=sb_s/(2\sqrt{w})\longrightarrow\infty. \tag*{(162)} V s = s b s / ( 2 w ) ⟶ ∞. (162)
Proof. Homodyne measurement of the pilot fraction gives X = t s ( s + a z ) A w ( u ) + Z / 2 X=\sqrt{t_s}(s+az)A_w(u)+Z/2 X = t s ( s + a z ) A w ( u ) + Z /2 , where Z ∼ N ( 0 , 1 ) Z\sim N(0,1) Z ∼ N ( 0 , 1 ) , A w ( u ) = ∫ 0 1 F ( ( t − u ) / w ) d t A_w(u)=\int_0^1F((t-u)/w)\,dt A w ( u ) = ∫ 0 1 F (( t − u ) / w ) d t and A w ′ ( u ) = − F ( ( 1 − u ) / w ) A_w'(u)=-F((1-u)/w) A w ′ ( u ) = − F (( 1 − u ) / w ) . Clip to the range on [ 0 , 1 − ϵ / 2 ] [0,1-\epsilon/2] [ 0 , 1 − ϵ /2 ] and invert A w A_w A w . On a good cell, ∣ u ^ − u ∣ ≤ q ϵ − 1 ( a L / s + ∣ Z ∣ / ( 2 s t s ) ) |\hat{u}-u|\leq q_\epsilon^{-1}(aL/s+|Z|/(2s\sqrt{t_s})) ∣ u ^ − u ∣ ≤ q ϵ − 1 ( a L / s + ∣ Z ∣/ ( 2 s t s )) ; integration over cell boundary strips gives P s P_s P s . The conditionally independent retained field is displaced by − 1 − t s s F ( ( t − u j ) / w ) -\sqrt{1-t_s}sF((t-u_j)/w) − 1 − t s s F (( t − u j ) / w ) at a reported good label. Taylor’s remainder and 1 − 1 − t s ≤ t s 1-\sqrt{1-t_s}\leq t_s 1 − 1 − t s ≤ t s give T s , H s T_s,H_s T s , H s . Wrong labels and bad positions contribute their probabilities. The reverse channel restores the supplied displacement with error T s T_s T s and assigns fixed states to bad labels. Pulling back complete POVMs gives the three risk bounds; their difference gives the gap bound.
Wide priors and integration over position
In a fixed fiber, take independent v ∼ Unif [ − V , V ] v\sim\operatorname{Unif}[-V,V] v ∼ Unif [ − V , V ] and z ∼ ν z\sim\nu z ∼ ν , with c K ( x − v ) = 2 min { 1 , ∣ x − v ∣ / K } c_K(x-v)=2\min\{1,|x-v|/K\} c K ( x − v ) = 2 min { 1 , ∣ x − v ∣/ K } . Translation by t t t displaces the first mode by κ t \kappa t κ t and preserves amplitude reports. Compactify the position reports by a failure point ∞ \infty ∞ of cost two.
Lemma 1 (Wide flat prior). As V → ∞ V\to\infty V → ∞ , the finite prior optimum converges to the complete covariant risk R λ , c o v ⋆ ( κ , d , c ) R^\star_{\lambda,\mathrm{cov}}(\kappa,d,c) R λ , cov ⋆ ( κ , d , c ) for every 0 ≤ λ ≤ 1 0\leq\lambda\leq1 0 ≤ λ ≤ 1 . Convergence is uniform on compact coefficient sets with κ , c > 0 \kappa,c>0 κ , c > 0 .
Proof. For the upper bound, replace covariant failure by a finite report and clip to [ − V , V ] [-V,V] [ − V , V ] . For the lower bound, average M V M_V M V with risk within V − 1 V^{-1} V − 1 of the optimum: M ‾ V ( B ) = ( 2 V ) − 1 ∫ − V V U t † M V ( B + t ) U t d t \overline{M}_V(B)=(2V)^{-1}\int_{-V}^{V}U_t^\dagger M_V(B+t)U_t\,dt M V ( B ) = ( 2 V ) − 1 ∫ − V V U t † M V ( B + t ) U t d t . Its risk at zero equals the finite prior risk, and translation by t t t changes a continuous test integral by at most ∣ t ∣ ∥ f ∥ ∞ / V |t|\|f\|_\infty/V ∣ t ∣∥ f ∥ ∞ / V . Apply the compactness and translated-average argument of Lemma J1.2 to paired reports ( R ∪ { ∞ } ) × [ − L , L ] (\mathbb{R}\cup\{\infty\})\times[-L,L] ( R ∪ { ∞ }) × [ − L , L ] . It gives a complete covariant limit with the limiting risk, since the cost field λ c K ( x ) Ξ 0 + ( 1 − λ ) ( y 2 Ξ 0 − 2 y Ξ 1 + Ξ 2 ) \lambda c_K(x)\Xi_0+(1-\lambda)(y^2\Xi_0-2y\Xi_1+\Xi_2) λ c K ( x ) Ξ 0 + ( 1 − λ ) ( y 2 Ξ 0 − 2 y Ξ 1 + Ξ 2 ) is trace-class valued and continuous on this compact space. Here Ξ j = ∫ z j ρ 0 , z ν ( d z ) \Xi_j=\int z^j\rho_{0,z}\nu(dz) Ξ j = ∫ z j ρ 0 , z ν ( d z ) ; the failure cost is included. For uniformity, set x = κ v x=\kappa v x = κ v . A change in κ \kappa κ changes the prior by ∣ κ − κ ′ ∣ / max { κ , κ ′ } |\kappa-\kappa'|/\max\{\kappa,\kappa'\} ∣ κ − κ ′ ∣/ max { κ , κ ′ } in total variation and the loss by twice this amount. A change in d , c d,c d , c changes the state distance by at most L ( d − d ′ ) 2 + ( c − c ′ ) 2 L\sqrt{(d-d')^2+(c-c')^2} L ( d − d ′ ) 2 + ( c − c ′ ) 2 , uniformly in x x x . Equicontinuity and a finite net give uniform convergence.
Theorem 2 (Global fixed width limit). At fixed width, the original full priors and losses (S.1) give
lim s → ∞ R λ , s ⋆ = ∫ 0 1 R λ , cov ⋆ ( κ ( u ) , d ( u ) , c ( u ) ) d u , (163) \lim_{s \to\infty} R^{\star}_{\lambda,s} = \int_{0}^{1} R^{\star}_{\lambda,\operatorname{cov}}(\kappa(u),d(u),c(u))\,du, \tag*{(163)} s → ∞ lim R λ , s ⋆ = ∫ 0 1 R λ , cov ⋆ ( κ ( u ) , d ( u ) , c ( u )) d u , (163) lim s → ∞ G λ , s = ∫ 0 1 g λ , w ( u ) d u . (164) \lim_{s \to\infty} G_{\lambda,s} = \int_{0}^{1} g_{\lambda,w}(u)\,du. \tag*{(164)} s → ∞ lim G λ , s = ∫ 0 1 g λ , w ( u ) d u . (164) The first identity at λ = 1 , 0 \lambda= 1,0 λ = 1 , 0 gives the marginal limits.
Proof. Clipping good cell reports gives local loss 2 min { 1 , q s ∣ v ~ − v ∣ / K } 2\min\{1,q_s|\widetilde{v}-v|/K\} 2 min { 1 , q s ∣ v − v ∣/ K } , where q s = s w / 1 + s 2 w → 1 q_s=s\sqrt{w}/\sqrt{1+s^2w}\to1 q s = s w / 1 + s 2 w → 1 ; its distance from c K cK cK is at most 2 ( 1 − q s ) 2(1-q_s) 2 ( 1 − q s ) . Optimal risks add over the labels. Lemma 1, uniformly on [ 0 , 1 − ϵ ] [0,1-\epsilon] [ 0 , 1 − ϵ ] , turns these sums into the fiber integral. The remaining contribution is bounded by B λ ( ϵ + b s ) B_{\lambda}(\epsilon+b_s) B λ ( ϵ + b s ) . The channel bounds, followed by s → ∞ s\to\infty s → ∞ and ϵ ↓ 0 \epsilon\downarrow0 ϵ ↓ 0 , prove the risk identity. Its cases λ = 1 , 0 \lambda=1,0 λ = 1 , 0 give the marginal limits and hence the gap identity. □ \square □
Proposition 1 (Residual amplitude optimum). The fiber amplitude optimum equals A L ( c ( u ) ) \mathcal{A}_{L}(c(u)) A L ( c ( u )) , the Bayes risk of the residual experiment
z ↦ ∣ c ( u ) z ⟩ , z ∼ ν . (165) z \mapsto|c(u)z\rangle,\qquad z\sim\nu. \tag*{(165)} z ↦ ∣ c ( u ) z ⟩ , z ∼ ν . (165)
Proof. Averaging the first mode over v v v gives the coherent mixture on [ − κ V + d z , κ V + d z ] [-\kappa V+dz,\kappa V+dz] [ − κV + d z , κV + d z ] . Replacing it by the mixture centered at zero costs at most min { 1 , ∣ d z ∣ / ( 2 κ V ) } \min\{1,|dz|/(2\kappa V)\} min { 1 , ∣ d z ∣/ ( 2 κV )} in trace distance. This replacement and discarding the first mode give
0 ≤ A L ( c ) − r amp , V ⋆ ≤ 4 L 2 E min { 1 , ∣ d z ∣ / ( 2 κ V ) } → 0. (166) 0\le\mathcal{A}_{L}(c)-r^{\star}_{\operatorname{amp},V}\le4L^{2}\mathbb{E}\min\{1,|dz|/(2\kappa V)\}\to0. \tag*{(166)} 0 ≤ A L ( c ) − r amp , V ⋆ ≤ 4 L 2 E min { 1 , ∣ d z ∣/ ( 2 κV )} → 0. (166)
Lemma 1 finishes the proof. □ \square □
Complete covariant decisions and loss control
On the support of Ω = ∫ ∣ c z ⟩ ⟨ c z ∣ ν ( d z ) \Omega=\int|cz\rangle\langle cz|\nu(dz) Ω = ∫ ∣ cz ⟩ ⟨ cz ∣ ν ( d z ) , set ψ z = ∣ c z ⟩ \psi_z=|cz\rangle ψ z = ∣ cz ⟩ , Ω j = ∫ z j ∣ ψ z ⟩ ⟨ ψ z ∣ ν ( d z ) \Omega_j=\int z^j|\psi_z\rangle\langle\psi_z|\nu(dz) Ω j = ∫ z j ∣ ψ z ⟩ ⟨ ψ z ∣ ν ( d z ) , Ω A + A Ω = 2 Ω 1 \Omega\mathcal{A}+\mathcal{A}\Omega=2\Omega_1 Ω A + A Ω = 2 Ω 1 and A = Tr Ω 2 − Tr ( Ω A 2 ) \mathcal{A}=\operatorname{Tr}\Omega_2-\operatorname{Tr}(\Omega\mathcal{A}^2) A = Tr Ω 2 − Tr ( Ω A 2 ) , with ∥ A ∥ ≤ L \|\mathcal{A}\|\le L ∥ A ∥ ≤ L . Use [ Q , P ] = i / 2 [Q,P]=i/2 [ Q , P ] = i /2 and φ ( p ) = ( 2 / π ) 1 / 4 e − p 2 \varphi(p)=(2/\pi)^{1/4}e^{-p^2} φ ( p ) = ( 2/ π ) 1/4 e − p 2 . Covariant position POVMs admit an isometry field [42 Theorem 2.4.3 and Corollary 2.4.5] . For a complete covariant paired POVM with real location reports, take a minimal joint Naimark dilation ( H ~ , J , P ) (\widetilde{\mathcal{H}},J,P) ( H , J , P ) . Covariance makes
U ~ t P ( B × C ) J ψ = P ( ( B + t ) × C ) J U t ψ \widetilde{U}_tP(B\times C)J\psi=P((B+t)\times C)JU_t\psi U t P ( B × C ) J ψ = P (( B + t ) × C ) J U t ψ
an isometry on a dense span, defining a strongly continuous unitary representation by compactly supported test functions. The imprimitivity theorem identifies the dilation with L 2 ( R , K ) L^2(\mathbb{R},\mathcal{K}) L 2 ( R , K ) , with position multiplication and translation shifts [42 Theorem 1.2.2] . The amplitude PVM commutes with position and translations, so it is a constant decomposable field I ⊗ F amp I\otimes F_{\operatorname{amp}} I ⊗ F amp . Thus Y = ∫ y F amp ( d y ) Y=\int yF_{\operatorname{amp}}(dy) Y = ∫ y F amp ( d y ) is self-adjoint with ∥ Y ∥ ≤ L \|Y\|\le L ∥ Y ∥ ≤ L , and the momentum intertwiner is a measurable isometry field W ( p ) : H amp → K W(p):\mathcal{H}_{\operatorname{amp}}\to\mathcal{K} W ( p ) : H amp → K . The location density is the squared norm of
κ / π ∫ e 2 i κ x p φ ( p ) e − 2 i d z p W ( p ) ψ z d p . (167) \sqrt{\kappa/\pi}\int e^{2i\kappa xp}\varphi(p)e^{-2idzp}W(p)\psi_z\,dp. \tag*{(167)} κ / π ∫ e 2 iκ x p φ ( p ) e − 2 i d z p W ( p ) ψ z d p . (167)
Writing P d ( ξ ) = ∫ e − 2 i d z ξ ∣ ψ z ⟩ ⟨ ψ z ∣ ν ( d z ) P_d(\xi)=\int e^{-2idz\xi}|\psi_z\rangle\langle\psi_z|\nu(dz) P d ( ξ ) = ∫ e − 2 i d z ξ ∣ ψ z ⟩ ⟨ ψ z ∣ ν ( d z ) and k b ( ξ ) = b sinc 2 ( b ξ ) / π k_b(\xi)=b\operatorname{sinc}^2(b\xi)/\pi k b ( ξ ) = b sinc 2 ( b ξ ) / π , b = κ K b=\kappa K b = κ K , Fourier transformation and Personick completion give
r loc ( W ) = 2 − 2 Re ∫ φ ( p ) φ ( q ) k b ( p − q ) Tr [ W ( p ) P d ( p − q ) W ( q ) † ] d p d q , (168) r_{\operatorname{loc}}(W)=2-2\operatorname{Re}\int\varphi(p)\varphi(q)k_b(p-q)\operatorname{Tr}[W(p)P_d(p-q)W(q)^\dagger]\,dp\,dq, \tag*{(168)} r loc ( W ) = 2 − 2 Re ∫ φ ( p ) φ ( q ) k b ( p − q ) Tr [ W ( p ) P d ( p − q ) W ( q ) † ] d p d q , (168)
r amp ( W , Y ) − A = ∫ φ ( p ) 2 ∥ Y W ( p ) − W ( p ) A ∥ Ω 2 d p . (169) r_{\operatorname{amp}}(W,Y)-\mathcal{A}=\int\varphi(p)^2\|YW(p)-W(p)\mathcal{A}\|_{\Omega}^{2}\,dp. \tag*{(169)} r amp ( W , Y ) − A = ∫ φ ( p ) 2 ∥ Y W ( p ) − W ( p ) A ∥ Ω 2 d p . (169)
These infima also include compactified failure reports. Disintegrate their invariant effects as F p ( d y ) F_p(dy) F p ( d y ) and partition [ − L , L ] [-L,L] [ − L , L ] into bins C j C_j C j of diameter η \eta η , with y j ∈ C j y_j \in C_j y j ∈ C j . The field V η ( p ) ψ = ⨁ j F p ( C j ) 1 / 2 ψ V_\eta(p)\psi=\bigoplus_j F_p(C_j)^{1/2}\psi V η ( p ) ψ = ⨁ j F p ( C j ) 1/2 ψ and Y η = ⨁ j y j I Y_\eta=\bigoplus_j y_j I Y η = ⨁ j y j I preserve amplitude bin probabilities by Plancherel. A real position report costs at most the failure cost two, and amplitude rounding costs at most 4 L η 4L\eta 4 L η . Direct sum with the real report part and let η ↓ 0 \eta\downarrow0 η ↓ 0 .
Theorem 3 (Quadratic compatibility). For quadratic location loss ( v ^ − v ) 2 (\hat v-v)^2 ( v ^ − v ) 2 , the fiber optimum is
r 2 ⋆ = 1 / 4 + d 2 A L ( c ) κ 2 . (170) r_2^\star=\frac{1/4+d^2 A_L(c)}{\kappa^2}. \tag*{(170)} r 2 ⋆ = κ 2 1/4 + d 2 A L ( c ) . (170) Joint measurement of first mode Q Q Q and residual mode A A A , reported as v ^ = ( X − d Y ) / κ \hat v=(X-dY)/\kappa v ^ = ( X − d Y ) / κ , z ^ = Y \hat z=Y z ^ = Y , attains both marginal optima in the complete covariant class, giving zero covariant gap for every weight.
Proof. By Plancherel, the quadratic risk equals ( 4 κ 2 ) − 1 ∬ ∥ ∂ p [ φ e − 2 i d z p W ψ z ] ∥ ν 2 ( d z ) d p (4\kappa^2)^{-1}\iint\|\partial_p[\varphi e^{-2idzp}W\psi_z]\|^2_{\nu}(dz)dp ( 4 κ 2 ) − 1 ∬ ∥ ∂ p [ φ e − 2 i d z p W ψ z ] ∥ ν 2 ( d z ) d p . Bounded z z z and positivity of φ \varphi φ on compact intervals give local weighted Sobolev regularity for finite risk. Isometry cancels the φ ′ \varphi' φ ′ cross term, ∫ ∣ φ ′ ∣ 2 = 1 \int|\varphi'|^2=1 ∫ ∣ φ ′ ∣ 2 = 1 , and the Sylvester equation gives
∫ ∥ ( W ′ − 2 i d z W ) ψ z ∥ ν 2 ( d z ) = 4 d 2 A + ∥ W ′ − 2 i d W A ∥ Ω 2 . (171) \int\|(W'-2idzW)\psi_z\|^2_{\nu}(dz)=4d^2A+\|W'-2idWA\|^2_{\Omega}. \tag*{(171)} ∫ ∥ ( W ′ − 2 i d z W ) ψ z ∥ ν 2 ( d z ) = 4 d 2 A + ∥ W ′ − 2 i d W A ∥ Ω 2 . (171)
On spectral projections P N P_N P N of Ω \Omega Ω , the compressed Sylvester equation and skew-adjointness of ( W P N ) † ( W P N ) ′ (WP_N)^\dagger(WP_N)' ( W P N ) † ( W P N ) ′ prove this identity. Bounded z , A z,A z , A control the cross terms in the Sobolev limit, yielding
r 2 ( W ) − r 2 ⋆ = 1 4 κ 2 ∫ φ 2 ∥ W ′ − 2 i d W A ∥ 2 d p . r_2(W)-r_2^\star=\frac{1}{4\kappa^2}\int\varphi^2\|W'-2idWA\|^2dp. r 2 ( W ) − r 2 ⋆ = 4 κ 2 1 ∫ φ 2 ∥ W ′ − 2 i d W A ∥ 2 d p .
The field W = e 2 i d p A W=e^{2idpA} W = e 2 i d p A makes both defects zero and yields ( Q − d A ) / κ (Q-dA)/\kappa ( Q − d A ) / κ . Here X = κ v + d z + N X=\kappa v+dz+N X = κ v + d z + N , with independent N ∼ N ( 0 , 1 / 4 ) N\sim N(0,1/4) N ∼ N ( 0 , 1/4 ) , so the location error is ( N + d ( z − Y ) ) / κ (N+d(z-Y))/\kappa ( N + d ( z − Y )) / κ .
Local bounds needed for the cubic width cost
Scalar coercivity and loss response
Set b = κ K b=\kappa K b = κ K and define
φ ( p ) = ( 2 / π ) 1 / 4 e − p 2 , μ ( d p ) = φ ( p ) 2 d p , k b ( x ) = b π sinc 2 ( b x ) , g b ( p , q ) = φ ( p ) φ ( q ) k b ( p − q ) . \varphi(p)=(2/\pi)^{1/4}e^{-p^2},\qquad\mu(dp)=\varphi(p)^2dp,\qquad k_b(x)=\frac{b}{\pi}\operatorname{sinc}^2(bx),\qquad g_b(p,q)=\varphi(p)\varphi(q)k_b(p-q). φ ( p ) = ( 2/ π ) 1/4 e − p 2 , μ ( d p ) = φ ( p ) 2 d p , k b ( x ) = π b sinc 2 ( b x ) , g b ( p , q ) = φ ( p ) φ ( q ) k b ( p − q ) .
For a real scalar field η \eta η , define
D b ( η ) = ∬ g b ( p , q ) ( η ( p ) − η ( q ) ) 2 d p d q , D_b(\eta)=\iint g_b(p,q)(\eta(p)-\eta(q))^2\,dp\,dq, D b ( η ) = ∬ g b ( p , q ) ( η ( p ) − η ( q ) ) 2 d p d q ,
( L b η ) ( p ) = ∫ g b ( p , q ) ( η ( p ) − η ( q ) ) d q , (L_b\eta)(p)=\int g_b(p,q)(\eta(p)-\eta(q))\,dq, ( L b η ) ( p ) = ∫ g b ( p , q ) ( η ( p ) − η ( q )) d q ,
f b ( p ) = ∫ g b ( p , q ) ( p − q ) 3 d q . (172) f_b(p)=\int g_b(p,q)(p-q)^3\,dq. \tag*{(172)} f b ( p ) = ∫ g b ( p , q ) ( p − q ) 3 d q . (172)
Then ∫ f b = 0 \int f_b=0 ∫ f b = 0 . Replacing the squared difference by ∥ H ( p ) − H ( q ) ∥ Ω 2 \|H(p)-H(q)\|^2_{\Omega} ∥ H ( p ) − H ( q ) ∥ Ω 2 defines D b , Ω ( H ) D_{b,\Omega}(H) D b , Ω ( H ) . Choose representatives of mean zero under μ \mu μ modulo constants.
Lemma 1 (Gaussian graph coercivity). For every b > 0 b>0 b > 0 , some c b > 0 c_b>0 c b > 0 satisfies D b ( η ) ≥ c b Var μ η D_b(\eta)\ge c_b\operatorname{Var}_{\mu}\eta D b ( η ) ≥ c b Var μ η . The degree measure m b ( p ) d p m_b(p)\,dp m b ( p ) d p , with m b ( p ) = ∫ g b ( p , q ) d q m_b(p)=\int g_b(p,q)\,dq m b ( p ) = ∫ g b ( p , q ) d q , also has a positive Poincare constant.
Proof. Define the two step symmetric kernel
g b ( 2 ) ( p , q ) = ∫ g b ( p , r ′ ) g b ( r ′ , q ) m b ( r ′ ) d r ′ . g_b^{(2)}(p,q)=\int\frac{g_b(p,r')g_b(r',q)}{m_b(r')} \,dr'. g b ( 2 ) ( p , q ) = ∫ m b ( r ′ ) g b ( p , r ′ ) g b ( r ′ , q ) d r ′ .
This kernel has degree m b m_b m b and form at most 4 D b 4\mathcal{D}_b 4 D b . With r b = π / ( 2 b ) r_b=\pi/(2b) r b = π / ( 2 b ) , boundedness of k b ∗ φ k_b * \varphi k b ∗ φ and integration over [ − r b , r b ] [-r_b,r_b] [ − r b , r b ] give
g b ( 2 ) ( p , q ) ≥ C b φ ( p ) ( ∣ p ∣ + r b ) 2 φ ( q ) ( ∣ q ∣ + r b ) 2 . g_b^{(2)}(p,q)\ge C_b\frac{\varphi(p)}{(|p|+r_b)^2}\frac{\varphi(q)}{(|q|+r_b)^2}. g b ( 2 ) ( p , q ) ≥ C b ( ∣ p ∣ + r b ) 2 φ ( p ) ( ∣ q ∣ + r b ) 2 φ ( q ) .
The sine square product averages to 1 / 4 + cos ( 2 b ( p − q ) ) / 8 ≥ 1 / 8 1/4+\cos(2b(p-q))/8\ge1/8 1/4 + cos ( 2 b ( p − q )) /8 ≥ 1/8 over this period. Boundedness of φ ( p ) ( ∣ p ∣ + r b ) 2 \varphi(p)(|p|+r_b)^2 φ ( p ) ( ∣ p ∣ + r b ) 2 makes this dominate μ ⊗ μ \mu\otimes\mu μ ⊗ μ ; its form is 2 Var μ η 2\operatorname{Var}_{\mu}\eta 2 Var μ η .
The period estimate also gives m b ( p ) ≳ φ ( p ) / ( 1 + p 2 ) m_b(p)\gtrsim\varphi(p)/(1+p^2) m b ( p ) ≳ φ ( p ) / ( 1 + p 2 ) . For the upper bound, split at ∣ q ∣ = ∣ p ∣ / 2 |q|=|p|/2 ∣ q ∣ = ∣ p ∣/2 , use k b ( p − q ) ≤ 4 / ( π b p 2 ) k_b(p-q)\le4/(\pi b p^2) k b ( p − q ) ≤ 4/ ( π b p 2 ) on the inner part and Gaussian decay outside; continuity covers bounded p p p . Hence m b ( p ) ≍ φ ( p ) / ( 1 + p 2 ) m_b(p)\asymp\varphi(p)/(1+p^2) m b ( p ) ≍ φ ( p ) / ( 1 + p 2 ) and μ ( d p ) ≤ C m b ( p ) d p \mu(dp)\le Cm_b(p)\,dp μ ( d p ) ≤ C m b ( p ) d p . The same kernel bound dominates m b ( p ) m b ( q ) m_b(p)m_b(q) m b ( p ) m b ( q ) , proving degree coercivity.
The equation L b χ b = f b \mathcal{L}_b\chi_b=f_b L b χ b = f b admits a unique mean zero weak solution in the Dirichlet space, characterized by
∬ g b ( p , q ) ( χ b ( p ) − χ b ( q ) ) ( η ( p ) − η ( q ) ) d p d q = 2 ∫ f b ( p ) η ( p ) d p . (173) \iint g_b(p,q)(\chi_b(p)-\chi_b(q))(\eta(p)-\eta(q))\,dp\,dq=2\int f_b(p)\eta(p)\,dp. \tag*{(173)} ∬ g b ( p , q ) ( χ b ( p ) − χ b ( q )) ( η ( p ) − η ( q )) d p d q = 2 ∫ f b ( p ) η ( p ) d p . (173)
The finite sixth kernel moment bounds the source in the Dirichlet norm. Coercivity embeds the mean zero completion in L 2 ( μ ) L^2(\mu) L 2 ( μ ) , so Riesz representation gives the solution.
Lemma 2 (Positive response variance). The response variance is finite, strictly positive and depends only on b b b :
V b : = Var μ χ b > 0. (174) V_b:=\operatorname{Var}_{\mu}\chi_b>0. \tag*{(174)} V b := Var μ χ b > 0. (174) Proof. Coercivity gives finite variance. The test field η ( p ) = p \eta(p)=p η ( p ) = p belongs to the Dirichlet space, and symmetry gives
∫ p f b ( p ) d p = 1 2 ∬ g b ( p , q ) ( p − q ) 4 d p d q > 0. (175) \int p f_b(p)\,dp=\frac{1}{2}\iint g_b(p,q)(p-q)^4\,dp\,dq>0. \tag*{(175)} ∫ p f b ( p ) d p = 2 1 ∬ g b ( p , q ) ( p − q ) 4 d p d q > 0. (175) A constant χ b \chi_b χ b would contradict this identity in the weak equation. Thus V b > 0 V_b>0 V b > 0 , and its defining objects depend only on b b b .
The complete local decision class
The uniform coherent residual
Use ν = Unif [ − L , L ] \nu=\operatorname{Unif}[-L,L] ν = Unif [ − L , L ] , L = 1 / 2 L=1/2 L = 1/2 , ψ z = ∣ c z ⟩ \psi_z=|c z\rangle ψ z = ∣ cz ⟩ , 0 < c ≤ a 0<c\le a 0 < c ≤ a , κ = κ F = 10 / 7 \kappa=\kappa_F=\sqrt{10/7} κ = κ F = 10/7 and K = 9 log 2 K=9\log2 K = 9 log 2 . On the support of Ω = ∫ ρ z ν ( d z ) \Omega=\int\rho_z\nu(dz) Ω = ∫ ρ z ν ( d z ) , set ρ z = ∣ ψ z ⟩ ⟨ ψ z ∣ \rho_z=|\psi_z\rangle\langle\psi_z| ρ z = ∣ ψ z ⟩ ⟨ ψ z ∣ , Ω j = ∫ z j ρ z ν ( d z ) \Omega_j=\int z^j\rho_z\nu(dz) Ω j = ∫ z j ρ z ν ( d z ) and S Ω ( X ) = { Ω , X } / 2 S_\Omega(X)=\{\Omega,X\}/2 S Ω ( X ) = { Ω , X } /2 . C1 gives S Ω ( A ) = Ω 1 S_\Omega(A)=\Omega_1 S Ω ( A ) = Ω 1 and ∥ A ∥ ≤ L \lVert A\rVert\le L ∥ A ∥ ≤ L . For the complete covariant experiment
∣ κ v + d z ⟩ ⊗ ∣ c z ⟩ , v ∈ R , (176) |\kappa v+d z\rangle\otimes|c z\rangle,\qquad v\in\mathbb{R}, \tag*{(176)} ∣ κ v + d z ⟩ ⊗ ∣ cz ⟩ , v ∈ R , (176)
the losses are 2 min { 1 , ∣ x − v ∣ / K } 2\min\{1,|x-v|/K\} 2 min { 1 , ∣ x − v ∣/ K } and ( y − z ) 2 (y-z)^2 ( y − z ) 2 , ∣ y ∣ ≤ L |y|\le L ∣ y ∣ ≤ L . Take d → 0 d\to0 d → 0 at fixed c c c and write g λ ( d ; c ) = g λ ( κ F , d , c ) g_\lambda(d;c)=g_\lambda(\kappa_F,d,c) g λ ( d ; c ) = g λ ( κ F , d , c ) . Note S gives A = Tr Ω 2 − Tr Ω A 2 \mathcal{A}=\operatorname{Tr}\Omega_2-\operatorname{Tr}\Omega A^2 A = Tr Ω 2 − Tr Ω A 2 , and Proposition C1.3 gives
B = Ω − 1 / 2 A Ω 1 / 2 , ∥ B ∥ ≤ R A . (177) B=\Omega^{-1/2}A\Omega^{1/2},\qquad\lVert B\rVert\le R_A. \tag*{(177)} B = Ω − 1/2 A Ω 1/2 , ∥ B ∥ ≤ R A . (177)
Define
R 3 = 2 ∫ { ( z I − A ) 3 , ρ z } ν ( d z ) , (178) \mathcal{R}_3 = 2\int\{(zI-A)^3,\rho_z\}\nu(dz), \tag*{(178)} R 3 = 2 ∫ {( z I − A ) 3 , ρ z } ν ( d z ) , (178)
C 3 = S Ω − 1 ( R 3 ) , J 3 = ∥ [ A , C 3 ] ∥ Ω 2 . (179) C_3 = \mathcal{S}_{\Omega}^{-1}(\mathcal{R}_3),\qquad J_3 = \lVert[A,C_3]\rVert_{\Omega}^{2}. \tag*{(179)} C 3 = S Ω − 1 ( R 3 ) , J 3 = ∥[ A , C 3 ] ∥ Ω 2 . (179)
The moment form of the cubic source is
R 3 = 4 Ω 3 − 6 { A , Ω 2 } + { A 3 , Ω } + 3 A { A , Ω } A . \mathcal{R}_3 = 4\Omega_3 - 6\{A,\Omega_2\} + \{A^3,\Omega\} + 3A\{A,\Omega\}A. R 3 = 4 Ω 3 − 6 { A , Ω 2 } + { A 3 , Ω } + 3 A { A , Ω } A .
Here Ω 1 = { Ω , A } / 2 \Omega_1=\{\Omega,A\}/2 Ω 1 = { Ω , A } /2 combines the ordered cubic products. With Z j = Ω − 1 / 2 Ω j Ω − 1 / 2 Z_j=\Omega^{-1/2}\Omega_j\Omega^{-1/2} Z j = Ω − 1/2 Ω j Ω − 1/2 and ∥ Z j ∥ ≤ L j \lVert Z_j\rVert\leq L^j ∥ Z j ∥ ≤ L j , this gives
Ω − 1 / 2 R 3 Ω − 1 / 2 = 4 Z 3 − 6 ( B Z 2 + Z 2 B † ) + 3 ( B 2 B † + B ( B † ) 2 ) + B 3 + ( B † ) 3 . (180) \Omega^{-1/2}\mathcal{R}_3\Omega^{-1/2}=4Z_3-6(BZ_2+Z_2B^\dagger)+3(B^2B^\dagger+B(B^\dagger)^2)+B^3+(B^\dagger)^3. \tag*{(180)} Ω − 1/2 R 3 Ω − 1/2 = 4 Z 3 − 6 ( B Z 2 + Z 2 B † ) + 3 ( B 2 B † + B ( B † ) 2 ) + B 3 + ( B † ) 3 . (180)
The relative source norm and ∥ C 3 ∥ \lVert C_3\rVert ∥ C 3 ∥ are at most M 3 = 4 ( L 3 + 3 R A L 2 + 2 R A 3 ) M_3=4(L^3+3R_A L^2+2R_A^3) M 3 = 4 ( L 3 + 3 R A L 2 + 2 R A 3 ) , by positivity of the inverse Sylvester map. Complete bounded Hermitian operators in ⟨ X , Y ⟩ Ω = Tr Ω ( X Y + Y X ) / 2 \langle X,Y\rangle_\Omega=\operatorname{Tr}\Omega(XY+YX)/2 ⟨ X , Y ⟩ Ω = Tr Ω ( X Y + Y X ) /2 to obtain h Ω \mathfrak{h}_\Omega h Ω . Spectral compressions are dense, so
∥ [ A , X ] ∥ Ω ≤ ( L + R A ) ∥ X ∥ Ω (181) \lVert[A,X]\rVert_\Omega\leq(L+R_A)\lVert X\rVert_\Omega \tag*{(181)} ∥[ A , X ] ∥ Ω ≤ ( L + R A ) ∥ X ∥ Ω (181)
extends the commutator continuously. For b = κ K b=\kappa K b = κ K , set
H ∗ = − χ b C 3 / 3 , E A ( H ) = ∫ ∥ [ A , H ( p ) − H ‾ ] ∥ Ω 2 μ ( d p ) , H ‾ = ∫ H d μ . H_*=-\chi_b C_3/3,\qquad\mathcal{E}_A(H)=\int\lVert[A,H(p)-\overline{H}]\rVert_\Omega^2\mu(dp),\qquad\overline{H}=\int H\,d\mu. H ∗ = − χ b C 3 /3 , E A ( H ) = ∫ ∥[ A , H ( p ) − H ] ∥ Ω 2 μ ( d p ) , H = ∫ H d μ .
The vector-valued Dirichlet form is denoted by D b , Ω D_{b,\Omega} D b , Ω , with fields identified modulo constants.
A common space for varying measurement dilations
Normalize m b ( p ) d p m_b(p)\,dp m b ( p ) d p to a probability measure m ^ b \widehat m_b m b . For fields Z : H → K Z:\mathcal{H}\to\mathcal{K} Z : H → K , write
∥ Z ∥ μ , Ω 2 = ∫ ∥ Z ( p ) ∥ Ω 2 μ ( d p ) , D b , Ω ( Z ) = ∬ g b ( p , q ) ∥ Z ( p ) − Z ( q ) ∥ Ω 2 d p d q . \lVert Z\rVert_{\mu,\Omega}^2=\int\lVert Z(p)\rVert_\Omega^2\mu(dp),\qquad D_{b,\Omega}(Z)=\iint g_b(p,q)\lVert Z(p)-Z(q)\rVert_\Omega^2\,dp\,dq. ∥ Z ∥ μ , Ω 2 = ∫ ∥ Z ( p ) ∥ Ω 2 μ ( d p ) , D b , Ω ( Z ) = ∬ g b ( p , q ) ∥ Z ( p ) − Z ( q ) ∥ Ω 2 d p d q .
Complete these fields in L 2 ( m ^ b ) L^2(\widehat m_b) L 2 ( m b ) and the Dirichlet seminorm. Scalar coercivity, applied coordinatewise, gives the Hilbert–Schmidt bounds and the embedding into L 2 ( μ ) L^2(\mu) L 2 ( μ ) .
Lemma 3 (Weighted dilation compactness). Let Ω \Omega Ω be faithful and trace class, with Tr Ω = 1 \operatorname{Tr}\Omega=1 Tr Ω = 1 . Let A = A † A=A^\dagger A = A † , ∥ A ∥ ≤ L \lVert A\rVert\leq L ∥ A ∥ ≤ L , and suppose Ω − 1 / 2 A Ω 1 / 2 \Omega^{-1/2}A\Omega^{1/2} Ω − 1/2 A Ω 1/2 has norm at most R R R . Let t k ≠ 0 t_k\neq0 t k = 0 tend to zero and let V k ( p ) : H → K k V_k(p):\mathcal{H}\to\mathcal{K}_k V k ( p ) : H → K k be measurable isometries with D b , Ω ( V k ) = O ( t k 2 ) D_{b,\Omega}(V_k)=O(t_k^2) D b , Ω ( V k ) = O ( t k 2 ) . After padding the separable output spaces and applying constant output unitaries, a subsequence and a common isometry J : H → K J:\mathcal{H}\to\mathcal{K} J : H → K satisfy
X k = ( V k − J ) / t k ⇀ X = J i H + N , J † N = 0 , X_k=(V_k-J)/t_k\rightharpoonup X=Ji_{\mathcal{H}}+N,\qquad J^\dagger N=0, X k = ( V k − J ) / t k ⇀ X = J i H + N , J † N = 0 , in the ambient weighted space, with H H H in the Hermitian weighted completion and
lim inf k t k − 2 D b , Ω ( V k ) ≥ D b , Ω ( H ) + D b , Ω ( N ) . (182) \liminf_k t_k^{-2}D_{b,\Omega}(V_k)\geq D_{b,\Omega}(H)+D_{b,\Omega}(N). \tag*{(182)} k lim inf t k − 2 D b , Ω ( V k ) ≥ D b , Ω ( H ) + D b , Ω ( N ) . (182) If, in addition, Y k = Y k † Y_k=Y_k^\dagger Y k = Y k † , ∥ Y k ∥ ≤ L \lVert Y_k\rVert\leq L ∥ Y k ∥ ≤ L , and ∥ Y k V k − V k A ∥ μ , Ω = O ( ∣ t k ∣ ) \lVert Y_kV_k-V_kA\rVert_{\mu,\Omega}=O(\lvert t_k\rvert) ∥ Y k V k − V k A ∥ μ , Ω = O (∣ t k ∣) , then a further subsequence admits a constant Hermitian weighted correction Y ∞ Y_\infty Y ∞ satisfying
lim inf k ∥ Y k V k − V k A ∥ μ , Ω 2 t k 2 ≥ ∫ ∥ Y ∞ + i [ A , H ( p ) ] ∥ Ω 2 μ ( d p ) ≥ E A ( H ) . (183) \liminf_k\frac{\lVert Y_kV_k-V_kA\rVert_{\mu,\Omega}^2}{t_k^2}\geq\int\lVert Y_\infty+i[A,H(p)]\rVert_\Omega^2\mu(dp)\geq\mathcal{E}_A(H). \tag*{(183)} k lim inf t k 2 ∥ Y k V k − V k A ∥ μ , Ω 2 ≥ ∫ ∥ Y ∞ + i [ A , H ( p )] ∥ Ω 2 μ ( d p ) ≥ E A ( H ) . (183) Both lower bounds hold jointly for every nonnegative weighted sum.
Proof. Let V ‾ k = ∫ V k d m ^ b \overline{V}_k=\int V_k\,\mathrm{d}\widehat{m}_b V k = ∫ V k d m b . Since each V k V_k V k is an isometry, 0 ⪯ V ‾ k ⪯ I 0\preceq\overline{V}_k\preceq I 0 ⪯ V k ⪯ I , and
∫ ∥ V k − V ‾ k ∥ Ω 2 d m ^ b = Tr Ω ( I − V ‾ k † V ‾ k ) = O ( t k 2 ) . \int\|V_k-\overline{V}_k\|_{\Omega}^{2}\mathrm{d}\widehat{m}_b=\operatorname{Tr}\Omega(I-\overline{V}_k^{\dagger}\overline{V}_k)=O(t_k^2). ∫ ∥ V k − V k ∥ Ω 2 d m b = Tr Ω ( I − V k † V k ) = O ( t k 2 ) .
The estimate follows from graph coercivity. Extend the polar part of V ‾ k \overline{V}_k V k to an isometry J k J_k J k ; then
∥ J k − V ‾ k ∥ Ω 2 = Tr Ω ( I − ∣ V ‾ k ∣ ) 2 ≤ Tr Ω ( I − ∣ V ‾ k ∣ 2 ) = O ( t k 2 ) . \|J_k-\overline{V}_k\|_{\Omega}^{2}=\operatorname{Tr}\Omega(I-|\overline{V}_k|)^2\leq\operatorname{Tr}\Omega(I-|\overline{V}_k|^2)=O(t_k^2). ∥ J k − V k ∥ Ω 2 = Tr Ω ( I − ∣ V k ∣ ) 2 ≤ Tr Ω ( I − ∣ V k ∣ 2 ) = O ( t k 2 ) .
After padding output complements to countably infinite dimension, constant unitaries identify J k J_k J k with J J J and conjugate Y k Y_k Y k . The resulting bounded fields X k X_k X k have a weak limit J i H + N JiH+N J i H + N , with J † N = 0 J^{\dagger}N=0 J † N = 0 .
To check that H H H is Hermitian, use V k † V k = I V_k^{\dagger}V_k=I V k † V k = I :
J † X k + X k † J = − t k X k † X k . J^{\dagger}X_k+X_k^{\dagger}J=-t_kX_k^{\dagger}X_k. J † X k + X k † J = − t k X k † X k .
For fixed eigenvectors of Ω \Omega Ω , the columns are bounded in L 2 ( m ^ b ) L^2(\widehat{m}_b) L 2 ( m b ) and the right side tends to zero in L 1 L^1 L 1 . Testing against bounded scalar functions makes J † X = i H J^{\dagger}X=iH J † X = i H skew-adjoint. Weak convergence of Dirichlet differences and orthogonality of the output components give (P.11).
Right multiplication by A A A is continuous in the weighted norm, because A Ω 1 / 2 = Ω 1 / 2 B A\Omega^{1/2}=\Omega^{1/2}B A Ω 1/2 = Ω 1/2 B with ∥ B ∥ ≤ R \|B\|\leq R ∥ B ∥ ≤ R . The triangle inequality, ∥ Y k ∥ ≤ L \|Y_k\|\leq L ∥ Y k ∥ ≤ L , and the L 2 ( μ ) L^2(\mu) L 2 ( μ ) bound on X k X_k X k therefore give
∥ Y k J − J A ∥ Ω = O ( ∣ t k ∣ ) . \|Y_kJ-JA\|_{\Omega}=O(|t_k|). ∥ Y k J − J A ∥ Ω = O ( ∣ t k ∣ ) .
Hence Z k = ( J † Y k J − A ) / t k Z_k=(J^{\dagger}Y_kJ-A)/t_k Z k = ( J † Y k J − A ) / t k has a Hermitian weak limit Y ∞ ∈ h Ω Y_{\infty}\in\mathfrak{h}_{\Omega} Y ∞ ∈ h Ω . Uniform operator bounds extend Y k J − J A → 0 Y_kJ-JA\to0 Y k J − J A → 0 from the dense range of Ω 1 / 2 \Omega^{1/2} Ω 1/2 to the whole space.
Set F k = J † Y k J − A F_k=J^{\dagger}Y_kJ-A F k = J † Y k J − A and E k = ( I − J J † ) Y k J E_k=(I-JJ^{\dagger})Y_kJ E k = ( I − J J † ) Y k J . Both operators converge strongly to zero and are uniformly bounded, with F k F_k F k self-adjoint. Projection of the scaled amplitude defect onto J H J\mathcal{H} J H gives
Z k + A J † X k − J † X k A + F k J † X k + E k † ( I − J J † ) X k . Z_k+AJ^{\dagger}X_k-J^{\dagger}X_kA+F_kJ^{\dagger}X_k+E_k^{\dagger}(I-JJ^{\dagger})X_k. Z k + A J † X k − J † X k A + F k J † X k + E k † ( I − J J † ) X k .
The last two terms vanish weakly by moving F k F_k F k , E k E_k E k onto fixed left vectors in finite rank tests, then using strong convergence, uniform L 2 L^2 L 2 bounds and density. Projection therefore gives Y ∞ + i [ A , H ] Y_{\infty}+i[A,H] Y ∞ + i [ A , H ] and the first amplitude inequality. The minimizing constant is − i [ A , ∫ H d μ ] -i[A,\int H\,\mathrm{d}\mu] − i [ A , ∫ H d μ ] by (P.10). Lower semicontinuity in the direct sum gives both bounds jointly. □
Lemma 4 (Attainable location correction). Bounded Hermitian fields H n = − χ n C 3 / 3 H_n=-\chi_nC_3/3 H n = − χ n C 3 /3 approximate H ∗ = − χ b C 3 / 3 H_*=-\chi_bC_3/3 H ∗ = − χ b C 3 /3 in the Dirichlet norm. The complete readouts W = e i d 3 H n e 2 i d p A W=e^{id^3H_n}e^{2idpA} W = e i d 3 H n e 2 i d p A attain Φ ( H n ) \Phi(H_n) Φ ( H n ) at order d 6 d^6 d 6 .
Proof. Coercivity places χ b \chi_b χ b in L 2 ( m b d p ) L^2(m_bdp) L 2 ( m b d p ) modulo constants. Bounded smooth approximations converge in energy since D b ( h ) ≤ 4 ∫ m b h 2 D_b(h)\leq4\int m_bh^2 D b ( h ) ≤ 4 ∫ m b h 2 . Bounded C 3 C_3 C 3 gives Hermitian H n H_n H n and continuity of the source functional. Gaussian moment bounds justify Taylor expansion for each H n H_n H n . Taking the approximation before d → 0 d\to0 d → 0 gives the benchmark upper limit by a diagonal choice. □
Lemma 5 (Sixth order local lower bound). For each fixed 0 < c ≤ a 0<c\leq a 0 < c ≤ a and 0 < λ < 1 0<\lambda<1 0 < λ < 1 , the gap over all complete covariant paired measurements satisfies
lim inf d → 0 g λ ( d ; c ) d 6 ≥ q λ ( c ) : = λ ( 1 − λ ) c b V b J 3 ( c ) 9 [ λ c b + ( 1 − λ ) ( L + R A ) 2 ] . (184) \liminf_{d\to0}\frac{g_{\lambda}(d;c)}{d^6}\geq q_{\lambda}(c):=\frac{\lambda(1-\lambda)c_bV_bJ_3(c)}{9[\lambda c_b+(1-\lambda)(L+R_A)^2]}. \tag*{(184)} d → 0 lim inf d 6 g λ ( d ; c ) ≥ q λ ( c ) := 9 [ λ c b + ( 1 − λ ) ( L + R A ) 2 ] λ ( 1 − λ ) c b V b J 3 ( c ) . (184) Here b = κ F K b=\kappa_FK b = κ F K , and the positive constants c b c_b c b and V b = Var μ χ b > 0 V_b=\operatorname{Var}_{\mu}\chi_b>0 V b = Var μ χ b > 0 depend only on b b b . Whenever J 3 ( c ) > 0 J_3(c)>0 J 3 ( c ) > 0 , the sixth order lower coefficient is strictly positive.
Proof. Use (S.13)–(S.14) and complete failure reports at mesh η ( d ) = o ( ∣ d ∣ 6 ) \eta(d)=o(|d|^{6}) η ( d ) = o ( ∣ d ∣ 6 ) . Write P d ( ξ ) = ∫ e − 2 i d z ξ ρ z ν ( d z ) P_d(\xi)=\int e^{-2\mathrm{i}dz\xi}\rho_z\nu(dz) P d ( ξ ) = ∫ e − 2 i d z ξ ρ z ν ( d z ) and e a m p = r a m p − A e_{\mathrm{amp}}=r_{\mathrm{amp}}-\mathcal{A} e amp = r amp − A . The reference W 0 ( p ) = e 2 i d p A W_0(p)=e^{2\mathrm{i}dpA} W 0 ( p ) = e 2 i d p A , Y = A Y=A Y = A , has zero amplitude excess and location risk r 0 ( d ) r_0(d) r 0 ( d ) . For W = V W 0 W=VW_0 W = V W 0 , set G d = W 0 ( p ) P d ( p − q ) W 0 ( q ) † G_d=W_0(p)P_d(p-q)W_0(q)^\dagger G d = W 0 ( p ) P d ( p − q ) W 0 ( q ) † , S d = ( G d + G d † ) / 2 S_d=(G_d+G_d^\dagger)/2 S d = ( G d + G d † ) /2 , T d = ( G d − G d † ) / 2 T_d=(G_d-G_d^\dagger)/2 T d = ( G d − G d † ) /2 . The difference in location risk is exactly
r l o c ( V W 0 ) − r 0 ( d ) = ∬ g b Tr [ ( V ( p ) − V ( q ) ) S d ( V ( p ) − V ( q ) ) † ] d p d q − 2 Re ∬ g b Tr [ V ( q ) † ( V ( p ) − V ( q ) ) T d ] d p d q . (185) r_{\mathrm{loc}}(VW_0)-r_0(d)=\iint g_b\operatorname{Tr}[(V(p)-V(q))S_d(V(p)-V(q))^\dagger]\,dpdq
-2\operatorname{Re}\iint g_b\operatorname{Tr}[V(q)^\dagger(V(p)-V(q))T_d]\,dpdq. \tag*{(185)} r loc ( V W 0 ) − r 0 ( d ) = ∬ g b Tr [( V ( p ) − V ( q )) S d ( V ( p ) − V ( q ) ) † ] d p d q − 2 Re ∬ g b Tr [ V ( q ) † ( V ( p ) − V ( q )) T d ] d p d q . (185)
The unital positive moment map and A Ω 1 / 2 = Ω 1 / 2 B A\Omega^{1/2}=\Omega^{1/2}B A Ω 1/2 = Ω 1/2 B give relative kernel e 2 i d p B Z d ( p − q ) e − 2 i d q B † e^{2\mathrm{i}dpB}Z_d(p-q)e^{-2\mathrm{i}dqB^\dagger} e 2 i d pB Z d ( p − q ) e − 2 i d q B † , where Z d = Ω − 1 / 2 P d Ω − 1 / 2 Z_d=\Omega^{-1/2}P_d\Omega^{-1/2} Z d = Ω − 1/2 P d Ω − 1/2 has derivatives bounded by ( 2 L ∣ p − q ∣ ) j (2L|p-q|)^j ( 2 L ∣ p − q ∣ ) j . On ∣ p ∣ , ∣ q ∣ ≤ P ( d ) = 40 log ( 1 / ∣ d ∣ ) |p|,|q|\le P(d)=\sqrt{40\log(1/|d|)} ∣ p ∣ , ∣ q ∣ ≤ P ( d ) = 40 log ( 1/∣ d ∣ ) , Taylor expansion gives
( 1 − η d ) Ω ⪯ S d ⪯ ( 1 + η d ) Ω , η d = O ( ∣ d ∣ P ( d ) ) = o ( 1 ) , (186) (1-\eta_d)\Omega\preceq S_d\preceq(1+\eta_d)\Omega,\qquad\eta_d=O(|d|P(d))=o(1), \tag*{(186)} ( 1 − η d ) Ω ⪯ S d ⪯ ( 1 + η d ) Ω , η d = O ( ∣ d ∣ P ( d )) = o ( 1 ) , (186)
T d = i d 3 ( p − q ) 3 3 R 3 + Ω 1 / 2 E d Ω 1 / 2 , ∥ E d ∥ = O ( d 4 ( 1 + ∣ p ∣ + ∣ q ∣ ) 4 ) . (187) T_d=\frac{\mathrm{i}d^3(p-q)^3}{3}\mathcal{R}_3+\Omega^{1/2}E_d\Omega^{1/2},\qquad\|E_d\|=O\left(d^4(1+|p|+|q|)^4\right). \tag*{(187)} T d = 3 i d 3 ( p − q ) 3 R 3 + Ω 1/2 E d Ω 1/2 , ∥ E d ∥ = O ( d 4 ( 1 + ∣ p ∣ + ∣ q ∣ ) 4 ) . (187)
The Personick equation cancels the first two anti-Hermitian orders. The anti-Hermitian part of e i t A ∫ e − i t z ρ z ν ( d z ) e^{\mathrm{i}tA}\int e^{-\mathrm{i}tz}\rho_z\nu(dz) e i t A ∫ e − i t z ρ z ν ( d z ) is i t 3 R 3 / 24 + O ( t 4 ) \mathrm{i}t^3\mathcal{R}_3/24+O(t^4) i t 3 R 3 /24 + O ( t 4 ) . Conjugating by e 2 i d q A e^{2\mathrm{i}dqA} e 2 i d q A and using t = 2 d ( p − q ) t=2d(p-q) t = 2 d ( p − q ) gives the stated i d 3 ( p − q ) 3 R 3 / 3 \mathrm{i}d^3(p-q)^3\mathcal{R}_3/3 i d 3 ( p − q ) 3 R 3 /3 term. Spectral tests and the bounded relative series justify the expansion. Gaussian tails, including all relative factors, contribute o ( d 12 ) o(d^{12}) o ( d 12 ) . Cauchy–Schwarz then gives
r l o c ( V W 0 ) − r 0 ( d ) ≥ ( 1 − o ( 1 ) ) D b , Ω ( V ) − C ∣ d ∣ 3 D b , Ω ( V ) 1 / 2 − o ( d 12 ) . r_{\mathrm{loc}}(VW_0)-r_0(d)\ge(1-o(1))\mathcal{D}_{b,\Omega}(V)-C|d|^3\mathcal{D}_{b,\Omega}(V)^{1/2}-o(d^{12}). r loc ( V W 0 ) − r 0 ( d ) ≥ ( 1 − o ( 1 )) D b , Ω ( V ) − C ∣ d ∣ 3 D b , Ω ( V ) 1/2 − o ( d 12 ) .
For a joint rule within o ( d 6 ) o(d^6) o ( d 6 ) of the optimum, comparison with the reference and positivity of both weights give D b , Ω ( V ) = O ( d 6 ) \mathcal{D}_{b,\Omega}(V)=O(d^6) D b , Ω ( V ) = O ( d 6 ) and e a m p = O ( d 6 ) e_{\mathrm{amp}}=O(d^6) e amp = O ( d 6 ) . Apply Lemma 3 with t = d 3 t=d^3 t = d 3 . For the amplitude defect, set D ( p ) = Y V ( p ) − V ( p ) A D(p)=YV(p)-V(p)A D ( p ) = Y V ( p ) − V ( p ) A . Since W 0 W_0 W 0 commutes with A A A ,
( Y W − W A ) Ω 1 / 2 = D Ω 1 / 2 e 2 i d p B . (YW-WA)\Omega^{1/2}=D\Omega^{1/2}e^{2\mathrm{i}dpB}. ( Y W − W A ) Ω 1/2 = D Ω 1/2 e 2 i d pB .
The last factor and its inverse are I + o ( 1 ) I+o(1) I + o ( 1 ) on the square; ∥ D ∥ ≤ 2 L \|D\|\le2L ∥ D ∥ ≤ 2 L controls the Gaussian exterior. In the source term, replacing V ( q ) † V(q)^\dagger V ( q ) † by the aligned J † J^\dagger J † has error, for bounded relative Z Z Z , at most
∥ Z ∥ ∥ V ( q ) − J ∥ Ω ∥ V ( p ) − V ( q ) ∥ Ω . \|Z\|\|V(q)-J\|_{\Omega}\|V(p)-V(q)\|_{\Omega}. ∥ Z ∥∥ V ( q ) − J ∥ Ω ∥ V ( p ) − V ( q ) ∥ Ω .
Its integral on the square is O ( d 6 P ( d ) 3 ) O(d^6P(d)^3) O ( d 6 P ( d ) 3 ) ; the cubic prefactor therefore gives o ( d 6 ) o(d^6) o ( d 6 ) , as does the Gaussian exterior. The relative bound and finite Gaussian moments make the source continuous. Joint lower semicontinuity gives
lim inf r l o c ( V W 0 ) − r 0 ( d ) d 6 ≥ Φ ( H ) = D b , Ω ( H ) + 4 3 ∫ f b ( p ) Tr [ H ( p ) R 3 ] d p , lim inf e a m p d 6 ≥ E A ( H ) . (188) \liminf\frac{r_{\mathrm{loc}}(VW_0)-r_0(d)}{d^6}\ge\Phi(H)=\mathcal{D}_{b,\Omega}(H)+\frac{4}{3}\int f_b(p)\operatorname{Tr}[H(p)\mathcal{R}_3]\,dp,\qquad\liminf\frac{e_{\mathrm{amp}}}{d^6}\ge\mathcal{E}_A(H). \tag*{(188)} lim inf d 6 r loc ( V W 0 ) − r 0 ( d ) ≥ Φ ( H ) = D b , Ω ( H ) + 3 4 ∫ f b ( p ) Tr [ H ( p ) R 3 ] d p , lim inf d 6 e amp ≥ E A ( H ) . (188)
Completing the square with S Ω ( C 3 ) = R 3 S_\Omega(C_3)=\mathcal{R}_3 S Ω ( C 3 ) = R 3 gives Φ ( H ) = D b , Ω ( H − H ∗ ) − D b , Ω ( H ∗ ) \Phi(H)=\mathcal{D}_{b,\Omega}(H-H_*)-\mathcal{D}_{b,\Omega}(H_*) Φ ( H ) = D b , Ω ( H − H ∗ ) − D b , Ω ( H ∗ ) . Lemma 4 supplies bounded approximations to H ∗ H_* H ∗ and complete readouts W = e i d 3 H W 0 W=e^{\mathrm{i}d^3H}W_0 W = e i d 3 H W 0 , giving
lim sup d → 0 r l o c ∗ ( d ) − r 0 ( d ) d 6 ≤ − D b , Ω ( H ∗ ) . (189) \limsup_{d\to0}\frac{r_{\mathrm{loc}}^*(d)-r_0(d)}{d^6}\le-\mathcal{D}_{b,\Omega}(H_*). \tag*{(189)} d → 0 lim sup d 6 r loc ∗ ( d ) − r 0 ( d ) ≤ − D b , Ω ( H ∗ ) . (189)
Subtract the location benchmark and exact amplitude optimum from the joint bound in (P.17) to obtain
lim inf d → 0 g λ ( d ; c ) d 6 ≥ inf H { λ D b , Ω ( H − H ∗ ) + ( 1 − λ ) E A ( H ) } . (190) \liminf_{d\to0}\frac{g_\lambda(d;c)}{d^6}\ge\inf_H\{\lambda\mathcal{D}_{b,\Omega}(H-H_*)+(1-\lambda)\mathcal{E}_A(H)\}. \tag*{(190)} d → 0 lim inf d 6 g λ ( d ; c ) ≥ H inf { λ D b , Ω ( H − H ∗ ) + ( 1 − λ ) E A ( H )} . (190)
Let T A T_A T A be the positive bounded operator associated with the commutator form, so that ∥ T A ∥ ≤ ( L + R A ) 2 \lVert T_A\rVert\le(L+R_A)^2 ∥ T A ∥ ≤ ( L + R A ) 2 . Graph coercivity gives
inf H { λ D b , Ω ( H − H ∗ ) + ( 1 − λ ) E A ( H ) } ≥ Var μ χ b 9 ⟨ C 3 , λ c b ( 1 − λ ) T A λ c b + ( 1 − λ ) T A C 3 ⟩ Ω . (191) \inf_H\{\lambda D_{b,\Omega}(H-H_*)+(1-\lambda)\mathcal{E}_A(H)\}\ge\frac{\operatorname{Var}_{\mu}\chi_b}{9}\left\langle C_3,\frac{\lambda c_b(1-\lambda)T_A}{\lambda c_b+(1-\lambda)T_A}C_3\right\rangle_{\Omega}. \tag*{(191)} H inf { λ D b , Ω ( H − H ∗ ) + ( 1 − λ ) E A ( H )} ≥ 9 Var μ χ b ⟨ C 3 , λ c b + ( 1 − λ ) T A λ c b ( 1 − λ ) T A C 3 ⟩ Ω . (191)
Since Var μ χ b = V b \operatorname{Var}_{\mu}\chi_b=V_b Var μ χ b = V b , bounded functional calculus gives (P.13).
Lemma 6 (Relative bound for the cubic observable). Uniformly for 0 ≤ c ≤ a 0\le c\le a 0 ≤ c ≤ a , the full residual and every unnormalized finite Fock section satisfy C1’s relative A A A bound and
∥ Ω − 1 / 2 C 3 Ω 1 / 2 ∥ ≤ R C : = 8 ( S 3 + 3 R A S 2 + 2 R A 3 ) . (192) \left\lVert\Omega^{-1/2}C_3\Omega^{1/2}\right\rVert\le R_C:=8(S_3+3R_AS_2+2R_A^3). \tag*{(192)} Ω − 1/2 C 3 Ω 1/2 ≤ R C := 8 ( S 3 + 3 R A S 2 + 2 R A 3 ) . (192) The constants S j , R A S_j,R_A S j , R A are those of (C1.4).
Proof. Equation (C1.10) of C1 gives ∥ Z j ∥ S c h ≤ S j \lVert Z_j\rVert_{\mathrm{Sch}}\le S_j ∥ Z j ∥ Sch ≤ S j , ∥ B ∥ S c h ≤ R A \lVert B\rVert_{\mathrm{Sch}}\le R_A ∥ B ∥ Sch ≤ R A . Submultiplicativity in (P.9) bounds the relative cubic source by 4 S 3 + 12 R A S 2 + 8 R A 3 4S_3+12R_AS_2+8R_A^3 4 S 3 + 12 R A S 2 + 8 R A 3 in this same norm. If ω i > 0 \omega_i>0 ω i > 0 are the support eigenvalues, the relative Sylvester solution has entries
( Ω − 1 / 2 C 3 Ω 1 / 2 ) i j = 2 ω j ω i + ω j ( Ω − 1 / 2 R 3 Ω − 1 / 2 ) i j . \left(\Omega^{-1/2}C_3\Omega^{1/2}\right)_{ij}=\frac{2\omega_j}{\omega_i+\omega_j}\left(\Omega^{-1/2}R_3\Omega^{-1/2}\right)_{ij}. ( Ω − 1/2 C 3 Ω 1/2 ) ij = ω i + ω j 2 ω j ( Ω − 1/2 R 3 Ω − 1/2 ) ij . The multiplier is at most two. Applying the same estimate to the compressed moments P N J j P N P_NJ^jP_N P N J j P N gives every unnormalized Fock section. At c = 0 c=0 c = 0 , A = C 3 = 0 A=C_3=0 A = C 3 = 0 on the vacuum support.
Returning the witness to the full residual
Let P N P_N P N project onto the first N N N Fock levels. Construct A N , C 3 , N A_N,C_{3,N} A N , C 3 , N from the unnormalized moments P N Ω j P N P_N\Omega_jP_N P N Ω j P N and extend them by zero; write Ω N = P N Ω P N \Omega_N=P_N\Omega P_N Ω N = P N Ω P N . Constants are uniform for c c c near zero.
Proposition 1 (Weighted witness return). For each fixed Fock cutoff N N N , the full and finite witnesses satisfy, as c ↓ 0 c\downarrow0 c ↓ 0 ,
∣ J 3 ( c ) − J 3 , N ( c ) ∣ = O ( c N / 4 ) , J 3 , N = ∥ [ A N , C 3 , N ] ∥ Ω N 2 . (193) \left|\sqrt{J_3(c)}-\sqrt{J_{3,N}(c)}\right|=O(c^{N/4}),\qquad J_{3,N}=\lVert[A_N,C_{3,N}]\rVert_{\Omega_N}^2. \tag*{(193)} J 3 ( c ) − J 3 , N ( c ) = O ( c N /4 ) , J 3 , N = ∥[ A N , C 3 , N ] ∥ Ω N 2 . (193) Proof. The coherent tail gives ∥ Ω j − P N Ω j P N ∥ 1 = O ( c N ) \lVert\Omega_j-P_N\Omega_jP_N\rVert_1=O(c^N) ∥ Ω j − P N Ω j P N ∥ 1 = O ( c N ) for j ≤ 3 j\le3 j ≤ 3 . The full Sylvester residual of A N A_N A N is therefore O ( c N ) O(c^N) O ( c N ) in trace norm; pairing it with A N − A A_N-A A N − A , whose norm is at most 2 L 2L 2 L , gives ∥ A N − A ∥ Ω = O ( c N / 2 ) \lVert A_N-A\rVert_{\Omega}=O(c^{N/2}) ∥ A N − A ∥ Ω = O ( c N /2 ) . The ordered cubic source, the uniform relative bounds above and
∥ A N j − A j ∥ Ω ≤ ∑ k = 0 j − 1 L j − 1 − k R A k ∥ A N − A ∥ Ω , j = 2 , 3 , \lVert A_N^j-A^j\rVert_{\Omega}\le\sum_{k=0}^{j-1}L^{j-1-k}R_A^k\lVert A_N-A\rVert_{\Omega},\qquad j=2,3, ∥ A N j − A j ∥ Ω ≤ k = 0 ∑ j − 1 L j − 1 − k R A k ∥ A N − A ∥ Ω , j = 2 , 3 , give a full cubic Sylvester residual O ( c N / 2 ) O(c^{N/2}) O ( c N /2 ) . Since both cubic observables have norm at most M 3 M_3 M 3 , the same energy pairing gives ∥ C 3 , N − C 3 ∥ Ω = O ( c N / 4 ) \lVert C_{3,N}-C_3\rVert_{\Omega}=O(c^{N/4}) ∥ C 3 , N − C 3 ∥ Ω = O ( c N /4 ) . The commutator difference expands as
A N ( C 3 , N − C 3 ) + ( A N − A ) C 3 − C 3 , N ( A N − A ) − ( C 3 , N − C 3 ) A . A_N(C_{3,N}-C_3)+(A_N-A)C_3-C_{3,N}(A_N-A)-(C_{3,N}-C_3)A. A N ( C 3 , N − C 3 ) + ( A N − A ) C 3 − C 3 , N ( A N − A ) − ( C 3 , N − C 3 ) A . The ordinary and relative bounds give error O ( c N / 4 ) O(c^{N/4}) O ( c N /4 ) . The finite commutator has identical Ω \Omega Ω and Ω N \Omega_N Ω N weights, so the triangle inequality proves (P.21).
Reference risks for the width law
Lemma 7 (Frequency envelope for the Personick reference). Let P ( t ) = ∫ e − 2 i t z ρ z ν ( d z ) P(t)=\int e^{-2itz}\rho_z\nu(dz) P ( t ) = ∫ e − 2 i t z ρ z ν ( d z ) and q A ( t ) = Re Tr ( e 2 i t A P ( t ) ) q_A(t)=\operatorname{Re}\operatorname{Tr}(e^{2itA}P(t)) q A ( t ) = Re Tr ( e 2 i t A P ( t )) . The complete field W A ( p ) = e 2 i d p A W_A(p)=e^{2idpA} W A ( p ) = e 2 i d p A has zero amplitude excess, and
r l o c ⋆ ≥ 2 − 2 ∫ k b ( ξ ) e − ξ 2 / 2 ∥ P ( d ξ ) ∥ 1 d ξ . (194) r_{\mathrm{loc}}^{\star}\ge2-2\int k_b(\xi)e^{-\xi^2/2}\lVert P(d\xi)\rVert_1d\xi. \tag*{(194)} r loc ⋆ ≥ 2 − 2 ∫ k b ( ξ ) e − ξ 2 /2 ∥ P ( d ξ ) ∥ 1 d ξ . (194) r l o c ( W A ) = 2 − 2 ∫ k b ( ξ ) e − ξ 2 / 2 q A ( d ξ ) d ξ . (195) r_{\mathrm{loc}}(W_A)=2-2\int k_b(\xi)e^{-\xi^2/2}q_A(d\xi)d\xi. \tag*{(195)} r loc ( W A ) = 2 − 2 ∫ k b ( ξ ) e − ξ 2 /2 q A ( d ξ ) d ξ . (195) With δ A ( t ) = ∥ P ( t ) ∥ 1 − q A ( t ) \delta_A(t)=\lVert P(t)\rVert_1-q_A(t) δ A ( t ) = ∥ P ( t ) ∥ 1 − q A ( t ) ,
0 ≤ δ A ( t ) ≤ 2 A L ( c ) t 2 , 0 ≤ e l o c ( W A ) ≤ 2 ∫ k b ( ξ ) e − ξ 2 / 2 δ A ( d ξ ) d ξ . 0\le\delta_A(t)\le2A_L(c)t^2,\qquad0\le e_{\mathrm{loc}}(W_A)\le2\int k_b(\xi)e^{-\xi^2/2}\delta_A(d\xi)d\xi. 0 ≤ δ A ( t ) ≤ 2 A L ( c ) t 2 , 0 ≤ e loc ( W A ) ≤ 2 ∫ k b ( ξ ) e − ξ 2 /2 δ A ( d ξ ) d ξ . Proof. In (S.13), W ( q ) † W ( p ) W(q)^\dagger W(p) W ( q ) † W ( p ) is a contraction. Trace duality and ∫ φ ( p ) φ ( p − ξ ) d p = e − ξ 2 / 2 \int\varphi(p)\varphi(p-\xi)dp=e^{-\xi^2/2} ∫ φ ( p ) φ ( p − ξ ) d p = e − ξ 2 /2 give the two risk bounds. Spectral calculus yields 1 − q A ( t ) ≤ 2 t 2 ∫ ∥ ( A − z ) ψ z ∥ 2 ν ( d z ) = 2 t 2 A L ( c ) 1-q_A(t)\le2t^2\int\lVert(A-z)\psi_z\rVert^2\nu(dz)=2t^2A_L(c) 1 − q A ( t ) ≤ 2 t 2 ∫ ∥( A − z ) ψ z ∥ 2 ν ( d z ) = 2 t 2 A L ( c ) . Use ∥ P ( t ) ∥ 1 ≤ 1 \lVert P(t)\rVert_1\le1 ∥ P ( t ) ∥ 1 ≤ 1 for the envelope, and subtract the location lower bound. The amplitude square in (S.14) vanishes because W A W_A W A commutes with A A A .
The uniform prior witness for the cubic width law
A noncommuting cubic response at weak residual sensitivity
For ν = Unif [ − 1 / 2 , 1 / 2 ] \nu=\operatorname{Unif}[-1/2,1/2] ν = Unif [ − 1/2 , 1/2 ] , the moments are
m j = ∫ z j ν ( d z ) , m 2 = 1 12 , m 4 = 1 80 , m 6 = 1 448 . m_j=\int z^j\nu(dz),\qquad m_2=\frac{1}{12},\qquad m_4=\frac{1}{80},\qquad m_6=\frac{1}{448}. m j = ∫ z j ν ( d z ) , m 2 = 12 1 , m 4 = 80 1 , m 6 = 448 1 .
Use Note P’s operators Ω j ( c ) \Omega_j(c) Ω j ( c ) , A ( c ) A(c) A ( c ) , C 3 ( c ) C_3(c) C 3 ( c ) on the average state support.
Theorem 1 (Uniform prior cubic witness). For the uniform residual prior, as c ↓ 0 c\downarrow0 c ↓ 0 ,
J 3 ( c ) : = ∥ [ A ( c ) , C 3 ( c ) ] ∥ Ω ( c ) 2 = 128 1575 2 c 4 + O ( c 6 ) . J_3(c):=\lVert[A(c),C_3(c)]\rVert_{\Omega(c)}^2=\frac{128}{1575^2}c^4+O(c^6). J 3 ( c ) := ∥[ A ( c ) , C 3 ( c )] ∥ Ω ( c ) 2 = 157 5 2 128 c 4 + O ( c 6 ) . For every fixed 0 < λ < 1 0<\lambda<1 0 < λ < 1 , Note P’s local lower coefficient q λ ( c ) q_\lambda(c) q λ ( c ) is strictly positive for all sufficiently small c > 0 c>0 c > 0 .
Proof. In the unnormalized N = 16 N=16 N = 16 Fock section, the moment matrices are analytic and Ω N > 0 \Omega_N>0 Ω N > 0 for c > 0 c>0 c > 0 : a nonzero polynomial has positive squared integral on the prior interval. Note P’s uniform bounds on A N , C 3 , N A_N,C_{3,N} A N , C 3 , N remove their meromorphic Sylvester singularities at zero.
Prior parity makes A N , C 3 , N A_N,C_{3,N} A N , C 3 , N odd in c c c and off-diagonal in Fock parity. Write A N = c A ( 1 ) + O ( c 3 ) A_N=cA^{(1)}+O(c^3) A N = c A ( 1 ) + O ( c 3 ) , C 3 , N = c C 3 ( 1 ) + O ( c 3 ) C_{3,N}=cC_3^{(1)}+O(c^3) C 3 , N = c C 3 ( 1 ) + O ( c 3 ) , and set a i j = A i j ( 1 ) a_{ij}=A^{(1)}_{ij} a ij = A ij ( 1 ) , c i j = ( C 3 ( 1 ) ) i j c_{ij}=(C_3^{(1)})_{ij} c ij = ( C 3 ( 1 ) ) ij . The moment expansion is
( Ω j ) k l = c k + l k ! l ! ∑ r = 0 ∞ ( − c 2 ) r r ! m k + l + j + 2 r . (\Omega_j)_{kl}=\frac{c^{k+l}}{\sqrt{k!l!}}\sum_{r=0}^{\infty}\frac{(-c^2)^r}{r!}m_{k+l+j+2r}. ( Ω j ) k l = k ! l ! c k + l r = 0 ∑ ∞ r ! ( − c 2 ) r m k + l + j + 2 r .
At order c c c the vacuum-row equations give a 01 = 2 m 2 a_{01}=2m_2 a 01 = 2 m 2 , c 01 = 8 ( m 4 − 3 m 2 2 ) c_{01}=8(m_4-3m_2^2) c 01 = 8 ( m 4 − 3 m 2 2 ) , and all a 0 j = c 0 j = 0 a_{0j}=c_{0j}=0 a 0 j = c 0 j = 0 for j ≥ 3 j\ge3 j ≥ 3 . The ( 1 , 2 ) (1,2) ( 1 , 2 ) Personick equation at order c 3 c^3 c 3 is
1 2 ( m 2 a 12 + m 2 a 01 2 ) = m 4 2 . \frac{1}{2}\left(m_2a_{12}+\frac{m_2a_{01}}{\sqrt{2}}\right)=\frac{m_4}{\sqrt{2}}. 2 1 ( m 2 a 12 + 2 m 2 a 01 ) = 2 m 4 .
The remaining a 1 j a_{1j} a 1 j , j ≥ 3 j \ge3 j ≥ 3 , vanish by the same equation and parity.
If r 12 ( 3 ) r^{(3)}_{12} r 12 ( 3 ) is the c 3 c^3 c 3 coefficient of the cubic right side, retaining its operator order gives
r 12 ( 3 ) = 4 m 6 2 − 6 ( a 01 m 4 2 + m 4 a 12 ) + 6 m 2 a 01 a 12 . r^{(3)}_{12} = \frac{4m_6}{\sqrt{2}} - 6\left(\frac{a_{01}m_4}{\sqrt{2}} + m_4a_{12}\right) + 6m_2a_{01}a_{12}. r 12 ( 3 ) = 2 4 m 6 − 6 ( 2 a 01 m 4 + m 4 a 12 ) + 6 m 2 a 01 a 12 .
The term involving A 3 Ω + Ω A 3 A^3\Omega+\Omega A^3 A 3 Ω + Ω A 3 starts at order c 5 c^5 c 5 in this entry. The cubic Sylvester equation gives
1 2 ( m 2 c 12 + m 2 c 01 / 2 ) = r 12 ( 3 ) . \frac{1}{2}\left(m_2c_{12}+m_2c_{01}/\sqrt{2}\right)=r^{(3)}_{12}. 2 1 ( m 2 c 12 + m 2 c 01 / 2 ) = r 12 ( 3 ) .
Substitution gives
a 01 = 1 6 , a 12 = 2 15 , c 01 = − 1 15 , c 12 = 2 2 525 . a_{01}=\frac{1}{6},\qquad a_{12}=\frac{\sqrt{2}}{15},\qquad c_{01}=-\frac{1}{15},\qquad c_{12}=\frac{2\sqrt{2}}{525}. a 01 = 6 1 , a 12 = 15 2 , c 01 = − 15 1 , c 12 = 525 2 2 .
The vanishing a 1 j a_{1j} a 1 j also gives c 1 j = 0 c_{1j}=0 c 1 j = 0 for j ≥ 4 j\ge4 j ≥ 4 .
The two adjacent Fock transitions give
[ A N , C 3 , N ] 02 = c 2 ( a 01 c 12 − c 01 a 12 ) + O ( c 4 ) = 8 2 1575 c 2 + O ( c 4 ) . \begin{aligned}
[A_N,C_{3,N}]_{02}=c^2(a_{01}c_{12}-c_{01}a_{12})+O(c^4) \\
&=\frac{8\sqrt{2}}{1575}c^2+O(c^4).
\end{aligned} [ A N , C 3 , N ] 02 = c 2 ( a 01 c 12 − c 01 a 12 ) + O ( c 4 ) = 1575 8 2 c 2 + O ( c 4 ) .
These exhaust the leading vacuum transitions. The vacuum weight is 1 + O ( c 2 ) 1+O(c^2) 1 + O ( c 2 ) and the other weights are O ( c 2 ) O(c^2) O ( c 2 ) , so parity gives
J 3 , N ( c ) = 128 1575 2 c 4 + O ( c 6 ) . J_{3,N}(c)=\frac{128}{1575^2}c^4+O(c^6). J 3 , N ( c ) = 157 5 2 128 c 4 + O ( c 6 ) .
The uniform relative bounds in Lemma 6 and Proposition 1 give ∣ J 3 − J 3 , 16 ∣ = O ( c 4 ) \left|\sqrt{J_3}-\sqrt{J_{3,16}}\right|=O(c^4) J 3 − J 3 , 16 = O ( c 4 ) . As J 3 , 16 = O ( c 2 ) \sqrt{J_{3,16}}=O(c^2) J 3 , 16 = O ( c 2 ) , this changes the witness by O ( c 6 ) O(c^6) O ( c 6 ) . Note P then gives q λ ( c ) > 0 q_\lambda(c)>0 q λ ( c ) > 0 for fixed 0 < λ < 1 0<\lambda<1 0 < λ < 1 and all sufficiently small c > 0 c>0 c > 0 .
Proof of the cubic width law under the full prior
Note S gives the limit as s → ∞ s\to\infty s → ∞ at fixed 0 < w < 1 0<w<1 0 < w < 1 . We match the lower bound from P and B with a complete upper readout, then let w ↓ 0 w\downarrow0 w ↓ 0 . For complete rises write g λ ( d ; c ) = g λ ( κ F , d , c ) g_\lambda(d;c)=g_\lambda(\kappa_F,d,c) g λ ( d ; c ) = g λ ( κ F , d , c ) , κ F = 10 / 7 \kappa_F=\sqrt{10/7} κ F = 10/7 ; g λ , w ( u ) g_{\lambda,w}(u) g λ , w ( u ) includes truncated rises.
A complete sixth order upper readout
Use Proposition C1.3’s bound ∥ Ω − 1 / 2 A Ω 1 / 2 ∥ ≤ R A \|\Omega^{-1/2}A\Omega^{1/2}\|\le R_A ∥ Ω − 1/2 A Ω 1/2 ∥ ≤ R A on c 2 L 2 ≤ log 2 c^2L^2\le\log2 c 2 L 2 ≤ log 2 , including c = 0 c=0 c = 0 , and write
P ( t ) = ∫ e − 2 i t z ρ z ν ( d z ) . P(t)=\int e^{-2itz}\rho_z\nu(dz). P ( t ) = ∫ e − 2 i t z ρ z ν ( d z ) .
Lemma 1 (Uniform sixth order reference regret). For L = 1 / 2 L=1/2 L = 1/2 and 0 ≤ c ≤ a 0\le c\le a 0 ≤ c ≤ a , a constant C = C ( L , R A ) C=C(L,R_A) C = C ( L , R A ) satisfies
δ A ( t ) = ∥ P ( t ) ∥ 1 − Re Tr ( e 2 i t A P ( t ) ) , \delta_A(t)=\|P(t)\|_1-\operatorname{Re}\operatorname{Tr}(e^{2itA}P(t)), δ A ( t ) = ∥ P ( t ) ∥ 1 − Re Tr ( e 2 i t A P ( t )) ,
0 ≤ δ A ( t ) ≤ C t 6 ( t ∈ R ) . 0\le\delta_A(t)\le Ct^6\qquad(t\in\mathbb{R}). 0 ≤ δ A ( t ) ≤ C t 6 ( t ∈ R ) .
The complete reference field W A = e 2 i d p A W_A=e^{2idp_A} W A = e 2 i d p A has e a m p = 0 e_{\mathrm{amp}}=0 e amp = 0 and e l o c ≤ 2 C J 6 ( κ K ) d 6 e_{\mathrm{loc}}\le2CJ_6(\kappa_K)d^6 e loc ≤ 2 C J 6 ( κ K ) d 6 , where J 6 ( b ) = ∫ k b ( ξ ) e − ξ 2 / 2 c 6 d ξ ≤ 16 / π J_6(b)=\int k_b(\xi)e^{-\xi^2/2c^6}d\xi\le16/\pi J 6 ( b ) = ∫ k b ( ξ ) e − ξ 2 /2 c 6 d ξ ≤ 16/ π . Proof. Set H A = 2 ( R A + L ) H_A = 2(R_A + L) H A = 2 ( R A + L ) , v A = L 2 / 3 v_A = L^2/3 v A = L 2 /3 and T = e 2 i t A P ( t ) = H R + i H I T = e^{2itA}P(t) = H_R + iH_I T = e 2 i t A P ( t ) = H R + i H I , with H R , H I H_R,H_I H R , H I self-adjoint. With x = H A ∣ t ∣ x = H_A|t| x = H A ∣ t ∣ , relative Taylor expansion and Personick cancellation give H R ⪰ ( 2 − e x ) Ω H_R \succeq(2-e^x)\Omega H R ⪰ ( 2 − e x ) Ω and ∥ Ω − 1 / 2 H I Ω − 1 / 2 ∥ ≤ x 3 e x / 6 \|\Omega^{-1/2}H_I\Omega^{-1/2}\| \le x^3e^x/6 ∥ Ω − 1/2 H I Ω − 1/2 ∥ ≤ x 3 e x /6 . For H R ⪰ η 0 Ω H_R \succeq\eta_0\Omega H R ⪰ η 0 Ω and H I = Ω 1 / 2 Z Ω 1 / 2 H_I=\Omega^{1/2}Z\Omega^{1/2} H I = Ω 1/2 Z Ω 1/2 , a contraction V V V has excess reward at most − η 0 ϵ V 2 / 2 + ϵ V ∥ Z ∥ -\eta_0\epsilon_V^2/2+\epsilon_V\|Z\| − η 0 ϵ V 2 /2 + ϵ V ∥ Z ∥ , with ϵ V = ∥ V − I ∥ Ω \epsilon_V=\|V-I\|_\Omega ϵ V = ∥ V − I ∥ Ω . Trace duality gives ∥ H R + i H I ∥ 1 − Tr H R ≤ ∥ Z ∥ 2 / ( 2 η 0 ) \|H_R+iH_I\|_1-\operatorname{Tr}H_R\le\|Z\|^2/(2\eta_0) ∥ H R + i H I ∥ 1 − Tr H R ≤ ∥ Z ∥ 2 / ( 2 η 0 ) . Thus δ A ( t ) ≤ H A 6 t 6 / 16 \delta_A(t)\le H_A^6t^6/16 δ A ( t ) ≤ H A 6 t 6 /16 for ∣ t ∣ ≤ log ( 3 / 2 ) / H A |t|\le\log(3/2)/H_A ∣ t ∣ ≤ log ( 3/2 ) / H A . Elsewhere δ A ( t ) ≤ 2 v A t 2 \delta_A(t)\le2v_At^2 δ A ( t ) ≤ 2 v A t 2 from Lemma 7. One may take C = max { H A 6 / 16 , 2 v A H A 4 / log ( 3 / 2 ) 4 } C=\max\{H_A^6/16,2v_AH_A^4/\log(3/2)^4\} C = max { H A 6 /16 , 2 v A H A 4 / log ( 3/2 ) 4 } . Integrating against Note P’s frequency envelope gives the regret bound. The kernel bound follows from k b ( ξ ) ≤ 1 / ( π ∣ ξ ∣ ) k_b(\xi)\le1/(\pi|\xi|) k b ( ξ ) ≤ 1/ ( π ∣ ξ ∣ ) and ∫ ∣ ξ ∣ 5 e − ξ 2 / 2 d ξ = 16 \int|\xi|^5e^{-\xi^2/2}\,d\xi=16 ∫ ∣ ξ ∣ 5 e − ξ 2 /2 d ξ = 16 , uniformly for every b > 0 b>0 b > 0 .
Integrating complete and terminal rises under the full prior
For the quintic profile, set
I F = ∫ 0 1 F 2 = 181 462 , κ F 2 = 10 7 , I_F=\int_0^1F^2=\frac{181}{462},\qquad\kappa_F^2=\frac{10}{7}, I F = ∫ 0 1 F 2 = 462 181 , κ F 2 = 7 10 ,
γ F = a 2 4 κ F 2 = 7 log 2 10 , ϑ = 1 − I F + 1 4 κ F 2 = 7237 9240 . \gamma_F=\frac{a^2}{4\kappa_F^2}=\frac{7\log2}{10},\qquad\vartheta=1-I_F+\frac{1}{4\kappa_F^2}=\frac{7237}{9240}. γ F = 4 κ F 2 a 2 = 10 7 log 2 , ϑ = 1 − I F + 4 κ F 2 1 = 9240 7237 .
Theorem 2 (Complete upper readout under the full prior). At every sufficiently small fixed w > 0 w>0 w > 0 , the strong signal limit admits a complete common readout with full prior excesses e a m p = 0 e_{\mathrm{amp}}=0 e amp = 0 and e l o c = O ( w 3 ) e_{\mathrm{loc}}=O(w^3) e loc = O ( w 3 ) , including terminally truncated rises.
Proof. Pair W A = e 2 i d p A W_A=e^{2idp_A} W A = e 2 i d p A with the exact Personick report in each fiber. For complete rises, u ≤ 1 − w u\le1-w u ≤ 1 − w ,
κ = κ F , d = − γ F w , c 2 = a 2 ( 1 − u − ϑ w ) , c 2 / 4 ≤ log 2. \kappa=\kappa_F,\qquad d=-\sqrt{\gamma_Fw},\qquad c^2=a^2(1-u-\vartheta w),\qquad c^2/4\le\log2. κ = κ F , d = − γ F w , c 2 = a 2 ( 1 − u − ϑw ) , c 2 /4 ≤ log 2.
The uniform R A R_A R A bound gives e l o c ≤ C 0 d 6 = C 0 γ F 3 w 3 e_{\mathrm{loc}}\le C_0d^6=C_0\gamma_F^3w^3 e loc ≤ C 0 d 6 = C 0 γ F 3 w 3 , with C 0 C_0 C 0 independent of u , w u,w u , w . For truncated rises, set τ = ( 1 − u ) / w \tau=(1-u)/w τ = ( 1 − u ) / w . Orthogonal decomposition gives
c 2 ≤ a 2 w I F , d 2 = a 2 w F ( τ ) 4 4 ∫ 0 τ ( F ′ ) 2 ≤ a 2 w I F . c^2\le a^2wI_F,\qquad d^2=\frac{a^2wF(\tau)^4}{4\int_0^\tau(F')^2}\le a^2wI_F. c 2 ≤ a 2 w I F , d 2 = 4 ∫ 0 τ ( F ′ ) 2 a 2 w F ( τ ) 4 ≤ a 2 w I F .
The last inequality is Cauchy–Schwarz applied to F ( τ ) 2 / 2 = ∫ 0 τ F F ′ F(\tau)^2/2=\int_0^\tau FF' F ( τ ) 2 /2 = ∫ 0 τ F F ′ . For small w w w , the uniform R A R_A R A and J 6 ≤ 16 / π J_6\le16/\pi J 6 ≤ 16/ π bounds give terminal regret at most C 0 ( a 2 I F ) 3 w 3 C_0(a^2I_F)^3w^3 C 0 ( a 2 I F ) 3 w 3 , also as κ ↓ 0 \kappa\downarrow0 κ ↓ 0 . Its region has mass w w w , hence integrated regret O ( w 4 ) O(w^4) O ( w 4 ) . The endpoint u = 1 u=1 u = 1 has zero mass; Note S integrates the complete fields or their measurable approximations for all three risks.
Corollary 3 (Ordered cubic width law under the full prior). For fixed 0 < λ < 1 0<\lambda<1 0 < λ < 1 , the full prior gap satisfies G λ ( w ) = Θ λ ( w 3 ) G_\lambda(w)=\Theta_\lambda(w^3) G λ ( w ) = Θ λ ( w 3 ) as w ↓ 0 w\downarrow0 w ↓ 0 .
Proof. Note B, Theorem 1, gives J 3 ( c ) = 128 c 4 / 1575 2 + O ( c 6 ) J_3(c)=128c^4/1575^2+O(c^6) J 3 ( c ) = 128 c 4 /157 5 2 + O ( c 6 ) . Choose c 0 > 0 c_0>0 c 0 > 0 such that J 3 ( c ) ≥ 64 c 4 / 1575 2 J_3(c)\ge64c^4/1575^2 J 3 ( c ) ≥ 64 c 4 /157 5 2 for 0 < c ≤ c 0 0<c\le c_0 0 < c ≤ c 0 , and fix 0 < c − < c + < min { c 0 , a } 0<c_-<c_+<\min\{c_0,a\} 0 < c − < c + < min { c 0 , a } . Set C = [ c − , c + ] C=[c_-,c_+] C = [ c − , c + ] . For all sufficiently small w w w , the positions
u w ( c ) = 1 − c 2 a 2 − ϑ w , c ∈ C , u_w(c)=1-\frac{c^2}{a^2}-\vartheta w,\qquad c\in C, u w ( c ) = 1 − a 2 c 2 − ϑw , c ∈ C ,
lie in ( 0 , 1 − w ) (0,1-w) ( 0 , 1 − w ) . This change of variables in the uniform position integral fixes the residual at c c c and has Jacobian 2 c / a 2 2c/a^2 2 c / a 2 . Note P applies to either sign of d d d , giving
lim inf w ↓ 0 g λ ( − γ F w ; c ) w 3 ≥ γ F 3 q λ ( c ) > 0. (196) \liminf_{w\downarrow0}\frac{g_\lambda(-\sqrt{\gamma_Fw};c)}{w^3}\ge\gamma_F^3q_\lambda(c)>0. \tag*{(196)} w ↓ 0 lim inf w 3 g λ ( − γ F w ; c ) ≥ γ F 3 q λ ( c ) > 0. (196)
For fixed b = κ F K b=\kappa_F K b = κ F K , the constants c b c_b c b , V b V_b V b are fixed. The uniform R A R_A R A bound and J 3 ( c ) ≥ 64 c 4 / 1575 2 J_3(c)\ge64c^4/1575^2 J 3 ( c ) ≥ 64 c 4 /157 5 2 in (P.13) give q λ ( c ) ≥ q ∗ > 0 q_\lambda(c)\ge q_*>0 q λ ( c ) ≥ q ∗ > 0 on C C C . By Note S, G ‾ λ ( w ) = ∫ 0 1 g λ , w ( u ) d u \overline{G}_\lambda(w)=\int_0^1 g_{\lambda,w}(u)\,du G λ ( w ) = ∫ 0 1 g λ , w ( u ) d u . Nonnegativity and Fatou’s lemma give
lim inf w → 0 G ‾ λ ( w ) w 3 ≥ γ F 3 ∫ C q λ ( c ) 2 c a 2 d c > 0. \liminf_{w\to0}\frac{\overline{G}_\lambda(w)}{w^3}\ge\gamma_F^3\int_C q_\lambda(c)\frac{2c}{a^2}\,dc>0. w → 0 lim inf w 3 G λ ( w ) ≥ γ F 3 ∫ C q λ ( c ) a 2 2 c d c > 0.
The complete upper readout gives G ‾ λ ≤ λ e l o c = O ( w 3 ) \overline{G}_\lambda\le\lambda e_{\mathrm{loc}}=O(w^3) G λ ≤ λ e loc = O ( w 3 ) .
The finite phase record certificate
The finite phase record certificate
Fix s = s c s=s_c s = s c , equal task weights and known L ∈ [ 7 / 16 , 9 / 16 ] L\in[7/16,9/16] L ∈ [ 7/16 , 9/16 ] , with the main text calibration and losses.
Certificate for the original record of 16 qubits
Physical primal and dual bounds
The record lies in the span of vacuum and the orthonormal vectors ∣ e j , l ⟩ = ∣ 0 ⟩ ⊗ ( j − 1 ) ∣ 1 ⟩ ∣ D l n − j ⟩ \lvert e_{j,l}\rangle=\lvert0\rangle^{\otimes(j-1)}\lvert1\rangle\lvert D_l^{n-j}\rangle ∣ e j , l ⟩ = ∣ 0 ⟩ ⊗ ( j − 1 ) ∣ 1 ⟩ ∣ D l n − j ⟩ , 1 ≤ j ≤ n 1\le j\le n 1 ≤ j ≤ n , 0 ≤ l ≤ n − j 0\le l\le n-j 0 ≤ l ≤ n − j , where D l m D_l^m D l m is the normalized Dicke vector with l l l excitations. For n = 16 n=16 n = 16 , the dimension is 1 + n ( n + 1 ) / 2 = 137 1+n(n+1)/2=137 1 + n ( n + 1 ) /2 = 137 . In the phase model, q 0 = 1 / 4 q_0=1/4 q 0 = 1/4 and α = θ / 2 = ( s + a z ) / n \alpha=\theta/2=(s+az)/\sqrt{n} α = θ /2 = ( s + a z ) / n . Grouping by the first excitation gives
⟨ 0 n ∣ Ψ k , α ⟩ = ( cos α ) n − k + 1 , ⟨ e j , l ∣ Ψ k , α ⟩ = { ( n − j l ) ( cos α ) n − k − l ( sin α ) l + 1 , j ≥ k , 0 , j < k . \langle0^n\vert\Psi_{k,\alpha}\rangle=(\cos\alpha)^{n-k+1},\qquad
\langle e_{j,l}\vert\Psi_{k,\alpha}\rangle=
\begin{cases}
\sqrt{\binom{n-j}{l}}(\cos\alpha)^{n-k-l}(\sin\alpha)^{l+1}, & j\ge k,\\
0, & j<k.
\end{cases} ⟨ 0 n ∣ Ψ k , α ⟩ = ( cos α ) n − k + 1 , ⟨ e j , l ∣ Ψ k , α ⟩ = { ( l n − j ) ( cos α ) n − k − l ( sin α ) l + 1 , 0 , j ≥ k , j < k .
Zeroth powers equal one.
For normalized Gauss quadrature with q = 32 q=32 q = 32 , put ω = 2 a n \omega=2a\sqrt{n} ω = 2 a n and C q = ( q ! ) 4 / [ ( 2 q + 1 ) ( ( 2 q ) ! ) 3 ] C_q=(q!)^4/[(2q+1)((2q)!)^3] C q = ( q ! ) 4 / [( 2 q + 1 ) (( 2 q )! ) 3 ] . The generator norm is at most a n a\sqrt{n} a n , hence ∥ ∂ z m ρ ∥ 1 ≤ ω m \|\partial_z^m\rho\|_1\le\omega^m ∥ ∂ z m ρ ∥ 1 ≤ ω m and, for r = 0 , 1 , 2 r=0,1,2 r = 0 , 1 , 2 ,
e r = C q ( 2 L ) 2 q ∑ j = 0 r ( 2 q j ) r ! L r − j ω 2 q − j ( r − j ) ! , E l o c = 2 e 0 , E a m p = L 2 e 0 + 2 L e 1 + e 2 . e_r=C_q(2L)^{2q}\sum_{j=0}^{r}\binom{2q}{j}\frac{r!L^{r-j}\omega^{2q-j}}{(r-j)!},\qquad E_{\mathrm{loc}}=2e_0,\qquad E_{\mathrm{amp}}=L^2e_0+2Le_1+e_2. e r = C q ( 2 L ) 2 q j = 0 ∑ r ( j 2 q ) ( r − j )! r ! L r − j ω 2 q − j , E loc = 2 e 0 , E amp = L 2 e 0 + 2 L e 1 + e 2 .
For a normalized dyadic contraction K K K into dimension D D D , let X X X be its successful weighted coordinates. If ∥ X − X c ∥ F ≤ e \|X-X_c\|_F\le e ∥ X − X c ∥ F ≤ e , ∣ z − z c ∣ ≤ e z |z-z_c|\le e_z ∣ z − z c ∣ ≤ e z and the location table error is e ℓ e_\ell e ℓ , the cost allowances are
ζ l o c = 2 e ( 2 + e ) + e ℓ ( 1 + e ) 2 , ζ a m p = 4 L 2 e ( 2 + e ) + 4 L e z ( 1 + e ) 2 . \zeta_{\mathrm{loc}}=2e(2+e)+e_\ell(1+e)^2,\qquad
\zeta_{\mathrm{amp}}=4L^2e(2+e)+4Le_z(1+e)^2. ζ loc = 2 e ( 2 + e ) + e ℓ ( 1 + e ) 2 , ζ amp = 4 L 2 e ( 2 + e ) + 4 L e z ( 1 + e ) 2 .
The lower risks include D ζ j + B j 2 δ + E j D\zeta_j+B_j\sqrt{2\delta}+E_j D ζ j + B j 2 δ + E j and the exact dual repairs, where δ \delta δ is compression failure mass. Joint fees are the weighted task fees.
Set S r = X c diag ( z c r ) X c † S_r=X_c\operatorname{diag}(z_c^r)X_c^\dagger S r = X c diag ( z c r ) X c † and E A = S 0 A ^ b + A ^ S 0 − 2 S 1 E_A=S_0\widehat{A}_b+\widehat{A}S_0-2S_1 E A = S 0 A b + A S 0 − 2 S 1 for a dyadic Hermitian A ^ \widehat{A} A . The amplitude dual Z = S 2 − A ^ S 0 A ^ − L ∥ E A ∥ I Z=S_2-\widehat{A}S_0\widehat{A}-L\|E_A\|I Z = S 2 − A S 0 A − L ∥ E A ∥ I is feasible for ∣ y ∣ ≤ L |y|\le L ∣ y ∣ ≤ L : its slack is ( y I − A ^ ) S 0 ( y I − A ^ ) + y E A + L ∥ E A ∥ I ⪰ 0 (yI-\widehat{A})S_0(yI-\widehat{A})+yE_A+L\|E_A\|I\succeq0 ( y I − A ) S 0 ( y I − A ) + y E A + L ∥ E A ∥ I ⪰ 0 . Exact integer residuals certify the joint dual at integer locations and 129 amplitude nodes. Subtracting ( 1 − λ ) Δ y 2 S 0 / 4 (1-\lambda)\Delta_y^2S_0/4 ( 1 − λ ) Δ y 2 S 0 /4 , Δ y = 2 L / 128 \Delta_y=2L/128 Δ y = 2 L /128 , extends feasibility to all amplitude reports. The position cost is concave between integer reports, so those actions suffice.
Normalize the upper frame by its exact Gram bound, pull it through K K K and add the positive complement with fallback amplitude zero. Starting from prior risk L 2 / 3 L^2/3 L 2 /3 , the allowance for effect improvements is 2 D ζ j + E j 2D\zeta_j+E_j 2 D ζ j + E j . The gap lower bound is R − λ r ‾ l o c − ( 1 − λ ) r ‾ a m p R-\lambda\overline{r}_{\mathrm{loc}}-(1-\lambda)\overline{r}_{\mathrm{amp}} R − λ r loc − ( 1 − λ ) r amp . The node endpoints use exact dyadic widths and actions and ball enclosures at 256 bit precision.
Interpolation in prior width
On I L = [ 7 / 16 , 9 / 16 ] I_L=[7/16,9/16] I L = [ 7/16 , 9/16 ] , put b n = sin 2 ( s c / n ) b_n=\sin^2(s_c/\sqrt{n}) b n = sin 2 ( s c / n ) , ξ = z / L \xi=z/L ξ = z / L and y = L η y=L\eta y = L η . Fixed POVM effects and normalized reports η \eta η give losses 2 min { 1 , b n ∣ k ^ − k ∣ / K } 2\min\{1,b_n|\hat{k}-k|/K\} 2 min { 1 , b n ∣ k ^ − k ∣/ K } and L 2 ( η − ξ ) 2 L^2(\eta-\xi)^2 L 2 ( η − ξ ) 2 on a common report space. For N = n − k + 1 N=n-k+1 N = n − k + 1 changed sites and α = ( s c + a L ξ ) / n \alpha=(s_c+aL\xi)/\sqrt{n} α = ( s c + a L ξ ) / n , the Pauli generator moments give ∥ ψ ′ ∥ 2 = a 2 ξ 2 N / n \|\psi'\|^2=a^2\xi^2N/n ∥ ψ ′ ∥ 2 = a 2 ξ 2 N / n and ∥ ψ ′ ′ ∥ 2 = a 4 ξ 4 ( 3 N 2 − 2 N ) / n 2 \|\psi''\|^2=a^4\xi^4(3N^2-2N)/n^2 ∥ ψ ′′ ∥ 2 = a 4 ξ 4 ( 3 N 2 − 2 N ) / n 2 . Thus
∥ ρ ′ ∥ 1 ≤ 2 a ∣ ξ ∣ N / n , ∥ ρ ′ ′ ∥ 1 ≤ 2 ( 1 + 3 ) a 2 ξ 2 N / n . (197) \|\rho'\|_1\le2a|\xi|\sqrt{N/n},\qquad\|\rho''\|_1\le2(1+\sqrt{3})a^2\xi^2N/n. \tag*{(197)} ∥ ρ ′ ∥ 1 ≤ 2 a ∣ ξ ∣ N / n , ∥ ρ ′′ ∥ 1 ≤ 2 ( 1 + 3 ) a 2 ξ 2 N / n . (197)
Primes denote L L L derivatives, which have trace zero. Set β 1 = n − 1 ∑ N = 1 n N / n \beta_1=n^{-1}\sum_{N=1}^n\sqrt{N/n} β 1 = n − 1 ∑ N = 1 n N / n , β 2 = ( n + 1 ) / ( 2 n ) \beta_2=(n+1)/(2n) β 2 = ( n + 1 ) / ( 2 n ) , C 2 = ( 1 + 3 ) a 2 β 2 / 3 C_2=(1+\sqrt{3})a^2\beta_2/3 C 2 = ( 1 + 3 ) a 2 β 2 /3 and L max = 9 / 16 L_{\max}=9/16 L m a x = 9/16 . The risk curvature bounds are
M l o c = 2 C 2 , M a m p = 8 + 8 L max a β 1 + 4 L max 2 C 2 , M λ = λ M l o c + ( 1 − λ ) M a m p . (198) M_{\mathrm{loc}}=2C_2,\qquad M_{\mathrm{amp}}=8+8L_{\max}a\beta_1+4L_{\max}^2C_2,\qquad M_\lambda=\lambda M_{\mathrm{loc}}+(1-\lambda)M_{\mathrm{amp}}. \tag*{(198)} M loc = 2 C 2 , M amp = 8 + 8 L m a x a β 1 + 4 L m a x 2 C 2 , M λ = λ M loc + ( 1 − λ ) M amp . (198)
Lemma 1 (Curvature and interpolation). Every fixed complete paired POVM satisfies ∣ r j ′ ′ ( L ) ∣ ≤ M j |r_j''(L)|\le M_j ∣ r j ′′ ( L ) ∣ ≤ M j on I L I_L I L . For [ L − , L + ] [L_-,L_+] [ L − , L + ] , put Δ L = L + − L − \Delta_L=L_+-L_- Δ L = L + − L − , τ = ( L − L − ) / Δ L \tau=(L-L_-)/\Delta_L τ = ( L − L − ) / Δ L and ϵ j = M j Δ L 2 / 8 \epsilon_j=M_j\Delta_L^2/8 ϵ j = M j Δ L 2 /8 . Then
r j ⋆ ( L ) ≥ ( 1 − τ ) r j ( L − ) + τ r j ( L + ) − ϵ j , (199) r_j^\star(L)\ge(1-\tau)r_j(L_-)+\tau r_j(L_+)-\epsilon_j, \tag*{(199)} r j ⋆ ( L ) ≥ ( 1 − τ ) r j ( L − ) + τ r j ( L + ) − ϵ j , (199) r j ⋆ ( L ) ≤ r ˉ j ( M ; L ) ≤ ( 1 − τ ) r ˉ j ( M ; L − ) + τ r ˉ j ( M ; L + ) + ϵ j . (200) r_j^\star(L)\le\bar r_j(M;L)\le(1-\tau)\bar r_j(M;L_-)+\tau\bar r_j(M;L_+)+\epsilon_j. \tag*{(200)} r j ⋆ ( L ) ≤ r ˉ j ( M ; L ) ≤ ( 1 − τ ) r ˉ j ( M ; L − ) + τ r ˉ j ( M ; L + ) + ϵ j . (200) For the joint risk use M λ M_\lambda M λ ; minimize the upper bound over complete strategies held fixed at both endpoints.
Proof. For Hermitian X X X of trace zero and 0 ⪯ Q ⪯ B I 0\preceq Q\preceq BI 0 ⪯ Q ⪯ B I , ∣ Tr X Q ∣ ≤ B ∥ X ∥ 1 / 2 |\operatorname{Tr}XQ|\le B\|X\|_1/2 ∣ Tr X Q ∣ ≤ B ∥ X ∥ 1 /2 . Taking B = 2 B=2 B = 2 and E ξ 2 = 1 / 3 \mathbb{E}\xi^2=1/3 E ξ 2 = 1/3 gives M l o c M_{\mathrm{loc}} M loc . For amplitude, put Q ξ = ( η − ξ ) 2 M ⪯ 4 I Q_\xi=(\eta-\xi)^2M\preceq4I Q ξ = ( η − ξ ) 2 M ⪯ 4 I and f = Tr ρ Q ξ f=\operatorname{Tr}\rho Q_\xi f = Tr ρ Q ξ . Equation (V.1) gives f ≤ 4 f\le4 f ≤ 4 , ∣ f ′ ∣ ≤ 4 a ∣ ξ ∣ N / n |f'|\le4a|\xi|\sqrt{N/n} ∣ f ′ ∣ ≤ 4 a ∣ ξ ∣ N / n and ∣ f ′ ′ ∣ ≤ 4 ( 1 + 3 ) a 2 ξ 2 N / n |f''|\le4(1+\sqrt{3})a^2\xi^2N/n ∣ f ′′ ∣ ≤ 4 ( 1 + 3 ) a 2 ξ 2 N / n . Average ( L 2 f ) ′ ′ = 2 f + 4 L f ′ + L 2 f ′ ′ (L^2f)''=2f+4Lf'+L^2f'' ( L 2 f ) ′′ = 2 f + 4 L f ′ + L 2 f ′′ , using E ∣ ξ ∣ = 1 / 2 \mathbb{E}|\xi|=1/2 E ∣ ξ ∣ = 1/2 , to obtain M a m p M_{\mathrm{amp}} M amp . These uniform bounds justify differentiation. Linear interpolation has error at most M j ( L − L − ) ( L + − L ) / 2 ≤ ϵ j M_j(L-L_-)(L_+-L)/2\le\epsilon_j M j ( L − L − ) ( L + − L ) /2 ≤ ϵ j . Take infima after this estimate for the lower bound; fix a strategy for the upper bound.
The nine nodes L i = ( 28 + i ) / 64 L_i=(28+i)/64 L i = ( 28 + i ) /64 , i = 0 , … , 8 i=0,\ldots,8 i = 0 , … , 8 , give eight cells. Subtract marginal envelopes from the joint envelope; divide by the positive marginal baseline for relative gaps. Comparisons at endpoints and affine crossings bound each cell. At a new width, keep an endpoint POVM, rescale its reports by known L / L i L/L_i L / L i , retain its complement and fallback zero, and apply the quadrature allowance at that width.
Numerical endpoints used by the width proof
The following node enclosures are the separately certified risks in the exact physical record. All decimal endpoints are rounded outward; the accompanying numerical certificate archive retains the exact integer inputs, continuous prior fees, complete upper-rule frames and all-report dual repairs.
Table 1 (PDF p. 51).
Supplementary references
The complete-readout curvature bounds in (V.2) are enclosed above by M l o c < 2.682760529 M_{\mathrm{loc}} < 2.682760529 M loc < 2.682760529 and M a m p < 14.904048650 M_{\mathrm{amp}} < 14.904048650 M amp < 14.904048650 ; M 1 / 2 = ( M l o c + M a m p ) / 2 M_{1/2} = (M_{\mathrm{loc}} + M_{\mathrm{amp}})/2 M 1/2 = ( M loc + M amp ) /2 . For cell length 1 / 64 1/64 1/64 , these give respective interpolation allowances M j / 32768 M_j/32768 M j /32768 . Applying the lower and fixed-strategy upper envelopes gives:
Table V.
The upper envelopes use the 75 complete physical strategy cross-evaluations at neighboring nodes, rather than a width-dependent mixture of incomplete success branches. Their exact affine coefficients and crossings are supplied in the archive’s width_interval_certificate.json; each endpoint risk includes the fallback probability and a fresh continuous-prior fee.
Theorem 1 (Conflict across prior widths). For every L ∈ [ 7 / 16 , 9 / 16 ] L \in[7/16,9/16] L ∈ [ 7/16 , 9/16 ] in the stated n = 16 n = 16 n = 16 experiment,
4.17 × 10 − 4 < G 1 / 2 ( L ) < 2.242 × 10 − 3 , 0.5847 % < G 1 / 2 ( L ) ( r l o c ⋆ ( L ) + r a m p ⋆ ( L ) ) / 2 < 2.759 % . (201) 4.17 \times10^{-4} < \mathcal{G}_{1/2}(L) < 2.242 \times10^{-3}, \qquad0.5847\% < \frac{\mathcal{G}_{1/2}(L)}{(r^\star_{\mathrm{loc}}(L) + r^\star_{\mathrm{amp}}(L))/2} < 2.759\%. \tag*{(201)} 4.17 × 1 0 − 4 < G 1/2 ( L ) < 2.242 × 1 0 − 3 , 0.5847% < ( r loc ⋆ ( L ) + r amp ⋆ ( L )) /2 G 1/2 ( L ) < 2.759%. (201) Both marginals have positive risk and improve on their priors:
r l o c ⋆ > 0.0901 , r a m p ⋆ > 0.0481 , r l o c n o d a t a − r l o c ⋆ > 0.342 , L 2 / 3 − r a m p ⋆ > 0.0147. r^\star_{\mathrm{loc}} > 0.0901,\qquad r^\star_{\mathrm{amp}} > 0.0481,\qquad r^{\mathrm{no\ data}}_{\mathrm{loc}} - r^\star_{\mathrm{loc}} > 0.342,\qquad L^{2/3} - r^\star_{\mathrm{amp}} > 0.0147. r loc ⋆ > 0.0901 , r amp ⋆ > 0.0481 , r loc no data − r loc ⋆ > 0.342 , L 2/3 − r amp ⋆ > 0.0147. Proof. The exact cell envelopes give the gap bounds, and the marginal envelopes give the remaining inequalities. The amplitude prior risk is L 2 / 3 L^{2/3} L 2/3 . For position, b n ≤ K / 16 b_n \le K/16 b n ≤ K /16 makes the loss unclipped, and the median absolute error of the uniform 16-site prior is four, giving prior risk 8 b n / K 8b_n/K 8 b n / K . The risk bounds include compression, rounding, quadrature and dual repairs, with positive complements in every upper measurement.
References