Introduction
We identify the principal ideals of alternating quadratic random operators, their response to fixed compact perturbations, and their attainable range under changes of dependence. Write I ( T ) \mathcal{I}(T) I ( T ) for the two-sided ideal generated by a compact operator T T T , with no norm closure, and I ( u ) = I ( D u ) \mathcal{I}(u) = \mathcal{I}(D_u) I ( u ) = I ( D u ) , where D u = diag u D_u = \operatorname{diag} u D u = diag u , for a decreasing nonnegative null sequence. Equality of principal ideals means mutual domination of singular-value sequences up to constants and finite index dilations. Each identity below therefore determines all two-sided ideal membership tests on one probability-one event for fixed input data and a fixed perturbation.
The random fields are centered countable arrays Z Z Z whose finite coordinate marginals have positive densities c F e − V F c_F e^{-V_F} c F e − V F satisfying
α I ≤ ∇ 2 V F ≤ β I , 0 < α ≤ β < ∞ , \alpha I \le\nabla^2 V_F \le\beta I,\qquad0 < \alpha\le\beta< \infty, α I ≤ ∇ 2 V F ≤ β I , 0 < α ≤ β < ∞ ,
with constants independent of the coordinates and their number. Coordinates may be dependent and conditional means may be nonzero. The lower curvature bound controls moments; the upper bound supplies conditional anti-concentration. We call these curvature arrays .
Let Q ≥ 0 Q \ge0 Q ≥ 0 be trace class with rank Q ≥ 2 Q \ge2 Q ≥ 2 , and let K ≠ 0 K \ne0 K = 0 be a skew-adjoint Hilbert–Schmidt operator. Put
a = s ( Q ) , q n = a n 2 , κ n = s n ( K ) , R n = ∑ j ≥ n q j , T n = ∑ j ≥ n κ j 2 . a = \sqrt{s(Q)},\qquad q_n = a_n^2,\qquad\kappa_n = s_n(K),\qquad R_n = \sum_{j\ge n} q_j,\qquad T_n = \sum_{j\ge n} \kappa_j^2. a = s ( Q ) , q n = a n 2 , κ n = s n ( K ) , R n = j ≥ n ∑ q j , T n = j ≥ n ∑ κ j 2 .
In particular a 2 > 0 a_2 > 0 a 2 > 0 . For finite-rank K K K , the raw evaluation is D a Z K Z ∗ D a D_a ZKZ^*D_a D a Z K Z ∗ D a . Centering gives a map that extends continuously to K ∈ S 2 K \in S_2 K ∈ S 2 in L 2 ( Ω ; S 2 ) L^2(\Omega;S_2) L 2 ( Ω ; S 2 ) ; denote the extension by W a ( K ) W_a(K) W a ( K ) . Theorem B constructs a deterministic compact skew operator E \mathcal{E} E such that, for every fixed compact B B B ,
r n 2 = R n 2 T n − R n + 1 2 T n + 1 , I ( W a ( K ) + B ) = I ( r ) + I ( B − E ) almost surely . (1) r_n^2 = R_n^2T_n - R_{n+1}^2T_{n+1},\qquad\mathcal{I}(W_a(K)+B)=\mathcal{I}(r)+\mathcal{I}(B-\mathcal{E})\quad\text{almost surely}. \tag*{(1)} r n 2 = R n 2 T n − R n + 1 2 T n + 1 , I ( W a ( K ) + B ) = I ( r ) + I ( B − E ) almost surely . (1)
The square tail R n 2 T n R_n^2T_n R n 2 T n records the three factors Q 1 / 2 Q^{1/2} Q 1/2 , K K K , Q 1 / 2 Q^{1/2} Q 1/2 in the independent Gaussian model. The compensator records the effect of skew covariance contractions.
2020 Mathematics Subject Classification. Primary 47B10, 60G15; Secondary 60B20, 60H05, 15A75.
Key words and phrases. Random operators, principal ideals, conditional covariance, strong log-concavity, exterior forms, Lévy area.
Its construction is explicit. Write K = ∑ ℓ ≥ 1 K ℓ K = \sum_{\ell\ge1} K_\ell K = ∑ ℓ ≥ 1 K ℓ in ordered two-dimensional skew spectral blocks, let M ℓ M_\ell M ℓ be the mean evaluation of K ℓ K_\ell K ℓ , and let P ℓ + 1 P_{\ell+1} P ℓ + 1 delete the first 2 ℓ 2\ell 2 ℓ spatial directions. Then
C = ∑ ℓ ≥ 1 P ℓ + 1 M ℓ P ℓ + 1 \mathcal{C} = \sum_{\ell\ge1} P_{\ell+1} M_\ell P_{\ell+1} C = ℓ ≥ 1 ∑ P ℓ + 1 M ℓ P ℓ + 1
converges absolutely in S 2 \mathcal{S}_2 S 2 , even when the raw means diverge. For fixed spectral choices this is a compact linear map of the full covariance, with explicit finite-dimensional error bounds. Its class modulo I ( r ) \mathcal{I}(r) I ( r ) is independent of those choices. For the fixed pair ( a , K ) (a,K) ( a , K ) , two centered evaluations have identical fixed-shift responses precisely when their compensators differ by an element of I ( r ) \mathcal{I}(r) I ( r ) . If the raw means converge to M M M , then A = W a ( K ) + M A = W_a(K) + M A = W a ( K ) + M satisfies
I ( A + D ) = I ( r ) + I ( Δ + D ) , Δ = M − C . \mathcal{I}(A + D) = \mathcal{I}(r) + \mathcal{I}(\Delta+ D), \qquad\Delta= M - \mathcal{C}. I ( A + D ) = I ( r ) + I ( Δ + D ) , Δ = M − C .
The distinction between M M M and Δ \Delta Δ matters even for finite-rank kernels: their raw evaluations have finite rank, but their means may have infinite rank. Theorem C determines the universal raw domain and the full attainable range. Along any prescribed coordinate exhaustion, raw evaluation converges in L 2 ( S 2 ) L^2(\mathcal{S}_2) L 2 ( S 2 ) for every curvature array if and only if K ∈ S 1 K \in\mathcal{S}_1 K ∈ S 1 . Failure outside this domain already occurs for Gaussian covariances arbitrarily close to the identity. For K ∈ S 1 K \in\mathcal{S}_1 K ∈ S 1 , write A Z A_Z A Z for the raw evaluation and put
S n = ∑ j ≥ n κ j , p n = R n S n − R n + 1 S n + 1 . S_n = \sum_{j \ge n} \kappa_j, \qquad p_n = R_n S_n - R_{n+1} S_{n+1}. S n = j ≥ n ∑ κ j , p n = R n S n − R n + 1 S n + 1 .
Then
{ J : I ( A Z ) = J a.s. for some curvature array Z } = { I ( u ) : I ( r ) ⊆ I ( u ) ⊆ I ( p ) } . (2) \{J : \mathcal{I}(A_Z) = J\ \text{a.s. for some curvature array } Z\} = \{\mathcal{I}(u) : \mathcal{I}(r) \subseteq\mathcal{I}(u) \subseteq\mathcal{I}(p)\}. \tag*{(2)} { J : I ( A Z ) = J a.s. for some curvature array Z } = { I ( u ) : I ( r ) ⊆ I ( u ) ⊆ I ( p )} . (2)
Here the arrays may have different dimension-free curvature constants. Every intermediate principal ideal is realized in any operator-norm neighborhood of the identity covariance by a Gaussian array whose individual rows remain standard Gaussian sequences. Example 5.1 gives a strict chain of such ideals for which every Schatten membership test has the same answer.
The linear tool is Theorem A , under finite-test curvature bounds. For conditional-residual tails ρ \rho ρ , optimal mean-square truncations η \eta η , and directional trace envelopes ϑ \vartheta ϑ , it gives
ρ ( T ) ≤ η ( T ) ≤ ϑ ( T ) , I ( ρ ( T ) ) = I ( ϑ ( T ) ) ⟹ I ( T + D ) = I ( η ( T ) ) + I ( D ) a.s. \rho(T) \le\eta(T) \le\vartheta(T), \qquad\mathcal{I}(\rho(T)) = \mathcal{I}(\vartheta(T)) \Longrightarrow\mathcal{I}(T + D) = \mathcal{I}(\eta(T)) + \mathcal{I}(D)\quad\text{a.s.} ρ ( T ) ≤ η ( T ) ≤ ϑ ( T ) , I ( ρ ( T )) = I ( ϑ ( T )) ⟹ I ( T + D ) = I ( η ( T )) + I ( D ) a.s.
This is a sufficient criterion, allowing singular Gaussian reference supports and nonlinear conditional prediction. It includes exponential lower tails and comparisons with every strict upper index dilation L > 2 L > 2 L > 2 and lower dilation λ > 1 \lambda> 1 λ > 1 . The quadratic output is not assumed to satisfy the linear curvature hypothesis. Zero skew contractions, and more generally finite skew defects, recover the unmodified sum law for exterior evaluation.
For Wiener area with covariance Q Q Q of rank at least two, the time spectrum gives
I ( A ) = I ( ( R n / n ) n ≥ 1 ) a.s. , I ( Q ) ⊆ I ( A ) . (3) \mathcal{I}(A) = \mathcal{I}\left((R_n/n)_{n \ge1}\right)\ \text{a.s.}, \qquad\mathcal{I}(Q) \subseteq\mathcal{I}(A). \tag*{(3)} I ( A ) = I ( ( R n / n ) n ≥ 1 ) a.s. , I ( Q ) ⊆ I ( A ) . (3)
Equality holds exactly when R n ≤ C n q n R_n \le C n q_n R n ≤ C n q n , the classical regularity-at-infinity condition. For a doubly monotone variance table v = ( v i j ) v = (v_{ij}) v = ( v ij ) with summable first row and first column, Theorem D and Corollary 7.1 give
h n 2 = ∑ j ≥ n v n j + ∑ i > n v i n , I ( T + D ) = I ( h ) + I ( D ) a.s. (4) h_n^2 = \sum_{j \ge n} v_{nj} + \sum_{i > n} v_{in}, \qquad\mathcal{I}(T + D) = \mathcal{I}(h) + \mathcal{I}(D)\quad\text{a.s.} \tag*{(4)} h n 2 = j ≥ n ∑ v nj + i > n ∑ v in , I ( T + D ) = I ( h ) + I ( D ) a.s. (4)
This holds for Gaussian arrays whose full covariance is comparable to the diagonal form with variances v i j v_{ij} v ij , and for weighted curvature arrays T i j = v i j Z i j T_{ij} = \sqrt{v_{ij}}Z_{ij} T ij = v ij Z ij . The exact boundedness condition h 1 < ∞ h_1 < \infty h 1 < ∞ also implies compactness and allows infinite total variance. Both applications include fixed compact perturbations.
Principal-ideal calculus, finite dilations, and regularity at infinity are classical [4 , 12 , 13 , 19 ] ; we use standard spectral variation [7 ] . Sharp norm and Schatten estimates for nonhomogeneous matrices [2 , 15 ] allow variance profiles beyond our ordered class. Our conclusion identifies algebraic principal ideals and fixed-shift responses. Rudelson and Zeitouni [17 ] obtain singular-value estimates under broad-connectedness assumptions. Fernandez V, Livshyts and Mui [8 ] prove sharp smallest-singular-value estimates for isotropic log-concave matrices in three settings: square unconditional matrices, independent columns, and sufficiently tall matrices. Our two-sided curvature assumption and conditional-residual criterion serve an infinite-dimensional ideal classification; we do not claim an improvement of their smallest-singular-value bounds.
The probability inputs are Brascamp–Lieb variance bounds [3 ] , Hargé’s convex comparison [10 ] , Gaussian concentration [21 ] , regression [11 ] , and conditional convergence [14 ] . The inverse-Gram step uses the negative second-moment identity [20 Lemma A.4] . Adamczak [1 ] studies concentration of dependent quadratic forms; here the quadratic conclusion identifies a covariance class modulo an algebraic operator ideal.
For area, [16 ] treats Banach rough paths, [9 ] gives double Wiener–Itô representations, and [6 ] studies finite-dimensional singular-value distributions and dimension asymptotics. We fix the spatial covariance and determine an infinite-dimensional ideal. The exterior-dilation estimates of [22 ] provide simultaneous Schatten bounds for 0 < s < 2 0 < s < 2 0 < s < 2 through one random dilation. Our qualitative identities do not supply that quantitative dilation; the logarithmic example exhibits ideal information beyond all Schatten membership tests.
Proof and organization
Conditional Laplace estimates and the inverse Gram identity give the linear lower bound; convex comparison gives the upper bound. Conditional contraction transfers the lower inclusion to exterior evaluations. Polarization, finite-rank deletion, and covariance compensation give the upper inclusion. Triangular correlations then realize the entire interval. Sections 2–3 develop the linear tools; Sections 4–5 prove compensation and realization. Sections 6–7 treat finite defects, Wiener area, and ordered matrices. Events may depend on each fixed shift; prescribed countable families share an event.
Conditional curvature and spectral ideals
All Hilbert spaces are real and separable. Write B , K \mathcal{B},\mathcal{K} B , K for bounded and compact operators, with the operator norm. Singular values are nonincreasing, include multiplicities and are zero-completed. They satisfy
s n ( T ) = inf rank R < n ∥ T − R ∥ . (5) s_n(T)=\inf_{\operatorname{rank} R<n}\lVert T-R\rVert. \tag*{(5)} s n ( T ) = rank R < n inf ∥ T − R ∥ . (5)
We compare rectangular operators after zero extension in a common infinite-dimensional Hilbert space; this and complexification preserve nonzero singular values. For 0 < p < ∞ 0 < p < \infty 0 < p < ∞ , write S p = { T ∈ K : ∑ n s n ( T ) p < ∞ } \mathcal{S}_p=\{T\in\mathcal{K}:\sum_n s_n(T)^p<\infty\} S p = { T ∈ K : ∑ n s n ( T ) p < ∞ } and ∥ T ∥ p = ( ∑ n s n ( T ) p ) 1 / p \lVert T\rVert_p=\left(\sum_n s_n(T)^p\right)^{1/p} ∥ T ∥ p = ( ∑ n s n ( T ) p ) 1/ p ; this is a quasi-norm when p < 1 p<1 p < 1 . In particular S 1 \mathcal{S}_1 S 1 and S 2 \mathcal{S}_2 S 2 are the trace-class and Hilbert–Schmidt classes. We also write ∥ T ∥ H S = ∥ T ∥ 2 \lVert T\rVert_{\mathrm{HS}}=\lVert T\rVert_2 ∥ T ∥ HS = ∥ T ∥ 2 and ( x ⊗ y ) z = x ⟨ y , z ⟩ (x\otimes y)z=x\langle y,z\rangle ( x ⊗ y ) z = x ⟨ y , z ⟩ . Unless indicated otherwise, ∥ T ∥ \lVert T\rVert ∥ T ∥ is the operator norm. We use [19 ]
s n ( A T B ) ≤ ∥ A ∥ ∥ B ∥ s n ( T ) , s j + k − 1 ( S + T ) ≤ s j ( S ) + s k ( T ) . (6) s_n(ATB)\leq\lVert A\rVert\lVert B\rVert s_n(T),\qquad s_{j+k-1}(S+T)\leq s_j(S)+s_k(T). \tag*{(6)} s n ( A T B ) ≤ ∥ A ∥ ∥ B ∥ s n ( T ) , s j + k − 1 ( S + T ) ≤ s j ( S ) + s k ( T ) . (6)
Thus ∣ s n ( S ) − s n ( T ) ∣ ≤ ∥ S − T ∥ |s_n(S)-s_n(T)|\leq\lVert S-T\rVert ∣ s n ( S ) − s n ( T ) ∣ ≤ ∥ S − T ∥ . The ideal I ( T ) \mathcal{I}(T) I ( T ) consists of all finite sums ∑ l A l T B l \sum_l A_lTB_l ∑ l A l T B l with bounded multipliers, and I ( a ) = I ( D a ) \mathcal{I}(a)=\mathcal{I}(D_a) I ( a ) = I ( D a ) , where D a = diag ( a n ) D_a=\operatorname{diag}(a_n) D a = diag ( a n ) . No closure is taken. Nonzero finite-rank operators all generate the finite-rank ideal. Random compact operators are Borel random elements of the separable space K \mathcal{K} K with its operator norm; operator-valued expectations in L p ( S 2 ) L^p(S_2) L p ( S 2 ) are Bochner expectations. Conditional distributions of countable arrays are taken on their canonical standard Borel coordinate spaces. Regular conditional laws and the martingale convergence statements used below may therefore be chosen in the usual sense [14 ] . A constant described as absolute is independent of the dimensions and of the deterministic coefficients.
Lemma 2.1 (Principal ideals and finite dilations). For nonnegative nonincreasing null sequences a , b , a,b, a , b ,
I ( a ) ⊆ I ( b ) ⟺ a n ≤ C b ⌈ n / M ⌉ ( n ≥ 1 ) for some C < ∞ , M ∈ N . (7) \mathcal{I}(a) \subseteq\mathcal{I}(b) \quad\Longleftrightarrow\quad a_n \le Cb_{\lceil n/M\rceil}\quad(n\ge1)\quad\text{for some }C<\infty,\ M\in\mathbb{N}. \tag*{(7)} I ( a ) ⊆ I ( b ) ⟺ a n ≤ C b ⌈ n / M ⌉ ( n ≥ 1 ) for some C < ∞ , M ∈ N . (7) Moreover, I ( T ) = I ( s ( T ) ) \mathcal{I}(T)=\mathcal{I}(s(T)) I ( T ) = I ( s ( T )) for every compact T T T , and
I ( a ) + I ( b ) = I ( max { a , b } ) , (8) \mathcal{I}(a)+\mathcal{I}(b)=\mathcal{I}(\max\{a,b\}), \tag*{(8)} I ( a ) + I ( b ) = I ( max { a , b }) , (8) where the maximum is coordinatewise. Also
I ( a ) ∩ I ( b ) = I ( min { a , b } ) . (9) \mathcal{I}(a)\cap\mathcal{I}(b)=\mathcal{I}(\min\{a,b\}). \tag*{(9)} I ( a ) ∩ I ( b ) = I ( min { a , b }) . (9) Thus principal ideals form a distributive lattice under sum and intersection, including the zero ideal.
Proof. If D a = ∑ j = 1 m A j D b B j D_a=\sum_{j=1}^{m}A_jD_bB_j D a = ∑ j = 1 m A j D b B j , the singular-value inequalities give
s m ( n − 1 ) + 1 ( D a ) ≤ ∑ j = 1 m ∥ A j ∥ ∥ B j ∥ b n , s_{m(n-1)+1}(D_a)\le\sum_{j=1}^{m}\lVert A_j\rVert\lVert B_j\rVert b_n, s m ( n − 1 ) + 1 ( D a ) ≤ j = 1 ∑ m ∥ A j ∥ ∥ B j ∥ b n ,
which proves the forward implication in (7) . Conversely, for ( Δ m b ) n = b ⌈ n / m ⌉ (\Delta_m b)_n=b_{\lceil n/m\rceil} ( Δ m b ) n = b ⌈ n / m ⌉ , the isometries V j e k = e m ( k − 1 ) + j V_je_k=e_{m(k-1)+j} V j e k = e m ( k − 1 ) + j satisfy D Δ m b = ∑ j = 1 m V j D b V j ∗ ∈ I ( b ) D_{\Delta_m b}=\sum_{j=1}^{m}V_jD_bV_j^*\in\mathcal{I}(b) D Δ m b = ∑ j = 1 m V j D b V j ∗ ∈ I ( b ) . The bound a ≤ C Δ m b a\le C\Delta_m b a ≤ C Δ m b then expresses D a D_a D a as a bounded diagonal multiple of D Δ m b D_{\Delta_m b} D Δ m b , with quotient zero at zero entries. Singular-value decomposition gives I ( T ) = I ( s ( T ) ) \mathcal{I}(T)=\mathcal{I}(s(T)) I ( T ) = I ( s ( T )) .
Splitting the coordinates according to a n ≥ b n a_n\ge b_n a n ≥ b n proves the sum formula. For the intersection, two bounds for s ( T ) s(T) s ( T ) may use the larger of their dilations and constants, yielding a bound by min { a ⌈ n / m ⌉ , b ⌈ n / m ⌉ } \min\{a_{\lceil n/m\rceil},b_{\lceil n/m\rceil}\} min { a ⌈ n / m ⌉ , b ⌈ n / m ⌉ } . The reverse inclusion is immediate. Pointwise minimum and maximum satisfy the distributive identities. These are classical consequences of Calkin’s correspondence [4 ] ; see also [13 Section 2] . □
Lemma 2.2 (Fixed-shift completion). Let X X X be a random compact operator, J J J a deterministic algebraic principal ideal and C C C a deterministic compact operator. Suppose X + C ∈ J X+C\in J X + C ∈ J almost surely and, for every fixed compact B B B , J ⊆ I ( X + B ) J\subseteq\mathcal{I}(X+B) J ⊆ I ( X + B ) almost surely. Then
I ( X + B ) = J + I ( B − C ) almost surely . (10) \mathcal{I}(X+B)=J+\mathcal{I}(B-C)\quad\text{almost surely}. \tag*{(10)} I ( X + B ) = J + I ( B − C ) almost surely . (10) The lower hypothesis may instead be verified after a fixed bounded linear map Λ \Lambda Λ satisfying I ( Λ S ) ⊆ I ( S ) \mathcal{I}(\Lambda S)\subseteq\mathcal{I}(S) I ( Λ S ) ⊆ I ( S ) for compact S S S : it suffices that J ⊆ I ( Λ X + F ) J\subseteq\mathcal{I}(\Lambda X+F) J ⊆ I ( Λ X + F ) for every fixed compact F F F . The least value of the deterministic response B ↦ I ( X + B ) B\mapsto\mathcal{I}(X+B) B ↦ I ( X + B ) is J J J , and
{ B : I ( X + B ) = J a.s. } = C + J . (11) \{B:\mathcal{I}(X+B)=J\ \text{a.s.}\}=C+J. \tag*{(11)} { B : I ( X + B ) = J a.s. } = C + J . (11) For an independent copy X ′ X' X ′ one also has I ( X − X ′ ) = J \mathcal{I}(X-X')=J I ( X − X ′ ) = J almost surely. In particular the response determines J J J and the compensator class C + J C+J C + J ; any two deterministic compensators differ by an element of J J J .
Proof. Fix B B B and intersect the two probability-one events in the hypotheses. Since X + B = ( X + C ) + ( B − C ) X+B=(X+C)+(B-C) X + B = ( X + C ) + ( B − C ) and X + C ∈ J X+C\in J X + C ∈ J , we have I ( X + B ) ⊆ J + I ( B − C ) \mathcal{I}(X+B)\subseteq J+\mathcal{I}(B-C) I ( X + B ) ⊆ J + I ( B − C ) . Conversely, the lower hypothesis gives J ⊆ I ( X + B ) J\subseteq\mathcal{I}(X+B) J ⊆ I ( X + B ) . Thus X + C ∈ I ( X + B ) X+C\in\mathcal{I}(X+B) X + C ∈ I ( X + B ) , and subtraction gives B − C ∈ I ( X + B ) B-C\in\mathcal{I}(X+B) B − C ∈ I ( X + B ) . This proves (10) . In the compressed formulation, take F = Λ B F=\Lambda B F = Λ B and use I ( Λ ( X + B ) ) ⊆ I ( X + B ) \mathcal{I}(\Lambda(X+B))\subseteq\mathcal{I}(X+B) I ( Λ ( X + B )) ⊆ I ( X + B ) .
The right side of (10) is always at least J J J and equals J J J precisely when B − C ∈ J B-C\in J B − C ∈ J . This proves (11) . For an independent copy X ′ X' X ′ , subtracting X + C X+C X + C and X ′ + C X'+C X ′ + C shows that X − X ′ ∈ J X-X'\in J X − X ′ ∈ J almost surely. Conditioning on X ′ X' X ′ , the lower hypothesis with the now fixed shift − X ′ -X' − X ′ gives the reverse ideal inclusion. To justify this conditioning, write J = I ( d ) J=\mathcal{I}(d) J = I ( d ) . The event J ⊆ I ( S ) J\subseteq\mathcal{I}(S) J ⊆ I ( S ) is
⋃ c , m ∈ N ⋂ n > 1 { d n ≤ c s ⌈ n / m ⌉ ( S ) } , \bigcup_{c,m\in\mathbb{N}}\bigcap_{n>1}\{d_n\le c s_{\lceil n/m\rceil}(S)\}, c , m ∈ N ⋃ n > 1 ⋂ { d n ≤ c s ⌈ n / m ⌉ ( S )} ,
which is Borel by continuity of singular values. Fubini’s theorem therefore applies to the independent pair. Finally, two compensators satisfy C − C ′ = ( X + C ) − ( X + C ′ ) ∈ J C-C'=(X+C)-(X+C')\in J C − C ′ = ( X + C ) − ( X + C ′ ) ∈ J on the intersection of their full-measure events. □
In the algebraic quotient K / J K/J K / J the upper hypothesis is [ X ] = − [ C ] [X] = -[C] [ X ] = − [ C ] almost surely. It does not imply the lower hypothesis; no quotient topology is used.
For U : H 0 → H 1 U : H_0 \to H_1 U : H 0 → H 1 random and compact with E ∥ U ∥ 2 < ∞ \mathbb{E}\lVert U\rVert^2 < \infty E ∥ U ∥ 2 < ∞ , set U P , Q = ( I − P ) U ( I − Q ) U_{P,Q} = (I-P)U(I-Q) U P , Q = ( I − P ) U ( I − Q ) , Γ L U ( P , Q ) = E ( U P , Q U P , Q ∗ ) \Gamma_L^U(P,Q) = \mathbb{E}(U_{P,Q}U_{P,Q}^{*}) Γ L U ( P , Q ) = E ( U P , Q U P , Q ∗ ) and Γ R U ( P , Q ) = E ( U P , Q ∗ U P , Q ) \Gamma_R^U(P,Q) = \mathbb{E}(U_{P,Q}^{*}U_{P,Q}) Γ R U ( P , Q ) = E ( U P , Q ∗ U P , Q ) . Let P p , q \mathcal{P}_{p,q} P p , q consist of deterministic orthogonal output/input projections of ranks at most p , q p,q p , q , and put
η n ( U ) 2 = inf ( P , Q ) ∈ P n − 1 , n − 1 max { ∥ Γ L U ( P , Q ) ∥ , ∥ Γ R U ( P , Q ) ∥ } . \eta_n(U)^2 = \inf_{(P,Q)\in\mathcal{P}_{n-1,n-1}} \max\{\lVert\Gamma_L^U(P,Q)\rVert,\lVert\Gamma_R^U(P,Q)\rVert\}. η n ( U ) 2 = ( P , Q ) ∈ P n − 1 , n − 1 inf max {∥ Γ L U ( P , Q )∥ , ∥ Γ R U ( P , Q )∥} .
All projections precede the sample. A centered random bilinear form U : H 0 × H 1 → L 2 ( Ω ) U : H_0 \times H_1 \to L^2(\Omega) U : H 0 × H 1 → L 2 ( Ω ) is tested against finite-rank matrices by finite linear combinations of its evaluations. Its full covariance form is C U C_U C U ; covariance comparison means comparison on every such test, not only on rank-one tests or mean Gram operators.
For a centered random bilinear form Z Z Z , call positive trace-class operators L L L on H 1 H_1 H 1 and R R R on H 0 H_0 H 0 directional upper envelopes if
E ∣ Z ( x , y ) ∣ 2 ≤ min { ∥ x ∥ 2 ⟨ L y , y ⟩ , ∥ y ∥ 2 ⟨ R x , x ⟩ } ( x ∈ H 0 , y ∈ H 1 ) . (12) \mathbb{E}|Z(x,y)|^2 \le\min\{\lVert x\rVert^2\langle Ly,y\rangle,\lVert y\rVert^2\langle Rx,x\rangle\}\qquad(x\in H_0,\ y\in H_1). \tag*{(12)} E ∣ Z ( x , y ) ∣ 2 ≤ min {∥ x ∥ 2 ⟨ L y , y ⟩ , ∥ y ∥ 2 ⟨ R x , x ⟩} ( x ∈ H 0 , y ∈ H 1 ) . (12)
With U P , Q ( x , y ) = U ( ( I − Q ) x , ( I − P ) y ) U_{P,Q}(x,y)=U((I-Q)x,(I-P)y) U P , Q ( x , y ) = U (( I − Q ) x , ( I − P ) y ) and empty infima equal to + ∞ +\infty + ∞ , define
ϑ n ( U ) 2 = inf ( P , Q ) ∈ P n − 1 , n − 1 L , R upper envelopes of U P , Q max { tr L , tr R } . (13) \vartheta_n(U)^2 = \inf_{\substack{(P,Q)\in\mathcal{P}_{n-1,n-1}\\ L,R\ \text{upper envelopes of }U_{P,Q}}}\max\{\operatorname{tr}L,\operatorname{tr}R\}. \tag*{(13)} ϑ n ( U ) 2 = ( P , Q ) ∈ P n − 1 , n − 1 L , R upper envelopes of U P , Q inf max { tr L , tr R } . (13)
For a compact U U U with square-integrable norm, a finite output space E E E and unit u ∈ E u\in E u ∈ E , define
F E , u U = σ ( U ∗ v : v ∈ E ∩ u ⊥ ) , S E , u U = Cov ( U ∗ u − E [ U ∗ u ∣ F E , u U ] ) . (14) \mathcal{F}_{E,u}^U=\sigma(U^*v:v\in E\cap u^\perp),\qquad S_{E,u}^U=\operatorname{Cov}(U^*u-\mathbb{E}[U^*u\mid\mathcal{F}_{E,u}^U]). \tag*{(14)} F E , u U = σ ( U ∗ v : v ∈ E ∩ u ⊥ ) , S E , u U = Cov ( U ∗ u − E [ U ∗ u ∣ F E , u U ]) . (14)
This is the unconditional covariance of the actual conditional-expectation residual. It equals the expectation of the conditional covariance, which need not be deterministic. It is not a linear-regression surrogate. For t n ( S ) = ∑ k ≥ n λ k ( S ) t_n(S)=\sum_{k\ge n}\lambda_k(S) t n ( S ) = ∑ k ≥ n λ k ( S ) , set
ρ n ( U ) 2 = sup dim E = n , U ∘ ∈ { U , U ∗ } u ∈ S ( E ) t n ( S E , u U ∘ ) . (15) \rho_n(U)^2=\sup_{\substack{\dim E=n,\ U^\circ\in\{U,U^*\}\\u\in S(E)}}t_n(S_{E,u}^{U^\circ}). \tag*{(15)} ρ n ( U ) 2 = d i m E = n , U ∘ ∈ { U , U ∗ } u ∈ S ( E ) sup t n ( S E , u U ∘ ) . (15)
Finite spaces are zero-extended; deterministic optimizers need not exist.
Finite-test curvature. Let T T T be a centered random bilinear form and G G G a centered Gaussian reference form, possibly on a different probability space. For every finite list of finite-rank tests M 1 , … , M d M_1,\ldots,M_d M 1 , … , M d , write W U = ( ⟨ U , M l ⟩ ) l ≤ d W_U=(\langle U,M_l\rangle)_{l\le d} W U = (⟨ U , M l ⟩ ) l ≤ d , C = Cov ( W G ) C=\operatorname{Cov}(W_G) C = Cov ( W G ) and F = ran C F=\operatorname{ran}C F = ran C . We assume that W T W_T W T is supported on F F F with positive density c e − V ce^{-V} c e − V there, and that fixed constants 0 < α ≤ β < ∞ 0<\alpha\le\beta<\infty 0 < α ≤ β < ∞ satisfy
α C F − 1 ≤ ∇ F 2 V ≤ β C F − 1 . (16) \alpha C_F^{-1}\le\nabla_F^2V\le\beta C_F^{-1}. \tag*{(16)} α C F − 1 ≤ ∇ F 2 V ≤ β C F − 1 . (16)
Only the restriction C F C_F C F is inverted. If F = { 0 } F=\{0\} F = { 0 } , the test is zero. The constants do not depend on the tests or their number. A C 2 C^2 C 2 potential suffices; we also allow C l o c 1 , 1 C_{\mathrm{loc}}^{1,1} C loc 1 , 1 potentials with these bounds almost everywhere, which permits finite-dimensional weak limits. No independence, conditional centering, or rotational invariance of T T T is assumed. This is a distributional condition, not merely covariance comparison.
Theorem A (Conditional-residual closure). Assume (H) and ϑ 1 ( G ) < ∞ \vartheta_1(G)<\infty ϑ 1 ( G ) < ∞ . Then T , G T,G T , G have unique compact realizations with square-integrable norms. Their three profiles are nonincreasing null sequences, and
ρ ( U ) ≤ η ( U ) ≤ ϑ ( U ) ( U = T , G ) , β − 1 / 2 f n ( G ) ≤ f n ( T ) ≤ α − 1 / 2 f n ( G ) ( f = ρ , η , ϑ ) . (17) \rho(U)\le\eta(U)\le\vartheta(U)\qquad(U=T,G),\qquad\beta^{-1/2}f_n(G)\le f_n(T)\le\alpha^{-1/2}f_n(G)\qquad(f=\rho,\eta,\vartheta). \tag*{(17)} ρ ( U ) ≤ η ( U ) ≤ ϑ ( U ) ( U = T , G ) , β − 1/2 f n ( G ) ≤ f n ( T ) ≤ α − 1/2 f n ( G ) ( f = ρ , η , ϑ ) . (17) For every fixed compact B B B , almost surely
I ( ρ ( T ) ) ⊆ I ( T + B ) , T ∈ I ( ϑ ( T ) ) . \mathcal{I}(\rho(T))\subseteq\mathcal{I}(T+B),\qquad T\in\mathcal{I}(\vartheta(T)). I ( ρ ( T )) ⊆ I ( T + B ) , T ∈ I ( ϑ ( T )) . Consequently,
I ( ρ ( T ) ) = I ( ϑ ( T ) ) ⟹ I ( T + B ) = I ( η ( T ) ) + I ( B ) . (18) \mathcal{I}(\rho(T))=\mathcal{I}(\vartheta(T))\quad\Longrightarrow\quad\mathcal{I}(T+B)=\mathcal{I}(\eta(T))+\mathcal{I}(B). \tag*{(18)} I ( ρ ( T )) = I ( ϑ ( T )) ⟹ I ( T + B ) = I ( η ( T )) + I ( B ) . (18) The closure condition is equivalent to I ( ρ ( G ) ) = I ( ϑ ( G ) ) \mathcal{I}(\rho(G))=\mathcal{I}(\vartheta(G)) I ( ρ ( G )) = I ( ϑ ( G )) , and I ( η ( T ) ) = I ( η ( G ) ) \mathcal{I}(\eta(T))=\mathcal{I}(\eta(G)) I ( η ( T )) = I ( η ( G )) . On one event there are finite positive random constants, for all L > 2 L>2 L > 2 and λ > 1 \lambda>1 λ > 1 , such that
s m ( T ) ≤ c ω , L ϑ ⌈ m / L ⌉ ( T ) , s ⌈ n / λ ⌉ ( T + B ) ≥ c ω , λ ρ n ( T ) . (19) s_m(T)\le c_{\omega,L}\vartheta_{\lceil m/L\rceil}(T),\qquad s_{\lceil n/\lambda\rceil}(T+B)\ge c_{\omega,\lambda}\rho_n(T). \tag*{(19)} s m ( T ) ≤ c ω , L ϑ ⌈ m / L ⌉ ( T ) , s ⌈ n / λ ⌉ ( T + B ) ≥ c ω , λ ρ n ( T ) . (19)
Either actual profile may be replaced by its reference profile. Thus ϑ ( T ) ≤ C ρ ( T ) \vartheta(T) \le C\rho(T) ϑ ( T ) ≤ C ρ ( T ) allows both to be replaced by η ( T ) \eta(T) η ( T ) . Zero and finite-support profiles are included. Events may depend on B B B ; countable prescribed shifts and independent random compact shifts are allowed.
The following quantitative bounds accompany the ideal inclusions. Put κ = β / α \kappa= \beta/\alpha κ = β / α . For 1 ≤ j < n 1 \le j < n 1 ≤ j < n , m ≥ 2 n − 1 m \ge2n-1 m ≥ 2 n − 1 and t > 0 t > 0 t > 0 , the proof gives
P { s j ( T + B ) 2 < ( n − j ) ρ n ( T ) 2 / ( 4 κ n ) } ≤ 2 − ⌊ ( n − j + 1 ) / 4 ⌋ , (20) \mathbb{P}\{s_j(T+B)^2 < (n-j)\rho_n(T)^2/(4\kappa n)\} \le2^{-\lfloor(n-j+1)/4\rfloor}, \tag*{(20)} P { s j ( T + B ) 2 < ( n − j ) ρ n ( T ) 2 / ( 4 κn )} ≤ 2 − ⌊( n − j + 1 ) /4 ⌋ , (20)
P { s m ( T ) > ( 2 2 + t ) κ ϑ n ( T ) } ≤ e − t 2 ( m − 2 n + 2 ) / 2 , (21) \mathbb{P}\{s_m(T) > (2\sqrt{2}+t)\sqrt{\kappa}\,\vartheta_n(T)\} \le e^{-t^2(m-2n+2)/2}, \tag*{(21)} P { s m ( T ) > ( 2 2 + t ) κ ϑ n ( T )} ≤ e − t 2 ( m − 2 n + 2 ) /2 , (21)
with the zero-profile convention proved below. For T = G T=G T = G and α = β = 1 \alpha=\beta=1 α = β = 1 , these are the corresponding Gaussian estimates. The closure condition is not necessary for a deterministic principal ideal. Indeed, if g g g is standard Gaussian and D ∈ S 2 D \in S_2 D ∈ S 2 has infinite rank, then T = g D T=gD T = g D satisfies (H) with reference G = T G=T G = T and ϑ 1 ( T ) < ∞ \vartheta_1(T)<\infty ϑ 1 ( T ) < ∞ . Nevertheless, I ( T ) = I ( D ) I(T)=I(D) I ( T ) = I ( D ) almost surely while ρ n ( T ) = 0 \rho_n(T)=0 ρ n ( T ) = 0 for every n ≥ 2 n \ge2 n ≥ 2 . Uniformity over sample-dependent shifts is impossible (B = − T B=-T B = − T ).
Curvature on conditional fibers
Lemma 2.3 (Finite-dimensional response). Let W W W have positive density c e − V ce^{-V} c e − V on a finite-dimensional support F F F , and let C > 0 C>0 C > 0 on F F F satisfy (H). For linear observation and target maps O O O , A A A , and g ∼ N F ( 0 , C ) g \sim N_F(0,C) g ∼ N F ( 0 , C ) , put S = Cov ( A g − E [ A g ∣ O g ] ) S=\operatorname{Cov}(Ag-\mathbb{E}[Ag\mid Og]) S = Cov ( A g − E [ A g ∣ O g ]) . Almost surely,
β − 1 S ≤ Cov ( A W ∣ O W ) ≤ α − 1 S . (22) \beta^{-1}S \le\operatorname{Cov}(AW\mid OW) \le\alpha^{-1}S. \tag*{(22)} β − 1 S ≤ Cov ( A W ∣ O W ) ≤ α − 1 S . (22) For an observation-measurable pair ( z , P ) (z,P) ( z , P ) , with P P P an orthogonal projection, and t > 0 t>0 t > 0 ,
E [ e − t ∥ P ( A W − z ) ∥ 2 / 2 ∣ O W ] ≤ det ( I + ( t / β ) P S P ) − 1 / 2 . (23) \mathbb{E}[e^{-t\lVert P(AW-z)\rVert^2/2}\mid OW] \le\det(I+(t/\beta)PSP)^{-1/2}. \tag*{(23)} E [ e − t ∥ P ( A W − z ) ∥ 2 /2 ∣ O W ] ≤ det ( I + ( t / β ) P S P ) − 1/2 . (23) If W W W is centered, every nonnegative convex test satisfies
E Φ ( W ) ≤ E Φ ( α − 1 / 2 g ) . (24) \mathbb{E}\Phi(W) \le\mathbb{E}\Phi(\alpha^{-1/2}g). \tag*{(24)} E Φ ( W ) ≤ E Φ ( α − 1/2 g ) . (24) For a Hilbert-space Gaussian ξ \xi ξ with mean a a a and covariance C 0 C_0 C 0 ,
E e − t ∥ ξ ∥ 2 / 2 = det ( I + t C 0 ) − 1 / 2 exp { − t 2 ⟨ a , ( I + t C 0 ) − 1 a ⟩ } . (25) \mathbb{E}e^{-t\lVert\xi\rVert^2/2}=\det(I+tC_0)^{-1/2}\exp\{-\tfrac{t}{2}\langle a,(I+tC_0)^{-1}a\rangle\}. \tag*{(25)} E e − t ∥ ξ ∥ 2 /2 = det ( I + t C 0 ) − 1/2 exp { − 2 t ⟨ a , ( I + t C 0 ) − 1 a ⟩} . (25) In particular it is at most det ( I + t R ) − 1 / 2 \det(I+tR)^{-1/2} det ( I + tR ) − 1/2 whenever 0 ≤ R ≤ C 0 0\le R\le C_0 0 ≤ R ≤ C 0 .
Proof. Work inside F F F and choose an isometry J J J onto ker O \ker O ker O . The conditional fiber is w y + J v w_y+Jv w y + J v , with potential V ( w y + J v ) V(w_y+Jv) V ( w y + J v ) ; the coarea Jacobian is constant on that fiber. If H = J ∗ C − 1 J H=J^*C^{-1}J H = J ∗ C − 1 J , its Hessian lies between α H \alpha H α H and β H \beta H β H , while completion of the Gaussian square gives
S = A J H − 1 J ∗ A ∗ . (26) S=AJH^{-1}J^*A^*. \tag*{(26)} S = A J H − 1 J ∗ A ∗ . (26)
A zero-dimensional fiber is deterministic. No inverse of a singular ambient covariance has been used.
For a smooth potential U U U with a I ≤ ∇ 2 U ≤ D 0 aI\le\nabla^2U\le D_0 a I ≤ ∇ 2 U ≤ D 0 , integration by parts gives
E ∇ U = 0 , E [ ( Z − E Z ) ( ∇ U ) ∗ ] = I , E [ ( ∇ U ) ( ∇ U ) ∗ ] = E ∇ 2 U . \mathbb{E}\nabla U=0,\qquad\mathbb{E}[(Z-\mathbb{E}Z)(\nabla U)^*]=I,\qquad\mathbb{E}[(\nabla U)(\nabla U)^*]=\mathbb{E}\nabla^2U. E ∇ U = 0 , E [( Z − E Z ) ( ∇ U ) ∗ ] = I , E [( ∇ U ) ( ∇ U ) ∗ ] = E ∇ 2 U .
The block covariance Schur complement yields the Cramér–Rao bound Cov ( Z ) ≥ ( E ∇ 2 U ) − 1 ≥ D 0 − 1 \operatorname{Cov}(Z)\ge(\mathbb{E}\nabla^2U)^{-1}\ge D_0^{-1} Cov ( Z ) ≥ ( E ∇ 2 U ) − 1 ≥ D 0 − 1 . For F ( t ) = E e − t ∥ D Z − z ∥ 2 / 2 F(t)=\mathbb{E}e^{-t\lVert DZ-z\rVert^2/2} F ( t ) = E e − t ∥ D Z − z ∥ 2 /2 , let ν t \nu_t ν t be the tilted probability measure with potential U t ( x ) = U ( x ) + t ∥ D x − z ∥ 2 / 2 U_t(x)=U(x)+t\lVert Dx-z\rVert^2/2 U t ( x ) = U ( x ) + t ∥ D x − z ∥ 2 /2 . Its Hessian is at most D 0 + t D ∗ D D_0+tD^*D D 0 + t D ∗ D . Applying the covariance lower bound to ν t \nu_t ν t gives
− d d t log F ( t ) = 1 2 E ν t ∥ D Z − z ∥ 2 = 1 2 tr ( D Cov ν t ( Z ) D ∗ ) + 1 2 ∥ E ν t Z − z ∥ 2 ≥ 1 2 tr ( D ( D 0 + t D ∗ D ) − 1 D ∗ ) . \begin{aligned}
-\frac{d}{dt}\log F(t)=\frac{1}{2}\mathbb{E}_{\nu_t}\lVert DZ-z\rVert^2 \\
&=\frac{1}{2}\operatorname{tr}(D\operatorname{Cov}_{\nu_t}(Z)D^*)+\frac{1}{2}\lVert\mathbb{E}_{\nu_t}Z-z\rVert^2 \\
&\ge\frac{1}{2}\operatorname{tr}\bigl(D(D_0+tD^*D)^{-1}D^*\bigr).
\end{aligned} − d t d log F ( t ) = 2 1 E ν t ∥ D Z − z ∥ 2 = 2 1 tr ( D Cov ν t ( Z ) D ∗ ) + 2 1 ∥ E ν t Z − z ∥ 2 ≥ 2 1 tr ( D ( D 0 + t D ∗ D ) − 1 D ∗ ) .
The trace on the last line is the derivative of log det ( I + t D D 0 − 1 D ∗ ) \log\det(I+tDD_0^{-1}D^*) log det ( I + t D D 0 − 1 D ∗ ) . Since F ( 0 ) = 1 F(0)=1 F ( 0 ) = 1 , integration yields F ( t ) ≤ det ( I + t D D 0 − 1 D ∗ ) − 1 / 2 F(t)\le\det(I+tDD_0^{-1}D^*)^{-1/2} F ( t ) ≤ det ( I + t D D 0 − 1 D ∗ ) − 1/2 . The nonnegative squared-mean term explains why the estimate is uniform in z z z ; covariance domination alone would not justify this Laplace bound.
Strong convexity supplies Gaussian tails and the upper Hessian bound gives linear gradient growth, justifying the integrations by cutoff. Use D 0 = β H D_0=\beta H D 0 = β H and D = P A J D=PAJ D = P A J on the fiber to obtain (23) and the lower half of (22) . Brascamp–Lieb’s variance inequality [3 ] , applied to fiberwise linear functions, gives the upper half.
The density of centered W W W relative to N F ( 0 , α − 1 C ) N_F(0,\alpha^{-1}C) N F ( 0 , α − 1 C ) is log-concave; Hargé’s theorem [10 ] gives (24) . These statements extend to C l o c 1 , 1 C_{\mathrm{loc}}^{1,1} C loc 1 , 1 potentials by convolution smoothing of the potential: both Hessian bounds persist, the normalizations converge with Gaussian domination, and centering the approximants gives the convex comparison. Fiberwise bounds hold for all centers before passage to the limit. Finally, diagonalization and bounded convergence prove (25) , including deterministic kernel coordinates; min–max gives its last assertion. □
Lemma 2.4 (Linear images and weak limits). A centered finite-dimensional law satisfying (H) relative to C C C retains that condition under a linear map L L L , relative to L C L ∗ LCL^* L C L ∗ on its range, with the same α , β \alpha,\beta α , β . On a fixed finite-dimensional space F F F , suppose centered laws μ j \mu_j μ j with positive densities satisfy the same bounds relative to C j > 0 C_j>0 C j > 0 . If C j → C > 0 C_j\to C>0 C j → C > 0 in operator norm and μ j ⇒ μ \mu_j\Rightarrow\mu μ j ⇒ μ , then μ \mu μ is centered and has a positive density with a C l o c 1 , 1 C_{\mathrm{loc}}^{1,1} C loc 1 , 1 potential satisfying those bounds relative to C C C . This also applies to tests with 0 ≤ C j ↑ C 0\le C_j\uparrow C 0 ≤ C j ↑ C : on F = ran C F=\operatorname{ran}C F = ran C , the reference covariances have full rank eventually.
Proof. For the linear-image assertion, whiten on the original reference support and split into orthogonal coordinates ( x , y ) (x,y) ( x , y ) so that the map depends only on x x x . The marginal potential obeys
W x x = E [ V x x ∣ x ] − Cov ( V x ∣ x ) , Cov ( V x ∣ x ) ≤ E [ V x y V y y − 1 V y x ∣ x ] . W_{xx}=\mathbb{E}[V_{xx}\mid x]-\operatorname{Cov}(V_x\mid x),\qquad\operatorname{Cov}(V_x\mid x)\le\mathbb{E}[V_{xy}V_{yy}^{-1}V_{yx}\mid x]. W xx = E [ V xx ∣ x ] − Cov ( V x ∣ x ) , Cov ( V x ∣ x ) ≤ E [ V x y V y y − 1 V y x ∣ x ] .
The second inequality is Brascamp–Lieb on the y y y -fiber. The Schur complement of ∇ 2 V ≥ α I \nabla^2V\ge\alpha I ∇ 2 V ≥ α I gives W x x ≥ α I W_{xx}\ge\alpha I W xx ≥ α I ; the first identity gives W x x ≤ β I W_{xx}\le\beta I W xx ≤ β I . Transform back on ran ( L C L ∗ ) \operatorname{ran}(LCL^*) ran ( L C L ∗ ) . Smooth approximation treats C 1 , 1 C^{1,1} C 1 , 1 potentials.
For weak closure, discard finitely many indices so C j C_j C j and C j − 1 C_j^{-1} C j − 1 are uniformly bounded. Normalize V j ( 0 ) = 0 V_j(0)=0 V j ( 0 ) = 0 . Integration by parts and convex comparison give
∥ ∇ V j ( 0 ) ∥ ≤ β ∥ C j − 1 ∥ E μ j ∥ x ∥ ≤ M < ∞ . \lVert\nabla V_j(0)\rVert\le\beta\lVert C_j^{-1}\rVert\mathbb{E}_{\mu_j}\lVert x\rVert\le M<\infty. ∥ ∇ V j ( 0 )∥ ≤ β ∥ C j − 1 ∥ E μ j ∥ x ∥ ≤ M < ∞.
The Hessian bounds now give locally equibounded, equicontinuous potentials and gradients, and a common lower bound V j ( x ) ≥ c ∥ x ∥ 2 − M ∥ x ∥ V_j(x)\ge c\lVert x\rVert^2-M\lVert x\rVert V j ( x ) ≥ c ∥ x ∥ 2 − M ∥ x ∥ . Arzelà–Ascoli gives subsequential local C 1 C^1 C 1 convergence. Quadratic domination passes the normalizations and first moments to the limit, identifying a positive centered density. Second-difference inequalities pass both Hessian bounds to its C l o c 1 , 1 C_{\mathrm{loc}}^{1,1} C loc 1 , 1 potential. Uniqueness of the weak limit identifies the law. If C j ↑ C C_j\uparrow C C j ↑ C , restrict to ran C \operatorname{ran}C ran C ; finite-dimensional eigenvalue convergence makes C j C_j C j positive there eventually. No singular ambient inverse is used. □
Trace envelopes and finite-rank deletion
We use Gaussian comparison for ordered increment variances and Gaussian Lipschitz concentration [21 Theorems 6.8 and 3.25] :
E e v ( f − E f ) ≤ e v 2 σ 2 / 2 , P { f − E f > t } ≤ e − t 2 / ( 2 σ 2 ) . (27) \mathbb{E}e^{v(f-\mathbb{E}f)}\le e^{v^2\sigma^2/2},\qquad\mathbb{P}\{f-\mathbb{E}f>t\}\le e^{-t^2/(2\sigma^2)}. \tag*{(27)} E e v ( f − E f ) ≤ e v 2 σ 2 /2 , P { f − E f > t } ≤ e − t 2 / ( 2 σ 2 ) . (27)
Lemma 2.5 (Trace envelopes and spectral concentration). A centered Gaussian bilinear form Z Z Z with envelopes L , R L,R L , R has a unique compact Gaussian realization with square-integrable norm, and
E ∥ Z ∥ ≤ 2 ( tr L + tr R ) . (28) \mathbb{E}\lVert Z\rVert\le\sqrt{2}(\sqrt{\operatorname{tr}L}+\sqrt{\operatorname{tr}R}). \tag*{(28)} E ∥ Z ∥ ≤ 2 ( tr L + tr R ) . (28) Under (H) and ϑ 1 ( G ) < ∞ \vartheta_1(G)<\infty ϑ 1 ( G ) < ∞ , T T T has a unique compact realization with square-integrable norm. For m ≥ 2 n − 1 m\ge2n-1 m ≥ 2 n − 1 , t > 0 t>0 t > 0 and ϑ n ( G ) < ∞ , \vartheta_n(G)<\infty, ϑ n ( G ) < ∞ ,
P { s m ( T ) > ( 2 2 + t ) α − 1 / 2 ϑ n ( G ) } ≤ exp { − t 2 ( m − 2 n + 2 ) / 2 } . (29) \mathbb{P}\{s_m(T)>(2\sqrt{2}+t)\alpha^{-1/2}\vartheta_n(G)\}\le\exp\{-t^2(m-2n+2)/2\}. \tag*{(29)} P { s m ( T ) > ( 2 2 + t ) α − 1/2 ϑ n ( G )} ≤ exp { − t 2 ( m − 2 n + 2 ) /2 } . (29) If ϑ n ( G ) = 0 \vartheta_n(G)=0 ϑ n ( G ) = 0 , then s 2 n − 1 ( T ) = 0 s_{2n-1}(T)=0 s 2 n − 1 ( T ) = 0 almost surely.
Proof. For unit vectors, the increment variances of Z ( x , y ) Z(x,y) Z ( x , y ) are bounded by 2 ⟨ L ( y − y ′ ) , y − y ′ ⟩ + 2 ⟨ R ( x − x ′ ) , x − x ′ ⟩ 2\langle L(y-y'),y-y'\rangle+2\langle R(x-x'),x-x'\rangle 2 ⟨ L ( y − y ′ ) , y − y ′ ⟩ + 2 ⟨ R ( x − x ′ ) , x − x ′ ⟩ . Comparison with 2 ( ⟨ ℓ , y ⟩ + ⟨ r , x ⟩ ) \sqrt{2}(\langle\ell,y\rangle+\langle r,x\rangle) 2 (⟨ ℓ , y ⟩ + ⟨ r , x ⟩) , for independent Gaussians of covariances L , R L,R L , R , gives (28) on finite compressions. Their increasing norms have bounded expectation, defining a bounded realization. Under nested finite-rank P N , Q N → I P_N,Q_N\to I P N , Q N → I , the residual envelopes are the compressions of L L L , R R R , with traces tending to zero. The decreasing residual norms vanish almost surely, and their complements have finite rank. This yields a compact Borel version. The weak variance σ ( Z ) 2 = sup ∥ x ∥ = ∥ y ∥ = 1 E ∣ Z ( x , y ) ∣ 2 \sigma(Z)^2 = \sup_{\lVert x\rVert=\lVert y\rVert=1} \mathbb{E}|Z(x,y)|^2 σ ( Z ) 2 = sup ∥ x ∥ = ∥ y ∥ = 1 E ∣ Z ( x , y ) ∣ 2 controls the Lipschitz constant of every finite maximum of symmetric bilinear tests. Concentration and passage to dense tests give E ∥ Z ∥ 2 < ∞ \mathbb{E}\lVert Z\rVert^2 < \infty E ∥ Z ∥ 2 < ∞ .
Apply (24) to the norm, its square, and finite tail compressions. The same argument gives the compact realization of T T T and E ∥ T ∥ 2 ≤ α − 1 E ∥ G ∥ 2 \mathbb{E}\lVert T\rVert^2 \le\alpha^{-1}\mathbb{E}\lVert G\rVert^2 E ∥ T ∥ 2 ≤ α − 1 E ∥ G ∥ 2 . Covariance comparison on scalar tests ensures L 2 L^2 L 2 agreement with the original forms and uniqueness on dense unions.
Choose a reference residual G P , Q G_{P,Q} G P , Q and envelopes with both traces at most a 2 a^2 a 2 , where a > ϑ n ( G ) a > \vartheta_n(G) a > ϑ n ( G ) . Put r = m − 2 n + 1 r=m-2n+1 r = m − 2 n + 1 and let F F F project onto r r r leading eigendirections of L L L . Write V U = ( I − F ) U P , Q V_U=(I-F)U_{P,Q} V U = ( I − F ) U P , Q for U = T , G U=T,G U = T , G . Then
rank ( T − V T ) ≤ 2 n − 2 + r = m − 1 , σ ( V G ) 2 ≤ λ r + 1 ( L ) ≤ a 2 r + 1 , E ∥ V G ∥ ≤ 2 2 a . \operatorname{rank}(T-V_T)\le2n-2+r=m-1,\qquad\sigma(V_G)^2\le\lambda_{r+1}(L)\le\frac{a^2}{r+1},\qquad\mathbb{E}\lVert V_G\rVert\le2\sqrt{2}a. rank ( T − V T ) ≤ 2 n − 2 + r = m − 1 , σ ( V G ) 2 ≤ λ r + 1 ( L ) ≤ r + 1 a 2 , E ∥ V G ∥ ≤ 2 2 a .
Convex comparison for the exponential norm, followed by (27) , gives
E e v ∥ V T ∥ ≤ exp { 2 2 v a / α + v 2 a 2 2 α ( r + 1 ) } . \mathbb{E}e^{v\lVert V_T\rVert}\le\exp\left\{2\sqrt{2}va/\sqrt{\alpha}+\frac{v^2a^2}{2\alpha(r+1)}\right\}. E e v ∥ V T ∥ ≤ exp { 2 2 v a / α + 2 α ( r + 1 ) v 2 a 2 } .
Chernoff optimization and a ↓ ϑ n ( G ) a\downarrow\vartheta_n(G) a ↓ ϑ n ( G ) prove (29) . Zero infimum is handled by the expected residual norm and s 2 n − 1 ( T ) ≤ ∥ T P , Q ∥ s_{2n-1}(T)\le\lVert T_{P,Q}\rVert s 2 n − 1 ( T ) ≤ ∥ T P , Q ∥ .
Conditional distances and inverse Gram estimates
For Gaussian Y : H → E Y:H\to E Y : H → E , linear prediction [11 ] gives
S E , u , x Y = inf M ∗ u = 0 Var ⟨ Y , u ⊗ x + M ⟩ 2 , M ∈ S 2 ( H , E ) . (30) S^Y_{E,u,x}=\inf_{M^*u=0}\operatorname{Var}\langle Y,u\otimes x+M\rangle_2,\qquad M\in S_2(H,E). \tag*{(30)} S E , u , x Y = M ∗ u = 0 inf Var ⟨ Y , u ⊗ x + M ⟩ 2 , M ∈ S 2 ( H , E ) . (30)
Finite-rank tests are dense in the Gaussian observation space. Thus the reference residual increases when observations are removed, is monotone and homogeneous in the full Gaussian covariance, and satisfies S E , u Y ≥ S S^Y_{E,u}\ge S S E , u Y ≥ S when Cov ( Y ) ≥ I E ⊗ S \operatorname{Cov}(Y)\ge I_E\otimes S Cov ( Y ) ≥ I E ⊗ S . Here I E ⊗ S I_E\otimes S I E ⊗ S is the form M ↦ ∥ M S 1 / 2 ∥ 2 2 M\mapsto\lVert MS^{1/2}\rVert_2^2 M ↦ ∥ M S 1/2 ∥ 2 2 .
Lemma 2.6 (Conditional distances on finite supports). Let Y Y Y have m m m finite-dimensional rows, with Gaussian reference rows g i g_i g i satisfying (H). Put S i = Cov ( g i − E [ g i ∣ ( g l ) l ≠ i ] ) S_i=\operatorname{Cov}(g_i-\mathbb{E}[g_i\mid(g_l)_{l\ne i}]) S i = Cov ( g i − E [ g i ∣ ( g l ) l = i ]) . After any fixed row shift, let d i d_i d i be the distance from row i i i to the span of the other rows. With F i \mathcal{F}_i F i generated by those other rows,
E [ e − t d i 2 / 2 ∣ F i ] ≤ ∏ k ≥ m ( 1 + t λ k ( S i ) / β ) − 1 / 2 , (31) \mathbb{E}[e^{-td_i^2/2}\mid\mathcal{F}_i]\le\prod_{k\ge m}(1+t\lambda_k(S_i)/\beta)^{-1/2}, \tag*{(31)} E [ e − t d i 2 /2 ∣ F i ] ≤ k ≥ m ∏ ( 1 + t λ k ( S i ) / β ) − 1/2 , (31) P { d i 2 < a ∣ F i } ≤ inf t > 0 e t a / 2 ∏ k ≥ m ( 1 + t λ k ( S i ) / β ) − 1 / 2 . (32) \mathbb{P}\{d_i^2<a\mid\mathcal{F}_i\}\le\inf_{t>0}e^{ta/2}\prod_{k\ge m}(1+t\lambda_k(S_i)/\beta)^{-1/2}. \tag*{(32)} P { d i 2 < a ∣ F i } ≤ t > 0 inf e t a /2 k ≥ m ∏ ( 1 + t λ k ( S i ) / β ) − 1/2 . (32) Proof. Gram–Schmidt makes the projection off the other shifted rows measurable; its codimension is at most m − 1 m-1 m − 1 . Lemma 2.3 applies on the conditional support, even if the joint row covariance is singular. Min–max gives λ k ( P S i P ) ≥ λ k + m − 1 ( S i ) \lambda_k(PS_iP)\ge\lambda_{k+m-1}(S_i) λ k ( P S i P ) ≥ λ k + m − 1 ( S i ) , proving the first bound. Conditional Markov proves the second. Fixed shifts change only the center, not the observation sigma fields. Rational t t t and continuity suffice.
Lemma 2.7 (Inverse distances and singular values). Let Y : H → R m Y:H\to\mathbb{R}^m Y : H → R m have rows ξ 1 , … , ξ m \xi_1,\ldots,\xi_m ξ 1 , … , ξ m , and put d i = dist ( ξ i , span { ξ l : l ≠ i } ) d_i=\operatorname{dist}(\xi_i,\operatorname{span}\{\xi_l:l\ne i\}) d i = dist ( ξ i , span { ξ l : l = i }) . For any nonempty set I I I of g g g indices with d i > 0 d_i>0 d i > 0 on I I I ,
s j ( Y ) 2 ≥ g − j + 1 ∑ i ∈ I d i − 2 ( 1 ≤ j ≤ g ) . (33) s_j(Y)^2\ge\frac{g-j+1}{\sum_{i\in I}d_i^{-2}}\qquad(1\le j\le g). \tag*{(33)} s j ( Y ) 2 ≥ ∑ i ∈ I d i − 2 g − j + 1 ( 1 ≤ j ≤ g ) . (33) In particular, if at most b b b distances have d i 2 < A / 2 d_i^2<A/2 d i 2 < A /2 , where A > 0 A>0 A > 0 and 0 ≤ b < m 0\le b<m 0 ≤ b < m is an integer, then
s j ( Y ) 2 ≥ A ( m − b − j + 1 ) 2 ( m − b ) ( 1 ≤ j ≤ m − b ) . s_j(Y)^2\ge\frac{A(m-b-j+1)}{2(m-b)}\qquad(1\le j\le m-b). s j ( Y ) 2 ≥ 2 ( m − b ) A ( m − b − j + 1 ) ( 1 ≤ j ≤ m − b ) .
Proof. The selected rows are linearly independent. Their compression Y I Y_I Y I has row distances e i ≥ d i e_i \ge d_i e i ≥ d i and positive definite Gram matrix G = Y I Y I ∗ G = Y_IY_I^* G = Y I Y I ∗ . The Schur complement gives ( G − 1 ) i i = e i − 2 (G^{-1})_{ii} = e_i^{-2} ( G − 1 ) ii = e i − 2 , hence the negative second-moment identity [20 Lemma A.4] yields
∑ l = 1 g s l ( Y I ) − 2 = tr ( G − 1 ) = ∑ i ∈ I e i − 2 ≤ ∑ i ∈ I d i − 2 . \sum_{l=1}^{g} s_l(Y_I)^{-2} = \operatorname{tr}(G^{-1}) = \sum_{i\in I} e_i^{-2} \le\sum_{i\in I} d_i^{-2}. l = 1 ∑ g s l ( Y I ) − 2 = tr ( G − 1 ) = i ∈ I ∑ e i − 2 ≤ i ∈ I ∑ d i − 2 .
The last g − j + 1 g-j+1 g − j + 1 terms are at least ( g − j + 1 ) s j ( Y I ) − 2 (g-j+1)s_j(Y_I)^{-2} ( g − j + 1 ) s j ( Y I ) − 2 . Together with s j ( Y ) ≥ s j ( Y I ) s_j(Y) \ge s_j(Y_I) s j ( Y ) ≥ s j ( Y I ) , this proves (33) . For the final assertion, select m − b m-b m − b indices with d i 2 ≥ A / 2 d_i^2 \ge A/2 d i 2 ≥ A /2 and bound the sum of inverse squared distances by 2 ( m − b ) / A 2(m-b)/A 2 ( m − b ) / A .
Proposition 2.8 (Negative moments from conditional spectral tails). Let Y : H → R m Y:H\to\mathbb{R}^m Y : H → R m have square-integrable Hilbert-space rows , with Gaussian reference rows g i g_i g i whose finite tests satisfy (H). Let S i = Cov ( g i − E [ g i ∣ ( g l ) l ≠ i ] ) S_i=\operatorname{Cov}(g_i-\mathbb{E}[g_i\mid(g_l)_{l\ne i}]) S i = Cov ( g i − E [ g i ∣ ( g l ) l = i ]) and τ i ( p ) = t m + 2 p ( S i ) \tau_i(p)=t_{m+2p}(S_i) τ i ( p ) = t m + 2 p ( S i ) , where p ≥ 1 p\ge1 p ≥ 1 is an integer. If every τ i ( p ) > 0 \tau_i(p)>0 τ i ( p ) > 0 , then for every fixed row shift B B B ,
∥ s j ( Y + B ) − 2 ∥ L p ≤ β m − j + 1 ∑ i = 1 m τ i ( p ) − 1 ( 1 ≤ j ≤ m ) . (34) \left\|s_j(Y+B)^{-2}\right\|_{L^p}\le\frac{\beta}{m-j+1}\sum_{i=1}^{m}\tau_i(p)^{-1}\qquad(1\le j\le m). \tag*{(34)} s j ( Y + B ) − 2 L p ≤ m − j + 1 β i = 1 ∑ m τ i ( p ) − 1 ( 1 ≤ j ≤ m ) . (34) Under Theorem A 's hypotheses, suppose dim E = n \dim E=n dim E = n and t n ( S E , u G ∘ ) ≥ τ t_n(S^{{G^\circ}}_{E,u})\ge\tau t n ( S E , u G ∘ ) ≥ τ for every unit u ∈ E u\in E u ∈ E , where G ∘ = G G^\circ=G G ∘ = G or G ∗ G^* G ∗ and τ ≥ 0 \tau\ge0 τ ≥ 0 . Then
P { s j ( T + B ) 2 < ( n − j ) τ 4 β n } ≤ 2 − ⌊ ( n − j + 1 ) / 4 ⌋ ( 1 ≤ j < n ) . (35) \mathbb{P}\left\{s_j(T+B)^2<\frac{(n-j)\tau}{4\beta n}\right\}\le2^{-\lfloor(n-j+1)/4\rfloor}\qquad(1\le j<n). \tag*{(35)} P { s j ( T + B ) 2 < 4 β n ( n − j ) τ } ≤ 2 − ⌊( n − j + 1 ) /4 ⌋ ( 1 ≤ j < n ) . (35) In particular, if ρ n ( G ) > 0 \rho_n(G)>0 ρ n ( G ) > 0 , then s n ( T + B ) > 0 s_n(T+B)>0 s n ( T + B ) > 0 almost surely.
Proof. First suppose the rows are finite-dimensional. For each i i i , put q = λ m + 2 p ( S i ) / β > 0 q=\lambda_{m+2p}(S_i)/\beta>0 q = λ m + 2 p ( S i ) / β > 0 , τ = τ i ( p ) / β \tau=\tau_i(p)/\beta τ = τ i ( p ) / β . In (31) , the terms with indices m , … , m + 2 p − 1 m,\ldots,m+2p-1 m , … , m + 2 p − 1 , divided by β \beta β , are at least q q q . For each term x ∈ [ 0 , q ] x\in[0,q] x ∈ [ 0 , q ] in the remaining tail, concavity gives log ( 1 + t x ) ≥ ( x / q ) log ( 1 + t q ) \log(1+tx)\ge(x/q)\log(1+tq) log ( 1 + t x ) ≥ ( x / q ) log ( 1 + tq ) . The tail sum is τ \tau τ , so the conditional Laplace transform is at most ( 1 + t q ) − p − τ / ( 2 q ) (1+tq)^{-p-\tau/(2q)} ( 1 + tq ) − p − τ / ( 2 q ) . This also tends to zero as t → ∞ t\to\infty t → ∞ , showing that each distance is positive almost surely. For x > 0 x>0 x > 0 , the identity x − p = ( 2 p Γ ( p ) ) − 1 ∫ 0 ∞ t p − 1 e − t x / 2 d t x^{-p}=(2^p\Gamma(p))^{-1}\int_0^\infty t^{p-1}e^{-tx/2}\,\mathrm{d}t x − p = ( 2 p Γ ( p ) ) − 1 ∫ 0 ∞ t p − 1 e − t x /2 d t and Tonelli’s theorem now give
E [ d i − 2 p ∣ F i ] ≤ 1 2 p Γ ( p ) ∫ 0 ∞ t p − 1 ( 1 + t q ) − p − τ / ( 2 q ) d t = ∏ l = 0 p − 1 ( τ + 2 l q ) − 1 ≤ τ − p . (36) \mathbb{E}[d_i^{-2p}\mid\mathcal{F}_i]\le\frac{1}{2^p\Gamma(p)}\int_0^\infty t^{p-1}(1+tq)^{-p-\tau/(2q)}\,\mathrm{d}t=\prod_{l=0}^{p-1}(\tau+2lq)^{-1}\le\tau^{-p}. \tag*{(36)} E [ d i − 2 p ∣ F i ] ≤ 2 p Γ ( p ) 1 ∫ 0 ∞ t p − 1 ( 1 + tq ) − p − τ / ( 2 q ) d t = l = 0 ∏ p − 1 ( τ + 2 l q ) − 1 ≤ τ − p . (36)
Positive distances, Lemma 2.7, and Minkowski prove (34) ; the distances need not be probabilistically independent.
For Hilbert-space rows take finite-rank P N ↑ I H P_N\uparrow I_H P N ↑ I H and define
S i , N = Cov ( P N g i − E [ P N g i ∣ ( P N g l ) l ≠ i ] ) . S_{i,N}=\operatorname{Cov}(P_Ng_i-\mathbb{E}[P_Ng_i\mid(P_Ng_l)_{l\ne i}]). S i , N = Cov ( P N g i − E [ P N g i ∣ ( P N g l ) l = i ]) .
Fewer observations and (30) give
S i , N ≥ P N S i P N , t m + 2 p ( P N S i P N ) ⟶ τ i ( p ) . (37) S_{i,N}\ge P_NS_iP_N,\qquad t_{m+2p}(P_NS_iP_N)\longrightarrow\tau_i(p). \tag*{(37)} S i , N ≥ P N S i P N , t m + 2 p ( P N S i P N ) ⟶ τ i ( p ) . (37)
The convergence follows from trace-norm convergence of the compressions. Apply the finite-dimensional estimate to ( Y + B ) P N (Y+B)P_N ( Y + B ) P N , using the right side of (37) as the lower bound for the spectral tail. Finite rows imply ( Y + B ) P N → Y + B (Y+B)P_N\to Y+B ( Y + B ) P N → Y + B almost surely in Hilbert–Schmidt norm. Singular-value continuity and Fatou prove (34) . No infinite-dimensional conditional density is required.
For (35) , τ = 0 \tau=0 τ = 0 is empty and n − j ≤ 2 n-j\le2 n − j ≤ 2 is trivial. Otherwise set m = ⌊ ( n + j ) / 2 ⌋ m=\lfloor(n+j)/2\rfloor m = ⌊( n + j ) /2 ⌋ , p = ⌊ ( n − j + 1 ) / 4 ⌋ ≥ 1 p=\lfloor(n-j+1)/4\rfloor\ge1 p = ⌊( n − j + 1 ) /4 ⌋ ≥ 1 ; then m + 2 p ≤ n m+2p\le n m + 2 p ≤ n . Compress T ∘ + B ∘ T^\circ+B^\circ T ∘ + B ∘ onto any deterministic m m m -dimensional E ′ ⊂ E E'\subset E E ′ ⊂ E and choose an orthonormal basis ( u i ) (u_i) ( u i ) . The reference residuals obey S i ≥ S E , u i G ∘ S_i\ge S^{G^\circ}_{E,u_i} S i ≥ S E , u i G ∘ , hence τ i ( p ) ≥ τ \tau_i(p)\ge\tau τ i ( p ) ≥ τ . The moment bound and Markov give failure probability at most 2 − p 2^{-p} 2 − p at squared threshold ( m − j + 1 ) τ / ( 2 β m ) (m-j+1)\tau/(2\beta m) ( m − j + 1 ) τ / ( 2 β m ) . Use ( m − j + 1 ) / ( 2 m ) ≥ ( n − j ) / ( 4 n ) (m-j+1)/(2m)\ge(n-j)/(4n) ( m − j + 1 ) / ( 2 m ) ≥ ( n − j ) / ( 4 n ) and compression.
If ρ n ( G ) > 0 \rho_n(G)>0 ρ n ( G ) > 0 , choose a window with positive inner infimum. Its n n n reference rows have λ n ( S i ) > 0 \lambda_n(S_i)>0 λ n ( S i ) > 0 . One sufficiently large N N N makes all λ n ( P N S i P N ) > 0 \lambda_n(P_NS_iP_N)>0 λ n ( P N S i P N ) > 0 . For this fixed finite compression, (2.26) tends to zero as t → ∞ t \to\infty t → ∞ for every row. Thus its shifted rows are almost surely linearly independent, so s n ( T + B ) > 0 s_n(T+B)>0 s n ( T + B ) > 0 , including singular ambient covariances. □
Actual residuals and closure
Lemma 2.9 (Comparison after infinite observation). Under (H) and ϑ 1 ( G ) < ∞ \vartheta_1(G)<\infty ϑ 1 ( G ) < ∞ , for each deterministic E , u E,u E , u , almost surely
β − 1 S E , u G ≤ Cov ( T ∗ u ∣ F E , u T ) ≤ α − 1 S E , u G . (38) \beta^{-1}S^G_{E,u} \leq\operatorname{Cov}(T^*u \mid\mathcal{F}^T_{E,u}) \leq\alpha^{-1}S^G_{E,u}. \tag*{(38)} β − 1 S E , u G ≤ Cov ( T ∗ u ∣ F E , u T ) ≤ α − 1 S E , u G . (38) Consequently, for f = ρ , η , ϑ f=\rho,\eta,\vartheta f = ρ , η , ϑ ,
β − 1 / 2 f n ( G ) ≤ f n ( T ) ≤ α − 1 / 2 f n ( G ) . (39) \beta^{-1/2}f_n(G) \leq f_n(T) \leq\alpha^{-1/2}f_n(G). \tag*{(39)} β − 1/2 f n ( G ) ≤ f n ( T ) ≤ α − 1/2 f n ( G ) . (39)
Proof. Fix finitely many target coordinates of T ∗ u T^*u T ∗ u , forming a vector U U U . Let F N \mathcal{F}_N F N be generated by the first N N N coordinates of the other rows in a fixed basis of E ∩ u ⊥ E\cap u^\perp E ∩ u ⊥ . Lemma 2.3 gives
β − 1 S N G ≤ Cov ( U ∣ F N ) ≤ α − 1 S N G . \beta^{-1}S^G_N \leq\operatorname{Cov}(U\mid\mathcal{F}_N)\leq\alpha^{-1}S^G_N. β − 1 S N G ≤ Cov ( U ∣ F N ) ≤ α − 1 S N G .
Conditional first and second moments converge almost surely by martingale convergence. Gaussian prediction converges in L 2 L^2 L 2 , so the deterministic S N G S^G_N S N G converge to the corresponding compression of S E , u G S^G_{E,u} S E , u G . Intersect over a countable dense set of target coordinates; the conditional second moment defines a positive trace-class operator almost surely. This proves (38) . No common conditional version for uncountably many E , u E,u E , u is asserted or needed.
Taking expectation gives β − 1 S E , u G ≤ S E , u T ≤ α − 1 S E , u G \beta^{-1}S^G_{E,u}\leq S^T_{E,u}\leq\alpha^{-1}S^G_{E,u} β − 1 S E , u G ≤ S E , u T ≤ α − 1 S E , u G . Spectral tails, direction infima and window suprema prove the ρ \rho ρ comparison. Empty observation in (22) gives β − 1 C G ≤ C T ≤ α − 1 C G \beta^{-1}C_G\leq C_T\leq\alpha^{-1}C_G β − 1 C G ≤ C T ≤ α − 1 C G . Sum directional variances to compare the two mean Grams of every deterministic residual, and then optimize, proving the η \eta η comparison. Scale each admissible envelope to prove the ϑ \vartheta ϑ comparison. □
Proof of Theorem A . Lemma 2.5 supplies the compact realizations and second moments. For any centered square-integrable random compact U U U , summing envelope inequalities gives η ( U ) ≤ ϑ ( U ) \eta(U)\leq\vartheta(U) η ( U ) ≤ ϑ ( U ) . If dim E = n \dim E=n dim E = n and rank P , rank Q < n \operatorname{rank}P,\operatorname{rank}Q<n rank P , rank Q < n , take a unit u ∈ E ∩ ker P u\in E\cap\ker P u ∈ E ∩ ker P . Orthogonal projection in L 2 L^2 L 2 gives
max { ∥ Γ L U ( P , Q ) ∥ , ∥ Γ R U ( P , Q ) ∥ } ≥ E ∥ ( I − Q ) U ∗ u ∥ 2 ≥ tr ( ( I − Q ) S E , u U ) ≥ t n ( S E , u U ) . \max\left\{\left\|\Gamma^U_L(P,Q)\right\|,\left\|\Gamma^U_R(P,Q)\right\|\right\}\geq\mathbb{E}\left\|(I-Q)U^*u\right\|^2\geq\operatorname{tr}\left((I-Q)S^U_{E,u}\right)\geq t_n(S^U_{E,u}). max { Γ L U ( P , Q ) , Γ R U ( P , Q ) } ≥ E ∥ ( I − Q ) U ∗ u ∥ 2 ≥ tr ( ( I − Q ) S E , u U ) ≥ t n ( S E , u U ) .
Interchange input/output and optimize to get ρ ≤ η \rho\leq\eta ρ ≤ η . Reducing observations increases the unconditional residual covariance; restricting windows proves ρ \rho ρ is nonincreasing. Larger admissible deletion ranks decrease η , ϑ \eta,\vartheta η , ϑ , and fixed-envelope tails give ϑ n → 0 \vartheta_n\to0 ϑ n → 0 . Lemma 2.9 completes (17) .
Proposition 2.8, on windows approaching the supremum, gives
P { s j ( T + B ) 2 < ( n − j ) ρ n ( G ) 2 / ( 4 β n ) } ≤ 2 − ⌊ ( n − j + 1 ) / 4 ⌋ . (40) \mathbb{P}\left\{s_j(T+B)^2<(n-j)\rho_n(G)^2/(4\beta n)\right\}\leq2^{-\lfloor(n-j+1)/4\rfloor}. \tag*{(40)} P { s j ( T + B ) 2 < ( n − j ) ρ n ( G ) 2 / ( 4 β n ) } ≤ 2 − ⌊( n − j + 1 ) /4 ⌋ . (40)
Strict sublevels justify this limit. For rational L > 2 L>2 L > 2 , take n = ⌊ m / L ⌋ n=\lfloor m/L\rfloor n = ⌊ m / L ⌋ and t = 1 t=1 t = 1 in (29) ; for rational λ > 1 \lambda>1 λ > 1 , take j = ⌈ n / λ ⌉ j=\lceil n/\lambda\rceil j = ⌈ n / λ ⌉ in (40) for all sufficiently large n n n . Both failure series converge. Borel–Cantelli gives eventual comparisons. A zero upper profile gives rank T ≤ 2 n − 2 \operatorname{rank}T\leq2n-2 rank T ≤ 2 n − 2 ; a positive lower profile gives s n ( T + B ) > 0 s_n(T+B)>0 s n ( T + B ) > 0 . Absorb finitely many initial ratios and intersect over rational parameters. Monotonicity extends to all real L > 2 , λ > 1 L>2,\lambda>1 L > 2 , λ > 1 , and (39) gives (19) and (20) –(21) .
The finite-deletion criterion gives the two ideal inclusions. If their bounds agree in an ideal J J J , then ρ ≤ η ≤ ϑ \rho\leq\eta\leq\vartheta ρ ≤ η ≤ ϑ identifies J = I ( η ( T ) ) J=\mathcal{I}(\eta(T)) J = I ( η ( T )) . Lemma 2.2, with C = 0 C=0 C = 0 , proves (18) . Reference and actual ideals agree by (39) . The ideal events are countable unions and intersections of singular-value inequalities. Thus prescribed countable shifts and independent random compact shifts follow by countable intersection and Fubini, without moment assumptions on the shifts. □
Covariance calculations and linear contractions
Finite positive covariance sums
Throughout this subsection, ρ ( Γ ) \rho(\Gamma) ρ ( Γ ) , η ( Γ ) \eta(\Gamma) η ( Γ ) , ϑ ( Γ ) \vartheta(\Gamma) ϑ ( Γ ) mean the profiles of the centered Gaussian form with covariance Γ \Gamma Γ . The actual non-Gaussian residual profile is not determined by covariance alone. Gaussian forms satisfy (H) with α = β = 1 \alpha=\beta=1 α = β = 1 , so Theorem A applies without changing any constants below. The regression variational formula makes the lower profile monotone in the full covariance Γ \Gamma Γ ; the upper profile is monotone by its definition. Both satisfy f n ( a Γ ) = a f n ( Γ ) f_n(a\Gamma)=\sqrt{a}\,f_n(\Gamma) f n ( a Γ ) = a f n ( Γ ) for f = ρ , ϑ f=\rho,\vartheta f = ρ , ϑ and a > 0 a>0 a > 0 .
Proposition 3.1 (Closure under finite sums and comparison). Let Γ = ∑ ℓ = 1 r Γ ℓ \Gamma=\sum_{\ell=1}^{r}\Gamma_\ell Γ = ∑ ℓ = 1 r Γ ℓ , r < ∞ r<\infty r < ∞ , be a sum of positive Gaussian covariance forms with ϑ 1 ( Γ ℓ ) < ∞ \vartheta_1(\Gamma_\ell)<\infty ϑ 1 ( Γ ℓ ) < ∞ . For integers n , n ℓ ≥ 1 n,n_\ell\ge1 n , n ℓ ≥ 1 ,
ρ n ( Γ ) ≥ max ℓ ρ n ( Γ ℓ ) , ϑ 1 + ∑ ℓ ( n ℓ − 1 ) ( Γ ) 2 ≤ ∑ ℓ ϑ n ℓ ( Γ ℓ ) 2 . (41) \rho_n(\Gamma)\ge\max_\ell\rho_n(\Gamma_\ell),\qquad
\vartheta_{1+\sum_\ell(n_\ell-1)}(\Gamma)^2\le\sum_\ell\vartheta_{n_\ell}(\Gamma_\ell)^2. \tag*{(41)} ρ n ( Γ ) ≥ ℓ max ρ n ( Γ ℓ ) , ϑ 1 + ∑ ℓ ( n ℓ − 1 ) ( Γ ) 2 ≤ ℓ ∑ ϑ n ℓ ( Γ ℓ ) 2 . (41) If I ( ρ ( Γ ℓ ) ) = I ( ϑ ( Γ ℓ ) ) = : J ℓ \mathcal I(\rho(\Gamma_\ell))=\mathcal I(\vartheta(\Gamma_\ell))=:J_\ell I ( ρ ( Γ ℓ )) = I ( ϑ ( Γ ℓ )) =: J ℓ for every ℓ \ell ℓ , then
I ( ρ ( Γ ) ) = I ( ϑ ( Γ ) ) = ∑ ℓ J ℓ . (42) \mathcal I(\rho(\Gamma))=\mathcal I(\vartheta(\Gamma))=\sum_\ell J_\ell. \tag*{(42)} I ( ρ ( Γ )) = I ( ϑ ( Γ )) = ℓ ∑ J ℓ . (42) The same common ideal is obtained for any centered Gaussian form of covariance Γ ~ \widetilde{\Gamma} Γ with a Γ ≤ Γ ~ ≤ b Γ a\Gamma\le\widetilde{\Gamma}\le b\Gamma a Γ ≤ Γ ≤ b Γ , 0 < a ≤ b < ∞ 0<a\le b<\infty 0 < a ≤ b < ∞ . Its realization therefore satisfies I ( X + D ) = ∑ ℓ J ℓ + I ( D ) \mathcal I(X+D)=\sum_\ell J_\ell+\mathcal I(D) I ( X + D ) = ∑ ℓ J ℓ + I ( D ) for each fixed compact D D D .
Proof. Covariance order proves the lower bound. For the upper bound choose near-optimal deletions and envelopes for each summand, and join the respective input and output deletion spaces. Their ranks are at most ∑ ℓ ( n ℓ − 1 ) \sum_\ell(n_\ell-1) ∑ ℓ ( n ℓ − 1 ) . Each new residual is a compression of its old one; its envelopes are the corresponding compressions of L ℓ L_\ell L ℓ , R ℓ R_\ell R ℓ , with no larger traces. Adding these envelopes and letting the approximation error vanish proves (41) .
With w n = max ℓ ϑ n ( Γ ℓ ) w_n=\max_\ell\vartheta_n(\Gamma_\ell) w n = max ℓ ϑ n ( Γ ℓ ) and all n ℓ = n n_\ell=n n ℓ = n , it follows that ϑ m ( Γ ) ≤ r w ⌈ m / r ⌉ \vartheta_m(\Gamma)\le\sqrt{r}\,w_{\lceil m/r\rceil} ϑ m ( Γ ) ≤ r w ⌈ m / r ⌉ . Thus
I ( ϑ ( Γ ) ) ⊆ ∑ ℓ J ℓ ⊆ I ( ρ ( Γ ) ) ⊆ I ( ϑ ( Γ ) ) . \mathcal I(\vartheta(\Gamma))\subseteq\sum_\ell J_\ell\subseteq\mathcal I(\rho(\Gamma))\subseteq\mathcal I(\vartheta(\Gamma)). I ( ϑ ( Γ )) ⊆ ℓ ∑ J ℓ ⊆ I ( ρ ( Γ )) ⊆ I ( ϑ ( Γ )) .
Covariance order and scaling give the same chain for Γ ~ \widetilde{\Gamma} Γ . Apply Theorem A . □ \square □
Only finite positive covariance sums are asserted, with the displayed dilation; arbitrary correlated summands need not have this sum covariance.
Tensor covariance and energy
Write F a ( n ) = ∑ k ≥ n a k 2 F_a(n)=\sum_{k\ge n}a_k^2 F a ( n ) = ∑ k ≥ n a k 2 . For T ∈ L 2 ( Ω ; S 2 ) T\in L^2(\Omega;S_2) T ∈ L 2 ( Ω ; S 2 ) , let
E p , q ( T ) = inf ( P , Q ) ∈ P p , q E ∥ T P , Q ∥ 2 2 . \mathcal E_{p,q}(T)=\inf_{(P,Q)\in\mathcal P_{p,q}}\mathbb E\lVert T_{P,Q}\rVert_2^2. E p , q ( T ) = ( P , Q ) ∈ P p , q inf E ∥ T P , Q ∥ 2 2 .
For Hilbert–Schmidt A : E 1 → H 1 A:E_1\to H_1 A : E 1 → H 1 and B : H 0 → E 0 B:H_0\to E_0 B : H 0 → E 0 , the notation A G B AGB A GB means the Gaussian contraction specified by
⟨ A G B , M ⟩ 2 = W ( A ∗ M B ∗ ) , M ∈ S 2 ( H 0 , H 1 ) , \langle AGB,M\rangle_2=\mathbb W(A^*MB^*),\qquad M\in S_2(H_0,H_1), ⟨ A GB , M ⟩ 2 = W ( A ∗ M B ∗ ) , M ∈ S 2 ( H 0 , H 1 ) ,
where W \mathbb W W is an isonormal process on S 2 ( E 0 , E 1 ) S_2(E_0,E_1) S 2 ( E 0 , E 1 ) . The construction below makes this an S 2 S_2 S 2 -valued random operator; G G G is not a samplewise bounded operator.
Corollary 3.2 (Gaussian contraction and its energy profile). For fixed Hilbert–Schmidt A , B A,B A , B , this contraction exists and satisfies
E ∥ A G B ∥ 2 2 = ∥ A ∥ 2 2 ∥ B ∥ 2 2 . (43) \mathbb E\lVert AGB\rVert_2^2=\lVert A\rVert_2^2\lVert B\rVert_2^2. \tag*{(43)} E ∥ A GB ∥ 2 2 = ∥ A ∥ 2 2 ∥ B ∥ 2 2 . (43) With a = s ( A ) a=s(A) a = s ( A ) , b = s ( B ) b=s(B) b = s ( B ) and h n 2 = a n 2 F b ( n ) + b n 2 F a ( n + 1 ) h_n^2=a_n^2F_b(n)+b_n^2F_a(n+1) h n 2 = a n 2 F b ( n ) + b n 2 F a ( n + 1 ) , one has
E p , q ( A G B ) = F a ( p + 1 ) F b ( q + 1 ) , h n 2 = E n − 1 , n − 1 ( A G B ) − E n , n ( A G B ) . (44) \mathcal E_{p,q}(AGB)=F_a(p+1)F_b(q+1),\qquad
h_n^2=\mathcal E_{n-1,n-1}(AGB)-\mathcal E_{n,n}(AGB). \tag*{(44)} E p , q ( A GB ) = F a ( p + 1 ) F b ( q + 1 ) , h n 2 = E n − 1 , n − 1 ( A GB ) − E n , n ( A GB ) . (44) For every fixed compact D D D ,
I ( A G B + D ) = I ( b ) + I ( D ) almost surely . (45) \mathcal I(AGB+D)=\mathcal I(b)+\mathcal I(D)\quad\text{almost surely}. \tag*{(45)} I ( A GB + D ) = I ( b ) + I ( D ) almost surely . (45) The construction and identity also hold conditionally for random ( A , B , D ) ∈ S 2 × S 2 × K (A,B,D)\in S_2\times S_2\times\mathcal K ( A , B , D ) ∈ S 2 × S 2 × K independent of W \mathbb W W , without unconditional moments. Finite singular-value lists are zero-completed.
Proof. In any orthonormal bases of E 1 , E 0 E_1,E_0 E 1 , E 0 , the same isonormal process gives
A G B = ∑ i , j g i j ( A e i ) ⊗ ( B ∗ f j ) , g i j = W ( e i ⊗ f j ) . (46) AGB=\sum_{i,j}g_{ij}(Ae_i)\otimes(B^*f_j),\qquad g_{ij}=\mathbb{W}(e_i\otimes f_j). \tag*{(46)} A GB = i , j ∑ g ij ( A e i ) ⊗ ( B ∗ f j ) , g ij = W ( e i ⊗ f j ) . (46)
The squared coefficient norms sum to ∥ A ∥ 2 2 ∥ B ∥ 2 2 \lVert A\rVert_2^2\lVert B\rVert_2^2 ∥ A ∥ 2 2 ∥ B ∥ 2 2 . Hilbert-space martingale convergence [5 ] proves the L 2 L^2 L 2 and almost sure square-section limits and isometry. Applied to differences, the same isometry gives joint continuity in ( A , B ) (A,B) ( A , B ) into L 2 ( Ω ; S 2 ) L^2(\Omega;S_2) L 2 ( Ω ; S 2 ) . Taking the square-section limit where it exists, and zero otherwise, gives a jointly Borel realization.
Set M n = max { a n 2 F b ( n ) , b n 2 F a ( n ) } M_n=\max\{a_n^2F_b(n),b_n^2F_a(n)\} M n = max { a n 2 F b ( n ) , b n 2 F a ( n )} . Delete the first n − 1 n-1 n − 1 left singular directions of A A A and right singular directions of B B B , obtaining A ′ , B ′ A',B' A ′ , B ′ . The covariance formula gives directional envelopes ∥ B ′ ∥ 2 2 A ′ A ′ ∗ \lVert B'\rVert_2^2A'A'^* ∥ B ′ ∥ 2 2 A ′ A ′ ∗ and ∥ A ′ ∥ 2 2 B ′ ∗ B ′ \lVert A'\rVert_2^2B'^*B' ∥ A ′ ∥ 2 2 B ′ ∗ B ′ ; their maximum trace is M n M_n M n . The first n n n left singular directions of A A A give a lower witness S = a n 2 B ∗ B S=a_n^2B^*B S = a n 2 B ∗ B ; the adjoint gives b n 2 A A ∗ b_n^2AA^* b n 2 A A ∗ . Thus
ρ n ( A G B ) 2 = η n ( A G B ) 2 = ϑ n ( A G B ) 2 = M n . \rho_n(AGB)^2=\eta_n(AGB)^2=\vartheta_n(AGB)^2=M_n. ρ n ( A GB ) 2 = η n ( A GB ) 2 = ϑ n ( A GB ) 2 = M n .
Zero values are covered by zero witnesses and deletion of the finite-rank factor. Since M n ≤ h n 2 ≤ 2 M n M_n\le h_n^2\le2M_n M n ≤ h n 2 ≤ 2 M n , Theorem A gives (45) . The expansion h n 2 = a n 2 b n 2 + a n 2 F b ( n + 1 ) + b n 2 F a ( n + 1 ) h_n^2=a_n^2b_n^2+a_n^2F_b(n+1)+b_n^2F_a(n+1) h n 2 = a n 2 b n 2 + a n 2 F b ( n + 1 ) + b n 2 F a ( n + 1 ) shows monotonicity.
The isometry applied to ( I − P ) A (I-P)A ( I − P ) A and B ( I − Q ) B(I-Q) B ( I − Q ) gives
E ∥ ( I − P ) A G B ( I − Q ) ∥ 2 2 = ∥ ( I − P ) A ∥ 2 2 ∥ B ( I − Q ) ∥ 2 2 . \mathbb{E}\lVert(I-P)AGB(I-Q)\rVert_2^2=\lVert(I-P)A\rVert_2^2\lVert B(I-Q)\rVert_2^2. E ∥( I − P ) A GB ( I − Q ) ∥ 2 2 = ∥( I − P ) A ∥ 2 2 ∥ B ( I − Q ) ∥ 2 2 .
Leading singular projections minimize the two factors independently, proving (44) . The profile and ideal events are Borel; Fubini gives the conditional claims. Unconditional L 2 L^2 L 2 convergence holds when E ( ∥ A ∥ 2 2 ∥ B ∥ 2 2 ) < ∞ \mathbb{E}(\lVert A\rVert_2^2\lVert B\rVert_2^2)<\infty E (∥ A ∥ 2 2 ∥ B ∥ 2 2 ) < ∞ . □
For nonnegative nonincreasing u , v ∈ ℓ 2 u,v\in\ell^2 u , v ∈ ℓ 2 , write u ⊙ v u\mathbin{\odot}v u ⊙ v for the energy-decrement profile just obtained:
( u ⊙ v ) n = F u ( n ) F v ( n ) − F u ( n + 1 ) F v ( n + 1 ) = u n 2 F v ( n ) + v n 2 F u ( n + 1 ) . (47) (u\mathbin{\odot}v)_n=F_u(n)F_v(n)-F_u(n+1)F_v(n+1)=u_n^2F_v(n)+v_n^2F_u(n+1). \tag*{(47)} ( u ⊙ v ) n = F u ( n ) F v ( n ) − F u ( n + 1 ) F v ( n + 1 ) = u n 2 F v ( n ) + v n 2 F u ( n + 1 ) . (47)
It is nonnegative and nonincreasing, and summation gives
F u ⊙ v ( n ) = F u ( n ) F v ( n ) , ∥ u ⊙ v ∥ 2 = ∥ u ∥ 2 ∥ v ∥ 2 . (48) F_{u\mathbin{\odot}v}(n)=F_u(n)F_v(n),\qquad\lVert u\mathbin{\odot}v\rVert_2=\lVert u\rVert_2\lVert v\rVert_2. \tag*{(48)} F u ⊙ v ( n ) = F u ( n ) F v ( n ) , ∥ u ⊙ v ∥ 2 = ∥ u ∥ 2 ∥ v ∥ 2 . (48)
Square tails determine their sequence, so ⊙ \odot ⊙ is commutative and associative. For ( Δ d u ) n = u ⌈ n / d ⌉ (\Delta_du)_n=u_{\lceil n/d\rceil} ( Δ d u ) n = u ⌈ n / d ⌉ and k = ⌈ n / d ⌉ k=\lceil n/d\rceil k = ⌈ n / d ⌉ , F Δ d u ( n ) ≤ d F u ( k ) F_{\Delta_du}(n)\le dF_u(k) F Δ d u ( n ) ≤ d F u ( k ) ; hence
Δ d u ⊙ Δ d v ≤ 2 d Δ d ( u ⊙ v ) . (49) \Delta_du\mathbin{\odot}\Delta_dv\le\sqrt{2d}\,\Delta_d(u\mathbin{\odot}v). \tag*{(49)} Δ d u ⊙ Δ d v ≤ 2 d Δ d ( u ⊙ v ) . (49)
Indeed its squared left side is at most d ( u k 2 F v ( k ) + v k 2 F u ( k ) ) ≤ 2 d ( u ⊙ v ) k 2 d(u_k^2F_v(k)+v_k^2F_u(k))\le2d(u\mathbin{\odot}v)_k^2 d ( u k 2 F v ( k ) + v k 2 F u ( k )) ≤ 2 d ( u ⊙ v ) k 2 . If u ≤ C Δ d w u\le C\Delta_dw u ≤ C Δ d w , monotonicity and v ≤ Δ d v v\le\Delta_dv v ≤ Δ d v imply u ⊙ v ≤ C 2 d Δ d ( w ⊙ v ) u\mathbin{\odot}v\le C\sqrt{2d}\,\Delta_d(w\mathbin{\odot}v) u ⊙ v ≤ C 2 d Δ d ( w ⊙ v ) . Thus either input may be replaced by any generator of the same principal ideal. This compatibility is what permits conditional iteration.
Corollary 3.3 (Finite positive tensor covariances). Let Γ 0 = ∑ ℓ = 1 r C ℓ ⊗ D ℓ \Gamma_0=\sum_{\ell=1}^r C_\ell\otimes D_\ell Γ 0 = ∑ ℓ = 1 r C ℓ ⊗ D ℓ with positive trace-class C ℓ , D ℓ C_\ell,D_\ell C ℓ , D ℓ , and let a centered Gaussian form X X X satisfy a Γ 0 ≤ C X ≤ b Γ 0 a\Gamma_0\le C_X\le b\Gamma_0 a Γ 0 ≤ C X ≤ b Γ 0 , 0 < a ≤ b < ∞ 0<a\le b<\infty 0 < a ≤ b < ∞ . Set h n ( ℓ ) = λ ( C ℓ ) ⊙ λ ( D ℓ ) h_n^{(\ell)}=\sqrt{\lambda(C_\ell)}\mathbin{\odot}\sqrt{\lambda(D_\ell)} h n ( ℓ ) = λ ( C ℓ ) ⊙ λ ( D ℓ ) and H n = max ℓ h n ( ℓ ) H_n=\max_\ell h_n^{(\ell)} H n = max ℓ h n ( ℓ ) . Then X X X has a compact realization and
I ( X + D ) = I ( H ) + I ( D ) = I ( η ( X ) ) + I ( D ) almost surely \mathcal{I}(X+D)=\mathcal{I}(H)+\mathcal{I}(D)=\mathcal{I}(\eta(X))+\mathcal{I}(D)\quad\text{almost surely} I ( X + D ) = I ( H ) + I ( D ) = I ( η ( X )) + I ( D ) almost surely for each fixed compact D D D . The factors need not commute. Different finite positive decompositions give the same ideal, not necessarily the same H H H .
Proof. The tensor C ℓ ⊗ D ℓ C_\ell\otimes D_\ell C ℓ ⊗ D ℓ is the covariance of C ℓ 1 / 2 G D ℓ 1 / 2 C_\ell^{1/2}GD_\ell^{1/2} C ℓ 1/2 G D ℓ 1/2 , whose profiles close by the preceding proof. Apply Proposition 3.1 and (8) . □
Curvature arrays and conditional contraction
Call a centered countable array Z Z Z a curvature array if each finite coordinate marginal has positive density c e − V ce^{-V} c e − V and
α I ≤ ∇ 2 V ≤ β I , 0 < α ≤ β < ∞ , (50) \alpha I \le\nabla^2 V \le\beta I,\qquad0 < \alpha\le\beta< \infty, \tag*{(50)} α I ≤ ∇ 2 V ≤ β I , 0 < α ≤ β < ∞ , (50)
with constants independent of the finite set. Potentials may be be C l o c 1 , 1 C_{\mathrm{loc}}^{1,1} C loc 1 , 1 , with Hessian inequalities almost everywhere. This is a verifiable special case of (H). Linear images and limits are understood on their reference supports, as in Lemma 2.4 ; changing orthonormal coordinates preserves these bounds, not the law or independence of the coordinates.
Lemma 3.4 (Linear contraction and conditional innovation). For a curvature array Z Z Z , deterministic Hilbert–Schmidt L , R L,R L , R and fixed compact F F F , the linear series L Z R LZR L Z R belongs to L 2 ( Ω ; S 2 ) L^2(\Omega;S_2) L 2 ( Ω ; S 2 ) and
E ∥ L Z R ∥ 2 2 ≤ α − 1 ∥ L ∥ 2 2 ∥ R ∥ 2 2 , I ( L Z R + F ) = I ( s ( L ) ∘ s ( R ) ) + I ( F ) a.s. (51) \mathbb{E}\lVert LZR\rVert_2^2 \le\alpha^{-1}\lVert L\rVert_2^2\lVert R\rVert_2^2,\qquad\mathcal{I}(LZR+F)=\mathcal{I}(s(L)\mathbin{\circ}s(R))+\mathcal{I}(F)\quad\text{a.s.} \tag*{(51)} E ∥ L Z R ∥ 2 2 ≤ α − 1 ∥ L ∥ 2 2 ∥ R ∥ 2 2 , I ( L Z R + F ) = I ( s ( L ) ∘ s ( R )) + I ( F ) a.s. (51) Now let ( U , V ) (U,V) ( U , V ) be disjoint coordinate subarrays of one curvature array, u , v ∈ ℓ 2 u,v\in\ell^2 u , v ∈ ℓ 2 nonnegative and nonincreasing, and K ∈ S 2 K\in S_2 K ∈ S 2 . Suppose the coordinate truncations of D u U K V ∗ D v D_uUKV^*D_v D u U K V ∗ D v along a fixed exhaustion, possibly after subtraction of deterministic Hilbert–Schmidt operators C N C_N C N , converge to B B B in L 2 ( Ω ; S 2 ) L^2(\Omega;S_2) L 2 ( Ω ; S 2 ) . The subtractions need not converge; C N = 0 C_N=0 C N = 0 is the raw case. With F = σ ( V ) F=\sigma(V) F = σ ( V ) and M = E [ B ∣ F ] M=\mathbb{E}[B\mid F] M = E [ B ∣ F ] , every fixed compact F F F satisfies
I ( B + F ) = I ( u ∘ s ( K ) ∘ v ) + I ( M + F ) a.s. (52) \mathcal{I}(B+F)=\mathcal{I}(u\mathbin{\circ}s(K)\mathbin{\circ}v)+\mathcal{I}(M+F)\quad\text{a.s.} \tag*{(52)} I ( B + F ) = I ( u ∘ s ( K ) ∘ v ) + I ( M + F ) a.s. (52) No conditional-centering assumption on U U U is made.
Proof. Linear contraction. For finite matrices, expanding the squared Hilbert–Schmidt norm and using Cov Z ≤ α − 1 I \operatorname{Cov} Z\le\alpha^{-1}I Cov Z ≤ α − 1 I gives
E ∥ L Z R ∥ 2 2 ≤ α − 1 ∥ L ∥ 2 2 ∥ R ∥ 2 2 . \mathbb{E}\lVert LZR\rVert_2^2\le\alpha^{-1}\lVert L\rVert_2^2\lVert R\rVert_2^2. E ∥ L Z R ∥ 2 2 ≤ α − 1 ∥ L ∥ 2 2 ∥ R ∥ 2 2 .
Applying this estimate to coefficient differences proves convergence of finite truncations and the stated L 2 L^2 L 2 estimate. Finite tests of L Z R LZR L Z R satisfy (H) relative to L G R LGR L GR with standard Gaussian G G G : apply the linear-image and weak-limit argument of Lemma 2.4 . The Gaussian profiles computed in the proof of Corollary 3.2 close at I ( s ( L ) ∘ s ( R ) ) \mathcal{I}(s(L)\mathbin{\circ}s(R)) I ( s ( L ) ∘ s ( R )) . Theorem A therefore proves (51) .
Conditional laws. The centered conditional field W = U − E [ U ∣ F ] W=U-\mathbb{E}[U\mid F] W = U − E [ U ∣ F ] is a curvature array for almost every environment. Fix a finite block U J U_J U J , exhaust V V V by V ( N ) V^{(N)} V ( N ) , and let μ N \mu_N μ N be the conditional law of U J U_J U J given V ( N ) V^{(N)} V ( N ) . Martingale convergence on a countable convergence-determining family of bounded Lipschitz functions, with tightness from the convergent conditional second moments, gives μ N ⇒ μ = L ( U J ∣ F ) \mu_N\Rightarrow\mu=\mathcal{L}(U_J\mid F) μ N ⇒ μ = L ( U J ∣ F ) almost surely. Also m N = E [ U J ∣ V ( N ) ] → m = E [ U J ∣ F ] m_N=\mathbb{E}[U_J\mid V^{(N)}]\to m=\mathbb{E}[U_J\mid F] m N = E [ U J ∣ V ( N ) ] → m = E [ U J ∣ F ] . Translating these measures by − m N → − m -m_N\to-m − m N → − m preserves weak convergence. Their centered potentials have the same Hessian bounds; the weak-limit part of Lemma 2.4 , with reference covariance I I I , retains them. Intersect over J J J . No infinite-dimensional density or interchange of a non-observation-measurable center with conditional expectation is used.
The innovation series. Set H = K V ∗ D v H=KV^*D_v H = K V ∗ D v . The first assertion applied to H ∗ H^* H ∗ gives H ∈ L 2 ( S 2 ) H\in L^2(S_2) H ∈ L 2 ( S 2 ) and I ( H ) = I ( s ( K ) ∘ v ) \mathcal{I}(H)=\mathcal{I}(s(K)\mathbin{\circ}v) I ( H ) = I ( s ( K ) ∘ v ) almost surely. Conditional covariance bounds imply, for every environment-measurable Hilbert–Schmidt H ′ H' H ′ ,
E [ ∥ D u W H ′ ∥ 2 2 ∣ F ] ≤ α − 1 ∥ u ∥ 2 2 ∥ H ′ ∥ 2 2 . \mathbb{E}\left[\lVert D_uWH'\rVert_2^2\mid F\right]\le\alpha^{-1}\lVert u\rVert_2^2\lVert H'\rVert_2^2. E [ ∥ D u W H ′ ∥ 2 2 ∣ F ] ≤ α − 1 ∥ u ∥ 2 2 ∥ H ′ ∥ 2 2 .
Thus the conditional series D u W H D_uWH D u W H is well defined. Let B N B_N B N be the finite evaluations with their specified deterministic subtractions. Those subtractions cancel exactly in B N − E [ B N ∣ F ] B_N-\mathbb{E}[B_N\mid F] B N − E [ B N ∣ F ] . The first limit is B − M B-M B − M because conditional expectation is an L 2 ( S 2 ) L^2(S_2) L 2 ( S 2 ) contraction. Let u ( N ) u^{(N)} u ( N ) be the finite row weights; the associated H N H_N H N converge to H H H in L 2 ( S 2 ) L^2(S_2) L 2 ( S 2 ) by the first linear estimate. Conditional covariance, first for simple environmental coefficients and then by L 2 L^2 L 2 approximation, gives
E ∥ D u ( N ) W H N − D u W H ∥ 2 2 ≤ 2 α − 1 ∥ u ∥ 2 2 E ∥ H N − H ∥ 2 2 + 2 α − 1 ∥ u ( N ) − u ∥ 2 2 E ∥ H ∥ 2 2 ⟶ 0. (53) \begin{aligned}
\mathbb{E}\lVert D_{u^{(N)}}WH_N-D_uWH\rVert_2^2
&\le2\alpha^{-1}\lVert u\rVert_2^2\mathbb{E}\lVert H_N-H\rVert_2^2\\
&\quad+2\alpha^{-1}\lVert u^{(N)}-u\rVert_2^2\mathbb{E}\lVert H\rVert_2^2\longrightarrow0. \tag*{(53)}
\end{aligned} E ∥ D u ( N ) W H N − D u W H ∥ 2 2 ≤ 2 α − 1 ∥ u ∥ 2 2 E ∥ H N − H ∥ 2 2 + 2 α − 1 ∥ u ( N ) − u ∥ 2 2 E ∥ H ∥ 2 2 ⟶ 0. (53)
Thus B = D u W H + M B=D_uWH+M B = D u W H + M .
Conditional ideal identity. Given F \mathcal{F} F , apply (51) with coefficients D u , H D_u,H D u , H and fixed shift M + F M+F M + F . Square-tail compatibility proves (52) . The events are countable singular-value comparisons, so integration against the regular conditional kernel is legitimate; no event uniform over all random coefficients is required. □ \square □
Corollary 3.5 (Independent Gaussian specialization). Let G i : E i → L 2 ( Ω ; H i ) G_i:E_i\to L^2(\Omega;H_i) G i : E i → L 2 ( Ω ; H i ) , i = 1 , 2 i=1,2 i = 1 , 2 , be independent centered Gaussian linear maps satisfying E ( G i f ⊗ G i g ) = ⟨ f , g ⟩ Q i \mathbb{E}(G_i f\otimes G_i g)=\langle f,g\rangle Q_i E ( G i f ⊗ G i g ) = ⟨ f , g ⟩ Q i for positive trace-class Q i Q_i Q i . Bilateral evaluation extends to T = G 1 K G 2 ∗ ∈ L 2 ( S 2 ) T=G_1KG_2^*\in L^2(S_2) T = G 1 K G 2 ∗ ∈ L 2 ( S 2 ) for every K ∈ S 2 ( E 2 , E 1 ) K\in S_2(E_2,E_1) K ∈ S 2 ( E 2 , E 1 ) . With u = s ( Q 1 1 / 2 ) u=s(Q_1^{1/2}) u = s ( Q 1 1/2 ) and v = s ( Q 2 1 / 2 ) v=s(Q_2^{1/2}) v = s ( Q 2 1/2 ) ,
E ∥ T ∥ 2 2 = tr Q 1 ∥ K ∥ 2 tr Q 2 , (54) \mathbb{E}\lVert T\rVert^2_2=\operatorname{tr}Q_1\lVert K\rVert^2\operatorname{tr}Q_2, \tag*{(54)} E ∥ T ∥ 2 2 = tr Q 1 ∥ K ∥ 2 tr Q 2 , (54) I ( T + F ) = I ( u ⊙ s ( K ) ⊙ v ) + I ( F ) a.s. (55) \mathcal{I}(T+F)=\mathcal{I}(u\odot s(K)\odot v)+\mathcal{I}(F)\quad\text{a.s.} \tag*{(55)} I ( T + F ) = I ( u ⊙ s ( K ) ⊙ v ) + I ( F ) a.s. (55) for each fixed compact F F F , including zero and finite-rank inputs.
Proof. Gaussian second moments give (54) for finite coefficients and their differences; hence finite-rank and covariance square-root truncations define the displayed L 2 L^2 L 2 limit. In covariance spectral bases, Lemma 3.4 applies with E [ T ∣ G 2 ] = 0 \mathbb{E}[T\mid\mathcal{G}_2]=0 E [ T ∣ G 2 ] = 0 by independence. Its prediction formula is exactly (55) . □ \square □
Exterior evaluation and covariance compensation
Spectral splitting, polarization and alternating evaluation
Lemma 4.1 (A reducing decomposition preserving the ideal). For a positive compact Q Q Q of rank at least two, there is an orthogonal projection P P P commuting with Q Q Q such that both restrictions Q + , Q − Q_{+},Q_{-} Q + , Q − to P H , ( I − P ) H PH,(I-P)H P H , ( I − P ) H are nonzero and
I ( Q + 1 / 2 ) = I ( Q 1 / 2 ) = I ( Q − 1 / 2 ) . \mathcal{I}(Q_{+}^{1/2})=\mathcal{I}(Q^{1/2})=\mathcal{I}(Q_{-}^{1/2}). I ( Q + 1/2 ) = I ( Q 1/2 ) = I ( Q − 1/2 ) .
Proof. Divide an ordered positive spectral basis alternately between the two subspaces, splitting eigenspaces when necessary. If a = s ( Q 1 / 2 ) a=s(Q^{1/2}) a = s ( Q 1/2 ) , their spectra are a n + = a 2 n − 1 a_{n}^{+}=a_{2n-1} a n + = a 2 n − 1 and a n − = a 2 n a_{n}^{-}=a_{2n} a n − = a 2 n , with zeros appended, and
a − ≤ a + ≤ a , a n ≤ max { 1 , a 1 / a 2 } a ⌈ n / 3 ⌉ . (56) a^{-}\leq a^{+}\leq a,\qquad a_{n}\leq\max\{1,a_{1}/a_{2}\}\,a_{\lceil n/3\rceil}. \tag*{(56)} a − ≤ a + ≤ a , a n ≤ max { 1 , a 1 / a 2 } a ⌈ n /3 ⌉ . (56)
For n ≥ 2 n\geq2 n ≥ 2 use 2 ⌈ n / 3 ⌉ ≤ n 2\lceil n/3\rceil\leq n 2 ⌈ n /3 ⌉ ≤ n ; the constant handles n = 1 n=1 n = 1 . The ideal criterion proves the assertion, including finite rank. □
A Hilbert two-form on a real separable E E E is a skew-adjoint Hilbert–Schmidt operator under f ∧ g ↦ f ⊗ g − g ⊗ f f\wedge g\mapsto f\otimes g-g\otimes f f ∧ g ↦ f ⊗ g − g ⊗ f ; for orthonormal increasing wedges, ∥ K ∥ 2 2 = 2 ∥ ω K ∥ 2 , E 2 \lVert K\rVert_{2}^{2}=2\lVert\omega_{K}\rVert_{2,E}^{2} ∥ K ∥ 2 2 = 2 ∥ ω K ∥ 2 , E 2 . Let Q ≥ 0 Q\geq0 Q ≥ 0 be trace class on H H H , and use a spectral basis of Q Q Q to write a n = q n a_{n}=\sqrt{q_{n}} a n = q n , q n = s n ( Q ) q_{n}=s_{n}(Q) q n = s n ( Q ) . For a curvature array Z Z Z , define Z f = D a Z f \mathcal{Z}f=D_{a}Zf Z f = D a Z f by L 4 ( H ) L^{4}(H) L 4 ( H ) limits. These exist by (24) . The weight Q Q Q need not be the actual output covariance, and Z Z Z is not asserted samplewise bounded. For finite-rank skew F F F , write
A a ( F ) = Z F Z ∗ , c a ( F ) = E A a ( F ) , W a ( F ) = A a ( F ) − c a ( F ) . A_{a}(F)=\mathcal{Z}F\mathcal{Z}^{*},\qquad c_{a}(F)=\mathbb{E}A_{a}(F),\qquad W_{a}(F)=A_{a}(F)-c_{a}(F). A a ( F ) = Z F Z ∗ , c a ( F ) = E A a ( F ) , W a ( F ) = A a ( F ) − c a ( F ) .
The raw evaluation is a finite sum of vector exterior products, so rank A a ( F ) ≤ rank F \operatorname{rank}A_{a}(F)\leq\operatorname{rank}F rank A a ( F ) ≤ rank F ; its mean may have infinite rank.
Fix a skew-adjoint K ∈ S 2 ( E ) K\in\mathcal{S}_{2}(E) K ∈ S 2 ( E ) and choose its ordered real spectral decomposition, completing it with zero blocks when needed:
K = ∑ ℓ ≥ 1 K ℓ , K ℓ = σ ℓ ( f ℓ ⊗ g ℓ − g ℓ ⊗ f ℓ ) , κ = s ( K ) = Δ 2 σ . (57) K=\sum_{\ell\geq1}K_{\ell},\qquad K_{\ell}=\sigma_{\ell}(f_{\ell}\otimes g_{\ell}-g_{\ell}\otimes f_{\ell}),\qquad\kappa=s(K)=\Delta_{2}\sigma. \tag*{(57)} K = ℓ ≥ 1 ∑ K ℓ , K ℓ = σ ℓ ( f ℓ ⊗ g ℓ − g ℓ ⊗ f ℓ ) , κ = s ( K ) = Δ 2 σ . (57)
Put F n = ∑ ℓ < n K ℓ F_{n}=\sum_{\ell<n}K_{\ell} F n = ∑ ℓ < n K ℓ , K ( n ) = K − F n K^{(n)}=K-F_{n} K ( n ) = K − F n , and let P n P_{n} P n project onto the spatial coordinates i ≥ 2 n − 1 i\geq2n-1 i ≥ 2 n − 1 . Thus F 1 = 0 F_{1}=0 F 1 = 0 , P 1 = I P_{1}=I P 1 = I and rank ( I − P n ) ≤ 2 n − 2 \operatorname{rank}(I-P_{n})\leq2n-2 rank ( I − P n ) ≤ 2 n − 2 . We use
R n = ∑ j ≥ n q j , T n = ∑ j ≥ n κ j 2 , c Q = ( tr Q ) 2 − tr ( Q 2 ) , (58) R_{n}=\sum_{j\geq n}q_{j},\qquad T_{n}=\sum_{j\geq n}\kappa_{j}^{2},\qquad c_{Q}=(\operatorname{tr}Q)^{2}-\operatorname{tr}(Q^{2}), \tag*{(58)} R n = j ≥ n ∑ q j , T n = j ≥ n ∑ κ j 2 , c Q = ( tr Q ) 2 − tr ( Q 2 ) , (58)
r = a ∘ κ ∘ a , r n 2 = R n 2 T n − R n + 1 2 T n + 1 . (59) r=a\mathbin{\circ}\kappa\mathbin{\circ}a,\qquad r_{n}^{2}=R_{n}^{2}T_{n}-R_{n+1}^{2}T_{n+1}. \tag*{(59)} r = a ∘ κ ∘ a , r n 2 = R n 2 T n − R n + 1 2 T n + 1 . (59)
Expanding the differences gives max { q n R n T n , κ n R n } ≤ r n ≤ 3 max { q n R n T n , κ n R n } \max\{\sqrt{q_{n}R_{n}T_{n}},\kappa_{n}R_{n}\}\leq r_{n}\leq\sqrt{3}\max\{\sqrt{q_{n}R_{n}T_{n}},\kappa_{n}R_{n}\} max { q n R n T n , κ n R n } ≤ r n ≤ 3 max { q n R n T n , κ n R n } .
For finite-rank skew kernels F F F , covariance domination on linear tests gives
c a ( F ) = D a B F D a , ∥ B F ∥ ≤ α − 1 ∥ F ∥ 1 , ∥ P n c a ( F ) P n ∥ ≤ α − 1 q 2 n − 1 ∥ F ∥ 1 . (60) c_{a}(F)=D_{a}B_{F}D_{a},\qquad\lVert B_{F}\rVert\leq\alpha^{-1}\lVert F\rVert_{1},\qquad\lVert P_{n}c_{a}(F)P_{n}\rVert\leq\alpha^{-1}q_{2n-1}\lVert F\rVert_{1}. \tag*{(60)} c a ( F ) = D a B F D a , ∥ B F ∥ ≤ α − 1 ∥ F ∥ 1 , ∥ P n c a ( F ) P n ∥ ≤ α − 1 q 2 n − 1 ∥ F ∥ 1 . (60)
Indeed a rank-one noise kernel pairs two spatial linear tests, whose covariance is bounded by α − 1 \alpha^{-1} α − 1 times their norms; sum a singular expansion of F F F . If F F F is skew, so is B F B_{F} B F .
Lemma 4.2 (Polarization comparison). Let X ∈ R d X\in\mathbb{R}^{d} X ∈ R d have a centered positive density c e − V ce^{-V} c e − V with ∇ 2 V ≥ α I \nabla^{2}V\geq\alpha I ∇ 2 V ≥ α I , and let B B B be a symmetric bilinear map into a finite-dimensional real normed space. For nonnegative convex Φ \Phi Φ ,
E Φ ( B ( X , X ) − E B ( X , X ) ) ≤ E Φ ( 2 α − 1 B ( g , h ) ) , (61) \mathbb{E}\Phi\bigl(B(X,X)-\mathbb{E}B(X,X)\bigr)\leq\mathbb{E}\Phi\bigl(2\alpha^{-1}B(g,h)\bigr), \tag*{(61)} E Φ ( B ( X , X ) − E B ( X , X ) ) ≤ E Φ ( 2 α − 1 B ( g , h ) ) , (61) where g , h g,h g , h are independent standard Gaussian vectors.
Proof. Take an independent copy X ′ X' X ′ and put U = ( X + X ′ ) / 2 U=(X+X')/\sqrt{2} U = ( X + X ′ ) / 2 , V 0 = ( X − X ′ ) / 2 V_{0}=(X-X')/\sqrt{2} V 0 = ( X − X ′ ) / 2 . Conditional Jensen and polarization bound the left side by E Φ ( 2 B ( U , V 0 ) ) \mathbb{E}\Phi(2B(U,V_{0})) E Φ ( 2 B ( U , V 0 )) . Given U = u U=u U = u , the potential of V 0 V_{0} V 0 is V ( ( u + v ) / 2 ) + V ( ( u − v ) / 2 ) V((u+v)/\sqrt{2})+V((u-v)/\sqrt{2}) V (( u + v ) / 2 ) + V (( u − v ) / 2 ) plus a constant. It is even in v v v and has Hessian at least α I \alpha I α I , so its mean is zero. For fixed u u u , the map v ↦ Φ ( 2 B ( u , v ) ) v\mapsto\Phi(2B(u,v)) v ↦ Φ ( 2 B ( u , v )) is convex. Hargé’s theorem therefore replaces the conditional variable V 0 V_0 V 0 by α − 1 / 2 h \alpha^{-1/2}h α − 1/2 h , with h h h independent of U U U . Next, independent convex-order comparison for X X X and X ′ X' X ′ gives
X + X ′ 2 ⪯ c x α − 1 / 2 g 1 + α − 1 / 2 g 2 2 = d α − 1 / 2 g . \frac{X+X'}{\sqrt{2}} \preceq_{\mathrm{cx}} \frac{\alpha^{-1/2}g_1+\alpha^{-1/2}g_2}{\sqrt{2}} \mathrel{\stackrel{d}{=}} \alpha^{-1/2}g. 2 X + X ′ ⪯ cx 2 α − 1/2 g 1 + α − 1/2 g 2 = d α − 1/2 g .
Indeed one replaces the first summand and then the second, keeping the other fixed in each convex test. Finally, u ↦ E h Φ ( 2 B ( u , α − 1 / 2 h ) ) u\mapsto\mathbb{E}_h\Phi(2B(u,\alpha^{-1/2}h)) u ↦ E h Φ ( 2 B ( u , α − 1/2 h )) is convex, so this second comparison gives (61) , with independent g , h g,h g , h . Smooth approximation gives the same statement for C l o c 1 , 1 C_{\mathrm{loc}}^{1,1} C loc 1 , 1 potentials; the norm-power tests used below have finite moments by Gaussian convex domination. □
Lemma 4.3 (Centered evaluation and residual moments). The map W a W_a W a extends uniquely and continuously to skew Hilbert–Schmidt kernels, with
E ∥ W a ( K ) ∥ 2 2 ≤ 2 α − 2 c Q ∥ K ∥ 2 2 . \mathbb{E}\lVert W_a(K)\rVert_2^2\le2\alpha^{-2}c_Q\lVert K\rVert_2^2. E ∥ W a ( K ) ∥ 2 2 ≤ 2 α − 2 c Q ∥ K ∥ 2 2 . For W n = P n W a ( K ( n ) ) P n W_n=P_nW_a(K^{(n)})P_n W n = P n W a ( K ( n ) ) P n and an absolute C C C ,
∥ W n ∥ L n ( o p ) ≤ C α − 1 r n ( n ≥ 1 ) . (62) \lVert W_n\rVert_{L^n(\mathrm{op})}\le C\alpha^{-1}r_n \qquad(n\ge1). \tag*{(62)} ∥ W n ∥ L n ( op ) ≤ C α − 1 r n ( n ≥ 1 ) . (62) Proof. Apply Lemma 4.2 to B ( U , V ) = 1 2 D d ( U K V ∗ + V K U ∗ ) D d B(U,V)=\frac{1}{2}D_d(UKV^*+VKU^*)D_d B ( U , V ) = 2 1 D d ( U K V ∗ + V K U ∗ ) D d . For C d = D d G 1 K G 2 ∗ D d C_d=D_dG_1KG_2^*D_d C d = D d G 1 K G 2 ∗ D d with independent standard Gaussian arrays,
∥ W d ( K ) ∥ L p ( o p ) ≤ 2 α − 1 ∥ C d ∥ L p ( o p ) ( p ≥ 1 ) . (63) \lVert W_d(K)\rVert_{L^p(\mathrm{op})}\le2\alpha^{-1}\lVert C_d\rVert_{L^p(\mathrm{op})}\qquad(p\ge1). \tag*{(63)} ∥ W d ( K ) ∥ L p ( op ) ≤ 2 α − 1 ∥ C d ∥ L p ( op ) ( p ≥ 1 ) . (63) The Hilbert–Schmidt comparison is more precisely with α − 1 ( C d − C d ∗ ) \alpha^{-1}(C_d-C_d^*) α − 1 ( C d − C d ∗ ) . For finite coefficients, E ∣ ( C d ) i j ∣ 2 = d i 2 d j 2 ∥ K ∥ 2 2 \mathbb{E}|(C_d)_{ij}|^2=d_i^2d_j^2\lVert K\rVert_2^2 E ∣ ( C d ) ij ∣ 2 = d i 2 d j 2 ∥ K ∥ 2 2 . If i ≠ j i\ne j i = j , the two entries ( C d ) i j (C_d)_{ij} ( C d ) ij and ( C d ) j i (C_d)_{ji} ( C d ) j i have zero covariance, while the diagonal of C d − C d ∗ C_d-C_d^* C d − C d ∗ vanishes. Consequently
E ∥ C d − C d ∗ ∥ 2 2 = 2 ∑ i ≠ j d i 2 d j 2 ∥ K ∥ 2 2 = 2 ( ( ∑ i d i 2 ) 2 − ∑ i d i 4 ) ∥ K ∥ 2 2 . \mathbb{E}\lVert C_d-C_d^*\rVert_2^2=2\sum_{i\ne j}d_i^2d_j^2\lVert K\rVert_2^2
=2\left(\left(\sum_i d_i^2\right)^2-\sum_i d_i^4\right)\lVert K\rVert_2^2. E ∥ C d − C d ∗ ∥ 2 2 = 2 i = j ∑ d i 2 d j 2 ∥ K ∥ 2 2 = 2 ( i ∑ d i 2 ) 2 − i ∑ d i 4 ∥ K ∥ 2 2 . Apply the same comparison to the difference of two skew kernels. It gives an L 2 ( S 2 ) L^2(\mathcal{S}_2) L 2 ( S 2 ) -Cauchy sequence whenever the kernels converge in S 2 \mathcal{S}_2 S 2 . For spatial truncations, apply the entrywise second-moment bound to the deleted entries and sum their squared weights; that sum tends to zero because d ∈ ℓ 2 d\in\ell^2 d ∈ ℓ 2 . Thus coordinate and spectral approximations give the same centered extension. For a finite-rank kernel, L 4 ( H ) L^4(H) L 4 ( H ) continuity of f ↦ Z f f\mapsto\mathcal{Z}f f ↦ Z f identifies this extension with the finite sum of vector exterior products minus its expectation.
For b = ∥ d ∥ ∞ b=\lVert d\rVert_\infty b = ∥ d ∥ ∞ , s = ∥ d ∥ 2 s=\lVert d\rVert_2 s = ∥ d ∥ 2 , h = ∥ L ∥ 2 h=\lVert L\rVert_2 h = ∥ L ∥ 2 , ℓ = ∥ L ∥ \ell=\lVert L\rVert ℓ = ∥ L ∥ , Gaussian comparison and concentration give
∥ D d G 1 L G 2 ∗ D d ∥ L p ( o p ) ≤ C ( b s h + s 2 ℓ + p b 2 h + p b s ℓ + p b 2 ℓ ) . (64) \lVert D_dG_1LG_2^*D_d\rVert_{L^p(\mathrm{op})}
\le C\left(bsh+s^2\ell+\sqrt{p}b^2h+\sqrt{p}b s\ell+pb^2\ell\right). \tag*{(64)} ∥ D d G 1 L G 2 ∗ D d ∥ L p ( op ) ≤ C ( b s h + s 2 ℓ + p b 2 h + p b s ℓ + p b 2 ℓ ) . (64) Indeed for H 0 = L G 2 ∗ D d H_0=LG_2^*D_d H 0 = L G 2 ∗ D d , [21 Theorems 6.8 and 3.25] give
∥ H 0 ∥ L p ( o p ) ≤ C ( h b + ℓ s + p ℓ b ) , ∥ H 0 ∥ L p ( S 2 ) ≤ h s + C p ℓ b . \lVert H_0\rVert_{L^p(\mathrm{op})}\le C(hb+\ell s+\sqrt{p}\,\ell b),\qquad
\lVert H_0\rVert_{L^p(\mathcal{S}_2)}\le hs+C\sqrt{p}\,\ell b. ∥ H 0 ∥ L p ( op ) ≤ C ( hb + ℓ s + p ℓ b ) , ∥ H 0 ∥ L p ( S 2 ) ≤ h s + C p ℓ b . Condition on G 2 G_2 G 2 and bound the first contraction by C { b ∥ H 0 ∥ L p ( S 2 ) + ( s + p b ) ∥ H 0 ∥ L p ( o p ) } C\{b\lVert H_0\rVert_{L^p(\mathcal{S}_2)}+(s+\sqrt{p}\,b)\lVert H_0\rVert_{L^p(\mathrm{op})}\} C { b ∥ H 0 ∥ L p ( S 2 ) + ( s + p b ) ∥ H 0 ∥ L p ( op ) } . Expanding the product gives the five displayed terms. Use p = n p=n p = n , d = P n a d=P_na d = P n a , L = K ( n ) L=K^{(n)} L = K ( n ) and n q 2 n − 1 ≤ R n nq_{2n-1}\le R_n n q 2 n − 1 ≤ R n . Its five terms are at most constant multiples of q n R n T n \sqrt{q_nR_nT_n} q n R n T n or κ n R n \kappa_nR_n κ n R n . Finite compressions, L 2 L^2 L 2 -convergent subsequences and Fatou justify (62) . □
Lemma 4.4 (Covariance compensation and finite-dimensional approximation). Set γ = α − 1 \gamma=\alpha^{-1} γ = α − 1 and M ℓ = c a ( K ℓ ) M_\ell=c_a(K_\ell) M ℓ = c a ( K ℓ ) . The deterministic skew-adjoint series
C = ∑ ℓ ≥ 1 P ℓ + 1 M ℓ P ℓ + 1 (65) \mathcal{C}=\sum_{\ell\ge1}P_{\ell+1}M_\ell P_{\ell+1} \tag*{(65)} C = ℓ ≥ 1 ∑ P ℓ + 1 M ℓ P ℓ + 1 (65) converges absolutely in S 2 \mathcal{S}_2 S 2 . More precisely,
∑ ℓ ≥ 1 ∥ P ℓ + 1 M ℓ P ℓ + 1 ∥ 2 ≤ γ ∥ K ∥ 2 ( ∑ i ≥ 1 ( i − 1 ) d i 2 ) 1 / 2 ≤ γ R 1 ∥ K ∥ 2 . (66) \sum_{\ell\ge1}\lVert P_{\ell+1}M_\ell P_{\ell+1}\rVert_2
\le\gamma\lVert K\rVert_2\left(\sum_{i\ge1}(i-1)d_i^2\right)^{1/2}
\le\gamma R_1\lVert K\rVert_2. \tag*{(66)} ℓ ≥ 1 ∑ ∥ P ℓ + 1 M ℓ P ℓ + 1 ∥ 2 ≤ γ ∥ K ∥ 2 ( i ≥ 1 ∑ ( i − 1 ) d i 2 ) 1/2 ≤ γ R 1 ∥ K ∥ 2 . (66)
The finite sums have the explicit Hilbert–Schmidt error bound
∥ C − ∑ ℓ < N P ℓ + 1 M ℓ P ℓ + 1 ∥ 2 ≤ 2 γ R 2 N + 1 T 2 N − 1 . (67) \left\|\mathcal{C}-\sum_{\ell<N}P_{\ell+1}M_{\ell}P_{\ell+1}\right\|_{2}\le\sqrt{2}\gamma R_{2N+1}\sqrt{T_{2N-1}}. \tag*{(67)} C − ℓ < N ∑ P ℓ + 1 M ℓ P ℓ + 1 2 ≤ 2 γ R 2 N + 1 T 2 N − 1 . (67)
The compressed remainder satisfies
∥ P n ( C − c a ( F n ) ) P n ∥ ≤ 2 γ r n . (68) \left\|P_n(\mathcal{C}-c_a(F_n))P_n\right\|\le\sqrt{2}\gamma r_n. \tag*{(68)} ∥ P n ( C − c a ( F n )) P n ∥ ≤ 2 γ r n . (68)
For fixed a a a , K K K and spectral choices, the same series defines a bounded real-linear map Ξ ↦ C Ξ \Xi\mapsto\mathcal{C}_{\Xi} Ξ ↦ C Ξ on bounded self-adjoint forms on ℓ 2 ⊗ E \ell^2\otimes E ℓ 2 ⊗ E . Define its finite-rank contractions by replacing E [ Z i ( f ) Z j ( g ) ] \mathbb{E}[Z_i(f)Z_j(g)] E [ Z i ( f ) Z j ( g )] with ⟨ Ξ ( e i ⊗ f ) , e j ⊗ g ⟩ \langle\Xi(e_i\otimes f),e_j\otimes g\rangle ⟨ Ξ ( e i ⊗ f ) , e j ⊗ g ⟩ . For any two such forms,
∥ C Ξ − C Ξ ~ ∥ 2 ≤ L a , K ∥ Ξ − Ξ ~ ∥ , L a , K = ∥ K ∥ 2 ( ∑ i ( i − 1 ) q i 2 ) 1 / 2 . (69) \|\mathcal{C}_{\Xi}-\mathcal{C}_{\widetilde{\Xi}}\|_2\le L_{a,K}\|\Xi-\widetilde{\Xi}\|,\qquad L_{a,K}=\|K\|^2\left(\sum_i(i-1)q_i^2\right)^{1/2}. \tag*{(69)} ∥ C Ξ − C Ξ ∥ 2 ≤ L a , K ∥Ξ − Ξ ∥ , L a , K = ∥ K ∥ 2 ( i ∑ ( i − 1 ) q i 2 ) 1/2 . (69)
For the actual covariance Γ \Gamma Γ , C = C Γ \mathcal{C}=\mathcal{C}_{\Gamma} C = C Γ . Let E L E_L E L project onto the first L L L spatial coordinates and put
C Ξ N , L = E L ∑ ℓ < N P ℓ + 1 c a , Ξ ( K ℓ ) P ℓ + 1 E L , ∥ C Ξ − C Ξ N , L ∥ 2 ≤ ∥ Ξ ∥ ϵ N , L , (70) \mathcal{C}_{\Xi}^{N,L}=E_L\sum_{\ell<N}P_{\ell+1}c_{a,\Xi}(K_\ell)P_{\ell+1}E_L,\qquad\|\mathcal{C}_{\Xi}-\mathcal{C}_{\Xi}^{N,L}\|_2\le\|\Xi\|\epsilon_{N,L}, \tag*{(70)} C Ξ N , L = E L ℓ < N ∑ P ℓ + 1 c a , Ξ ( K ℓ ) P ℓ + 1 E L , ∥ C Ξ − C Ξ N , L ∥ 2 ≤ ∥Ξ∥ ϵ N , L , (70)
where
ϵ N , L = 2 R 2 N + 1 T 2 N − 1 + ∥ F N ∥ 1 2 R 1 R L + 1 . (71) \epsilon_{N,L}=\sqrt{2}R_{2N+1}\sqrt{T_{2N-1}}+\|F_N\|_1\sqrt{2R_1R_{L+1}}. \tag*{(71)} ϵ N , L = 2 R 2 N + 1 T 2 N − 1 + ∥ F N ∥ 1 2 R 1 R L + 1 . (71)
For the fixed kernel and spectral choices, this covariance map is compact and, on norm-bounded sets of forms, continuous from the weak operator topology to the S 2 S_2 S 2 norm.
Proof. By (60) , M ℓ = D a B ℓ D a M_\ell=D_aB_\ell D_a M ℓ = D a B ℓ D a with B ℓ ∗ = − B ℓ B_\ell^*=-B_\ell B ℓ ∗ = − B ℓ and ∥ B ℓ ∥ ≤ 2 γ σ ℓ \|B_\ell\|\le2\gamma\sigma_\ell ∥ B ℓ ∥ ≤ 2 γ σ ℓ . Schatten Hölder gives ∥ D d B D d ∥ 2 ≤ ∥ D d ∥ 4 2 ∥ B ∥ = ( ∑ i d i 4 ) 1 / 2 ∥ B ∥ \|D_dBD_d\|_2\le\|D_d\|_4^2\|B\|=(\sum_i d_i^4)^{1/2}\|B\| ∥ D d B D d ∥ 2 ≤ ∥ D d ∥ 4 2 ∥ B ∥ = ( ∑ i d i 4 ) 1/2 ∥ B ∥ . Taking d i = a i 1 { i ≥ 2 ℓ + 1 } d_i=a_i\mathbf{1}_{\{i\ge2\ell+1\}} d i = a i 1 { i ≥ 2 ℓ + 1 } therefore gives
∥ P ℓ + 1 M ℓ P ℓ + 1 ∥ 2 ≤ 2 γ σ ℓ ( ∑ i ≥ 2 ℓ + 1 q i 2 ) 1 / 2 . \|P_{\ell+1}M_\ell P_{\ell+1}\|_2\le2\gamma\sigma_\ell\left(\sum_{i\ge2\ell+1}q_i^2\right)^{1/2}. ∥ P ℓ + 1 M ℓ P ℓ + 1 ∥ 2 ≤ 2 γ σ ℓ ( i ≥ 2 ℓ + 1 ∑ q i 2 ) 1/2 .
Cauchy–Schwarz, 2 ∑ ℓ σ ℓ 2 = ∥ K ∥ 1 2 2\sum_\ell\sigma_\ell^2=\|K\|_1^2 2 ∑ ℓ σ ℓ 2 = ∥ K ∥ 1 2 and ∑ ℓ ∑ i ≥ 2 ℓ + 1 q i 2 ≤ 1 2 ∑ i ( i − 1 ) q i 2 \sum_\ell\sum_{i\ge2\ell+1}q_i^2\le\frac12\sum_i(i-1)q_i^2 ∑ ℓ ∑ i ≥ 2 ℓ + 1 q i 2 ≤ 2 1 ∑ i ( i − 1 ) q i 2 prove the first bound. Monotonicity gives ∑ i ( i − 1 ) q i 2 ≤ ∑ j < i q i q j ≤ R 1 2 \sum_i(i-1)q_i^2\le\sum_{j<i}q_iq_j\le R_1^2 ∑ i ( i − 1 ) q i 2 ≤ ∑ j < i q i q j ≤ R 1 2 . Starting the same estimate at ℓ = N \ell=N ℓ = N , the second squared factor is at most ∑ j ≥ 1 j q 2 N + j 2 ≤ R 2 N + 1 2 \sum_{j\ge1}j q_{2N+j}^2\le R_{2N+1}^2 ∑ j ≥ 1 j q 2 N + j 2 ≤ R 2 N + 1 2 ; this proves (67) .
For ℓ < n \ell<n ℓ < n , nested projections give P n P ℓ + 1 = P n P_nP_{\ell+1}=P_n P n P ℓ + 1 = P n ; for ℓ ≥ n \ell\ge n ℓ ≥ n , they give P n P ℓ + 1 = P ℓ + 1 P_nP_{\ell+1}=P_{\ell+1} P n P ℓ + 1 = P ℓ + 1 . Therefore
P n ( C − c a ( F n ) ) P n = ∑ ℓ ≥ n P ℓ + 1 M ℓ P ℓ + 1 . P_n(\mathcal{C}-c_a(F_n))P_n=\sum_{\ell\ge n}P_{\ell+1}M_\ell P_{\ell+1}. P n ( C − c a ( F n )) P n = ℓ ≥ n ∑ P ℓ + 1 M ℓ P ℓ + 1 .
Its norm is at most
2 γ ∑ ℓ ≥ n σ ℓ q 2 ℓ + 1 ≤ 2 γ T 2 n − 1 q 2 n + 1 R 2 n + 1 ≤ 2 γ T n q n R n ≤ 2 γ r n . 2\gamma\sum_{\ell\ge n}\sigma_\ell q_{2\ell+1}\le\sqrt{2}\gamma\sqrt{T_{2n-1}q_{2n+1}R_{2n+1}}\le\sqrt{2}\gamma\sqrt{T_nq_nR_n}\le\sqrt{2}\gamma r_n. 2 γ ℓ ≥ n ∑ σ ℓ q 2 ℓ + 1 ≤ 2 γ T 2 n − 1 q 2 n + 1 R 2 n + 1 ≤ 2 γ T n q n R n ≤ 2 γ r n .
All sums are justified by absolute convergence. The contraction estimate uses only ∥ Ξ ∥ \|\Xi\| ∥Ξ∥ , not positivity, so all these bounds hold with γ = ∥ Ξ ∥ \gamma=\|\Xi\| γ = ∥Ξ∥ . Applying the absolute-sum bound to Ξ − Ξ ~ \Xi-\widetilde{\Xi} Ξ − Ξ proves (69) and linearity.
Write D N = ∑ ℓ < N P ℓ + 1 c a , Ξ ( K ℓ ) P ℓ + 1 D_N=\sum_{\ell<N}P_{\ell+1}c_{a,\Xi}(K_\ell)P_{\ell+1} D N = ∑ ℓ < N P ℓ + 1 c a , Ξ ( K ℓ ) P ℓ + 1 . The entry bound ∣ ( D N ) i j ∣ ≤ ∥ Ξ ∥ ∥ F N ∥ 1 a i a j |(D_N)_{ij}|\le\|\Xi\|\|F_N\|_1a_ia_j ∣ ( D N ) ij ∣ ≤ ∥Ξ∥∥ F N ∥ 1 a i a j gives
∥ D N − E L D N E L ∥ 2 2 ≤ ∥ Ξ ∥ 2 ∥ F N ∥ 1 2 ( 2 R 1 R L + 1 − R L + 1 2 ) . \|D_N-E_LD_NE_L\|_2^2\le\|\Xi\|^2\|F_N\|_1^2(2R_1R_{L+1}-R_{L+1}^2). ∥ D N − E L D N E L ∥ 2 2 ≤ ∥Ξ ∥ 2 ∥ F N ∥ 1 2 ( 2 R 1 R L + 1 − R L + 1 2 ) .
Combine this with (67) . For fixed N N N , L L L , the finite-dimensional map depends only on the finitely many covariance pairings on { e i ⊗ f ℓ , e i ⊗ g ℓ : i ≤ L , ℓ < N } \{e_i\otimes f_\ell,e_i\otimes g_\ell:i\le L,\ell<N\} { e i ⊗ f ℓ , e i ⊗ g ℓ : i ≤ L , ℓ < N } and has finite-dimensional range. First choose N N N and then L L L to make (71) small. Thus finite-rank maps approximate the covariance map in operator norm. They are weak-operator continuous; uniform approximation on each bounded set proves the last assertion as well as compactness. □
The same construction uses finite tables in the original noise basis. If Θ m \Theta_m Θ m is its first-m m m projection, replace each K ℓ K_\ell K ℓ in (70) by Θ m K ℓ Θ m \Theta_mK_\ell\Theta_m Θ m K ℓ Θ m . The additional S 2 S_2 S 2 error is at most ∥ Ξ ∥ R 1 ∑ ℓ < N ∥ K ℓ − Θ m K ℓ Θ m ∥ 1 \|\Xi\|R_1\sum_{\ell<N}\|K_\ell-\Theta_mK_\ell\Theta_m\|_1 ∥Ξ∥ R 1 ∑ ℓ < N ∥ K ℓ − Θ m K ℓ Θ m ∥ 1 , which tends to zero for fixed N N N . Together with (71) , this gives finite-dimensional approximations in the prescribed basis with explicit error bounds whenever the coefficient tails are available.
Theorem B (Exterior classification by one deterministic compensator). Assume ( H 0 ) (\mathrm{H}_0) ( H 0 ) , rank Q ≥ 2 Q \ge2 Q ≥ 2 and K ≠ 0 K \ne0 K = 0 . For every fixed compact operator B B B ,
I ( W a ( K ) + B ) = I ( r ) + I ( B − C ) almost surely . (72) \mathcal{I}(W_a(K)+B)=\mathcal{I}(r)+\mathcal{I}(B-\mathcal{C})\qquad\text{almost surely}. \tag*{(72)} I ( W a ( K ) + B ) = I ( r ) + I ( B − C ) almost surely . (72) In particular Y 0 = W a ( K ) + C Y_0=W_a(K)+\mathcal{C} Y 0 = W a ( K ) + C has
∥ s 6 n − 5 ( Y 0 ) ∥ L n ≤ C α − 1 r n , I ( Y 0 + D ) = I ( r ) + I ( D ) a . s . (73) \|s_{6n-5}(Y_0)\|_{L^n}\le C\alpha^{-1}r_n,\qquad\mathcal{I}(Y_0+D)=\mathcal{I}(r)+\mathcal{I}(D)\quad a.s. \tag*{(73)} ∥ s 6 n − 5 ( Y 0 ) ∥ L n ≤ C α − 1 r n , I ( Y 0 + D ) = I ( r ) + I ( D ) a . s . (73) for every fixed compact D D D . Each prescribed countable family of shifts shares a probability-one event.
Proof. First we prove, for every fixed compact B B B , the universal lower inclusion
I ( r ) ⊆ I ( W a ( K ) + B ) almost surely . (74) \mathcal{I}(r)\subseteq\mathcal{I}(W_a(K)+B)\qquad\text{almost surely}. \tag*{(74)} I ( r ) ⊆ I ( W a ( K ) + B ) almost surely . (74) Use the odd/even split of Lemma 4.1, with u n = a 2 n − 1 u_n=a_{2n-1} u n = a 2 n − 1 , v n = a 2 n v_n=a_{2n} v n = a 2 n , and let V V V be that block of W a ( K ) W_a(K) W a ( K ) . Its finite evaluations have only deterministic subtractions, so Lemma 3.4 applies even when the raw means diverge. For F = σ ( Z − ) \mathcal{F}=\sigma(Z_-) F = σ ( Z − ) , U = Z + − E [ Z + ∣ F ] U=Z_+-\mathbb{E}[Z_+\mid\mathcal{F}] U = Z + − E [ Z + ∣ F ] and H 0 = K Z − ∗ − D v H_0=KZ_-^*-D_v H 0 = K Z − ∗ − D v , its conditional innovation identity is
V = D u U H 0 + E [ V ∣ F ] . (75) V=D_uUH_0+\mathbb{E}[V\mid\mathcal{F}]. \tag*{(75)} V = D u U H 0 + E [ V ∣ F ] . (75) The same lemma gives I ( u κ v ) ⊆ I ( V + B + − ) \mathcal{I}(u\mathbin{\kappa}v)\subseteq\mathcal{I}(V+B_{+-}) I ( u κ v ) ⊆ I ( V + B +− ) . Square-tail compatibility and (4.1) identify the left side as I ( r ) \mathcal{I}(r) I ( r ) ; compression proves (4.19). On a common probability-one event, for every n ≥ 1 n\ge1 n ≥ 1 , write
Y 0 = R n f i n + W n + P n ( C − c a ( F n ) ) P n , rank R n f i n ≤ 6 n − 6. (76) Y_0=R_n^{\mathrm{fin}}+W_n+P_n(\mathcal{C}-c_a(F_n))P_n,\qquad\operatorname{rank}R_n^{\mathrm{fin}}\le6n-6. \tag*{(76)} Y 0 = R n fin + W n + P n ( C − c a ( F n )) P n , rank R n fin ≤ 6 n − 6. (76) Indeed R n f i n = Y 0 − P n Y 0 P n + P n A a ( F n ) P n R_n^{\mathrm{fin}}=Y_0-P_nY_0P_n+P_nA_a(F_n)P_n R n fin = Y 0 − P n Y 0 P n + P n A a ( F n ) P n . The identity Y 0 − P n Y 0 P n = ( I − P n ) Y 0 + P n Y 0 ( I − P n ) Y_0-P_nY_0P_n=(I-P_n)Y_0+P_nY_0(I-P_n) Y 0 − P n Y 0 P n = ( I − P n ) Y 0 + P n Y 0 ( I − P n ) bounds the rank of this spatial difference by 4 n − 4 4n-4 4 n − 4 . The remaining term P n A a ( F n ) P n P_nA_a(F_n)P_n P n A a ( F n ) P n has rank at most 2 n − 2 2n-2 2 n − 2 , because it is the raw evaluation of a kernel of that rank. Its centered version need not have this property: the deterministic mean can have infinite rank. Lemmas 4.3 and 4.4 prove the moment bound in (4.18). Markov at e C α − 1 r n eC\alpha^{-1}r_n e C α − 1 r n and Borel–Cantelli give Y 0 ∈ I ( r ) Y_0\in\mathcal{I}(r) Y 0 ∈ I ( r ) ; zero tails give the corresponding rank bounds directly. Together with (4.19), Lemma 2.2 applied to X = W a ( K ) X=W_a(K) X = W a ( K ) , C = C C=\mathcal{C} C = C and J = I ( r ) J=\mathcal{I}(r) J = I ( r ) proves both fixed-perturbation identities. □
Raw coordinate truncations. Let E m E_m E m retain the first m m m spatial coordinates and Π n \Pi_n Π n the first n n n coordinates in the prescribed noise basis. Define the finite polynomial
A m , n = E m D a Z Π n K Π n Z ∗ D a E m , A N = A m N , n N , m N , n N ↑ ∞ . (77) A_{m,n}=E_mD_aZ\Pi_nK\Pi_nZ^*D_aE_m,\qquad A_N=A_{m_N,n_N},\qquad m_N,n_N\uparrow\infty. \tag*{(77)} A m , n = E m D a Z Π n K Π n Z ∗ D a E m , A N = A m N , n N , m N , n N ↑ ∞. (77) Only the m × n m\times n m × n coordinate block of Z Z Z occurs in this formula. Throughout, an exhaustion is a fixed pair of nondecreasing sequences m N , n N → ∞ m_N,n_N\to\infty m N , n N → ∞ , with possibly different growth rates. The centered limit is independent of these rates; raw convergence below refers to the chosen exhaustion.
Corollary 4.5 (Raw existence, covariance and the obstruction class). Under the hypotheses of Theorem B, fix an exhaustion in (4.22). The operators A N A_N A N converge in L 2 ( S 2 ) L^2(\mathcal{S}_2) L 2 ( S 2 ) if and only if their deterministic means converge in S 2 \mathcal{S}_2 S 2 . If E A N → M \mathbb{E}A_N\to M E A N → M , set A = W a ( K ) + M A=W_a(K)+M A = W a ( K ) + M and Δ = M − C \Delta=M-\mathcal{C} Δ = M − C . Then
I ( A + D ) = I ( r ) + I ( Δ + D ) almost surely (78) \mathcal{I}(A+D)=\mathcal{I}(r)+\mathcal{I}(\Delta+D)\qquad\text{almost surely} \tag*{(78)} I ( A + D ) = I ( r ) + I ( Δ + D ) almost surely (78) for each fixed compact D D D . Moreover
I ( A ) = I ( r ) a . s . ⟺ Δ ∈ I ( r ) ⟺ I ( A + D ) = I ( A ) + I ( D ) a . s . for every fixed D . (79) \mathcal{I}(A)=\mathcal{I}(r)\ a.s.\quad\Longleftrightarrow\quad\Delta\in\mathcal{I}(r)\quad\Longleftrightarrow\quad\mathcal{I}(A+D)=\mathcal{I}(A)+\mathcal{I}(D)\ a.s.\ \text{for every fixed }D. \tag*{(79)} I ( A ) = I ( r ) a . s . ⟺ Δ ∈ I ( r ) ⟺ I ( A + D ) = I ( A ) + I ( D ) a . s . for every fixed D . (79) The centered classification, and raw existence and classification along the same exhaustion, depend only on covariance within the curvature class. In the algebraic quotient,
[ W a ( K ) ] = − [ C ] in K ( H ) / I ( r ) almost surely . (80) [W_a(K)]=-[\mathcal{C}]\quad\text{in }\mathcal{K}(H)/\mathcal{I}(r)\quad\text{almost surely}. \tag*{(80)} [ W a ( K )] = − [ C ] in K ( H ) / I ( r ) almost surely . (80) Different allowed spectral choices give the same compensator class. More precisely, for two curvature arrays with the same a , K a,K a , K , put J = I ( r ) J=\mathcal{I}(r) J = I ( r ) , W j = W Z j ( K ) W_j=W_{Z_j}(K) W j = W Z j ( K ) and C j = C r j C_j=\mathcal{C}_{r_j} C j = C r j . Then
[ I ( W 1 + B ) = I ( W 2 + B ) a . s . for each fixed B ] ⟺ C 1 − C 2 ∈ J . (81) [\mathcal{I}(W_1+B)=\mathcal{I}(W_2+B)\ a.s.\ \text{for each fixed }B]\quad\Longleftrightarrow\quad C_1-C_2\in J. \tag*{(81)} [ I ( W 1 + B ) = I ( W 2 + B ) a . s . for each fixed B ] ⟺ C 1 − C 2 ∈ J . (81)
The left side compares deterministic almost-sure ideals. On any prescribed coupling, the same condition is equivalent to
W 1 − W 2 ∈ J almost surely . (82) W_1-W_2\in J\quad\text{almost surely}. \tag*{(82)} W 1 − W 2 ∈ J almost surely . (82)
Equal finite-rank skew covariance contractions are sufficient, not necessary. An independent copy W ′ W' W ′ satisfies I ( W − W ′ ) = J \mathcal{I}(W-W')=J I ( W − W ′ ) = J almost surely.
Proof. For finite sections, A m , n − E A m , n = E m W a ( Π n K Π n ) E m A_{m,n}-\mathbb{E}A_{m,n}=E_mW_a(\Pi_nK\Pi_n)E_m A m , n − E A m , n = E m W a ( Π n K Π n ) E m . Lemma 4.3 gives
∥ A m , n − E A m , n − W a ( K ) ∥ L 2 ( S 2 ) ≤ 2 c Q α − 1 ∥ Π n K Π n − K ∥ 2 + ∥ E m W a ( K ) E m − W a ( K ) ∥ L 2 ( S 2 ) → 0 \lVert A_{m,n}-\mathbb{E}A_{m,n}-W_a(K)\rVert_{L^2(\mathcal{S}_2)}
\le\sqrt{2}c_Q\alpha^{-1}\lVert\Pi_nK\Pi_n-K\rVert_2
+\lVert E_mW_a(K)E_m-W_a(K)\rVert_{L^2(\mathcal{S}_2)}\to0 ∥ A m , n − E A m , n − W a ( K ) ∥ L 2 ( S 2 ) ≤ 2 c Q α − 1 ∥ Π n K Π n − K ∥ 2 + ∥ E m W a ( K ) E m − W a ( K ) ∥ L 2 ( S 2 ) → 0
as m , n → ∞ m,n\to\infty m , n → ∞ independently. The second term tends to zero by dominated convergence in L 2 ( S 2 ) L^2(\mathcal{S}_2) L 2 ( S 2 ) . Expectation is continuous there, so raw convergence along the chosen exhaustion is equivalent to convergence of its means. This proves the existence assertion and (78) . The first equivalence in (79) is immediate. If Δ ∈ I ( r ) \Delta\in\mathcal{I}(r) Δ ∈ I ( r ) , adding it inside any ideal sum with I ( r ) \mathcal{I}(r) I ( r ) has no effect. Conversely, apply the asserted sum law to the fixed D = − Δ D=-\Delta D = − Δ ; (78) then forces Δ ∈ I ( r ) \Delta\in\mathcal{I}(r) Δ ∈ I ( r ) . Each M ℓ M_\ell M ℓ is a covariance contraction. Equal covariances give equal means and equal compensators; the Gaussian field with this covariance also satisfies the curvature bounds, since β − 1 I ≤ Cov Z ≤ α − 1 I \beta^{-1}I\le\operatorname{Cov}Z\le\alpha^{-1}I β − 1 I ≤ Cov Z ≤ α − 1 I . The membership W a ( K ) + C ∈ I ( r ) W_a(K)+\mathcal{C}\in\mathcal{I}(r) W a ( K ) + C ∈ I ( r ) gives (80) . Lemma 2.2 identifies the compensator class, the independent-copy ideal and the response equivalence: necessity in (81) needs only the shift B = C 1 B=C_1 B = C 1 . Subtracting W j + C j ∈ J W_j+C_j\in J W j + C j ∈ J gives the equivalent coupling criterion (82) . If rank Q ≤ 1 \operatorname{rank}Q\le1 rank Q ≤ 1 or K = 0 K=0 K = 0 , alternation instead gives W = A = 0 W=A=0 W = A = 0 .
Remark (Finite-rank kernels and centering). If K K K has N N N nonzero skew blocks, then M = c a ( K ) M=c_a(K) M = c a ( K ) and P N + 1 ( M − C ) P N + 1 = 0 P_{N+1}(M-\mathcal{C})P_{N+1}=0 P N + 1 ( M − C ) P N + 1 = 0 , so M − C M-\mathcal{C} M − C has finite rank. Thus I ( W a ( K ) + B ) = F + I ( B − M ) \mathcal{I}(W_a(K)+B)=\mathcal{F}+\mathcal{I}(B-M) I ( W a ( K ) + B ) = F + I ( B − M ) , where F \mathcal{F} F is the finite-rank ideal. The raw evaluation has finite rank; centering can change that ideal. For example, use the paired construction underlying (97) below with K = J 0 K=J_0 K = J 0 on one noise block and c i 1 = c ∈ ( 0 , 1 ) c_{i1}=c\in(0,1) c i 1 = c ∈ ( 0 , 1 ) . Then rank A ≤ 2 \operatorname{rank}A\le2 rank A ≤ 2 , but M = 2 c ⨁ i a 2 i − 1 a 2 i J 0 M=2c\bigoplus_i a_{2i-1}a_{2i}J_0 M = 2 c ⨁ i a 2 i − 1 a 2 i J 0 may have infinite rank. Use independent centered scalar curvature coordinates in all remaining entries. This also satisfies ( A C P ) (\mathrm{ACP}) ( ACP ) for the projection onto the first noise block, without requiring its mean to have finite rank.
The raw domain and attainable principal ideals
The classification computes each fixed covariance response. We now determine both the kernel domain valid for every curvature array and its full range. Here the law ranges over all centered arrays admitting some fixed 0 < α ≤ β < ∞ 0<\alpha\le\beta<\infty 0 < α ≤ β < ∞ in ( H 0 ) (\mathrm{H}_0) ( H 0 ) ; these constants are uniform over finite tests of that law, not prescribed for the whole range. Fix a 2 > 0 a_2>0 a 2 > 0 and nonzero skew K ∈ S 2 K\in\mathcal{S}_2 K ∈ S 2 . When K ∈ S 1 K\in\mathcal{S}_1 K ∈ S 1 , put
S n = ∑ j ≥ n κ j , p n = R n S n − R n + 1 S n + 1 = q n S n + κ n R n + 1 . (83) S_n=\sum_{j\ge n}\kappa_j,\qquad p_n=R_nS_n-R_{n+1}S_{n+1}=q_nS_n+\kappa_nR_{n+1}. \tag*{(83)} S n = j ≥ n ∑ κ j , p n = R n S n − R n + 1 S n + 1 = q n S n + κ n R n + 1 . (83)
Thus p = ( a ∘ κ ) 2 p=(a\circ\sqrt{\kappa})^2 p = ( a ∘ κ ) 2 is decreasing and summable. Expansion gives r ≤ p r\le p r ≤ p and κ n R n ≤ r n \kappa_nR_n\le r_n κ n R n ≤ r n .
Theorem C (The raw domain and the entire principal-ideal interval). Let Q ≥ 0 Q\ge0 Q ≥ 0 be trace class with rank Q ≥ 2 \operatorname{rank}Q\ge2 rank Q ≥ 2 , put a = s ( Q ) a=\sqrt{s(Q)} a = s ( Q ) , and let 0 ≠ K ∈ S 2 ( E ) 0\ne K\in\mathcal{S}_2(E) 0 = K ∈ S 2 ( E ) be skew-adjoint. For any prescribed exhaustion (77) , raw evaluation converges in L 2 ( S 2 ) L^2(\mathcal{S}_2) L 2 ( S 2 ) for every centered curvature array if and only if K ∈ S 1 K\in\mathcal{S}_1 K ∈ S 1 . If K ∉ S 1 K\notin\mathcal{S}_1 K ∈ / S 1 , failure already occurs for Gaussian covariances arbitrarily close to the identity.
Suppose K ∈ S 1 K\in\mathcal{S}_1 K ∈ S 1 . Every such array has A Z ∈ L 2 ( S 1 ) A_Z\in L^2(\mathcal{S}_1) A Z ∈ L 2 ( S 1 ) . Its deterministic representative Δ = E A Z − C \Delta=\mathbb{E}A_Z-\mathcal{C} Δ = E A Z − C satisfies
s 4 n − 3 ( Δ ) ≤ 2 α − 1 q 2 n − 1 S 2 n − 1 , I ( Δ ) ⊂ I ( ( q n S n ) n ) . (84) s_{4n-3}(\Delta)\le2\alpha^{-1}q_{2n-1}S_{2n-1},\qquad\mathcal{I}(\Delta)\subset\mathcal{I}((q_nS_n)_n). \tag*{(84)} s 4 n − 3 ( Δ ) ≤ 2 α − 1 q 2 n − 1 S 2 n − 1 , I ( Δ ) ⊂ I (( q n S n ) n ) . (84) Consequently its principal ideal is deterministic and
I ( r ) ⊂ I ( A Z ) ⊂ I ( p ) almost surely . (85) \mathcal{I}(r)\subset\mathcal{I}(A_Z)\subset\mathcal{I}(p)\quad\text{almost surely}. \tag*{(85)} I ( r ) ⊂ I ( A Z ) ⊂ I ( p ) almost surely . (85)
Conversely, for this fixed trace-class pair let u u u be any nonnegative decreasing null sequence with I ( r ) ⊆ I ( u ) ⊆ I ( p ) \mathcal{I}(r) \subseteq\mathcal{I}(u) \subseteq\mathcal{I}(p) I ( r ) ⊆ I ( u ) ⊆ I ( p ) . For every 0 < ε < 1 0<\varepsilon<1 0 < ε < 1 there is a centered Gaussian array Z u , ε Z^{u,\varepsilon} Z u , ε whose covariance, on all square-summable tests, satisfies
( 1 − ε ) I ≤ Cov Z u , ε ≤ ( 1 + ε ) I , I ( A Z u , ε ) = I ( u ) a.s. (86) (1-\varepsilon)I \leq\operatorname{Cov} Z^{u,\varepsilon} \leq(1+\varepsilon)I,\qquad\mathcal{I}(AZ^{u,\varepsilon})=\mathcal{I}(u)\quad\text{a.s.} \tag*{(86)} ( 1 − ε ) I ≤ Cov Z u , ε ≤ ( 1 + ε ) I , I ( A Z u , ε ) = I ( u ) a.s. (86)
Each individual row remains a standard Gaussian sequence. The constructed field has curvature bounds ( 1 + ε ) − 1 (1+\varepsilon)^{-1} ( 1 + ε ) − 1 and ( 1 − ε ) − 1 (1-\varepsilon)^{-1} ( 1 − ε ) − 1 . The compensator in (65) is exactly zero and the mean is the explicit compact skew operator
M u = ε ⨁ i ≥ 1 min { u i , δ i } J 0 , J 0 = ( 0 1 − 1 0 ) , δ i = a 2 i − 1 a 2 i S 2 i − 1 . (87) M_u=\varepsilon\bigoplus_{i\geq1}\min\{u_i,\delta_i\}J_0,\qquad J_0=\begin{pmatrix}0&1\\-1&0\end{pmatrix},\qquad\delta_i=a_{2i-1}a_{2i}S_{2i-1}. \tag*{(87)} M u = ε i ≥ 1 ⨁ min { u i , δ i } J 0 , J 0 = ( 0 − 1 1 0 ) , δ i = a 2 i − 1 a 2 i S 2 i − 1 . (87)
For every fixed compact D D D it has the full response
I ( A Z u , ε + D ) = I ( r ) + I ( M u + D ) a.s. (88) \mathcal{I}(AZ^{u,\varepsilon}+D)=\mathcal{I}(r)+\mathcal{I}(M_u+D)\quad\text{a.s.} \tag*{(88)} I ( A Z u , ε + D ) = I ( r ) + I ( M u + D ) a.s. (88)
Hence the attainable ideals form the entire order interval between I ( r ) \mathcal{I}(r) I ( r ) and I ( p ) \mathcal{I}(p) I ( p ) in the principal-ideal lattice. This interval is not asserted to be a chain. Events are for fixed u , ε , D u,\varepsilon,D u , ε , D , not uniform over these choices or sample-dependent shifts. The smallness in (86) is covariance operator norm, not finite rank or a Schatten norm.
Proof. The universal raw domain. Convex comparison gives E ∥ D d Z f ∥ 4 ≤ 3 α − 2 ∥ d ∥ 4 ∥ f ∥ 4 \mathbb{E}\lVert D_dZf\rVert^4\leq3\alpha^{-2}\lVert d\rVert^4\lVert f\rVert^4 E ∥ D d Z f ∥ 4 ≤ 3 α − 2 ∥ d ∥ 4 ∥ f ∥ 4 . For a trace-class kernel L = ∑ j λ j f j ⊗ g j L=\sum_j\lambda_j f_j\otimes g_j L = ∑ j λ j f j ⊗ g j with unit singular vectors, each raw rank-one evaluation is ( D d Z f j ) ⊗ ( D e Z g j ) (D_dZf_j)\otimes(D_eZg_j) ( D d Z f j ) ⊗ ( D e Z g j ) . Its L 2 ( S 1 ) L^2(S_1) L 2 ( S 1 ) norm is bounded by
( E ∥ D d Z f j ∥ 4 ) 1 / 4 ( E ∥ D e Z g j ∥ 4 ) 1 / 4 ≤ 3 α − 1 ∥ d ∥ 2 ∥ e ∥ 2 . \left(\mathbb{E}\lVert D_dZf_j\rVert^4\right)^{1/4}\left(\mathbb{E}\lVert D_eZg_j\rVert^4\right)^{1/4}\leq\sqrt{3}\alpha^{-1}\lVert d\rVert_2\lVert e\rVert_2. ( E ∥ D d Z f j ∥ 4 ) 1/4 ( E ∥ D e Z g j ∥ 4 ) 1/4 ≤ 3 α − 1 ∥ d ∥ 2 ∥ e ∥ 2 .
Minkowski’s inequality and ∑ j λ j = ∥ L ∥ 1 \sum_j\lambda_j=\lVert L\rVert_1 ∑ j λ j = ∥ L ∥ 1 therefore yield
∥ D d Z L Z ∗ D e ∥ L 2 ( S 1 ) ≤ 3 α − 1 ∥ d ∥ 2 ∥ e ∥ 2 ∥ L ∥ 1 . (89) \lVert D_dZLZ^*D_e\rVert_{L^2(S_1)}\leq\sqrt{3}\alpha^{-1}\lVert d\rVert_2\lVert e\rVert_2\lVert L\rVert_1. \tag*{(89)} ∥ D d Z L Z ∗ D e ∥ L 2 ( S 1 ) ≤ 3 α − 1 ∥ d ∥ 2 ∥ e ∥ 2 ∥ L ∥ 1 . (89)
Applied to differences, this proves existence and agreement of coordinate and spectral limits. Corollary 4.5 then applies.
For necessity retain the prescribed noise projections Π n N \Pi_{n_N} Π n N and write K N = Π n N K Π n N K_N=\Pi_{n_N}K\Pi_{n_N} K N = Π n N K Π n N . If K ∉ S 1 K\notin S_1 K ∈ / S 1 , then sup N ∥ K N ∥ 1 = ∞ \sup_N\lVert K_N\rVert_1=\infty sup N ∥ K N ∥ 1 = ∞ , since uniformly bounded trace norms and singular-value convergence would force K ∈ S 1 K\in S_1 K ∈ S 1 . On the real Banach space of bounded skew-adjoint operators, the functionals ℓ N ( Γ ) = tr ( K N ∗ Γ ) \ell_N(\Gamma)=\operatorname{tr}(K_N^*\Gamma) ℓ N ( Γ ) = tr ( K N ∗ Γ ) have norms ∥ K N ∥ 1 \lVert K_N\rVert_1 ∥ K N ∥ 1 ; the skew polar partial isometry of K N K_N K N attains this norm. The uniform boundedness principle [18 Chapter 2] supplies a skew Γ \Gamma Γ with sup N ∣ ℓ N ( Γ ) ∣ = ∞ \sup_N|\ell_N(\Gamma)|=\infty sup N ∣ ℓ N ( Γ ) ∣ = ∞ ; scale it so ∥ Γ ∥ < ε < 1 \lVert\Gamma\rVert<\varepsilon<1 ∥ Γ ∥ < ε < 1 . Give the first two Gaussian rows covariance
( I Γ − Γ I ) \begin{pmatrix}I&\Gamma\\-\Gamma&I\end{pmatrix} ( I − Γ Γ I )
and add independent standard rows. Its full covariance is within ε \varepsilon ε of I I I , hence it is a curvature array. The first off-diagonal raw mean is a 1 a 2 ℓ N ( Γ ) a_1a_2\ell_N(\Gamma) a 1 a 2 ℓ N ( Γ ) once m N ≥ 2 m_N\geq2 m N ≥ 2 . It is unbounded, precluding L 2 ( S 2 ) L^2(S_2) L 2 ( S 2 ) convergence in the original coordinates. No change of the prescribed exhaustion is used.
A deterministic bound for the obstruction. Put M = c α ( K ) = ∑ ℓ M ℓ M=c_\alpha(K)=\sum_\ell M_\ell M = c α ( K ) = ∑ ℓ M ℓ and γ = α − 1 \gamma=\alpha^{-1} γ = α − 1 . These means and the compensator series converge absolutely in trace norm: ∥ M ℓ ∥ 1 ≤ 2 γ σ ℓ R 1 \lVert M_\ell\rVert_1\leq2\gamma\sigma_\ell R_1 ∥ M ℓ ∥ 1 ≤ 2 γ σ ℓ R 1 . Nested projections cancel every term with ℓ < n \ell<n ℓ < n , giving
P n Δ P n = ∑ ℓ ≥ n ( P n M ℓ P n − P ℓ + 1 M ℓ P ℓ + 1 ) . (90) P_n\Delta P_n=\sum_{\ell\geq n}\left(P_nM_\ell P_n-P_{\ell+1}M_\ell P_{\ell+1}\right). \tag*{(90)} P n Δ P n = ℓ ≥ n ∑ ( P n M ℓ P n − P ℓ + 1 M ℓ P ℓ + 1 ) . (90)
By (60) each summand has norm at most 4 γ σ ℓ q 2 n − 1 4\gamma\sigma_\ell q_{2n-1} 4 γ σ ℓ q 2 n − 1 . Thus ∥ P n Δ P n ∥ ≤ 2 γ q 2 n − 1 S 2 n − 1 \lVert P_n\Delta P_n\rVert\leq2\gamma q_{2n-1}S_{2n-1} ∥ P n Δ P n ∥ ≤ 2 γ q 2 n − 1 S 2 n − 1 . Since rank ( Δ − P n Δ P n ) ≤ 4 n − 4 \operatorname{rank}(\Delta-P_n\Delta P_n)\leq4n-4 rank ( Δ − P n Δ P n ) ≤ 4 n − 4 , this proves (84) . Now r ≤ p r\leq p r ≤ p and q n S n ≤ p n q_nS_n\leq p_n q n S n ≤ p n ; (78) yields (85) .
The available deterministic part. Write w i = a 2 i − 1 a 2 i w_i=a_{2i-1}a_{2i} w i = a 2 i − 1 a 2 i . Then δ i = w i S 2 i − 1 \delta_i=w_iS_{2i-1} δ i = w i S 2 i − 1 decreases, and
δ n ≤ q n S n , q n S n ≤ max { 1 , a 1 / a 2 } δ ⌈ n / 3 ⌉ . \delta_n\leq q_nS_n,\qquad q_nS_n\leq\max\{1,a_1/a_2\}\,\delta_{\lceil n/3\rceil}. δ n ≤ q n S n , q n S n ≤ max { 1 , a 1 / a 2 } δ ⌈ n /3 ⌉ .
For n ≥ 2 n\geq2 n ≥ 2 the latter follows from 2 ⌈ n / 3 ⌉ ≤ n 2\lceil n/3\rceil\leq n 2 ⌈ n /3 ⌉ ≤ n ; n = 1 n=1 n = 1 accounts for the constant. Since q n S n ≤ p n ≤ q n S n + r n q_nS_n\leq p_n\leq q_nS_n+r_n q n S n ≤ p n ≤ q n S n + r n and r n ≤ p n r_n\leq p_n r n ≤ p n , we obtain
I ( δ ) = I ( ( q n S n ) n ) , I ( p ) = I ( r ) + I ( δ ) . (91) \mathcal{I}(\delta)=\mathcal{I}((q_nS_n)_n),\qquad\mathcal{I}(p)=\mathcal{I}(r)+\mathcal{I}(\delta). \tag*{(91)} I ( δ ) = I (( q n S n ) n ) , I ( p ) = I ( r ) + I ( δ ) . (91)
Put t = min { u , δ } t=\min\{u,\delta\} t = min { u , δ } . The lattice identities of Lemma 2.1 give
I ( u ) = I ( u ) ∩ ( I ( r ) + I ( δ ) ) = I ( r ) + I ( t ) . (92) \mathcal{I}(u)=\mathcal{I}(u)\cap\bigl(\mathcal{I}(r)+\mathcal{I}(\delta)\bigr)=\mathcal{I}(r)+\mathcal{I}(t). \tag*{(92)} I ( u ) = I ( u ) ∩ ( I ( r ) + I ( δ ) ) = I ( r ) + I ( t ) . (92)
Here I ( r ) ⊆ I ( u ) ⊆ I ( p ) \mathcal{I}(r)\subseteq\mathcal{I}(u)\subseteq\mathcal{I}(p) I ( r ) ⊆ I ( u ) ⊆ I ( p ) and (91) justify the first equality and absorption. No same-index domination of the target is assumed.
Triangular correlations. In the canonical noise blocks K = ⨁ ℓ σ ℓ J 0 K=\bigoplus_\ell\sigma_\ell J_0 K = ⨁ ℓ σ ℓ J 0 , take independent standard Gaussian row vectors X i ℓ , Y i ℓ ∈ R 2 X_{i\ell},Y_{i\ell}\in\mathbb{R}^2 X i ℓ , Y i ℓ ∈ R 2 and set
Z 2 i − 1 , ℓ = X i ℓ , Z 2 i , ℓ = c i ℓ X i ℓ J 0 + 1 − c i ℓ 2 Y i ℓ , c i ℓ = λ i 1 { ℓ ≥ i } , (93) Z_{2i-1,\ell}=X_{i\ell},\qquad Z_{2i,\ell}=c_{i\ell}X_{i\ell}J_0+\sqrt{1-c_{i\ell}^{2}}Y_{i\ell},\qquad c_{i\ell}=\lambda_i\mathbf{1}_{\{\ell\ge i\}}, \tag*{(93)} Z 2 i − 1 , ℓ = X i ℓ , Z 2 i , ℓ = c i ℓ X i ℓ J 0 + 1 − c i ℓ 2 Y i ℓ , c i ℓ = λ i 1 { ℓ ≥ i } , (93)
where λ i = ε t i / δ i \lambda_i=\varepsilon t_i/\delta_i λ i = ε t i / δ i for δ i > 0 \delta_i>0 δ i > 0 and λ i = 0 \lambda_i=0 λ i = 0 otherwise. On ker K \ker K ker K use independent standard noise. Each covariance block has eigenvalues 1 ± c i ℓ 1\pm c_{i\ell} 1 ± c i ℓ , proving (86) and the curvature hypothesis. The bounded Gaussian covariance transports the construction to the prescribed noise basis, and (89) identifies its raw evaluation there. Direct covariance contraction gives
M ℓ = 2 σ ℓ ⨁ i w i c i ℓ J 0 , P ℓ + 1 M ℓ P ℓ + 1 = 0 , ∑ ℓ M ℓ = M u . (94) M_\ell=2\sigma_\ell\bigoplus_i w_i c_{i\ell}J_0,\qquad P_{\ell+1}M_\ell P_{\ell+1}=0,\qquad\sum_\ell M_\ell=M_u. \tag*{(94)} M ℓ = 2 σ ℓ i ⨁ w i c i ℓ J 0 , P ℓ + 1 M ℓ P ℓ + 1 = 0 , ℓ ∑ M ℓ = M u . (94)
The middle equality is the point of the construction: M ℓ M_\ell M ℓ is supported on the first ℓ \ell ℓ spatial pairs, which P ℓ + 1 P_{\ell+1} P ℓ + 1 deletes. The sum converges in trace norm by (89) , and 2 ∑ ℓ ≥ i σ ℓ = S 2 i − 1 2\sum_{\ell\ge i}\sigma_\ell=S_{2i-1} 2 ∑ ℓ ≥ i σ ℓ = S 2 i − 1 verifies its value. Consequently C = 0 \mathcal{C}=0 C = 0 . Corollary 4.5 gives (88) ; taking D = 0 D=0 D = 0 , duplicating each singular value and using (92) gives I ( A Z ) = I ( u ) \mathcal{I}(A_Z)=\mathcal{I}(u) I ( A Z ) = I ( u ) . For u = p u=p u = p , t = δ t=\delta t = δ and I ( Δ ) = I ( δ ) \mathcal{I}(\Delta)=\mathcal{I}(\delta) I ( Δ ) = I ( δ ) , attaining the ideal bound in (84) . □
Example 5.1. (Distinct principal ideals with identical Schatten membership). Fix q n = n − 2 q_n=n^{-2} q n = n − 2 and a skew kernel with paired singular values
κ 2 ℓ − 1 = κ 2 ℓ = 1 ℓ [ log ( e ℓ ) ] 2 . \kappa_{2\ell-1}=\kappa_{2\ell}=\frac{1}{\ell[\log(e\ell)]^2}. κ 2 ℓ − 1 = κ 2 ℓ = ℓ [ log ( e ℓ ) ] 2 1 . The kernel is trace class. Integral comparison, and for the square tail a sum over ℓ \ell ℓ between m m m and 2 m 2m 2 m , give
R n ≍ n − 1 , S n ≍ [ log ( e n ) ] − 1 , T n ≍ 1 n [ log ( e n ) ] 4 . R_n\asymp n^{-1},\qquad S_n\asymp[\log(en)]^{-1},\qquad T_n\asymp\frac{1}{n[\log(en)]^4}. R n ≍ n − 1 , S n ≍ [ log ( e n ) ] − 1 , T n ≍ n [ log ( e n ) ] 4 1 . Using the exact expansion r n 2 = q n ( R n + R n + 1 ) T n + κ n 2 R n + 1 2 r_n^2=q_n(R_n+R_{n+1})T_n+\kappa_{n}^{2}R_{n+1}^{2} r n 2 = q n ( R n + R n + 1 ) T n + κ n 2 R n + 1 2 , rather than differencing asymptotic equivalents, and p n = q n S n + κ n R n + 1 p_n=q_nS_n+\kappa_nR_{n+1} p n = q n S n + κ n R n + 1 , we obtain
r n ≍ 1 n 2 [ log ( e n ) ] 2 , p n ≍ 1 n 2 log ( e n ) , u n ( θ ) = 1 n 2 [ log ( e n ) ] 2 − θ , 0 ≤ θ ≤ 1. (95) r_n\asymp\frac{1}{n^2[\log(en)]^2},\qquad p_n\asymp\frac{1}{n^2\log(en)},\qquad u_n^{(\theta)}=\frac{1}{n^2[\log(en)]^{2-\theta}} ,\qquad0\le\theta\le1. \tag*{(95)} r n ≍ n 2 [ log ( e n ) ] 2 1 , p n ≍ n 2 log ( e n ) 1 , u n ( θ ) = n 2 [ log ( e n ) ] 2 − θ 1 , 0 ≤ θ ≤ 1. (95) For θ 1 < θ 2 \theta_1<\theta_2 θ 1 < θ 2 and every fixed integer M ≥ 1 M\ge1 M ≥ 1 , u n ( θ 2 ) / u ⌈ n / M ⌉ ( θ 1 ) → ∞ u_n^{(\theta_2)}/u_{\lceil n/M\rceil}^{(\theta_1)}\to\infty u n ( θ 2 ) / u ⌈ n / M ⌉ ( θ 1 ) → ∞ . The finite-dilation criterion therefore gives the strict chain I ( u ( θ 1 ) ) ⊊ I ( u ( θ 2 ) ) \mathcal{I}(u^{(\theta_1)})\subsetneq\mathcal{I}(u^{(\theta_2)}) I ( u ( θ 1 ) ) ⊊ I ( u ( θ 2 ) ) . Theorem C realizes each of these ideals for this same a a a , K K K , with covariance arbitrarily close to I I I and unchanged standard Gaussian row marginals. Yet for every 0 < s < ∞ 0<s<\infty 0 < s < ∞ ,
∑ n ( u n ( θ ) ) s < ∞ ⟺ s > 1 2 ( 0 ≤ θ ≤ 1 ) . (96) \sum_n\bigl(u_n^{(\theta)}\bigr)^s<\infty\quad\Longleftrightarrow\quad s>\frac12\qquad(0\le\theta\le1). \tag*{(96)} n ∑ ( u n ( θ ) ) s < ∞ ⟺ s > 2 1 ( 0 ≤ θ ≤ 1 ) . (96) At s = 1 / 2 s=1/2 s = 1/2 the series is ∑ n n − 1 [ log ( e n ) ] − 1 + θ / 2 \sum_n n^{-1}[\log(en)]^{-1+\theta/2} ∑ n n − 1 [ log ( e n ) ] − 1 + θ /2 , which diverges; off that endpoint the power of n n n decides convergence. Finite dilations preserve ℓ s \ell^s ℓ s membership, so the realized operators have exactly the same Schatten membership answers despite pairwise distinct principal ideals. This chain illustrates, but does not exhaust, the full order interval.
For a preassigned curvature window [ α , β ] [\alpha,\beta] [ α , β ] , the above Gaussian construction is admissible whenever α ≤ ( 1 + ε ) − 1 \alpha\le(1+\varepsilon)^{-1} α ≤ ( 1 + ε ) − 1 and ( 1 − ε ) − 1 ≤ β (1-\varepsilon)^{-1}\le\beta ( 1 − ε ) − 1 ≤ β . In particular, the centered class α = β = 1 \alpha=\beta=1 α = β = 1 contains only the independent standard Gaussian field and admits no nontrivial covariance perturbation.
Remark (Arbitrary signed correlations). More generally, use (93) with any real array ∣ c i ℓ ∣ ≤ c ∗ < 1 |c_{i\ell}| \le c_* < 1 ∣ c i ℓ ∣ ≤ c ∗ < 1 , without the triangular restriction. The same covariance calculation and the compensator series give the exact raw representative
Δ c = ⨁ i ( 2 w i ∑ ℓ ≥ i c i ℓ ) J 0 , I ( A Z + D ) = I ( r ) + I ( Δ c + D ) a.s. (97) \Delta_c=\bigoplus_i\left(2w_i\sum_{\ell\ge i}c_{i\ell}\right)J_0,\qquad\mathcal{I}(AZ+D)=\mathcal{I}(r)+\mathcal{I}(\Delta_c+D)\quad\text{a.s.} \tag*{(97)} Δ c = i ⨁ ( 2 w i ℓ ≥ i ∑ c i ℓ ) J 0 , I ( A Z + D ) = I ( r ) + I ( Δ c + D ) a.s. (97)
Indeed the full mean uses all ℓ \ell ℓ , whereas C \mathcal{C} C uses precisely ℓ < i \ell<i ℓ < i in the i i i th spatial pair. Absolute trace-norm convergence follows from K ∈ S 1 K\in\mathcal{S}_1 K ∈ S 1 and ∑ i w i < ∞ \sum_i w_i<\infty ∑ i w i < ∞ . This computes the correction in (78) explicitly. The Gaussian coordinates can also be replaced by independent copies of any centered variance-one non-Gaussian scalar law with two-sided curvature bounds α 0 , β 0 \alpha_0,\beta_0 α 0 , β 0 . Each block transformation has squared singular values 1 ± ∣ c i ℓ ∣ 1\pm|c_{i\ell}| 1 ± ∣ c i ℓ ∣ ; hence the transformed field has bounds α 0 / ( 1 + c ∗ ) \alpha_0/(1+c_*) α 0 / ( 1 + c ∗ ) and β 0 / ( 1 − c ∗ ) \beta_0/(1-c_*) β 0 / ( 1 − c ∗ ) and the same covariance. Marginal and weak closure transfer this assertion to the prescribed noise basis. All formulas, including the full interval realization, follow by Corollary 4.5. For instance one may standardize the density c exp ( − x 2 / 2 − log cosh x ) c\exp(-x^2/2-\log\cosh x) c exp ( − x 2 /2 − log cosh x ) .
Corollary 5.2 (Preservation throughout the curvature class). For the fixed trace-class pair ( a , K ) (a,K) ( a , K ) , the following are equivalent: I ( p ) = I ( r ) \mathcal{I}(p)=\mathcal{I}(r) I ( p ) = I ( r ) ; every centered curvature array has I ( A Z ) = I ( r ) \mathcal{I}(AZ)=\mathcal{I}(r) I ( A Z ) = I ( r ) almost surely; and every such array has I ( A Z + D ) = I ( r ) + I ( D ) \mathcal{I}(AZ+D)=\mathcal{I}(r)+\mathcal{I}(D) I ( A Z + D ) = I ( r ) + I ( D ) almost surely for each fixed compact D D D . If I ( r ) ⊊ I ( p ) \mathcal{I}(r)\subsetneq\mathcal{I}(p) I ( r ) ⊊ I ( p ) , every principal ideal in between is attained by arbitrarily small Gaussian covariance perturbations. For each realized ideal strictly larger than I ( r ) \mathcal{I}(r) I ( r ) , the deterministic shift D = − M u D=-M_u D = − M u in (88) reduces it exactly to I ( r ) \mathcal{I}(r) I ( r ) , so the unmodified sum law fails.
Proof. The bounds prove the first implication, and (79) gives the fixed-shift formula. The upper endpoint realization proves the converse. The final assertion follows from (88) and (92) . □
The interval concerns raw principal ideals, not arbitrary quotient classes. For Gaussian covariances Γ t = I + t ( Γ − I ) \Gamma_t=I+t(\Gamma-I) Γ t = I + t ( Γ − I ) , with Γ \Gamma Γ bounded uniformly positive and 0 < t ≤ 1 0<t\le1 0 < t ≤ 1 , covariance linearity gives
M N ( t ) = t M N ( 1 ) , C Γ t = t C Γ , Δ t = t Δ 1 . (98) M_N(t)=tM_N(1),\qquad C_{\Gamma_t}=tC_\Gamma,\qquad\Delta_t=t\Delta_1. \tag*{(98)} M N ( t ) = t M N ( 1 ) , C Γ t = t C Γ , Δ t = t Δ 1 . (98)
The last identity holds when the raw mean exists. Raw existence along the same exhaustion is equivalent for every fixed 0 < t ≤ 1 0<t\le1 0 < t ≤ 1 ; when it holds, I ( A t + D ) = I ( r ) + I ( Δ 1 + D / t ) \mathcal{I}(A_t+D)=\mathcal{I}(r)+\mathcal{I}(\Delta_1+D/t) I ( A t + D ) = I ( r ) + I ( Δ 1 + D / t ) . Thus the unshifted ideal is constant for 0 < t ≤ 1 0<t\le1 0 < t ≤ 1 , even arbitrarily close to I I I , but can change at t = 0 t=0 t = 0 . This remains valid for Hilbert–Schmidt kernels with convergent raw means; events are for fixed t , D t,D t , D . Thus continuity of the compact readout in Lemma 4.4 does not imply continuity of the algebraic ideal. Each fixed singular value has an explicit approximation error bound; arbitrary ideal membership is not thereby a finite-precision decision problem.
If I ( r ) ⊊ I ( p ) \mathcal{I}(r)\subsetneq\mathcal{I}(p) I ( r ) ⊊ I ( p ) , the triangular constructions for u = p u=p u = p with opposite correlations both have unshifted ideals I ( p ) \mathcal{I}(p) I ( p ) and identical covariance-block eigenvalues. Their means are M p M_p M p and − M p -M_p − M p ; the fixed shift D = − M p D=-M_p D = − M p gives respectively I ( r ) \mathcal{I}(r) I ( r ) and I ( p ) \mathcal{I}(p) I ( p ) . Thus neither the unshifted ideal nor the covariance spectrum determines the shift response.
Finite skew defects, tail regularity and Wiener area
Finite-dimensional skew contractions
Let P P P be a deterministic finite-rank orthogonal projection on the noise space E E E . The zero-contraction and finite-defect conditions are, respectively,
E Z i k Z j ℓ = E Z i ℓ Z j k ( i , j , k , ℓ ≥ 1 ) , \mathbb{E}Z_{ik}Z_{j\ell}=\mathbb{E}Z_{i\ell}Z_{jk}\qquad(i,j,k,\ell\ge1), E Z ik Z j ℓ = E Z i ℓ Z j k ( i , j , k , ℓ ≥ 1 ) ,
E [ Z i ( f ) Z j ( g ) ] = E [ Z i ( g ) Z j ( f ) ] ( f , g ∈ ( I − P ) E , i , j ≥ 1 ) , d = rank P < ∞ . \mathbb{E}[Z_i(f)Z_j(g)]=\mathbb{E}[Z_i(g)Z_j(f)]\qquad(f,g\in(I-P)E,\ i,j\ge1),\qquad d=\operatorname{rank}P<\infty. E [ Z i ( f ) Z j ( g )] = E [ Z i ( g ) Z j ( f )] ( f , g ∈ ( I − P ) E , i , j ≥ 1 ) , d = rank P < ∞.
Here P P P acts on the noise space, unlike the spatial P n P_n P n above. These conditions concern skew covariance, not conditional means.
Corollary 6.1 (Recovery with finite skew defects). Under ( H 0 ) (\mathrm{H}_0) ( H 0 ) and ( A C P ) (\mathrm{AC}_P) ( AC P ) , raw exterior evaluation extends uniquely and continuously to skew K ∈ S 2 ( E ) K\in\mathcal{S}_2(E) K ∈ S 2 ( E ) in L 2 ( Ω ; S 2 ) L^2(\Omega;\mathcal{S}_2) L 2 ( Ω ; S 2 ) , with
β − 2 c Q ∥ K ∥ 2 2 ≤ E ∥ A ∥ 2 2 ≤ ( 2 + 2 d ) α − 2 c Q ∥ K ∥ 2 2 . (99) \beta^{-2}c_{\mathcal Q}\lVert K\rVert_2^2\le\mathbb{E}\lVert A\rVert_2^2\le(2+2d)\alpha^{-2}c_{\mathcal Q}\lVert K\rVert_2^2. \tag*{(99)} β − 2 c Q ∥ K ∥ 2 2 ≤ E ∥ A ∥ 2 2 ≤ ( 2 + 2 d ) α − 2 c Q ∥ K ∥ 2 2 . (99) :::`
If rank Q ≥ 2 \operatorname{rank} Q \ge2 rank Q ≥ 2 , for every fixed compact D D D ,
I ( A + D ) = I ( r ) + I ( D ) = I ( max { r , s ( D ) } ) almost surely , (100) \mathcal{I}(A+D)=\mathcal{I}(r)+\mathcal{I}(D)=\mathcal{I}(\max\{r,s(D)\}) \quad\text{almost surely}, \tag*{(100)} I ( A + D ) = I ( r ) + I ( D ) = I ( max { r , s ( D )}) almost surely , (100)
and, for an absolute constant C C C ,
∥ s 6 n + 2 d − 5 ( A ) ∥ L n ≤ C α − 1 r n ( n ≥ 1 ) . (101) \left\|s_{6n+2d-5}(A)\right\|_{L^n}\le C\alpha^{-1}r_n \qquad(n\ge1). \tag*{(101)} ∥ s 6 n + 2 d − 5 ( A ) ∥ L n ≤ C α − 1 r n ( n ≥ 1 ) . (101)
For P = 0 P=0 P = 0 the mean vanishes. For d > 0 d>0 d > 0 it can have infinite rank. The zero and countable-event conventions above apply.
Proof. Put γ = α − 1 \gamma=\alpha^{-1} γ = α − 1 and Q 0 = I − P Q_0=I-P Q 0 = I − P . For the cross-row covariance C i j C_{ij} C ij and its skew part Γ i j \Gamma_{ij} Γ ij , (ACP) gives
Q 0 Γ i j Q 0 = 0 , rank Γ i j ≤ 2 d , ∥ Γ i j ∥ 2 ≤ 2 d γ . (102) Q_0\Gamma_{ij}Q_0=0,\qquad\operatorname{rank}\Gamma_{ij}\le2d,\qquad\|\Gamma_{ij}\|_2\le\sqrt{2d}\,\gamma. \tag*{(102)} Q 0 Γ ij Q 0 = 0 , rank Γ ij ≤ 2 d , ∥ Γ ij ∥ 2 ≤ 2 d γ . (102)
Pairing a skew kernel with Γ i j \Gamma_{ij} Γ ij yields
∥ c α ( F ) ∥ 2 2 ≤ 2 d γ 2 c Q ∥ F ∥ 2 2 . (103) \|c_{\alpha}(F)\|_2^2\le2d\gamma^2c_Q\|F\|_2^2. \tag*{(103)} ∥ c α ( F ) ∥ 2 2 ≤ 2 d γ 2 c Q ∥ F ∥ 2 2 . (103)
Thus the mean map extends continuously to S 2 S_2 S 2 ; together with Lemma 4.3 it constructs the raw evaluation. Orthogonality of its centered and deterministic parts gives the upper second moment. Conditional covariance of row i i i given row j j j , followed by the marginal covariance lower bound, gives E ∣ Z i K Z j ∗ ∣ 2 2 ≥ β − 2 ∥ K ∥ 2 2 \mathbb{E}|Z_iKZ_j^*|_2^2\ge\beta^{-2}\|K\|_2^2 E ∣ Z i K Z j ∗ ∣ 2 2 ≥ β − 2 ∥ K ∥ 2 2 for i ≠ j i\ne j i = j . Finite-coordinate approximation and summation prove the lower bound.
For the ideal formula set Δ = c α ( K ) − C \Delta=c_{\alpha}(K)-C Δ = c α ( K ) − C . With K ( n ) = K − F n K^{(n)}=K-F_n K ( n ) = K − F n , the defect condition and continuity give c α ( K ( n ) ) = c α ( P K ( n ) + Q 0 K ( n ) P ) c_{\alpha}(K^{(n)})=c_{\alpha}(PK^{(n)}+Q_0K^{(n)}P) c α ( K ( n ) ) = c α ( P K ( n ) + Q 0 K ( n ) P ) . The kernel on the right has rank at most 2 d 2d 2 d and trace norm at most 4 d κ 2 n − 1 4d\kappa^2n^{-1} 4 d κ 2 n − 1 . Equations (60) and (68) therefore give
∥ P n Δ P n ∥ ≤ ( 4 d + 2 ) γ r n , s 4 n − 3 ( Δ ) ≤ ( 4 d + 2 ) γ r n . (104) \|P_n\Delta P_n\|\le(4d+\sqrt{2})\gamma r_n,\qquad s_{4n-3}(\Delta)\le(4d+\sqrt{2})\gamma r_n. \tag*{(104)} ∥ P n Δ P n ∥ ≤ ( 4 d + 2 ) γ r n , s 4 n − 3 ( Δ ) ≤ ( 4 d + 2 ) γ r n . (104)
For the second inequality, Δ − P n Δ P n \Delta-P_n\Delta P_n Δ − P n Δ P n has rank at most 4 n − 4 4n-4 4 n − 4 . Thus Δ ∈ I ( r ) \Delta\in\mathcal{I}(r) Δ ∈ I ( r ) and (79) proves (100) .
For (101) , write K = K 0 + F K=K_0+F K = K 0 + F , where K 0 = Q 0 K Q 0 K_0=Q_0KQ_0 K 0 = Q 0 K Q 0 , F = P K + Q 0 K P F=PK+Q_0KP F = P K + Q 0 K P and rank F ≤ 2 d \operatorname{rank}F\le2d rank F ≤ 2 d . On Q 0 E Q_0E Q 0 E the skew contractions vanish, so the mean and compensator vanish and (73) applies to A ( K 0 ) A(K_0) A ( K 0 ) . Since s ( K 0 ) ≤ s ( K ) s(K_0)\le s(K) s ( K 0 ) ≤ s ( K ) , its profile is bounded by r r r . The raw difference A ( F ) A(F) A ( F ) has rank at most 2 d 2d 2 d ; the singular-value rank inequality proves (101) . □
Remark (Gaussian recovery and exact normalization). For independent standard Gaussian coordinates, the map is the isonormal map
⟨ G f , h ⟩ = W ( f ⊗ Q 1 / 2 h ) , E ( G f ⊗ G g ) = ⟨ f , g ⟩ Q . (105) \langle Gf,h\rangle=\mathcal{W}(f\otimes Q^{1/2}h),\qquad\mathbb{E}(Gf\otimes Gg)=\langle f,g\rangle_Q. \tag*{(105)} ⟨ G f , h ⟩ = W ( f ⊗ Q 1/2 h ) , E ( G f ⊗ G g ) = ⟨ f , g ⟩ Q . (105)
The curvature and zero-defect hypotheses hold with α = β = 1 \alpha=\beta=1 α = β = 1 , so the full ideal and fixed-shift formulas are special cases of Corollary 6.1. The exact Gaussian normalization, not claimed for general noise, is
E ∥ G K G ∗ ∥ 2 2 = c Q ∥ K ∥ 2 2 . (106) \mathbb{E}\|GKG^*\|_2^2=c_Q\|K\|_2^2. \tag*{(106)} E ∥ G K G ∗ ∥ 2 2 = c Q ∥ K ∥ 2 2 . (106)
Indeed a canonical block has mean-square norm 2 c Q 2c_Q 2 c Q , distinct blocks have zero cross moments, and ∥ K ∥ 2 2 = 2 ∑ l σ l 2 \|K\|_2^2=2\sum_l\sigma_l^2 ∥ K ∥ 2 2 = 2 ∑ l σ l 2 . If J f , J g J_f,J_g J f , J g embed the two canonical spans, the familiar identity remains
A = C − C ∗ , C = ( G J f ) D σ ( G J g ) ∗ . (107) A=C-C^*,\qquad C=(GJ_f)D_\sigma(GJ_g)^*. \tag*{(107)} A = C − C ∗ , C = ( G J f ) D σ ( G J g ) ∗ . (107)
This independent representation is unnecessary for the general proof. For rank Q ≥ 2 \operatorname{rank}Q\ge2 rank Q ≥ 2 and K ≠ 0 K\ne0 K = 0 , the evaluation is nonzero finite rank almost surely exactly when one input has finite rank, by (100) . No event covers arbitrary sample-dependent shifts (D = − A D=-A D = − A ).
Remark (Nonzero conditional prediction). For independent pairs of spatial rows, use in each noise coordinate the centered density p ε ( x , y ) ∝ exp [ − ( x 2 + y 2 ) / 2 − ε log cosh ( x − y ) ] p_\varepsilon(x,y)\propto\exp[-(x^2+y^2)/2-\varepsilon\log\cosh(x-y)] p ε ( x , y ) ∝ exp [ − ( x 2 + y 2 ) /2 − ε log cosh ( x − y )] . Its Hessian is between I I I and ( 1 + 2 ε ) I (1+2\varepsilon)I ( 1 + 2 ε ) I , and cross-row covariances are scalar operators, so (AC) holds. Yet with s = ( x + y ) / 2 s=(x+y)/\sqrt{2} s = ( x + y ) / 2 and d = ( x − y ) / 2 d=(x-y)/\sqrt{2} d = ( x − y ) / 2 , one has E d 2 + 2 ε E [ d tanh ( 2 d ) ] = 1 \mathbb{E}d^2+\sqrt{2}\varepsilon\mathbb{E}[d\tanh(\sqrt{2}d)]=1 E d 2 + 2 ε E [ d tanh ( 2 d )] = 1 , whence Cov ( x , y ) > 0 \operatorname{Cov}(x,y)>0 Cov ( x , y ) > 0 and E [ x ∣ y ] ≠ 0 \mathbb{E}[x\mid y]\ne0 E [ x ∣ y ] = 0 . Thus zero conditional row means are not necessary.
The universal raw-domain assertion of Theorem C does not exclude a specified non-trace-class kernel under (AC) or (ACP); these are structural restrictions on the field.
The following tail computations apply for each fixed curvature array satisfying (AC), including the Gaussian array. In the non-Gaussian case Q Q Q remains a reference weight; covariance terminology in these profiles refers to that Gaussian reference. Wiener area below uses only the Gaussian specialization and its exact isometry.
Tail averages and regularity at infinity
For a nonnegative nonincreasing sequence ξ ∈ ℓ 1 \xi\in\ell^{1} ξ ∈ ℓ 1 , put
U n = ∑ j ≥ n ξ j , ( H ∞ ξ ) n = U n / n . (108) U_n=\sum_{j\ge n}\xi_j,\qquad(H_{\infty}\xi)_n=U_n/n. \tag*{(108)} U n = j ≥ n ∑ ξ j , ( H ∞ ξ ) n = U n / n . (108)
The tail average decreases to zero but need not be summable. A nonzero sequence is regular at infinity if
U n ≤ C n ξ n ( n ≥ 1 ) (109) U_n\le Cn\xi_n\qquad(n\ge1) \tag*{(109)} U n ≤ C n ξ n ( n ≥ 1 ) (109)
for some C < ∞ C<\infty C < ∞ . For infinite support this is [12 Definition 4.11 and Theorem 4.12] : our inclusive tail only adds ξ n / n \xi_n/n ξ n / n to the strict-tail average. It also includes nonzero finite support without losing the last entry.
Lemma 6.2 (Removing dilation from a tail estimate). For every nonnegative nonincreasing ξ ∈ ℓ 1 \xi\in\ell^{1} ξ ∈ ℓ 1 ,
I ( ξ ) ⊆ I ( H ∞ ξ ) . (110) \mathcal{I}(\xi)\subseteq\mathcal{I}(H_{\infty}\xi). \tag*{(110)} I ( ξ ) ⊆ I ( H ∞ ξ ) . (110) Equality holds exactly under (109) . Indeed, even for some C < ∞ C<\infty C < ∞ , M ∈ N M\in\mathbb{N} M ∈ N , the estimate
U n / n ≤ C ξ ⌈ n / M ⌉ ( n ≥ 1 ) (111) U_n/n\le C\xi_{\lceil n/M\rceil}\qquad(n\ge1) \tag*{(111)} U n / n ≤ C ξ ⌈ n / M ⌉ ( n ≥ 1 ) (111) already implies (109) . Zero and finite-support sequences are included.
Proof. For m = ⌈ n / 2 ⌉ m=\lceil n/2\rceil m = ⌈ n /2 ⌉ , U m ≥ ( n − m + 1 ) ξ n ≥ m ξ n U_m\ge(n-m+1)\xi_n\ge m\xi_n U m ≥ ( n − m + 1 ) ξ n ≥ m ξ n , which proves (110) by the finite-dilation criterion. Equality of the ideals gives (111) ; evaluating it at n = M k n=Mk n = M k yields
U k = ∑ j = k M k − 1 ξ j + U M k ≤ ( M − 1 + C M ) k ξ k . U_k=\sum_{j=k}^{Mk-1}\xi_j+U_{Mk}\le(M-1+CM)k\xi_k. U k = j = k ∑ M k − 1 ξ j + U M k ≤ ( M − 1 + C M ) k ξ k . This proves the tail condition, whose converse is immediate. The first sum is empty if M = 1 M=1 M = 1 , and no sequence entry is divided out. For support in { 1 , … , N } \{1,\ldots,N\} { 1 , … , N } , U n ≤ N ξ n ≤ N n ξ n U_n\le N\xi_n\le Nn\xi_n U n ≤ N ξ n ≤ N n ξ n ; the zero case is immediate. □ \square □
Universal simplification of the exterior profile
Separating the covariance and form tails gives
r n 2 = q n ( R n + R n + 1 ) T n + κ n 2 R n + 1 2 . (112) r_n^2=q_n(R_n+R_{n+1})T_n+\kappa_n^2R_{n+1}^2. \tag*{(112)} r n 2 = q n ( R n + R n + 1 ) T n + κ n 2 R n + 1 2 . (112)
Put x n = q n R n T n x_n=\sqrt{q_nR_nT_n} x n = q n R n T n and y n = κ n R n y_n=\kappa_nR_n y n = κ n R n , both nonnegative, nonincreasing and square summable. The first term is at least x n 2 x_n^2 x n 2 , while T n ≥ κ n 2 T_n\ge\kappa_n^2 T n ≥ κ n 2 gives r n 2 ≥ y n 2 r_n^2\ge y_n^2 r n 2 ≥ y n 2 ; the same formula gives r n 2 ≤ 2 x n 2 + y n 2 r_n^2\le2x_n^2+y_n^2 r n 2 ≤ 2 x n 2 + y n 2 . Hence
max { x n , y n } ≤ r n ≤ 3 max { x n , y n } ( n ≥ 1 ) , (113) \max\{x_n,y_n\}\le r_n\le\sqrt{3}\max\{x_n,y_n\}\qquad(n\ge1), \tag*{(113)} max { x n , y n } ≤ r n ≤ 3 max { x n , y n } ( n ≥ 1 ) , (113)
and for rank Q ≥ 2 Q\ge2 Q ≥ 2 ,
I ( A ) = I ( ( q n R n T n ) n ) + I ( ( κ n R n ) n ) almost surely . (114) \mathcal{I}(A)=\mathcal{I}\bigl((\sqrt{q_nR_nT_n})_n\bigr)+\mathcal{I}\bigl((\kappa_nR_n)_n\bigr)\quad\text{almost surely}. \tag*{(114)} I ( A ) = I ( ( q n R n T n ) n ) + I ( ( κ n R n ) n ) almost surely . (114)
The same decomposition holds in (100) . Absorption of the first ideal for every Q Q Q is detected by the single test Q 0 = diag ( n − 2 ) Q_0=\operatorname{diag}(n^{-2}) Q 0 = diag ( n − 2 ) .
Corollary 6.3 (Universal collapse of the exterior profile). Fix a curvature array satisfying ( H 0 ) (H_0) ( H 0 ) and ( A C ) (AC) ( A C ) , and a skew-adjoint K ∈ S 2 ( E ) K\in S_2(E) K ∈ S 2 ( E ) . Put κ n = s n ( K ) \kappa_n=s_n(K) κ n = s n ( K ) and T n = ∑ j ≥ n κ j 2 T_n=\sum_{j\ge n}\kappa_j^2 T n = ∑ j ≥ n κ j 2 . For each positive trace-class Q Q Q , let A Q A_Q A Q be its raw exterior evaluation with weights s n ( Q ) \sqrt{s_n(Q)} s n ( Q ) . The following are equivalent:
(i) There is C < ∞ C<\infty C < ∞ such that
T n ≤ C n κ n 2 ( n ≥ 1 ) . (115) T_n\le Cn\kappa_n^2\qquad(n\ge1). \tag*{(115)} T n ≤ C n κ n 2 ( n ≥ 1 ) . (115)
(ii) For every fixed trace-class covariance Q Q Q of rank at least two,
I ( A Q ) = I ( ( κ n R n ) n ) almost surely , R n = ∑ j ≥ n s j ( Q ) . (116) \mathcal{I}(A_Q)=\mathcal{I}((\kappa_nR_n)_n)\quad\text{almost surely},\qquad R_n=\sum_{j\ge n}s_j(Q). \tag*{(116)} I ( A Q ) = I (( κ n R n ) n ) almost surely , R n = j ≥ n ∑ s j ( Q ) . (116)
(iii) Identity (116) holds almost surely for Q 0 = diag ( n − 2 ) Q_0=\operatorname{diag}(n^{-2}) Q 0 = diag ( n − 2 ) on ℓ 2 \ell^2 ℓ 2 .
For K ≠ 0 K\ne0 K = 0 , condition (i) is regularity at infinity of κ 2 \kappa^2 κ 2 . The zero and finite-rank cases are included. The event in (ii) may depend on the fixed Q Q Q .
Proof. Assume (i) with C ≥ 1 C\ge1 C ≥ 1 . For m = ⌈ n / 2 ⌉ m=\lceil n/2\rceil m = ⌈ n /2 ⌉ , R m ≥ ∑ j = m n q j ≥ m q n R_m\ge\sum_{j=m}^n q_j\ge mq_n R m ≥ ∑ j = m n q j ≥ m q n , and therefore
x n 2 = q n R n T n ≤ C n q n R n κ n 2 ≤ 2 C κ m 2 R m 2 = 2 C y m 2 . x_n^2=q_nR_nT_n\le Cnq_nR_n\kappa_n^2\le2C\kappa_m^2R_m^2=2Cy_m^2. x n 2 = q n R n T n ≤ C n q n R n κ n 2 ≤ 2 C κ m 2 R m 2 = 2 C y m 2 .
Together with (112) , this gives
y n ≤ r n ≤ 1 + 4 C y ⌈ n / 2 ⌉ . y_n\le r_n\le\sqrt{1+4C y_{\lceil n/2\rceil}}. y n ≤ r n ≤ 1 + 4 C y ⌈ n /2 ⌉ .
The ideal criterion and Corollary 6.1 prove (ii), which immediately implies (iii).
Conversely, take q n = n − 2 q_n=n^{-2} q n = n − 2 , so 1 / n ≤ R n ≤ 2 / n 1/n\le R_n\le2/n 1/ n ≤ R n ≤ 2/ n . Condition (iii) and Corollary 6.1 give I ( r ) = I ( y ) \mathcal{I}(r)=\mathcal{I}(y) I ( r ) = I ( y ) ; hence r n ≤ B y m r_n\le By_m r n ≤ B y m for some B > 0 B>0 B > 0 , M ∈ N M\in\mathbb{N} M ∈ N and m = ⌈ n / M ⌉ m=\lceil n/M\rceil m = ⌈ n / M ⌉ . Since r n ≥ x n r_n\ge x_n r n ≥ x n ,
T n n 3 ≤ x n 2 ≤ r n 2 ≤ B 2 κ m 2 R m 2 ≤ 4 B 2 M 2 n 2 κ m 2 . \frac{T_n}{n^3}\le x_n^2\le r_n^2\le B^2\kappa_m^2R_m^2\le\frac{4B^2M^2}{n^2}\kappa_m^2. n 3 T n ≤ x n 2 ≤ r n 2 ≤ B 2 κ m 2 R m 2 ≤ n 2 4 B 2 M 2 κ m 2 .
Thus T n / n ≤ 4 B 2 M 2 κ ⌈ n / M ⌉ 2 T_n/n\le4B^2M^2\kappa_{\lceil n/M\rceil}^2 T n / n ≤ 4 B 2 M 2 κ ⌈ n / M ⌉ 2 . Lemma 6.2, applied to ξ = κ 2 \xi=\kappa^2 ξ = κ 2 , removes the dilation and proves (i), including zero and finite support. □
Under these conditions, (100) becomes
I ( A Q + D ) = I ( ( κ n R n ) n ) + I ( D ) almost surely \mathcal{I}(A_Q+D)=\mathcal{I}((\kappa_nR_n)_n)+\mathcal{I}(D)\quad\text{almost surely} I ( A Q + D ) = I (( κ n R n ) n ) + I ( D ) almost surely
for every fixed Q Q Q of rank at least two and compact D D D . The condition holds when κ n ≍ n − β \kappa_n\asymp n^{-\beta} κ n ≍ n − β with β > 1 / 2 \beta>1/2 β > 1/2 , since T n ≍ n 1 − 2 β T_n\asymp n^{1-2\beta} T n ≍ n 1 − 2 β ; this includes the Wiener time operator below.
Remark. (Why both exterior tails are needed.) Take κ n ≍ [ n 1 / 2 log ( n + 1 ) ] − 1 \kappa_n\asymp[n^{1/2}\log(n+1)]^{-1} κ n ≍ [ n 1/2 log ( n + 1 ) ] − 1 , with paired multiplicities, and q n = n − 2 q_n=n^{-2} q n = n − 2 . Then r n ≍ n − 3 / 2 / log ( n + 1 ) r_n\asymp n^{-3/2}/\sqrt{\log(n+1)} r n ≍ n − 3/2 / log ( n + 1 ) , whereas κ n R n ≍ n − 3 / 2 / log ( n + 1 ) \kappa_nR_n\asymp n^{-3/2}/\log(n+1) κ n R n ≍ n − 3/2 / log ( n + 1 ) . No fixed index dilation absorbs the logarithmic gap, so both terms in (114) matter. In contrast, with q n = 2 − n q_n=2^{-n} q n = 2 − n , one has I ( r ) = I ( ( κ n R n ) n ) \mathcal{I}(r)=\mathcal{I}((\kappa_nR_n)_n) I ( r ) = I (( κ n R n ) n ) for this same K K K , despite failure of (115) . Here R n / ( n q n ) = 2 / n R_n/(nq_n)=2/n R n / ( n q n ) = 2/ n , so the test covariance is itself regular at infinity. Thus regularity of a test covariance does not suffice to replace the prescribed Q 0 = diag ( n − 2 ) Q_0=\operatorname{diag}(n^{-2}) Q 0 = diag ( n − 2 ) .
Wiener area as covariance tail averaging
Let W = ( W t ) 0 ≤ t ≤ 1 W=(W_t)_{0\le t\le1} W = ( W t ) 0 ≤ t ≤ 1 be a Wiener process in a real separable Hilbert space with trace-class covariance Q ≥ 0 Q\ge0 Q ≥ 0 . Its skew area is
A = 1 2 ∫ 0 1 ( W t ⊗ d W t − d W t ⊗ W t ) . (117) A=\frac{1}{2}\int_0^1(W_t\otimes dW_t-dW_t\otimes W_t). \tag*{(117)} A = 2 1 ∫ 0 1 ( W t ⊗ d W t − d W t ⊗ W t ) . (117)
The Hilbert–Schmidt-valued Itô integral is well defined, since E ∥ ∫ 0 1 W t ⊗ d W t ∥ H S 2 = ( tr Q ) 2 / 2 \mathbb{E}\lVert\int_0^1W_t\otimes dW_t\rVert_{\mathrm{HS}}^2=(\operatorname{tr}Q)^2/2 E ∥ ∫ 0 1 W t ⊗ d W t ∥ HS 2 = ( tr Q ) 2 /2 and the other integral is its adjoint; see [5 Chapter 4] . Write q 1 ≥ q 2 ≥ ⋯ ≥ 0 q_1\ge q_2\ge\cdots\ge0 q 1 ≥ q 2 ≥ ⋯ ≥ 0 for the covariance eigenvalues, padded by zeros in finite rank, and put
R n = ∑ k ≥ n q k , b n = ( H ∞ q ) n = R n n . R_n=\sum_{k\ge n}q_k,\qquad b_n=(H_\infty q)_n=\frac{R_n}{n}. R n = k ≥ n ∑ q k , b n = ( H ∞ q ) n = n R n .
Corollary 6.4 (Wiener area and the covariance ideal). Suppose rank Q ≥ 2 Q\ge2 Q ≥ 2 . For every fixed deterministic compact operator D D D , without a skew-adjointness assumption,
I ( A + D ) = I ( b ) + I ( D ) almost surely . (118) \mathcal{I}(A+D)=\mathcal{I}(b)+\mathcal{I}(D)\quad\text{almost surely}. \tag*{(118)} I ( A + D ) = I ( b ) + I ( D ) almost surely . (118) In particular, on one event of probability one,
I ( A ) = I ( b ) , I ( Q ) ⊆ I ( A ) , (119) \mathcal{I}(A)=\mathcal{I}(b),\qquad\mathcal{I}(Q)\subseteq\mathcal{I}(A), \tag*{(119)} I ( A ) = I ( b ) , I ( Q ) ⊆ I ( A ) , (119) and
I ( A ) = I ( Q ) ⟺ R n ≤ C n q n ( n ≥ 1 ) for some C < ∞ . (120) \mathcal{I}(A)=\mathcal{I}(Q)\quad\Longleftrightarrow\quad R_n\le Cnq_n\quad(n\ge1)\text{ for some }C<\infty. \tag*{(120)} I ( A ) = I ( Q ) ⟺ R n ≤ C n q n ( n ≥ 1 ) for some C < ∞. (120) This includes finite rank. For rank Q ≤ 1 Q\le1 Q ≤ 1 , A = 0 A=0 A = 0 almost surely and I ( A + D ) = I ( D ) \mathcal{I}(A+D)=\mathcal{I}(D) I ( A + D ) = I ( D ) .
Proof. The Wiener integral G f = ∫ 0 1 f ( t ) d W t Gf = \int_0^1 f(t)\,\mathrm{d}W_t G f = ∫ 0 1 f ( t ) d W t is a Gaussian linear map on L 2 ( 0 , 1 ) L^2(0,1) L 2 ( 0 , 1 ) of covariance Q Q Q . For a partition π = ( t i ) \pi=(t_i) π = ( t i ) , set
K π = 1 2 ∑ i ( 1 [ 0 , t i ] ⊗ 1 ( t i , t i + 1 ] − 1 ( t i , t i + 1 ] ⊗ 1 [ 0 , t i ] ) . K_\pi=\frac{1}{2}\sum_i\left(\mathbf{1}_{[0,t_i]}\otimes\mathbf{1}_{(t_i,t_{i+1}]}-\mathbf{1}_{(t_i,t_{i+1}]}\otimes\mathbf{1}_{[0,t_i]}\right). K π = 2 1 i ∑ ( 1 [ 0 , t i ] ⊗ 1 ( t i , t i + 1 ] − 1 ( t i , t i + 1 ] ⊗ 1 [ 0 , t i ] ) .
Its evaluation is exactly the left Itô sum in (117) . As the mesh tends to zero, K π → K K_\pi\to K K π → K in S 2 S_2 S 2 , where ( K f ) ( s ) = 1 2 ∫ 0 1 sgn ( t − s ) f ( t ) d t (Kf)(s)=\frac{1}{2}\int_0^1\operatorname{sgn}(t-s)f(t)\,\mathrm{d}t ( K f ) ( s ) = 2 1 ∫ 0 1 sgn ( t − s ) f ( t ) d t . The evaluation and Itô isometries identify A = G K G ∗ A=GKG^* A = G K G ∗ .
On the complexification, the functions e i π ( 2 j + 1 ) t e^{i\pi(2j+1)t} e iπ ( 2 j + 1 ) t , j ∈ Z j\in\mathbb{Z} j ∈ Z , form an orthonormal basis, and direct integration gives, for α = π ( 2 j + 1 ) \alpha=\pi(2j+1) α = π ( 2 j + 1 ) ,
K ( e i α ⋅ ) ( s ) = e i α + 1 − 2 e i α s 2 i α = i α e i α s . K(e^{i\alpha\cdot})(s)=\frac{e^{i\alpha}+1-2e^{i\alpha s}}{2i\alpha}=\frac{i}{\alpha}e^{i\alpha s}. K ( e i α ⋅ ) ( s ) = 2 i α e i α + 1 − 2 e i α s = α i e i α s .
Consequently the full singular-value sequence, with multiplicities, is
κ n = 1 π ( 2 ⌈ n / 2 ⌉ − 1 ) ≍ n − 1 , ∑ k ≥ n κ k 2 ≍ n κ n 2 . \kappa_n=\frac{1}{\pi(2\lceil n/2\rceil-1)}\asymp n^{-1},\qquad\sum_{k\ge n}\kappa_k^2\asymp n\kappa_n^2. κ n = π ( 2 ⌈ n /2 ⌉ − 1 ) 1 ≍ n − 1 , k ≥ n ∑ κ k 2 ≍ n κ n 2 .
Thus κ 2 \kappa^2 κ 2 is regular at infinity, and Corollary 6.3 gives I ( A ) = I ( ( κ n R n ) n ) = I ( b ) \mathcal{I}(A)=\mathcal{I}((\kappa_nR_n)_n)=\mathcal{I}(b) I ( A ) = I (( κ n R n ) n ) = I ( b ) and its fixed-perturbation formula (118) . Lemma 6.2, applied to q q q , now gives I ( Q ) ⊆ I ( A ) \mathcal{I}(Q)\subseteq\mathcal{I}(A) I ( Q ) ⊆ I ( A ) and (120) , including finite rank. If rank Q ≤ 1 \operatorname{rank}Q\le1 rank Q ≤ 1 , the path lies in a fixed line and (117) vanishes.
Brownian scaling gives the same formulas for every fixed interval [ s , t ] ⊆ [ 0 , 1 ] [s,t]\subseteq[0,1] [ s , t ] ⊆ [ 0 , 1 ] , s < t s<t s < t , with the path recentered at s s s ; a common event may be chosen for any prescribed countable family of intervals and fixed compact perturbations.
For example, let q n ≍ [ n ( log ( e n ) ) α ] − 1 q_n\asymp[n(\log(en))^\alpha]^{-1} q n ≍ [ n ( log ( e n ) ) α ] − 1 with α > 1 \alpha>1 α > 1 . The integral test gives R n ≍ ( log ( e n ) ) 1 − α R_n\asymp(\log(en))^{1-\alpha} R n ≍ ( log ( e n ) ) 1 − α , hence R n / ( n q n ) ≍ log ( e n ) → ∞ R_n/(nq_n)\asymp\log(en)\to\infty R n / ( n q n ) ≍ log ( e n ) → ∞ and
I ( Q ) ⊊ I ( A ) almost surely . (121) \mathcal{I}(Q)\subsetneq\mathcal{I}(A)\quad\text{almost surely}. \tag*{(121)} I ( Q ) ⊊ I ( A ) almost surely . (121)
Ordered covariance envelopes: exact evaluation
Fix orthonormal input and output bases, zero-extended in ℓ 2 \ell^2 ℓ 2 . Here we compute the covariance profiles and the exact existence domain.
A table of finite nonnegative variances v = ( v i j ) v=(v_{ij}) v = ( v ij ) is doubly monotone if
v i + 1 , j ≤ v i j , v i , j + 1 ≤ v i j . (122) v_{i+1,j}\le v_{ij},\qquad v_{i,j+1}\le v_{ij}. \tag*{(122)} v i + 1 , j ≤ v ij , v i , j + 1 ≤ v ij . (122)
Let X = ( X i j ) X=(X_{ij}) X = ( X ij ) be a centered jointly Gaussian array. Assume that, for some 0 < c − ≤ c + < ∞ 0<c_-\le c_+<\infty 0 < c − ≤ c + < ∞ and every finitely supported real table u u u ,
c − ∑ i , j v i j u i j 2 ≤ E ( ∑ i , j u i j X i j ) 2 ≤ c + ∑ i , j v i j u i j 2 . (123) c_-\sum_{i,j}v_{ij}u_{ij}^2\le\mathbb{E}\left(\sum_{i,j}u_{ij}X_{ij}\right)^2\le c_+\sum_{i,j}v_{ij}u_{ij}^2. \tag*{(123)} c − i , j ∑ v ij u ij 2 ≤ E ( i , j ∑ u ij X ij ) 2 ≤ c + i , j ∑ v ij u ij 2 . (123)
The diagonal-covariance case c − = c + = 1 c_-=c_+=1 c − = c + = 1 is precisely X i j = v i j g i j X_{ij}=\sqrt{v_{ij}}g_{ij} X ij = v ij g ij with independent standard Gaussian g i j g_{ij} g ij . Define
A n = ∑ j ≥ n v n j , C n = ∑ i ≥ n v i n , h n 2 = A n + C n − v n n = ∑ min ( i , j ) = n v i j . (124) A_n=\sum_{j\ge n}v_{nj},\qquad C_n=\sum_{i\ge n}v_{in},\qquad h_n^2=A_n+C_n-v_{nn}=\sum_{\min(i,j)=n}v_{ij}. \tag*{(124)} A n = j ≥ n ∑ v nj , C n = i ≥ n ∑ v in , h n 2 = A n + C n − v nn = m i n ( i , j ) = n ∑ v ij . (124)
Initially X X X is only an array; a bounded realization has these entries. The total variance ∑ n h n 2 \sum_n h_n^2 ∑ n h n 2 may be infinite.
For the variational identities, set
α p , q ( T ) = inf ( P , Q ) ∈ P p , q ∥ L T ( P , Q ) ∥ , β p , q ( T ) = inf ( P , Q ) ∈ P p , q ∥ R T ( P , Q ) ∥ . \alpha_{p,q}(T)=\inf_{(P,Q)\in\mathcal{P}_{p,q}}\left\|L^T(P,Q)\right\|,\qquad\beta_{p,q}(T)=\inf_{(P,Q)\in\mathcal{P}_{p,q}}\left\|R^T(P,Q)\right\|. α p , q ( T ) = ( P , Q ) ∈ P p , q inf L T ( P , Q ) , β p , q ( T ) = ( P , Q ) ∈ P p , q inf R T ( P , Q ) .
Theorem D (Ordered covariance envelopes and optimal truncation). Assume (122) and (C). A bounded realization exists almost surely exactly when
h 1 2 = ∑ j ≥ 1 v 1 j + ∑ i ≥ 2 v i 1 < ∞ . (125) h_1^2=\sum_{j\ge1}v_{1j}+\sum_{i\ge2}v_{i1}<\infty. \tag*{(125)} h 1 2 = j ≥ 1 ∑ v 1 j + i ≥ 2 ∑ v i 1 < ∞. (125)
It is then unique and compact, and h h h is a nonincreasing null sequence. For every fixed compact B B B , there is an event of probability one on which, for all L > 2 L>2 L > 2 and λ > 1 \lambda>1 λ > 1 , finite positive constants C ω , L C_{\omega,L} C ω , L and c ω , λ c_{\omega,\lambda} c ω , λ satisfy
s m ( X ) ≤ C ω , L h ⌈ m / L ⌉ , s ⌈ n / λ ⌉ ( X ) ≥ c ω , λ h n ( m , n ≥ 1 ) , (126) s_m(X) \le C_{\omega,L}h_{\lceil m/L\rceil}, \qquad s_{\lceil n/\lambda\rceil}(X) \ge c_{\omega,\lambda}h_n \qquad(m,n\ge1), \tag*{(126)} s m ( X ) ≤ C ω , L h ⌈ m / L ⌉ , s ⌈ n / λ ⌉ ( X ) ≥ c ω , λ h n ( m , n ≥ 1 ) , (126)
I ( X ) = I ( h ) = I ( η ( X ) ) , (127) \mathcal{I}(X)=\mathcal{I}(h)=\mathcal{I}(\eta(X)), \tag*{(127)} I ( X ) = I ( h ) = I ( η ( X )) , (127)
I ( X + B ) = I ( h ) + I ( B ) = I ( max { h , s ( B ) } ) a.s. (128) \mathcal{I}(X+B)=\mathcal{I}(h)+\mathcal{I}(B)=\mathcal{I}(\max\{h,s(B)\}) \quad\text{a.s.} \tag*{(128)} I ( X + B ) = I ( h ) + I ( B ) = I ( max { h , s ( B )}) a.s. (128)
The lower comparison holds on the same event with X + B X+B X + B in place of X X X . The event conventions of Theorem A apply. All mean Grams are compact, and for ( P , Q ) ∈ P p , q (P,Q)\in\mathcal{P}_{p,q} ( P , Q ) ∈ P p , q ,
λ k ( Γ L X ( P , Q ) ) ≥ c − ∑ j > q v p + k , j , λ k ( Γ R X ( P , Q ) ) ≥ c − ∑ i > p v i , q + k ( k ≥ 1 ) . \lambda_k(\Gamma_L^X(P,Q))\ge c_{-}\sum_{j>q}v_{p+k,j}, \qquad\lambda_k(\Gamma_R^X(P,Q))\ge c_{-}\sum_{i>p}v_{i,q+k} \qquad(k\ge1). λ k ( Γ L X ( P , Q )) ≥ c − j > q ∑ v p + k , j , λ k ( Γ R X ( P , Q )) ≥ c − i > p ∑ v i , q + k ( k ≥ 1 ) .
Writing M n = max { A n , C n } M_n=\max\{A_n,C_n\} M n = max { A n , C n } , their optimal values obey
c − M n ≤ ρ n 2 ≤ η n 2 ≤ ϑ n 2 ≤ c + M n , c_{-}M_n\le\rho_n^2\le\eta_n^2\le\vartheta_n^2\le c_{+}M_n, c − M n ≤ ρ n 2 ≤ η n 2 ≤ ϑ n 2 ≤ c + M n ,
c − h n 2 ≤ α n − 1 , n − 1 ( X ) + β n , n − 1 ( X ) ≤ c + h n 2 . c_{-}h_n^2\le\alpha_{n-1,n-1}(X)+\beta_{n,n-1}(X)\le c_{+}h_n^2. c − h n 2 ≤ α n − 1 , n − 1 ( X ) + β n , n − 1 ( X ) ≤ c + h n 2 .
In diagonal covariance, leading projections attain both entire mean-Gram spectra simultaneously, and
ρ n 2 = η n 2 = ϑ n 2 = M n , h n 2 = α n − 1 , n − 1 ( X ) + β n , n − 1 ( X ) , η n ≤ h n ≤ 2 η n . \rho_n^2=\eta_n^2=\vartheta_n^2=M_n,\qquad h_n^2=\alpha_{n-1,n-1}(X)+\beta_{n,n-1}(X),\qquad\eta_n\le h_n\le\sqrt{2}\eta_n. ρ n 2 = η n 2 = ϑ n 2 = M n , h n 2 = α n − 1 , n − 1 ( X ) + β n , n − 1 ( X ) , η n ≤ h n ≤ 2 η n .
These formulas include infinite total variance. If (125) fails, almost surely there is no bounded realization. The shift assertions extend to independent compact random B B B without moments.
Covariance estimates and proof of Theorem D
Proof of Theorem D. Let X 0 X^0 X 0 be the independent Gaussian reference with variances v i j v_{ij} v ij , P k − 1 P_{k-1} P k − 1 the first k − 1 k-1 k − 1 coordinate projection, and Q k = I − P k − 1 Q_k=I-P_{k-1} Q k = I − P k − 1 . Condition (C) bounds the residual covariance between c − c_{-} c − and c + c_{+} c + times that of X 0 X^0 X 0 . Under h 1 < ∞ h_1<\infty h 1 < ∞ , double monotonicity makes h h h nonincreasing and gives
h k 2 ≤ ∑ j ≥ k v 1 j + ∑ i ≥ k v i 1 ⟶ 0. h_k^2\le\sum_{j\ge k}v_{1j}+\sum_{i\ge k}v_{i1}\longrightarrow0. h k 2 ≤ j ≥ k ∑ v 1 j + i ≥ k ∑ v i 1 ⟶ 0.
The corner Q k X Q k Q_kXQ_k Q k X Q k has trace envelopes L k = c + diag ( v i k 1 i ≥ k ) L_k=c_{+}\operatorname{diag}(v_{ik}1_{i\ge k}) L k = c + diag ( v ik 1 i ≥ k ) and R k = c + diag ( v k j 1 j ≥ k ) R_k=c_{+}\operatorname{diag}(v_{kj}1_{j\ge k}) R k = c + diag ( v k j 1 j ≥ k ) . Lemma 2.5 constructs the compact X X X , and
E ∥ Q k X Q k ∥ ≤ 2 2 c + h k ⟶ 0. (129) \mathbb{E}\lVert Q_kXQ_k\rVert\le2\sqrt{2c_{+}}h_k\longrightarrow0. \tag*{(129)} E ∥ Q k X Q k ∥ ≤ 2 2 c + h k ⟶ 0. (129)
The deletion identity
X − Q k X Q k = P k − 1 X + Q k X P k − 1 , rank ( X − Q k X Q k ) ≤ 2 k − 2 (130) X-Q_kXQ_k=P_{k-1}X+Q_kXP_{k-1},\qquad\operatorname{rank}(X-Q_kXQ_k)\le2k-2 \tag*{(130)} X − Q k X Q k = P k − 1 X + Q k X P k − 1 , rank ( X − Q k X Q k ) ≤ 2 k − 2 (130)
gives the rank of the finite approximation. Finite compressions converge in operator norm and all ideal comparison events are Borel.
For the lower covariance estimate, use the first n n n output directions and S n = c − diag ( v n j ) S_n=c_{-}\operatorname{diag}(v_{nj}) S n = c − diag ( v nj ) . The full covariance order (C) yields Cov ( P E n X ) ≥ I E n ⊗ S n \operatorname{Cov}(P_{E_n}X)\ge I_{E_n}\otimes S_n Cov ( P E n X ) ≥ I E n ⊗ S n ; transposition gives the column counterpart. Thus ρ n ( X ) 2 ≥ c − max ( A n , C n ) = c − M n \rho_n(X)^2\ge c_{-}\max(A_n,C_n)=c_{-}M_n ρ n ( X ) 2 ≥ c − max ( A n , C n ) = c − M n . Deleting the first n − 1 n-1 n − 1 input and output directions and using the above trace envelopes gives ϑ n ( X ) 2 ≤ c + M n \vartheta_n(X)^2\le c_{+}M_n ϑ n ( X ) 2 ≤ c + M n . By Theorem A ,
c − M n ≤ ρ n ( X ) 2 ≤ η n ( X ) 2 ≤ ϑ n ( X ) 2 ≤ c + M n , M n ≤ h n 2 ≤ 2 M n . c_{-}M_n\le\rho_n(X)^2\le\eta_n(X)^2\le\vartheta_n(X)^2\le c_{+}M_n,\qquad M_n\le h_n^2\le2M_n. c − M n ≤ ρ n ( X ) 2 ≤ η n ( X ) 2 ≤ ϑ n ( X ) 2 ≤ c + M n , M n ≤ h n 2 ≤ 2 M n .
These bounds prove the ideal and index formulas, including fixed perturbations and zero profiles. When the covariance is diagonal, the same leading projections are optimal on both sides, and the covariance witnesses attain the lower bounds; hence all three squared profiles equal M n M_n M n .
For any deterministic P , Q P,Q P , Q of ranks at most p , q p,q p , q , the space E p + k ∩ ker P E_{p+k}\cap\ker P E p + k ∩ ker P has dimension at least k k k . Each unit u u u in it satisfies
⟨ Γ L X ( P , Q ) u , u ⟩ ≥ tr ( ( I − Q ) S p + k ) ≥ c − ∑ j > q v p + k , j . \langle\Gamma_L^X(P,Q)u,u\rangle\ge\operatorname{tr}((I-Q)S_{p+k})\ge c_{-}\sum_{j>q}v_{p+k,j}. ⟨ Γ L X ( P , Q ) u , u ⟩ ≥ tr (( I − Q ) S p + k ) ≥ c − j > q ∑ v p + k , j .
The min–max principle proves the claimed spectrum of Γ L \Gamma_L Γ L ; transpose for Γ R \Gamma_R Γ R . Summing covariance tests gives the upper mean-Gram bounds. Leading coordinate deletions are optimal for X 0 X^0 X 0 , so c − h n 2 ≤ α n − 1 , n − 1 ( X ) + β n , n − 1 ( X ) ≤ c + h n 2 c_-h_n^2 \le\alpha_{n-1,n-1}(X)+\beta_{n,n-1}(X) \le c_+h_n^2 c − h n 2 ≤ α n − 1 , n − 1 ( X ) + β n , n − 1 ( X ) ≤ c + h n 2 , with equality at c − = c + = 1 c_-=c_+=1 c − = c + = 1 . When total variance is finite, taking traces additionally gives
E p , q ( X 0 ) = ∑ i > p , j > q v i j , h n 2 = E n − 1 , n − 1 ( X 0 ) − E n , n ( X 0 ) . \mathcal{E}_{p,q}(X^0)=\sum_{i>p,j>q}v_{ij},\qquad h_n^2=\mathcal{E}_{n-1,n-1}(X^0)-\mathcal{E}_{n,n}(X^0). E p , q ( X 0 ) = i > p , j > q ∑ v ij , h n 2 = E n − 1 , n − 1 ( X 0 ) − E n , n ( X 0 ) .
No difference of infinite traces is used.
If h 1 = ∞ h_1=\infty h 1 = ∞ , either the first row or column weight sum diverges. For the first row, covariance domination and Lemma 2.3 give
E exp ( − ∑ j ≤ N X 1 j 2 ) ≤ ∏ j ≤ N ( 1 + 2 c − v 1 j ) − 1 / 2 ⟶ 0. \mathbb{E}\exp\left(-\sum_{j\le N}X_{1j}^2\right)\le\prod_{j\le N}(1+2c_-v_{1j})^{-1/2}\longrightarrow0. E exp − j ≤ N ∑ X 1 j 2 ≤ j ≤ N ∏ ( 1 + 2 c − v 1 j ) − 1/2 ⟶ 0.
Bounded convergence forces infinite row norm almost surely; the column argument is identical. Finally, the explicit Gaussian estimates follow from the same covariance estimates: since
h n 2 ≥ ( k − n + 1 ) v k k ( k ≥ n ) , (131) h_n^2\ge(k-n+1)v_{kk}\qquad(k\ge n), \tag*{(131)} h n 2 ≥ ( k − n + 1 ) v k k ( k ≥ n ) , (131)
Gaussian concentration gives
P { ∥ Q k X Q k ∥ > c + ( 2 2 + 1 ) h n } ≤ e − ( k − n + 1 ) / 2 ( k ≥ n , h n > 0 ) . (132) \mathbb{P}\left\{\lVert Q_kXQ_k\rVert>\sqrt{c_+}(2\sqrt{2}+1)h_n\right\}\le e^{-(k-n+1)/2}\qquad(k\ge n,\ h_n>0). \tag*{(132)} P { ∥ Q k X Q k ∥ > c + ( 2 2 + 1 ) h n } ≤ e − ( k − n + 1 ) /2 ( k ≥ n , h n > 0 ) . (132)
The larger of the row and column witnesses has t n ( S ) ≥ c − h n 2 / 2 t_n(S)\ge c_-h_n^2/2 t n ( S ) ≥ c − h n 2 /2 . Proposition 2.8 yields
P { s j ( X + B ) < c − ( n − j ) 8 n h n } ≤ 2 − ⌊ ( n − j + 1 ) / 4 ⌋ ( 1 ≤ j < n ) . \mathbb{P}\left\{s_j(X+B)<\sqrt{\frac{c_-(n-j)}{8n}}h_n\right\}\le2^{-\lfloor(n-j+1)/4\rfloor}\qquad(1\le j<n). P { s j ( X + B ) < 8 n c − ( n − j ) h n } ≤ 2 − ⌊( n − j + 1 ) /4 ⌋ ( 1 ≤ j < n ) .
Theorem A converts these estimates into the asserted strict-dilation comparisons, including the finite initial indices. □
No endpoint L = 2 L=2 L = 2 or λ = 1 \lambda=1 λ = 1 is asserted. For fixed B B B , equality I ( X + B ) = I ( X ) \mathcal{I}(X+B)=\mathcal{I}(X) I ( X + B ) = I ( X ) holds exactly when B ∈ I ( h ) B\in\mathcal{I}(h) B ∈ I ( h ) , with strict inclusion otherwise. On that same event, for every two-sided ideal J \mathcal{J} J ,
X + B ∈ J ⟺ D h ∈ J and B ∈ J . (133) X+B\in\mathcal{J}\quad\Longleftrightarrow\quad D_h\in\mathcal{J}\text{ and }B\in\mathcal{J}. \tag*{(133)} X + B ∈ J ⟺ D h ∈ J and B ∈ J . (133)
Remark (The upper index dilation). For v i j = 1 { i + j ≤ 2 m + 1 } v_{ij}=\mathbf{1}_{\{i+j\le2m+1\}} v ij = 1 { i + j ≤ 2 m + 1 } , one has h n = 0 h_n=0 h n = 0 for n > m n>m n > m , while reversing the first 2 m 2m 2 m columns makes the nonzero block triangular with nonzero diagonal almost surely. Hence rank X = 2 m \operatorname{rank}X=2m rank X = 2 m , ruling out every upper dilation L < 2 L<2 L < 2 . This does not settle L = 2 L=2 L = 2 .
Corollary 7.1 (Ordered curvature models). Let Z = ( Z i j ) Z=(Z_{ij}) Z = ( Z ij ) be a centered countable field such that every finite coordinate vector has positive density c e − V ce^{-V} c e − V and α I ≤ ∇ 2 V ≤ β I \alpha I\le\nabla^2V\le\beta I α I ≤ ∇ 2 V ≤ β I , with constants independent of the coordinates. For a doubly monotone nonnegative table v v v , put T i j = v i j Z i j T_{ij}=\sqrt{v_{ij}}Z_{ij} T ij = v ij Z ij and let h h h be (7.2). A bounded realization exists almost surely exactly when h 1 < ∞ h_1<\infty h 1 < ∞ . It is then unique and compact, and for every fixed compact B B B ,
I ( T + B ) = I ( h ) + I ( B ) = I ( η ( T ) ) + I ( B ) almost surely . (134) \mathcal{I}(T+B)=\mathcal{I}(h)+\mathcal{I}(B)=\mathcal{I}(\eta(T))+\mathcal{I}(B)\quad\text{almost surely}. \tag*{(134)} I ( T + B ) = I ( h ) + I ( B ) = I ( η ( T )) + I ( B ) almost surely . (134) All strict-index comparisons (126) hold for T T T , with the lower bound also valid for T + B T+B T + B . No coordinate or row independence is assumed.
Proof. Use G i j = v i j g i j G_{ij}=\sqrt{v_{ij}}g_{ij} G ij = v ij g ij with independent standard Gaussian g i j g_{ij} g ij . Finite linear images satisfy (H) by Lemma 2.4 , with zero weights treated on the image support. If h 1 < ∞ h_1<\infty h 1 < ∞ , convex comparison on finite corners and the trace envelopes of G G G first construct the compact T T T with square-integrable norm, as in Lemma 2.5 . Arbitrary finite-rank tests are L 2 L^2 L 2 limits of coordinate truncations. Their Gaussian covariances increase to the limiting covariance; Lemma 2.4 passes the curvature bounds to these tests. Theorem D gives ρ n ( G ) 2 = η n ( G ) 2 = ϑ n ( G ) 2 = M n \rho_n(G)^2=\eta_n(G)^2=\vartheta_n(G)^2=M_n ρ n ( G ) 2 = η n ( G ) 2 = ϑ n ( G ) 2 = M n and M n ≤ h n 2 ≤ 2 M n M_n\le h_n^2\le2M_n M n ≤ h n 2 ≤ 2 M n ; Theorem A proves the assertions. If the first row sum diverges, Lemma 2.3 with empty observation gives
E exp ( − ∑ j ≤ N T 1 j 2 ) ≤ ∏ j ≤ N ( 1 + 2 v 1 j / β ) − 1 / 2 ⟶ 0. \mathbb{E}\exp\left(-\sum_{j\le N}T_{1j}^2\right)\le\prod_{j\le N}(1+2v_{1j}/\beta)^{-1/2}\longrightarrow0. E exp − j ≤ N ∑ T 1 j 2 ≤ j ≤ N ∏ ( 1 + 2 v 1 j / β ) − 1/2 ⟶ 0.
The row squared norm is infinite almost surely, precluding boundedness. The same argument applies to a divergent first column sum. □
Declaration of AI Use
The author used GPT-6.0 (OpenAI) interactively in developing and refining the mathematical arguments and the presentation of this paper. The author takes full responsibility for the correctness and mathematical content of the manuscript.
References
School of Mathematical Sciences, University of Science and Technology of China, Hefei 230026, China
Email address: zhaoguangqian@mail.ustc.edu.cn