betaFeedback

Reading v2View latest version

WorksMathematics

Gaussian Matrices and Exterior Forms: Deterministic Principal Ideals

Abstract

For an infinite Gaussian matrix with independent entries and doubly monotone variances, we identify its algebraic principal ideal by a deterministic shell profile. The exact boundedness condition also implies compactness. A lower singular-value estimate uniform in the mean gives an exact sum law under each fixed compact perturbation. Conditional products and spatial compression yield the corresponding classification for Gaussian evaluations of Hilbert two-forms. For Wiener area of covariance rank at least two, the generator is the covariance tail divided by its index. The area and covariance generate the same ideal precisely under regularity at infinity.

Introduction

We identify the principal ideals of Gaussian matrices with doubly monotone variances by their diagonal shell masses. Each identity below determines all algebraic two-sided ideal membership tests on one probability-one event for fixed input data and a fixed perturbation. Conditional iteration gives the ideals of Gaussian exterior evaluations; Wiener area follows by inserting its time spectrum. Write I(T)I(T) for the two-sided ideal generated by a compact operator TT, with no norm closure, and I(d)=I(diag⁡(dn))I(d)=I(\operatorname{diag}(d_n)) for a decreasing nonnegative null sequence dd. Equality of principal ideals means mutual domination of singular-value sequences up to constants and finite index dilations.

Let Xij=vijgijX_{ij}=\sqrt{v_{ij}}g_{ij}, where the gijg_{ij} are independent standard real Gaussian variables and vi+1,j≤vijv_{i+1,j}\le v_{ij}, vi,j+1≤vijv_{i,j+1}\le v_{ij}. A bounded realization exists almost surely exactly when the first row and first column of vv are summable. It is then compact, and Theorem A gives, for every fixed deterministic compact DD,

hn2=∑j≥nvnj+∑i≥nvin,I(X+D)=I(h)+I(D)almost surely.(1)h_n^2=\sum_{j\ge n}v_{nj}+\sum_{i\ge n}v_{in},\qquad I(X+D)=I(h)+I(D)\quad\text{almost surely}. \tag*{(1)}

Here hn2h_n^2 is the variance in the nnth diagonal shell. The sum is generated by max⁡{h,s(D)}\max\{h,s(D)\}: a prescribed perturbation cannot cancel either ideal contribution. The classification includes nonseparable variances and compact realizations outside the Hilbert--Schmidt class.

For separable variances vij=ai2bj2v_{ij}=a_i^2b_j^2, the tail sums of h2h^2 are products of those of a2a^2 and b2b^2. Conditioning allows this rule to be iterated. Let Q≥0Q\ge0 be trace class, let KK be skew-adjoint and Hilbert--Schmidt, and write qn=sn(Q)q_n=s_n(Q), κn=sn(K)\kappa_n=s_n(K), with zeros appended in finite rank. The Gaussian evaluation A=GKG∗A=GKG^* is defined in L2(Ω;S2)L^2(\Omega;S_2), where E(Gf⊗Gg)=⟨f,g⟩Q\mathbb{E}(Gf\otimes Gg)=\langle f,g\rangle_Q; no samplewise bounded map GG is required. For Rn=∑j≥nqjR_n=\sum_{j\ge n}q_j, Tn=∑j≥nκj2T_n=\sum_{j\ge n}\kappa_j^2, and rank⁡Q≥2\operatorname{rank}Q\ge2, Theorem 3.1 gives

rn2=Rn2Tn−Rn+12Tn+1,I(A+D)=I(r)+I(D)almost surely.(2)r_n^2=R_n^2T_n-R_{n+1}^2T_{n+1},\qquad I(A+D)=I(r)+I(D)\quad\text{almost surely}. \tag*{(2)}

The square tail Rn2TnR_n^2T_n records the factors Q1/2Q^{1/2}, KK, Q1/2Q^{1/2}. The canonical form of KK gives the upper ideal inclusion; an independent odd--even covariance block gives the reverse. Rank at most one gives zero alternating evaluation. A deterministic tail estimate then shows that I(r)=I((κnRn)n)I(r)=I((\kappa_nR_n)_n) for every covariance of rank at least two precisely when Tn≤Cnκn2T_n\le Cn\kappa_n^2 for some finite CC, the regularity-at-infinity condition for κ2\kappa^2.

2020 Mathematics Subject Classification. Primary 47B10, 60G15; Secondary 60B20, 60H05, 15A75.

Key words and phrases. Gaussian matrices, principal ideals, monotone variances, Gaussian exterior forms, regularity at infinity, Lévy area.

The Wiener time operator has κn≍n−1\kappa_n \asymp n^{-1} and satisfies this condition. Hence

I(A)=I((Rn/n)n≥1)almost surely,rank⁡Q≥2.(3)\mathcal{I}(A)=\mathcal{I}\left((R_n/n)_{n\ge1}\right)\quad\text{almost surely},\qquad\operatorname{rank}Q\ge2. \tag*{(3)}

Moreover, I(Q)⊆I(A)\mathcal{I}(Q)\subseteq\mathcal{I}(A), with equality exactly when Rn≤CnqnR_n\le Cnq_n. For example, if qn=1/[n(log⁡(en))3]q_n=1/[n(\log(en))^3], then Rn/(nqn)∼log⁡(en)→∞R_n/(nq_n)\sim\log(en)\to\infty, so I(Q)⊊I(A)\mathcal{I}(Q)\subsetneq\mathcal{I}(A) almost surely.

Calkin’s correspondence [2] describes operator ideals by singular-value sequences; principal ideals and finite dilations are treated in [8], and Schatten classes in [12]. The tail-average criterion used here belongs to the arithmetic-mean theory and regularity at infinity of Kaftal and Weiss [7]. Our use of that criterion concerns the generators produced by Gaussian evaluation.

Bandeira and van Handel [1] prove sharp nonasymptotic operator-norm estimates for matrices with independent entries. Latała, van Handel and Youssef [9] give dimension-free Schatten-norm, moment and tail comparisons for 2≤p≤∞2\le p\le\infty, including a characterization of bounded infinite Gaussian matrices and nonsymmetric extensions. Their variance profiles are more general; our boundedness criterion specializes that theory. We identify the whole singular-value sequence up to finite dilation and determine its principal ideal under fixed compact perturbations. Rudelson and Zeitouni [11] obtain lower singular-value estimates under broad-connectedness assumptions. Here monotonicity controls the covariance spectrum after projection and gives lower estimates uniform in the deterministic mean.

For stochastic area, Ledoux, Lyons and Qian [10] develop Wiener rough paths in Banach spaces under tensor-norm exactness assumptions. Ferreiro-Castilla and Utzet [5] construct Gaussian area through double Wiener–Itô integrals and study its characteristic function. Dela Cruz and Oberhauser [4] study singular-value distributions of finite-dimensional Brownian-area matrices and their high-dimensional asymptotics. We fix the spatial covariance and time interval and determine the resulting compact operator’s principal ideal.

Gaussian exterior dilation [15] compares a random regularized determinant with its expectation through one random dilation, uniformly over nonnegative parameters, and proves a universal weak-L1L^1 bound with optimal tail exponent. Its Wiener-area application includes the Schatten membership criteria and the covariance-entropy condition at trace class. Here we determine principal-ideal generators, their sums under fixed compact perturbations, and ideal-preservation conditions.

Proof and organization

Gaussian comparison and concentration control tail-corner norms, giving upper singular-value bounds after deleting finitely many rows and columns. For the lower bounds, covariance eigenvalue interlacing and a Gaussian Laplace estimate control the distances of conditioned rows from the spans of the other rows, uniformly in the means. The inverse Gram identity converts these distances into spectral bounds.

Section 22 contains the matrix theorem, its proof, and the Gaussian-product consequences. Sections 33 and 44 treat exterior evaluation and Wiener area. Covariance comparison is proved in Appendix A.

Gaussian matrices

All Hilbert spaces and Gaussian variables are real. We use H=ℓ2(N)H=\ell^2(\mathbb{N}) with its standard basis (en)(e_n), and write B(H)\mathcal{B}(H) and K(H)\mathcal{K}(H) for the bounded and compact operators, with the operator norm. The singular values of a compact operator are listed in nonincreasing order, with multiplicities and zero completion in finite rank. They satisfy

sn(T)=inf⁡{∥T−R∥:rank⁡R<n}.(4)s_n(T)=\inf\{\lVert T-R\rVert:\operatorname{rank}R<n\}. \tag*{(4)}

Operators between different separable Hilbert spaces are compared after embedding the spaces in HH and extending by zero. This and complexification preserve the nonzero singular values and their multiplicities. The Hilbert–Schmidt class is S2={T:∑nsn(T)2<∞}\mathcal{S}_2=\{T:\sum_n s_n(T)^2<\infty\}, with norm ∥T∥HS=∥T∥2=(∑nsn(T)2)1/2\lVert T\rVert_{\mathrm{HS}}=\lVert T\rVert_2=(\sum_n s_n(T)^2)^{1/2}. We use (x⊗y)z=x⟨y,z⟩(x\otimes y)z=x\langle y,z\rangle and the standard inequalities [12]

sn(ATB)≤∥A∥∥B∥sn(T),sj+k−1(S+T)≤sj(S)+sk(T).(5)s_n(ATB)\le\lVert A\rVert\lVert B\rVert s_n(T),\qquad s_{j+k-1}(S+T)\le s_j(S)+s_k(T). \tag*{(5)}

In particular, ∣sn(S)−sn(T)∣≤∥S−T∥|s_n(S)-s_n(T)| \le\lVert S-T\rVert.

For a nonnegative nonincreasing null sequence aa, write Da=diag⁡(an)D_a=\operatorname{diag}(a_n) and I(a)=I(Da)\mathcal{I}(a)=\mathcal{I}(D_a), where

I(T)={∑l=1rAlTBl:r<∞, Al,Bl∈B(H)}.\mathcal{I}(T)=\left\{\sum_{l=1}^{r} A_lTB_l:r<\infty,\ A_l,B_l\in\mathcal{B}(H)\right\}.

No closure is taken. All nonzero finite-rank operators generate the finite-rank ideal. Throughout, a two-sided ideal is closed under finite sums and multiplication on either side by bounded operators, but need not be topologically closed.

Lemma 2.1 (Principal ideals and finite dilations). For nonnegative nonincreasing null sequences a,ba,b,

I(a)⊆I(b)⟺an≤Cb⌈n/M⌉ (n≥1)for some C<∞, M∈N.(6)\mathcal{I}(a)\subseteq\mathcal{I}(b)\quad\Longleftrightarrow\quad a_n\le Cb_{\lceil n/M\rceil}\ (n\ge1)\quad\text{for some }C<\infty,\ M\in\mathbb{N}. \tag*{(6)}

Moreover, I(T)=I(s(T))\mathcal{I}(T)=\mathcal{I}(s(T)) for every compact TT, and

I(a)+I(b)=I(max⁡{a,b}),(7)\mathcal{I}(a)+\mathcal{I}(b)=\mathcal{I}(\max\{a,b\}), \tag*{(7)}

where the maximum is coordinatewise.

Proof. These are the principal-ideal consequences of Calkin’s correspondence [2, 8]. For completeness, a sum of MM bounded left–right multiples of DbD_b has sM(n−1)+1≤Cbns_{M(n-1)+1}\le Cb_n by (5). Conversely, the stated domination expresses DaD_a as the sum of MM coordinate copies of DbD_b multiplied by bounded diagonal quotients; the quotients are zero where bn=0b_n=0. A singular-value decomposition gives the assertion for TT, using partial isometries on its positive singular subspaces. Finally, with PP the coordinate projection onto {n:an≥bn}\{n:a_n\ge b_n\}, one has Dmax⁡(a,b)=PDa+(I−P)DbD_{\max(a,b)}=PD_a+(I-P)D_b, proving (7). □

A table of finite nonnegative variances v=(vij)v=(v_{ij}) is doubly monotone if

vi+1,j≤vij,vi,j+1≤vij.(8)v_{i+1,j}\le v_{ij},\qquad v_{i,j+1}\le v_{ij}. \tag*{(8)}

Let Xij=vijgijX_{ij}=\sqrt{v_{ij}}g_{ij} for independent standard Gaussian gijg_{ij}, and define

An=∑j≥nvnj,Cn=∑i≥nvin,hn2=An+Cn−vnn=∑min⁡(i,j)=nvij.(9)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*{(9)}

The total variance ∑nhn2\sum_n h_n^2 may be infinite. Initially XX is only an array; a bounded realization is a bounded operator with these matrix entries.

Theorem A (Spectral classification by the shell profile). Suppose that vv is doubly monotone. The matrix XX has a bounded realization almost surely if and only if

h12=∑j≥1v1j+∑i≥2vi1<∞.(10)h_1^2=\sum_{j\ge1}v_{1j}+\sum_{i\ge2}v_{i1}<\infty. \tag*{(10)}

Under this condition hh is a nonincreasing null sequence, and the realization is unique and compact almost surely. Moreover:

  1. For each fixed vv satisfying (10), one event of probability one supports, for every L>2L>2 and λ>1\lambda>1, finite positive constants Cω,LC_{\omega,L} and cω,λc_{\omega,\lambda} such that

sm(X)≤Cω,Lh⌈m/L⌉,s⌈n/λ⌉(X)≥cω,λhn(m,n≥1).(11)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*{(11)}

In particular,

I(X)=I(h).(12)\mathcal{I}(X)=\mathcal{I}(h). \tag*{(12)}
  1. For every fixed deterministic compact operator BB, almost surely

I(X+B)=I(h)+I(B)=I(max⁡{h,s(B)}).(13)\mathcal{I}(X+B)=\mathcal{I}(h)+\mathcal{I}(B)=\mathcal{I}(\max\{h,s(B)\}). \tag*{(13)}

On an event that may depend on BB, the lower comparison in (11) also holds for X+BX+B, simultaneously for all λ>1\lambda>1.

If (10) fails, almost surely no bounded operator has the specified matrix entries.

An event may depend on the fixed variance table and perturbation; a prescribed countable family of perturbations can share one event. Uniformity over all compact perturbations is generally impossible, as the sample-dependent choice B=−XB=-X shows. The case h=0h=0 is trivial, and no assertion is made at L=2L=2 or λ=1\lambda=1.

For each fixed compact BB, almost surely I(X+B)=I(X)\mathcal{I}(X+B)=\mathcal{I}(X) if and only if B∈I(h)B\in\mathcal{I}(h); otherwise I(X)⊂I(X+B)\mathcal{I}(X)\subset\mathcal{I}(X+B). On this same event, simultaneously for every two-sided ideal J\mathcal{J},

X+B∈J⟺Dh∈J and B∈J.(14)X+B\in\mathcal{J}\quad\Longleftrightarrow\quad D_h\in\mathcal{J}\ \text{and}\ B\in\mathcal{J}. \tag*{(14)}

Remark (A nonseparable example). For vij=(i+j)−γv_{ij}=(i+j)^{-\gamma}, γ>1\gamma>1,

hn2=(2n)−γ+2∑k>2nk−γ∼22−γγ−1n1−γ.h_n^2=(2n)^{-\gamma}+2\sum_{k>2n}k^{-\gamma}\sim\frac{2^{2-\gamma}}{\gamma-1}n^{1-\gamma}.

Thus

sn(X)≲ωn−(γ−1)/2.(15)s_n(X)\lesssim_{\omega} n^{-(\gamma-1)/2}. \tag*{(15)}

The operator is compact but not Hilbert–Schmidt when 1<γ≤21<\gamma\le2; for γ≤1\gamma\le1 it has no bounded realization.

Remark (The upper index dilation). For vij=1{i+j≤2m+1}v_{ij}=\mathbf{1}_{\{i+j\le2m+1\}}, one has hn=0h_n=0 for n>mn>m, while reversing the first 2m2m columns makes the nonzero block triangular with nonzero diagonal almost surely. Hence rank⁡X=2m\operatorname{rank}X=2m, ruling out every upper dilation L<2L<2. This does not settle L=2L=2.

Proof of the matrix theorem

Construction and upper bound. We use Gaussian comparison for ordered increment variances and the concentration inequality

P{f−Ef>t}≤e−t2/(2σ2)(16)\mathbb{P}\{f-\mathbb{E}f>t\}\le e^{-t^2/(2\sigma^2)} \tag*{(16)}

for a σ\sigma-Lipschitz function of a finite standard Gaussian vector; see [14 Theorems 6.8 and 3.25].

Lemma 2.2. For a finite rectangular matrix Z=(wijgij)Z=(\sqrt{w_{ij}}g_{ij}), where the gijg_{ij} are independent standard Gaussian variables, put ri=max⁡jwijr_i=\max_j w_{ij} and cj=max⁡iwijc_j=\max_i w_{ij}. Then

E∥Z∥≤2(∑iri+∑jcj).(17)\mathbb{E}\lVert Z\rVert\le\sqrt{2}\left(\sqrt{\sum_i r_i}+\sqrt{\sum_j c_j}\right). \tag*{(17)}

Proof. For unit vectors s,ts,t, the process Us,t=⟨Zt,s⟩U_{s,t}=\langle Zt,s\rangle satisfies

E∣Us,t−Us′,t′∣2≤2∑iri(si−si′)2+2∑jcj(tj−tj′)2.\mathbb{E}\lvert U_{s,t}-U_{s',t'}\rvert^2\le2\sum_i r_i(s_i-s_i')^2+2\sum_j c_j(t_j-t_j')^2.

Compare it with 2∑iriξisi+2∑jcjηjtj\sqrt{2}\sum_i\sqrt{r_i}\xi_i s_i+\sqrt{2}\sum_j\sqrt{c_j}\eta_jt_j, with independent standard Gaussian ξ,η\xi,\eta. Its supremum is a sum of two Euclidean norms; Jensen’s inequality proves the bound. Comparison on the spheres follows by finite dense approximations. □

Let PnP_n project onto the first nn coordinates, P0=0P_0=0, and Qk=I−Pk−1Q_k=I-P_{k-1}. Assume (10). For the leading finite sections X(N)X^{(N)}, Lemma 2.2 bounds their increasing norms by E∥X(N)∥≤22h1\mathbb{E}\lVert X^{(N)}\rVert\le2\sqrt{2}h_1. Hence sup⁡N∥X(N)∥<∞\sup_N\lVert X^{(N)}\rVert<\infty almost surely. The associated bilinear form on finitely supported vectors is bounded and extends uniquely, giving a bounded operator XX with the prescribed entries. Increasing the finite corners and using monotone convergence yields

E∥QkXQk∥≤2(Ck+Ak)≤22hk.(18)\mathbb{E}\lVert Q_kXQ_k\rVert\le\sqrt{2}\left(\sqrt{C_k}+\sqrt{A_k}\right)\le2\sqrt{2}h_k. \tag*{(18)}

Double monotonicity makes hh nonincreasing, and

hn2≤∑j≥nv1j+∑i>nvi1⟶0.h_n^2\le\sum_{j\ge n}v_{1j}+\sum_{i>n}v_{i1}\longrightarrow0.

The corner norms decrease with kk, so (18) and Fatou’s lemma imply ∥QkXQk∥→0\lVert Q_kXQ_k\rVert\to0 almost surely. Since

X−QkXQk=Pk−1X+QkXPk−1,rank⁡(X−QkXQk)≤2k−2,(19)X-Q_kXQ_k=P_{k-1}X+Q_kXP_{k-1},\qquad\operatorname{rank}(X-Q_kXQ_k)\le2k-2, \tag*{(19)}

XX is compact. Consequently X(N)=PNXPN→XX^{(N)}=P_NXP_N\to X in norm. The norm-Cauchy event of these finite sections is Borel; taking their limit there and zero elsewhere defines a Borel K(H)\mathcal{K}(H)-valued random variable. The comparison events in (6) are Borel as well, because singular values are continuous and CC, MM may be restricted to integers.

If (10) fails, the first row or column has an infinite variance sum. For wj≥0w_j\ge0 with ∑jwj=∞\sum_j w_j=\infty,

Ee−∑jwjgj2=∏j(1+2wj)−1/2=0,\mathbb{E}e^{-\sum_j w_jg_j^2}=\prod_j(1+2w_j)^{-1/2}=0,

since ∑jlog⁡(1+2wj)=∞\sum_j\log(1+2w_j)=\infty. Thus that row or column is almost surely not in ℓ2\ell^2, excluding a bounded realization.

A finite corner starting at kk is vkk\sqrt{v_{kk}}-Lipschitz in its Gaussian coordinates. For k≥nk\ge n,

hn2≥∑j=nkvnj≥(k−n+1)vkk.(20)h_n^2\ge\sum_{j=n}^{k}v_{nj}\ge(k-n+1)v_{kk}. \tag*{(20)}

Concentration, (18), and increasing finite corners therefore give, with K0=22+1K_0=2\sqrt{2}+1,

P{∥QkXQk∥>K0hn}≤e−(k−n+1)/2(k≥n, hn>0).(21)\mathbb{P}\{\lVert Q_kXQ_k\rVert>K_0h_n\}\le e^{-(k-n+1)/2}\qquad(k\ge n,\ h_n>0). \tag*{(21)}

A zero vkkv_{kk} makes the corner zero. Fix L>2L>2 and put nm=⌈m/L⌉n_m=\lceil m/L\rceil, km=⌊(m+1)/2⌋k_m=\lfloor(m+1)/2\rfloor. Then sm(X)≤∥QkmXQkm∥s_m(X)\le\lVert Q_{k_m}XQ_{k_m}\rVert by (19), and km−nm≥(1/2−1/L)m−1k_m-n_m\ge(1/2-1/L)m-1 for large mm. Borel–Cantelli gives the eventual upper comparison. If hnm=0h_{n_m}=0, then rank⁡X≤2nm−2<m\operatorname{rank}X\le2n_m-2<m, so sm(X)=0s_m(X)=0; the remaining initial ratios are absorbed in the constant. Intersect over rational L>2L>2 and use h⌈m/r⌉≤h⌈m/L⌉h_{\lceil m/r\rceil}\le h_{\lceil m/L\rceil} for 2<r<L2<r<L to obtain all real L>2L>2 on one event.

Conditional distances and fixed perturbations. For a Hilbert-space Gaussian vector with mean mm, its covariance is C=E[(ξ−m)⊗(ξ−m)]C=\mathbb{E}[(\xi-m)\otimes(\xi-m)], a positive trace-class operator. The order C≥RC\ge R is the usual quadratic-form order.

Lemma 2.3 (A Gaussian Laplace bound). For such a vector and t>0t>0,

Ee−t∥ξ∥2/2=det⁡(I+tC)−1/2exp⁡{−t2⟨m,(I+tC)−1m⟩}.(22)\mathbb{E}e^{-t\lVert\xi\rVert^2/2}=\det(I+tC)^{-1/2}\exp\left\{-\frac{t}{2}\langle m,(I+tC)^{-1}m\rangle\right\}. \tag*{(22)}

In particular, if 0≤R≤C0\le R\le C, then

Ee−t∥ξ∥2/2≤det⁡(I+tR)−1/2.(23)\mathbb{E}e^{-t\lVert\xi\rVert^2/2}\le\det(I+tR)^{-1/2}. \tag*{(23)}

Here det⁡(I+tC)=∏k(1+tλk(C))\det(I+tC)=\prod_k(1+t\lambda_k(C)) is the Fredholm determinant.

Proof. Diagonalize CC and multiply the one-dimensional Gaussian integrals; bounded convergence handles the infinite product, including deterministic coordinates in ker⁡C\ker C. The second claim follows from the min–max inequality λk(C)≥λk(R)\lambda_k(C)\ge\lambda_k(R) and the mean factor being at most one. □

Lemma 2.4 (Gaussian distances and covariance eigenvalues). Let ξ1,…,ξm\xi_1,\ldots,\xi_m be jointly Gaussian Hilbert-space vectors with arbitrary deterministic means and joint covariance C≥Im⊗RC\ge I_m\otimes R, where R≥0R\ge0 is trace class. Write (rk)(r_k) for its decreasing eigenvalues, completed by zeros, and set di=dist⁡(ξi,span⁡{ξl:l≠i})d_i=\operatorname{dist}(\xi_i,\operatorname{span}\{\xi_l:l\ne i\}). For every t,a>0t,a>0, almost surely

E[e−tdi2/2∣(ξl)l≠i]≤∏k≥m(1+trk)−1/2,(24)\mathbb{E}[e^{-td_i^2/2}\mid(\xi_l)_{l\ne i}]\le\prod_{k\ge m}(1+tr_k)^{-1/2}, \tag*{(24)}
P{di2<a∣(ξl)l≠i}≤inf⁡t>0eta/2∏k≥m(1+trk)−1/2.(25)\mathbb{P}\{d_i^2<a\mid(\xi_l)_{l\ne i}\}\le\inf_{t>0}e^{ta/2}\prod_{k\ge m}(1+tr_k)^{-1/2}. \tag*{(25)}

Proof. For any orthogonal projection PP of codimension at most m−1m-1, the min–max principle gives λk(PRP)≥rk+m−1\lambda_k(PRP)\ge r_{k+m-1}: when the right side is positive, intersect the first k+m−1k+m-1 eigendirections with ran⁡P\operatorname{ran}P; otherwise the inequality is immediate. Thus det⁡(I+tPRP)≥∏k≥m(1+trk)\det(I+tPRP)\ge\prod_{k\ge m}(1+tr_k).

Realize the joint law as ξi=μi+ηi+ζi\xi_i=\mu_i+\eta_i+\zeta_i, where μi=Eξi\mu_i=\mathbb{E}\xi_i, the ηi\eta_i are independent centered Gaussian vectors with covariance RR, and the independent centered Gaussian vector ζ\zeta has covariance C−Im⊗RC-I_m\otimes R. Now let PP project off the other rows’ span; it has codimension at most m−1m-1 and is measurable by Gram–Schmidt with zero residuals discarded. Since ηi\eta_i is independent of the other rows and ζi\zeta_i jointly, conditioning on them makes PξiP\xi_i Gaussian with a fixed mean and covariance PRPPRP. The preceding determinant bound, Lemma 2.3, and the tower property give the first inequality. Markov gives the second, first with each positive rational tt, then with the infimum over all t>0t>0 by continuity. This includes zero and finite-rank RR.

Lemma 2.5 (The inverse Gram identity). Let Y:H→RmY:H\to\mathbb{R}^m have rows ξ1,…,ξm\xi_1,\ldots,\xi_m. If the preceding row distances satisfy di2≥A/2>0d_i^2\ge A/2>0, then

sj(Y)2≥A(m−j+1)2m(1≤j≤m).(26)s_j(Y)^2\ge\frac{A(m-j+1)}{2m}\qquad(1\le j\le m). \tag*{(26)}

Proof. The rows are linearly independent. The Schur complement formula gives ((YY∗)−1)ii=di−2\left((YY^*)^{-1}\right)_{ii}=d_i^{-2}. Taking traces gives the negative second moment identity [13 Lemma A.4] and the bound

∑r=1msr(Y)−2=∑i=1mdi−2≤2m/A.\sum_{r=1}^{m}s_r(Y)^{-2}=\sum_{i=1}^{m}d_i^{-2}\le2m/A.

The last m−j+1m-j+1 terms are at least sj(Y)−2s_j(Y)^{-2}.

Proposition 2.6 (A lower tail bound uniform in the mean). Assume (2.5) and (2.7). For every fixed compact BB and integers 1≤j<n1\le j<n,

P{sj(X+B)<n−j8n hn}≤ne−(n−j)/8.(27)\mathbb{P}\left\{s_j(X+B)<\sqrt{\frac{n-j}{8n}}\,h_n\right\}\le ne^{-(n-j)/8}. \tag*{(27)}

Proof. If q=vnn=0q=v_{nn}=0, then hn=0h_n=0 and the event is empty. Otherwise set

Rn=diag⁡(vnk)k≥1.(28)R_n=\operatorname{diag}(v_{nk})_{k\ge1}. \tag*{(28)}

For j≤m<nj\le m<n, the first mm rows of X+BX+B are independent Hilbert-space Gaussian vectors with fixed means and covariances diag⁡(vik)≥Rn\operatorname{diag}(v_{ik})\ge R_n, so their joint covariance dominates Im⊗RnI_m\otimes R_n. Their coordinate series converge in L2(Ω;ℓ2)L^2(\Omega;\ell^2) since ∑kvik≤∑kv1k<∞\sum_k v_{ik}\le\sum_k v_{1k}<\infty. The eigenvalues of RnR_n are vnkv_{nk} in their given order. Since vnk≥qv_{nk}\ge q for m≤k<nm\le k<n and vnk≤qv_{nk}\le q for k≥nk\ge n,

∑k≥mlog⁡(1+vnk/q)≥n−m2+An2q.\sum_{k\ge m}\log(1+v_{nk}/q)\ge\frac{n-m}{2}+\frac{A_n}{2q}.

using log⁡(1+x)≥12min⁡{x,1}\log(1+x)\ge\frac12\min\{x,1\}. Lemma 2.4 with t=1/qt=1/q and a=An/2a=A_n/2 gives

P{di2<An/2∣other rows}≤e−(n−m)/4.\mathbb{P}\{d_i^2<A_n/2\mid\text{other rows}\}\le e^{-(n-m)/4}.

The union bound, Lemma 2.5, and sj(Pm(X+B))≤sj(X+B)s_j(P_m(X+B))\le s_j(X+B) imply

P{sj(X+B)2<(m−j+1)An2m}≤me−(n−m)/4.\mathbb{P}\left\{s_j(X+B)^2<\frac{(m-j+1)A_n}{2m}\right\}\le me^{-(n-m)/4}.

Transposition replaces AnA_n by CnC_n. Choose the larger one; it is at least hn2/2h_n^2/2. Taking m=⌈(n+j)/2⌉m=\lceil(n+j)/2\rceil and using n−m≥(n−j)/2n-m\ge(n-j)/2, m−j+1≥(n−j)/2m-j+1\ge(n-j)/2, m≤nm\le n proves (27).

Completion of the proof of Theorem A. Fix BB and λ>1\lambda>1, and put δ=1−1/λ\delta=1-1/\lambda. For n≥2/δn\ge2/\delta, Proposition 2.6 gives

P{s⌈n/λ⌉(X+B)<δ hn/4}≤ne−δn/16.\mathbb{P}\{s_{\lceil n/\lambda\rceil}(X+B)<\sqrt{\delta}\,h_n/4\}\le ne^{-\delta n/16}.

Borel–Cantelli proves the eventual lower comparison. When hn>0h_n>0, every variance in the leading n×nn\times n block is positive. Its determinant after adding the fixed block of BB is a nonzero polynomial in independent continuous variables, and is nonzero almost surely. Thus sn(X+B)>0s_n(X+B)>0 whenever hn>0h_n>0, simultaneously for all nn, so the finite initial ratios can be absorbed in a positive constant. Intersect over rational λ>1\lambda>1; for real λ>1\lambda>1 choose rational 1<r<λ1<r<\lambda and use s⌈n/λ⌉≥s⌈n/r⌉s_{\lceil n/\lambda\rceil}\ge s_{\lceil n/r\rceil}.

For B=0B=0, combine the upper and lower events. Taking L=4L=4 and λ=2\lambda=2 in (2.8) proves I(X)=I(h)\mathcal{I}(X)=\mathcal{I}(h) by (2.3). For fixed BB, the lower bound gives I(h)⊂I(X+B)\mathcal{I}(h)\subset\mathcal{I}(X+B). On the same event X∈I(h)X\in\mathcal{I}(h), so B=(X+B)−X∈I(X+B)B=(X+B)-X\in\mathcal{I}(X+B). This proves one inclusion in (2.10); the other follows from X+B∈I(h)+I(B)X+B\in\mathcal{I}(h)+\mathcal{I}(B). (7) completes the proof.

Gaussian products

The separable case of Theorem A applies conditionally to independent Gaussian products. Their spectral profiles compose by multiplication of square tails.

For nonnegative nonincreasing u,v∈ℓ2u,v\in\ell^2, put Fu(n)=∑k≥nuk2F_u(n)=\sum_{k\ge n}u_k^2 and define

(u⊙v)n2=Fu(n)Fv(n)−Fu(n+1)Fv(n+1)=un2Fv(n)+vn2Fu(n+1).(29)(u\odot v)_n^2=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*{(29)}

This is the squared energy of u⊗vu\otimes v in the layer {(i,j):min⁡(i,j)=n}\{(i,j):\min(i,j)=n\}. Expanding it as un2vn2+un2Fv(n+1)+vn2Fu(n+1)u_n^2v_n^2+u_n^2F_v(n+1)+v_n^2F_u(n+1) proves that u⊙vu\odot v is nonnegative and nonincreasing. Summation gives

Fu⊙v(n)=Fu(n)Fv(n),∥u⊙v∥2=∥u∥2∥v∥2.(30)F_{u\odot v}(n)=F_u(n)F_v(n),\qquad\lVert u\odot v\rVert_2=\lVert u\rVert_2\lVert v\rVert_2. \tag*{(30)}

Since square tails determine the sequence, ⊙\odot is commutative and associative.

The operation also respects equality of principal ideals. Put (Δdu)n=u⌈n/d⌉(\Delta_d u)_n=u_{\lceil n/d\rceil} for an integer d≥1d\ge1. For k=⌈n/d⌉k=\lceil n/d\rceil, one has FΔdu(n)≤dFu(k)F_{\Delta_d u}(n)\le dF_u(k), and similarly for vv. Hence

(Δdu⊙Δdv)n2≤d(uk2Fv(k)+vk2Fu(k))≤2d(u⊙v)k2,(\Delta_d u\odot\Delta_d v)_n^2\le d\left(u_k^2F_v(k)+v_k^2F_u(k)\right)\le2d(u\odot v)_k^2,

which gives

Δdu⊙Δdv≤2d Δd(u⊙v).(31)\Delta_d u\odot\Delta_d v\le\sqrt{2d}\,\Delta_d(u\odot v). \tag*{(31)}

If u≤CΔdwu\le C\Delta_d w, monotonicity of the defining formula and v≤Δdvv\le\Delta_d v give u⊙v≤C2d Δd(w⊙v)u\odot v\le C\sqrt{2d}\,\Delta_d(w\odot v). The principal-ideal criterion therefore allows either input to be replaced by any sequence generating the same ideal. Thus a deterministic generator for the ideal of an intermediate random factor can be used in the next conditional application.

Corollary 2.7 (Gaussian products). Let G=(gij)G=(g_{ij}) have independent standard real Gaussian entries, and let A,BA,B be fixed Hilbert–Schmidt operators. The series

AGB=∑i,j≥1gij(Aei)⊗(B∗ej)(32)AGB=\sum_{i,j\ge1}g_{ij}(Ae_i)\otimes(B^*e_j) \tag*{(32)}

converges in L2(Ω;S2)L^2(\Omega;S_2) and almost surely along square partial sums, with

E∥AGB∥22=∥A∥22∥B∥22.(33)\mathbb{E}\lVert AGB\rVert_2^2=\lVert A\rVert_2^2\lVert B\rVert_2^2. \tag*{(33)}

Put a=s(A)a=s(A), b=s(B)b=s(B), and h=a⊙bh=a\odot b, so that

hn2=an2∑k≥nbk2+bn2∑k>nak2.(34)h_n^2=a_n^2\sum_{k\ge n}b_k^2+b_n^2\sum_{k>n}a_k^2. \tag*{(34)}

For every fixed compact DD with the same source and target as AGBAGB,

I(AGB+D)=I(h)+I(D)almost surely.(35)\mathcal{I}(AGB+D)=\mathcal{I}(h)+\mathcal{I}(D)\qquad\text{almost surely}. \tag*{(35)}

In particular, the matrix (aigijbj)(a_ig_{ij}b_j) generates I(a⊙b)\mathcal{I}(a\odot b). The construction and ideal identity remain valid conditionally for random (A,B,D)(A,B,D) in S2×S2×K(H)S_2\times S_2\times K(H) jointly independent of GG, with hh calculated from their singular values. No moment assumption on these inputs is needed for the almost sure assertions. Finite singular-value lists are completed by zeros.

Proof. Orthogonality of the Gaussian coefficients gives

∑i,j∥(Aei)⊗(B∗ej)∥22=∥A∥22∥B∥22.\sum_{i,j}\lVert(Ae_i)\otimes(B^*e_j)\rVert_2^2=\lVert A\rVert_2^2\lVert B\rVert_2^2.

Thus the square partial sums form an L2L^2-bounded S2S_2-valued martingale. Hilbert-space martingale convergence proves the asserted convergence and (33). The symbol GG denotes this series construction; it need not be a samplewise bounded operator.

Write the positive-singular-value decompositions A=∑kakuk⊗vkA=\sum_k a_ku_k\otimes v_k and B=∑lblrl⊗slB=\sum_l b_lr_l\otimes s_l. The L2L^2-defined variables

g^kl=∑i,j⟨ei,vk⟩⟨rl,ej⟩gij\widehat{g}_{kl}=\sum_{i,j}\langle e_i,v_k\rangle\langle r_l,e_j\rangle g_{ij}

are independent standard Gaussians, because their coefficient vectors are orthonormal in ℓ2(N2)\ell^2(\mathbb{N}^2). For the singular-value truncations ArA_r, BsB_s, ArGBs=∑k≤r,l≤sakblgkluk⊗slA_rGB_s = \sum_{k\le r,l\le s} a_kb_lg_{kl}u_k\otimes s_l. This identity passes to the limit by the Gaussian isometry and

∥AGB−A′GB′∥L2(Ω;S2)≤∥A−A′∥2∥B∥2+∥A′∥2∥B−B′∥2.\lVert AGB-A'GB'\rVert_{L^2(\Omega;S_2)}\le\lVert A-A'\rVert_2\lVert B\rVert_2+\lVert A'\rVert_2\lVert B-B'\rVert_2.

Accordingly, AGB=UYV∗AGB=UYV^* and Y=U∗AGBVY=U^*AGBV, where Y=(akgklbl)Y=(a_k g_{kl} b_l) and U,VU,V are the coordinate partial isometries for (uk),(sl)(u_k),(s_l). Zero extensions and unused Gaussian variables handle finite singular-value lists. The variance table ak2bl2a_k^2b_l^2 is doubly monotone with shell profile hh, so Theorem A gives I(AGB)=I(Y)=I(h)\mathcal{I}(AGB)=\mathcal{I}(Y)=\mathcal{I}(h). Moreover,

U∗(AGB+D)V=Y+U∗DV.U^*(AGB+D)V=Y+U^*DV.

Applying that theorem with the fixed compact perturbation U∗DVU^*DV gives I(h)⊆I(AGB+D)\mathcal{I}(h)\subseteq\mathcal{I}(AGB+D). Together with D=(AGB+D)−AGBD=(AGB+D)-AGB and the opposite inclusion obtained by addition, this proves (35).

For the conditional assertion, finite partial sums are jointly Borel in (A,B)(A,B) and the countable Gaussian coordinates. Convergence in S2S_2 is a Borel condition; set the limit to zero when it fails. Each coordinate of hh is Borel, since singular values are continuous and its tail sums are limits of finite sums. The ideal equality is Borel by (6). Independence and Fubini therefore apply the fixed-input result to the joint law of (A,B,D)(A,B,D) and GG. The L2L^2 assertion is conditional; unconditional L2L^2 convergence follows if E(∥A∥22∥B∥22)<∞\mathbb{E}(\lVert A\rVert_2^2\lVert B\rVert_2^2)<\infty. □

Lemma 2.8 (Independent bilateral Gaussian contraction). Let u,vu,v be nonnegative nonincreasing sequences in ℓ2\ell^2, let KK be fixed and Hilbert–Schmidt, and let G1,G2G_1,G_2 be independent arrays of independent standard real Gaussians. The finite-dimensional products defining

B=DuG1KG2∗DvB=D_uG_1KG_2^*D_v

converge in L2(Ω;S2)L^2(\Omega;S_2) as the weights are truncated and KK is approximated in Hilbert–Schmidt norm. Their limit satisfies

E∥B∥22=∥u∥22∥K∥22∥v∥22.(36)\mathbb{E}\lVert B\rVert_2^2=\lVert u\rVert_2^2\lVert K\rVert_2^2\lVert v\rVert_2^2. \tag*{(36)}

For every fixed compact FF with the same source and target as BB,

I(B+F)=I(u∘s(K)∘v)+I(F)almost surely.(37)\mathcal{I}(B+F)=\mathcal{I}(u\circ s(K)\circ v)+\mathcal{I}(F)\qquad\text{almost surely}. \tag*{(37)}

The event may depend on the fixed inputs and on FF; a prescribed countable family of perturbations can be treated on one event. Zero and zero-completed finite-rank factors are allowed.

Proof. Set H=KG2∗DvH=KG_2^*D_v. Corollary 2.7 gives

E∥H∥22=∥K∥22∥v∥22,I(H)=I(s(K)∘v)almost surely.\mathbb{E}\lVert H\rVert_2^2=\lVert K\rVert_2^2\lVert v\rVert_2^2,\qquad\mathcal{I}(H)=\mathcal{I}(s(K)\circ v)\quad\text{almost surely}.

Since G1G_1 is independent of HH, the conditional assertion defines B=DuG1HB=D_uG_1H and gives

E[∥B∥22∣H]=∥u∥22∥H∥22,I(B+F)=I(u∘s(H))+I(F)almost surely.\mathbb{E}[\lVert B\rVert_2^2\mid H]=\lVert u\rVert_2^2\lVert H\rVert_2^2,\qquad\mathcal{I}(B+F)=\mathcal{I}(u\circ s(H))+\mathcal{I}(F)\quad\text{almost surely}.

Taking expectations proves (36). Compatibility now gives I(u∘s(H))=I(u∘(s(K)∘v))\mathcal{I}(u\circ s(H))=\mathcal{I}(u\circ(s(K)\circ v)); associativity proves (37). The same isometries applied to differences in u,K,vu,K,v prove continuity in ℓ2×S2×ℓ2\ell^2\times S_2\times\ell^2, identifying this construction with the stated L2L^2 limit of finite-dimensional products. A countable intersection handles the prescribed family of perturbations. □

Gaussian evaluation of a Hilbert two-form

A Hilbert two-form on a real separable Hilbert space EE is identified with a skew-adjoint Hilbert–Schmidt operator by f∧g↦f⊗g−g⊗ff\wedge g\mapsto f\otimes g-g\otimes f. With increasing orthonormal wedges normalized to have norm ∥K∥22=2∥ωK∥AE2\lVert K\rVert_2^2=2\lVert\omega_K\rVert_{A^2_E}.

Let Q≥0Q\ge0 be trace class on a real separable Hilbert space HH, with positive eigenvalues qiq_i in decreasing order and eigenvectors eie_i. Complete a finite eigenvalue list by zeros and put ai=qia_i=\sqrt{q_i}. For independent isonormal Gaussian processes ξi\xi_i over EE [6], define

Gf=∑qi>0aiξi(f)eiin L2(Ω;H).(38)\mathcal{G}f=\sum_{q_i>0}a_i\xi_i(f)e_i\quad\text{in }L^2(\Omega;H). \tag*{(38)}

Then E(Gf⊗Gg)=⟨f,g⟩Q\mathbb{E}(\mathcal{G}f \otimes\mathcal{G}g)=\langle f,g\rangle_Q. Evaluations act on ran⁡Q‾\overline{\operatorname{ran} Q} and are extended by zero on ker⁡Q\ker Q; finite-dimensional spaces are embedded in ℓ2\ell^2 when their principal ideals are compared.

For finite-rank skew-adjoint KK, define A=GKG∗A=\mathcal{G}K\mathcal{G}^{*} by sending f⊗g−g⊗ff\otimes g-g\otimes f to Gf⊗Gg−Gg⊗Gf\mathcal{G}f\otimes\mathcal{G}g-\mathcal{G}g\otimes\mathcal{G}f. This is representation-independent by linear algebra on the finite-dimensional subspace involved; the notation does not require a samplewise bounded map G:E→H\mathcal{G}:E\to H.

Theorem 3.1 (Classification of Gaussian exterior evaluation). For fixed QQ as above, finite-rank evaluation extends uniquely and continuously to every skew-adjoint K∈S2(E)K\in S_2(E) in L2(Ω;S2)L^2(\Omega;S_2), with

E∥A∥22=((tr⁡Q)2−tr⁡(Q2))∥K∥22.(39)\mathbb{E}\lVert A\rVert_2^2=\left((\operatorname{tr}Q)^2-\operatorname{tr}(Q^2)\right)\lVert K\rVert_2^2. \tag*{(39)}

Set

κn=sn(K),Rn=∑k≥nqk,Tn=∑k≥nκk2,(40)\kappa_n=s_n(K),\qquad R_n=\sum_{k\ge n}q_k,\qquad T_n=\sum_{k\ge n}\kappa_k^2, \tag*{(40)}

and

r=a⊙κ⊙a,rn2=Rn2Tn−Rn+12Tn+1.(41)r=a\mathbin{\odot}\kappa\mathbin{\odot}a,\qquad r_n^2=R_n^2T_n-R_{n+1}^2T_{n+1}. \tag*{(41)}

If rank Q≥2Q\ge2, then for every fixed compact operator DD on HH, without a symmetry or Hilbert–Schmidt assumption,

I(A+D)=I(r)+I(D)=I(max⁡{r,s(D)})almost surely.(42)\mathcal{I}(A+D)=\mathcal{I}(r)+\mathcal{I}(D)=\mathcal{I}\left(\max\{r,s(D)\}\right)\quad\text{almost surely}. \tag*{(42)}

In particular, I(A)=I(r)\mathcal{I}(A)=\mathcal{I}(r) almost surely. For fixed Q,KQ,K, one event accommodates any prescribed countable family of deterministic compact DD. If rank Q≤1Q\le1 or K=0K=0, then A=0A=0 almost surely and I(A+D)=I(D)\mathcal{I}(A+D)=\mathcal{I}(D) for every compact $D.

Thus

rn2=∑min⁡(i,j,k)=nqiκj2qkr_n^2=\sum_{\min(i,j,k)=n}q_i\kappa_j^2q_k

is the squared energy of a⊗κ⊗aa\otimes\kappa\otimes a on the shell {(i,j,k):min⁡(i,j,k)=n}\{(i,j,k):\min(i,j,k)=n\}.

Proof. For finite-rank skew-adjoint KK, the diagonal entries of its evaluation vanish, while independence gives variance qiqj∥K∥22q_iq_j\lVert K\rVert_2^2 for entry (i,j)(i,j) when i≠ji\ne j. Summation proves (39); applying the same identity to differences gives the continuous extension and the zero cases. Henceforth assume q2>0q_2>0 and K≠0K\ne0, and fix DD.

For the upper inclusion, use the canonical form

K=∑lσl(fl⊗gl−gl⊗fl),σ1≥σ2≥⋯>0,(43)K=\sum_l\sigma_l(f_l\otimes g_l-g_l\otimes f_l),\qquad\sigma_1\ge\sigma_2\ge\cdots>0, \tag*{(43)}

with mutually orthonormal fl,glf_l,g_l, Hilbert–Schmidt convergence, and κ=Δ2σ\kappa=\Delta_2\sigma. The arrays Xil=ξi(fl)X_{il}=\xi_i(f_l) and Yil=ξi(gl)Y_{il}=\xi_i(g_l) are independent standard Gaussian arrays, and the evaluation and product isometries give

A=C−C∗,C=DaXDσY∗Dain L2(Ω;S2).(44)A=C-C^{*},\qquad C=D_aXD_\sigma Y^{*}D_a\quad\text{in }L^2(\Omega;S_2). \tag*{(44)}

Zero padding handles finite coordinate lists. Lemma 2.8 and compatibility of ⊙\odot with principal ideals yield I(C)=I(a⊙σ⊙a)=I(r)\mathcal{I}(C)=\mathcal{I}(a\odot\sigma\odot a)=\mathcal{I}(r) almost surely. Closure under adjoints and finite sums therefore gives A∈I(r)A\in\mathcal{I}(r).

For the lower inclusion, separate the even and odd covariance coordinates:

Aoe=DaoGoKGe∗Dae,ano=a2n−1,ane=a2n.(45)A_{oe}=D_{a^o}G_oKG_e^{*}D_{a^e},\qquad a_n^o=a_{2n-1},\qquad a_n^e=a_{2n}. \tag*{(45)}

The disjoint row families are independent, and the two L2L^2 isometries extend this identity from finite-rank KK. Both weight sequences retain I(a)\mathcal{I}(a), since ae≤ao≤aa^e\le a^o\le a and

an≤max⁡{1,a1/a2}a⌈n/3⌉e(n≥1).(46)a_n\le\max\{1,a_1/a_2\}a_{\lceil n/3\rceil}^e\qquad(n\ge1). \tag*{(46)}

For n≥2n\ge2, use 2⌈n/3⌉≤n2\lceil n/3\rceil\le n; the constant handles n=1n=1. This also covers finite support. For the same compact block DoeD_{oe} of DD, (37) and the compatibility of ⊙\odot give

I(r)+I(Doe)=I(Aoe+Doe)⊂I(A+D)almost surely.\mathcal{I}(r)+\mathcal{I}(D_{oe})=\mathcal{I}(A_{oe}+D_{oe})\subset\mathcal{I}(A+D)\quad\text{almost surely}.

Since A∈I(r)A\in\mathcal{I}(r), we also have D=(A+D)−A∈I(A+D)D=(A+D)-A\in\mathcal{I}(A+D). This proves I(r)+I(D)⊂I(A+D)\mathcal{I}(r)+\mathcal{I}(D)\subset\mathcal{I}(A+D); the reverse inclusion follows from A∈I(r)A\in\mathcal{I}(r). The maximum rule for principal-ideal sums and a countable intersection finish the proof.

Rank one must be excluded: A=0A=0 although r≠0r\ne0 when K≠0K\ne0. For rank Q≥2Q\ge2 and K≠0K\ne0, the profile shows that AA is almost surely nonzero finite rank if either QQ or KK has finite rank, and infinite rank otherwise. The event in (42) cannot cover all compact perturbations, since the sample-dependent choice D=−AD=-A would contradict the identity.

Tail averages and regularity at infinity

For a nonnegative nonincreasing sequence ξ∈ℓ1\xi\in\ell^1, put

Un=∑j≥nξj,(H∞ξ)n=Un/n.(47)U_n=\sum_{j\ge n}\xi_j,\qquad(H_\infty\xi)_n=U_n/n. \tag*{(47)}

The tail average decreases to zero but need not be summable. A nonzero sequence is regular at infinity if

Un≤Cnξn(n≥1)(48)U_n\le Cn\xi_n\qquad(n\ge1) \tag*{(48)}

for some C<∞C<\infty. For infinite support this is the classical condition in [7 Definition 4.11 and Theorem 4.12]: replacing the strict tail there by our inclusive tail adds ξn/n\xi_n/n to the average and does not change the condition. We also include nonzero finite-support sequences; the inclusive convention preserves their last nonzero entry.

Lemma 3.2 (Removing dilation from a tail estimate). For every nonnegative nonincreasing ξ∈ℓ1\xi\in\ell^1,

I(ξ)⊆I(H∞ξ).(49)\mathcal{I}(\xi)\subseteq\mathcal{I}(H_\infty\xi). \tag*{(49)}

Equality holds if and only if (48) holds. In particular, a finite dilation does not weaken that tail condition: for some C<∞C<\infty and M∈NM\in\mathbb{N}, the estimate

Un/n≤Cξ⌈n/M⌉(n≥1)(50)U_n/n\le C\xi_{\lceil n/M\rceil}\qquad(n\ge1) \tag*{(50)}

already implies (48). Zero and finite-support sequences are included.

Proof. For m=⌈n/2⌉m=\lceil n/2\rceil, Um≥(n−m+1)ξn≥mξnU_m\ge(n-m+1)\xi_n\ge m\xi_n, which proves (49) by the finite-dilation criterion. Equality of the ideals gives (50); evaluating it at n=Mkn=Mk yields

Uk=∑j=kMk−1ξj+UMk≤(M−1+CM)kξk.U_k=\sum_{j=k}^{Mk-1}\xi_j+U_{Mk}\le(M-1+CM)k\xi_k.

This proves the tail condition, whose converse is immediate. The first sum is empty if M=1M=1, and no sequence entry is divided out. For support in {1,…,N}\{1,\ldots,N\}, Un≤Nξn≤NnξnU_n\le N\xi_n\le Nn\xi_n; the zero case is immediate. □\square

Universal simplification of the exterior profile

The profile separates according to changes in the covariance and form tails:

rn2=qn(Rn+Rn+1)Tn+κn2Rn+12.(51)r_n^2=q_n(R_n+R_{n+1})T_n+\kappa_n^2R_{n+1}^2. \tag*{(51)}

Put xn=qnRnTnx_n=\sqrt{q_nR_nT_n} and yn=κnRny_n=\kappa_nR_n, both nonnegative, nonincreasing and square summable. The first term is at least xn2x_n^2, while Tn≥κn2T_n\ge\kappa_n^2 gives rn2≥yn2r_n^2\ge y_n^2; the same formula gives rn2≤2xn2+yn2r_n^2\le2x_n^2+y_n^2. Hence

max⁡{xn,yn}≤rn≤3max⁡{xn,yn}(n≥1),(52)\max\{x_n,y_n\}\le r_n\le\sqrt{3}\max\{x_n,y_n\}\qquad(n\ge1), \tag*{(52)}

and for rank Q≥2Q\ge2,

I(A)=I((qnRnTn)n)+I((κnRn)n)almost surely.(53)\mathcal{I}(A)=\mathcal{I}\left(\left(\sqrt{q_nR_nT_n}\right)_n\right)+\mathcal{I}\left(\left(\kappa_nR_n\right)_n\right)\qquad\text{almost surely}. \tag*{(53)}

These two ideals also replace I(r)\mathcal{I}(r) in (42). The next equivalence makes the tail condition (54) necessary and sufficient for the first ideal to be absorbed by the second for every fixed QQ of rank at least two. The prescribed test Q0=diag⁡(n−2)Q_0=\operatorname{diag}(n^{-2}) already detects the same condition.

Corollary 3.3 (Universal collapse of the exterior profile). Fix a skew-adjoint K∈S2(E)K\in\mathcal{S}_2(E), with κn=sn(K)\kappa_n=s_n(K) and Tn=∑j≥nκj2T_n=\sum_{j\ge n}\kappa_j^2. The following are equivalent:

(i) There is C<∞C<\infty such that

Tn≤Cnκn2(n≥1).(54)T_n\le Cn\kappa_n^2\qquad(n\ge1). \tag*{(54)}

:::`

(ii) For every fixed trace-class covariance QQ of rank at least two,

I(AQ)=I((κnRn)n)almost surely,Rn=∑j≥nsj(Q).(55)\mathcal{I}(A_Q)=\mathcal{I}((\kappa_n R_n)_n)\quad\text{almost surely},\qquad R_n=\sum_{j\ge n}s_j(Q). \tag*{(55)}

(iii) Identity (55) holds almost surely for Q0=diag⁡(n−2)Q_0=\operatorname{diag}(n^{-2}) on ℓ2\ell^2.

For K≠0K\ne0, condition (i) is regularity at infinity of κ2\kappa^2. The zero and finite-rank cases are included. The event in (ii) may depend on the fixed QQ.

Proof. Assume (i) with C≥1C\ge1. For m=⌈n/2⌉m=\lceil n/2\rceil, Rm≥∑j=mnqj≥mqnR_m\ge\sum_{j=m}^{n}q_j\ge mq_n, and therefore

xn2=qnRnTn≤CnqnRnκn2≤2Cκm2Rm2=2Cym2.x_n^2=q_nR_nT_n\le Cnq_nR_n\kappa_n^2\le2C\kappa_m^2R_m^2=2Cy_m^2.

Together with (51), this gives

yn≤rn≤1+4C y⌈n/2⌉.y_n\le r_n\le\sqrt{1+4C\,y_{\lceil n/2\rceil}}.

The ideal criterion and Theorem 3.1 prove (ii), which immediately implies (iii).

Conversely, take qn=n−2q_n=n^{-2}, so 1/n≤Rn≤2/n1/n\le R_n\le2/n. Condition (iii) and Theorem 3.1 give I(r)=I(y)\mathcal{I}(r)=\mathcal{I}(y); hence rn≤Bymr_n\le By_m for some B>0B>0, M∈NM\in\mathbb{N} and m=⌈n/M⌉m=\lceil n/M\rceil. Since rn≥xnr_n\ge x_n,

Tnn3≤xn2≤rn2≤B2κm2Rm2≤4B2M2n2κm2.\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.

Thus Tn/n≤4B2M2κ⌈n/M⌉2T_n/n\le4B^2M^2\kappa_{\lceil n/M\rceil}^2. Lemma 3.2, applied to ξ=κ2\xi=\kappa^2, removes the dilation and proves (i), including zero and finite support. □

Under these conditions, (42) becomes

I(AQ+D)=I((κnRn)n)+I(D)almost surely\mathcal{I}(A_Q+D)=\mathcal{I}((\kappa_nR_n)_n)+\mathcal{I}(D)\quad\text{almost surely}

for every fixed QQ of rank at least two and compact DD. The condition holds when κn≍n−β\kappa_n\asymp n^{-\beta} with β>1/2\beta>1/2, since Tn≍n1−2βT_n\asymp n^{1-2\beta}; this includes the Wiener time operator below.

Remark. (One form and two covariance regimes.) Take the canonical coefficients σl=[llog⁡(l+1)]−1\sigma_l=[\sqrt{l}\log(l+1)]^{-1} in (43). They are square summable, and

κn=σ⌈n/2⌉≍1nlog⁡(n+1),Tn≍1log⁡(n+1).\kappa_n=\sigma_{\lceil n/2\rceil}\asymp\frac{1}{\sqrt{n}\log(n+1)},\qquad T_n\asymp\frac{1}{\log(n+1)}.

For qn=n−2q_n=n^{-2}, one has Rn≍n−1R_n\asymp n^{-1} and

rn≍xn≍1n3/2log⁡(n+1),yn≍1n3/2log⁡(n+1).r_n\asymp x_n\asymp\frac{1}{n^{3/2}\sqrt{\log(n+1)}},\qquad y_n\asymp\frac{1}{n^{3/2}\log(n+1)}.

Thus rn/y⌈n/M⌉→∞r_n/y_{\lceil n/M\rceil}\to\infty for every fixed MM, and I(y)⊊I(r)\mathcal{I}(y)\subsetneq\mathcal{I}(r): the first contribution in (53) cannot be omitted.

For the same KK but qn=2−nq_n=2^{-n}, Rn=21−nR_n=2^{1-n} and

rn≍xn≍2−nlog⁡(n+1),yn≍2−nnlog⁡(n+1).r_n\asymp x_n\asymp\frac{2^{-n}}{\sqrt{\log(n+1)}},\qquad y_n\asymp\frac{2^{-n}}{\sqrt{n}\log(n+1)}.

Now rn/y⌈n/2⌉≲2−n/2nlog⁡(n+1)→0r_n/y_{\lceil n/2\rceil}\lesssim2^{-n/2}\sqrt{n\log(n+1)}\to0, so I(r)=I(y)\mathcal{I}(r)=\mathcal{I}(y) despite failure of (54). This covariance is itself regular at infinity, since Rn/(nqn)=2/nR_n/(nq_n)=2/n. Thus regularity of a test covariance alone does not make it a valid substitute for the prescribed Q0Q_0 in Corollary 3.3.

Wiener area as covariance tail averaging

Let W=(Wt)0≤t≤1W=(W_t)_{0\le t\le1} be a Wiener process in a real separable Hilbert space with trace-class covariance Q≥0Q\ge0. Its skew area is

A=12∫01(Wt⊗dWt−dWt⊗Wt).(56)A=\frac{1}{2}\int_0^1(W_t\otimes dW_t-dW_t\otimes W_t). \tag*{(56)}

The Hilbert–Schmidt-valued Itô integral is well defined, since E∥∫01Wt⊗dWt∥HS2=(tr⁡Q)2/2\mathbb{E}\left\|\int_0^1 W_t\otimes dW_t\right\|_{\mathrm{HS}}^2=(\operatorname{tr}Q)^2/2 and the other integral is its adjoint; see [3 Chapter 4]. Write q1≥q2≥⋯≥0q_1\ge q_2\ge\cdots\ge0 for the covariance eigenvalues, padded by zeros in finite rank, and put

Rn=∑k≥nqk,bn=(H∞q)n=Rnn.R_n=\sum_{k\ge n}q_k,\qquad b_n=(H_\infty q)_n=\frac{R_n}{n}.

The time kernel reduces the exterior-form classification to this tail average.

Corollary 4.1 (Wiener area and the covariance ideal). Suppose rank Q≥2Q\ge2. For every fixed deterministic compact operator DD, without a skew-adjointness assumption,

I(A+D)=I(b)+I(D)almost surely.(57)\mathcal{I}(A+D)=\mathcal{I}(b)+\mathcal{I}(D)\qquad\text{almost surely.} \tag*{(57)}

In particular, on one event of probability one,

I(A)=I(b),I(Q)⊆I(A),(58)\mathcal{I}(A)=\mathcal{I}(b),\qquad\mathcal{I}(Q)\subseteq\mathcal{I}(A), \tag*{(58)}

and

I(A)=I(Q)⟺Rn≤Cnqn(n≥1) for some C<∞.(59)\mathcal{I}(A)=\mathcal{I}(Q)\quad\Longleftrightarrow\quad R_n\le Cnq_n\quad(n\ge1)\text{ for some }C<\infty. \tag*{(59)}

Thus regularity at infinity is exactly the condition that covariance tail averaging preserves the principal ideal. It includes finite-rank covariances of rank at least two. If rank Q≤1Q\le1, then A=0A=0 almost surely and I(A+D)=I(D)\mathcal{I}(A+D)=\mathcal{I}(D).

Proof. Write Wt=∑iqi βi(t)eiW_t=\sum_i\sqrt{q_i}\,\beta_i(t)e_i in an eigenbasis of QQ. The independent isonormal processes ξi(f)=∫01f(t) dβi(t)\xi_i(f)=\int_0^1 f(t)\,d\beta_i(t) act on L2(0,1)L^2(0,1). For the skew-adjoint Hilbert–Schmidt operator

(Kf)(s)=12∫01sgn⁡(t−s)f(t) dt,(Kf)(s)=\frac{1}{2}\int_0^1\operatorname{sgn}(t-s)f(t)\,dt,

the off-diagonal entries of AA are

qiqj∫(0,1)212sgn⁡(t−s) dβi(s) dβj(t).\sqrt{q_iq_j}\int_{(0,1)^2}\frac{1}{2}\operatorname{sgn}(t-s)\,d\beta_i(s)\,d\beta_j(t).

Thus AA is the Gaussian skew compression of KK: the identity holds first for step kernels and then by the Itô isometry, with the mixed integral defined by the isometric extension of f⊗g↦ξi(f)ξj(g)f\otimes g\mapsto\xi_i(f)\xi_j(g) for i≠ji\ne j.

On the complexification, the functions eiπ(2j+1)te^{i\pi(2j+1)t}, j∈Zj\in\mathbb{Z}, form an orthonormal basis, and direct integration gives, for α=π(2j+1)\alpha=\pi(2j+1),

K(eiα⋅)(s)=eiα+1−2eiαs2iα=iαeiαs.K(e^{i\alpha\cdot})(s)=\frac{e^{i\alpha}+1-2e^{i\alpha s}}{2i\alpha}=\frac{i}{\alpha}e^{i\alpha s}.

Consequently the full singular-value sequence, with multiplicities, is

κn=1π(2⌊n/2⌋−1)≍n−1,∑k≥nκk2≍nκn2.\kappa_n=\frac{1}{\pi(2\lfloor n/2\rfloor-1)}\asymp n^{-1},\qquad\sum_{k\ge n}\kappa_k^2\asymp n\kappa_n^2.

Thus κ2\kappa^2 is regular at infinity, and Corollary 3.3 gives I(A)=I((κnRn)n)=I(b)\mathcal{I}(A)=\mathcal{I}((\kappa_nR_n)_n)=\mathcal{I}(b) and its fixed-perturbation formula (57). Lemma 3.2, applied to qq, now gives I(Q)⊆I(A)\mathcal{I}(Q)\subseteq\mathcal{I}(A) and (59), including finite rank. If rank Q≤1Q\le1, the path lies in a fixed line and (56) vanishes. □

Brownian scaling gives the same formulas for every fixed interval [s,t]⊆[0,1][s,t]\subseteq[0,1], s<ts<t, with the path recentered at ss; a common event may be chosen for any prescribed countable family of intervals and fixed compact perturbations.

For example, let qn∼[n(log⁡(en))α]−1q_n\sim[n(\log(en))^\alpha]^{-1} with α>1\alpha>1. The integral test gives Rn∼(log⁡(en))1−αR_n\sim(\log(en))^{1-\alpha}, hence Rn/(nqn)∼log⁡(en)→∞R_n/(nq_n)\sim\log(en)\to\infty and

I(Q)⊊I(A)almost surely.(60)\mathcal{I}(Q)\subsetneq\mathcal{I}(A)\qquad\text{almost surely.} \tag*{(60)}

Comparison of full Gaussian covariance forms

The matrix classification is stable under two-sided comparison of the full Gaussian covariance form. The proof transfers the upper and mean-uniform lower estimates from Subsection 2.1.

Corollary A.1 (Full covariance comparability). Let vv satisfy (8), and let Z=(Zij)Z=(Z_{ij}) be a centered jointly Gaussian array. Suppose there are 0<a≤b<∞0<a\le b<\infty such that, for every finitely supported real coefficient table uu,

a∑i,jvijuij2≤E(∑i,juijZij)2≤b∑i,jvijuij2.(61)a\sum_{i,j}v_{ij}u_{ij}^{2}\le\mathbb{E}\left(\sum_{i,j}u_{ij}Z_{ij}\right)^{2}\le b\sum_{i,j}v_{ij}u_{ij}^{2}. \tag*{(61)}

Then ZZ has a bounded realization almost surely if and only if (10) holds. If that condition fails, almost surely there is no bounded realization. If it holds, the realization is unique and compact almost surely, I(Z)=I(h)\mathcal{I}(Z)=\mathcal{I}(h), and all comparisons in (11) hold with ZZ in place of XX on one event. For every fixed compact BB,

I(Z+B)=I(h)+I(B)almost surely.(62)\mathcal{I}(Z+B)=\mathcal{I}(h)+\mathcal{I}(B)\qquad\text{almost surely}. \tag*{(62)}

The simultaneous ideal tests in (14) hold as well. The same assertions concerning Z+BZ+B hold for a compact random BB independent of ZZ, without a moment assumption on $B.

Proof. Assume (10). We compare covariance forms on finite coefficient arrays, so no finite total variance is required. For each finite corner starting at kk, Gaussian comparison with bX\sqrt{b}X gives expected norm at most 22b hk2\sqrt{2b}\,h_k. Its norm is bvkk\sqrt{bv_{kk}}-Lipschitz in standard Gaussian coordinates, since it is the supremum of unit bilinear forms whose variances are at most bvkkbv_{kk}. Thus the finite-compression construction and (21) carry over with K0K_0 replaced by bK0\sqrt{b}K_0, giving a unique compact Borel realization and the upper comparisons for every L>2L>2.

Fix a compact BB. For m<nm<n, the first mm rows of Z+BZ+B are jointly Gaussian vectors in HmH^m: their total centered second moment is at most bm∑kv1k<∞bm\sum_k v_{1k}<\infty. Their covariance dominates Im⊗(aRn)I_m\otimes(aR_n), with RnR_n as in (28), by (A.1) and double monotonicity. After division by a\sqrt{a}, Lemma 2.4 therefore gives the same row-distance bound used in Proposition 2.6. Its inverse Gram and transposition argument yields, uniformly in BB,

P{sj(Z+B)<a(n−j)8n hn}≤ne−(n−j)/8(1≤j<n),\mathbb{P}\left\{s_j(Z+B)<\sqrt{\frac{a(n-j)}{8n}}\,h_n\right\}\le ne^{-(n-j)/8}\qquad(1\le j<n),

including the trivial case hn=0h_n=0.

The completion argument for Theorem A now applies. Indeed, if hn>0h_n>0, the covariance of the leading n×nn\times n block dominates avnnav_{nn} times the identity, so its determinant after addition of the fixed block of BB is nonzero almost surely. This handles the initial lower indices; Borel–Cantelli and the same rational-parameter intersections give, for each fixed BB, one event supporting the upper comparisons for ZZ and the lower comparisons for Z+BZ+B for every L>2L>2 and λ>1\lambda>1. For B=0B=0 the comparisons give I(Z)=I(h)\mathcal{I}(Z)=\mathcal{I}(h); for general BB, the lower inclusion I(h)⊂I(Z+B)\mathcal{I}(h)\subset\mathcal{I}(Z+B) and B=(Z+B)−ZB=(Z+B)-Z prove (A.2); the ideal tests follow algebraically.

If (10) fails, the first row or column has infinite variance sum. In the row case, write wj=v1jw_j=v_{1j}. Its first NN coordinates have covariance at least adiag⁡(w1,…,wN)a\operatorname{diag}(w_1,\ldots,w_N), so Lemma 2.3 gives

Eexp⁡{−∑j=1NZ1j2}≤∏j=1N(1+2awj)−1/2⟶0.\mathbb{E}\exp\left\{-\sum_{j=1}^{N}Z_{1j}^{2}\right\}\le\prod_{j=1}^{N}(1+2aw_j)^{-1/2}\longrightarrow0.

Bounded convergence implies ∑jZ1j2=∞\sum_j Z_{1j}^{2}=\infty almost surely, which excludes a bounded realization. The column case is identical.

Finally, under (10), the assertions for a fixed BB define a Borel event in (Z,B)∈K(H)2(Z,B)\in\mathcal{K}(H)^2: singular values are continuous, and the finite-dilation criterion (6) and the comparisons use only countably many integer constants and rational parameters. Fubini’s theorem therefore gives the same conclusions for an independent compact random BB, without a moment assumption. ∎

Covariance information alone cannot replace joint Gaussianity. Let a,ba,b be strictly positive nonincreasing sequences in ℓ2\ell^2, and let ξ,η\xi,\eta be independent standard Gaussian sequences. The array Tij=aibjξiηjT_{ij}=a_i b_j \xi_i\eta_j is almost surely rank one, yet

E(TijTkl)=ai2bj21{i=k, j=l}.\mathbb{E}(T_{ij}T_{kl})=a_i^2b_j^2\mathbf{1}_{\{i=k,\ j=l\}}.

It has the same full covariance form as the independent-entry Gaussian matrix with variances ai2bj2a_i^2b_j^2, whose shell profile is strictly positive at every index and whose principal ideal therefore has an infinite-rank generator.

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: [email protected]

References

  1. [1]Bandeira, A. S. and van Handel, R. (2016). Sharp nonasymptotic bounds on the norm of random matrices with independent entries. Annals of Probability 44, no. 4, 2479–2506. doi:10.1214/15-AOP1025.
  2. [2]Calkin, J. W. (1941). Two-sided ideals and congruences in the ring of bounded operators in Hilbert space. Annals of Mathematics (2) 42, no. 4, 839–873. doi:10.2307/1968771.DOI
  3. [3]Da Prato, G. and Zabczyk, J. (2014). Stochastic Equations in Infinite Dimensions, 2nd ed. Encyclopedia of Mathematics and its Applications, vol. 152. Cambridge University Press, Cambridge. doi:10.1017/CBO9781107295513.DOI
  4. [4]Dela Cruz, Danilo Jr and Oberhauser, H. (2026). The singular values of Lévy’s area matrix. Preprint, arXiv:2606.08928. https://arxiv.org/abs/2606.08928.
  5. [5]Ferreiro-Castilla, A. and Utzet, F. (2011). Lévy area for Gaussian processes: A double Wiener–Itô integral approach. Statistics & Probability Letters 81, no. 9, 1380–1391. doi:10.1016/j.spl.2011.04.015.
  6. [6]Janson, S. (1997). Gaussian Hilbert Spaces. Cambridge Tracts in Mathematics, vol. 129. Cambridge University Press, Cambridge. doi:10.1017/CBO9780511526169.DOI
  7. [7]Kaftal, V. and Weiss, G. (2010). Traces on operator ideals and arithmetic means. Journal of Operator Theory 63, no. 1, 3–46. https://jot.theta.ro/jot/archive/2010-063-001/2010-063-001-001.pdf.
  8. [8]Kaftal, V. and Weiss, G. (2011). B(H) lattices, density and arithmetic mean ideals. Houston Journal of Mathematics 37, no. 1, 233–283. https://www.math.uh.edu/~hjm/Vol37-1.html. Author preprint: arXiv:0707.3115.
  9. [9]Latała, R., van Handel, R. and Youssef, P. (2018). The dimension-free structure of nonhomogeneous random matrices. Inventiones Mathematicae 214, no. 3, 1031–1080. doi:10.1007/s00222-018-0817-x.
  10. [10]Ledoux, M., Lyons, T. and Qian, Z. (2002). Lévy area of Wiener processes in Banach spaces. Annals of Probability 30, no. 2, 546–578. doi:10.1214/aop/1023481002.DOI
  11. [11]Rudelson, M. and Zeitouni, O. (2016). Singular values of Gaussian matrices and permanent estimators. Random Structures & Algorithms 48, no. 1, 183–212. doi:10.1002/rsa.20564.
  12. [12]Simon, B. (2005). Trace Ideals and Their Applications, 2nd ed. Mathematical Surveys and Monographs, vol. 120. American Mathematical Society, Providence, RI. doi:10.1090/surv/120.
  13. [13]Tao, T. and Vu, V. (2010). Random matrices: Universality of ESDs and the circular law. With an appendix by M. Krishnapur. Annals of Probability 38, no. 5, 2023–2065. doi:10.1214/10-AOP534.DOI
  14. [14]van Handel, R. (2016). Probability in High Dimension. APC 550 Lecture Notes, Princeton University, version December 21, 2016. https://web.math.princeton.edu/~rvan/APC550.pdf.
  15. [15]Zhao, G. (2026). Gaussian exterior dilation and Schatten classes of Lévy area. Preprint, arXiv:2610.34984. https://arxiv.org/abs/2610.34984.

Paper details

vlnc:W_4a6fdp2n
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 1

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Mathematics

GPT-6.0 (OpenAI)

View statement
Looking for related papers…
Contents