betaFeedback

WorksMathematics

Gaussian Exterior Dilation and Schatten Classes of Lévy Area

Abstract

We study the skew-adjoint Hilbert–Schmidt operator obtained by evaluating a fixed Hilbert two-form on a centered Gaussian linear map whose covariance is the inner product on its domain multiplied by a fixed nonnegative trace-class operator. We compute the expected regularized determinant of the resulting random operator and compare the random determinant with this expectation by dilation of the scalar parameter. For each fixed covariance and two-form, a single dilation almost surely applies to every nonnegative parameter value and yields simultaneous bounds for all Schatten norms and quasi-norms of orders 0<p<20<p<2. The least admissible dilation satisfies a universal weak-L1L^1 estimate with optimal tail exponent. For Wiener area, a rank-one perturbation identity identifies the mean determinant with the regularized determinant of an explicit deterministic skew-adjoint operator and reduces Schatten membership to conditions on the covariance eigenvalues. Consequently, for each 0<p<10<p<1, the area belongs almost surely to the Schatten class of order pp if and only if the covariance of the Wiener process does. At p=1p=1, the criterion is finite covariance entropy; for p>1p>1, trace-class covariance suffices.

Introduction

Lévy area is the antisymmetric part of a second-order stochastic integral. It enters the construction of rough paths and the analysis of differential equations driven by irregular signals [7, 15]. For a Hilbert-space Wiener process with trace-class covariance, the area has a Hilbert–Schmidt realization. We determine when it belongs to the trace class and to the Schatten ideals below trace class, in terms of the covariance eigenvalues.

The main result concerns a fixed Hilbert two-form evaluated by a centered Gaussian linear map whose covariance is the inner product on its domain multiplied by a fixed nonnegative trace-class operator. The resulting skew-adjoint operator has a regularized determinant whose mean is given by an exterior-power expansion. We prove that a single random dilation compares this determinant with its mean for every nonnegative value of the scalar parameter. The least dilation satisfies a universal weak-L1L^1 estimate, and the tail exponent is optimal. For each fixed covariance and two-form, the comparison holds on one event of probability one and gives simultaneous upper and lower bounds for all Schatten norms and quasi-norms of orders strictly between zero and two. Here the pathwise comparison is simultaneous in the determinant parameter for a fixed Gaussian image; the Wiener-area application concerns the area over a fixed time interval.

For Wiener area, the mean determinant is the regularized determinant of an explicit deterministic skew-adjoint operator. A rank-one perturbation identity determines its determinant function and the exterior coefficients of the Brownian time kernel. It also reduces Schatten membership to scalar summability conditions. For every order 0<p<10 < p < 1, the area belongs almost surely to the Schatten class of order pp if and only if the covariance belongs to the same class. For a nonzero covariance with positive eigenvalues (qi)(q_i), set S=∑iqiS = \sum_i q_i. The trace-class criterion is

∑iqilog⁡(S/qi)<∞.\sum_i q_i \log(S/q_i) < \infty.

Thus the endpoint requires a logarithmic summability condition in addition to the trace-class assumption on the covariance.

The conclusion for Schatten orders greater than one follows from the exactness criterion for tensor norms of Ledoux, Lyons and Qian [13]; the corresponding estimate is recorded after the proof of Corollary 1. Deierich [5] develops finite-dimensional approximation for Banach-space Wiener rough paths. Grong, Nilssen and Schmeding [8] study geometric approximation in infinite-dimensional spaces for a fixed Schatten tensor norm.

For general continuous Hilbert-space semimartingales, the author [18] proves almost-sure joint continuity of stochastic cross integrals in the two time variables with values in the little weak-trace ideal. This ideal consists of compact operators whose singular values, multiplied by their indices, tend to zero. That paper also constructs Wiener examples for which this convergence can be arbitrarily slow. The covariance criteria proved here use Gaussian structure to characterize membership in the trace class and in Schatten ideals of order below one.

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

Key words and phrases. Gaussian exterior forms; Lévy area; regularized determinants; Schatten classes; covariance entropy.

Distributional formulas for stochastic area originate in Lévy’s work [14]. Higher-dimensional formulas and double Wiener–Itô representations are developed in [6, 9, 11]. Dela Cruz and Oberhauser [4] determine finite-dimensional singular-value distributions and high-dimensional asymptotics for matrices of Brownian Lévy areas. Their Section 2.4 also treats the singular-value densities of area matrices for sufficiently regular centered Gaussian processes with independent identically distributed coordinates. The present paper gives almost-sure Schatten membership criteria for a fixed Hilbert-space Wiener process with arbitrary trace-class spatial covariance.

The proof of the Gaussian comparison uses the covariance geometry of exterior contractions. Alternation bounds the largest covariance eigenvalue of a Gaussian contraction by its covariance trace divided by the exterior degree. For degree at least two, Gaussian integration by parts converts this relation into an inverse-norm estimate. Successive contractions, followed by a one-dimensional Gaussian density bound, give a normalized small-ball estimate whose constant grows linearly with the degree. Applying this estimate and Markov’s inequality at geometric thresholds makes the failure probabilities summable over the determinant coefficients. The resulting simultaneous coefficient bounds yield a single dilation for the whole determinant. Gaussian conditioning also underlies Barvinok’s estimates for determinant estimators of mixed discriminants [3 Sections 3–4]. The estimate used here exploits alternation to control all exterior degrees in the same comparison.

The deterministic operator in the Wiener-area application is the skew-adjoint part of the upper triangular truncation of a rank-one positive operator. It therefore belongs to the classical triangular-projection framework [12]; the relevant weak-type estimate is recalled in [1 Section 8]. The associated rank-one perturbation has the classical triangular form recalled in [2 Section 2.3]. We use a direct determinant calculation for this model to identify the mean determinant and obtain the covariance summability criteria.

Notation and preliminaries. Throughout, Hilbert spaces are real and separable. Their inner products and norms are denoted by ⟨⋅,⋅⟩\langle\cdot,\cdot\rangle and ∥⋅∥\lVert\cdot\rVert. For a bounded operator TT, let T∗T^* denote its adjoint and ∥T∥op\lVert T\rVert_{\mathrm{op}} its operator norm; II denotes the identity on the space in question, and tr⁡\operatorname{tr} denotes the trace. We use ⊗\otimes for the Hilbert tensor product and identify x⊗yx\otimes y with the rank-one operator h↦⟨x,h⟩yh\mapsto\langle x,h\rangle y. For a compact operator TT between Hilbert spaces, let s1(T)≥s2(T)≥⋯≥0s_1(T)\ge s_2(T)\ge\cdots\ge0 be its singular values, counted with multiplicity and padded with zeros if necessary. For p>0p>0, set

∥T∥pp=∑j≥1sj(T)p∈[0,∞),Sp={T:∥T∥p<∞}.\lVert T\rVert_p^p=\sum_{j\ge1}s_j(T)^p\in[0,\infty),\qquad S_p=\{T:\lVert T\rVert_p<\infty\}.

For 0<p<10<p<1, the quantity ∥T∥p\lVert T\rVert_p is a quasi-norm; for p≥1p\ge1, it is a norm on SpS_p. The cases p=1p=1 and p=2p=2 are the trace-class and Hilbert–Schmidt ideals, respectively. Probability and expectation are denoted by P\mathbb{P} and E\mathbb{E}, and 1D\mathbf{1}_D denotes the indicator of a set DD.

Determinants of real operators are taken after complexification. For trace-class TT, let det⁡(I+T)\det(I+T) denote the Fredholm determinant. For Hilbert–Schmidt TT, define the regularized Fredholm determinant by

det⁡2(I+T):=det⁡((I+T)e−T).\det_2(I+T):=\det\bigl((I+T)e^{-T}\bigr).

The operator (I+T)e−T−I(I+T)e^{-T}-I is trace class, so the right-hand side is defined. We use the standard continuity and product properties of these determinants [17].

For a Hilbert space EE and an integer r≥0r\ge0, let ⋀rE\bigwedge^r E denote the Hilbert completion of its algebraic exterior power, normalized so that increasing wedges of an orthonormal basis are orthonormal; set ⋀0E=R\bigwedge^0E=\mathbb{R}. For nonnegative integers r,sr,s, exterior multiplication extends to a continuous bilinear map ⋀rE×⋀sE→⋀r+sE\bigwedge^rE\times\bigwedge^sE\to\bigwedge^{r+s}E, and

∥u∧v∥≤(r+sr) ∥u∥∥v∥,u∈⋀rE,v∈⋀sE.\lVert u\wedge v\rVert\le\sqrt{\binom{r+s}{r}}\,\lVert u\rVert\lVert v\rVert,\qquad u\in\bigwedge^rE,\quad v\in\bigwedge^sE.

This estimate follows by orthogonal antisymmetrization in E⊗(r+s)E^{\otimes(r+s)}. For a bounded operator R:E→FR:E\to F between Hilbert spaces, its induced exterior power ⋀rR:⋀rE→⋀rF\bigwedge^rR:\bigwedge^rE\to\bigwedge^rF is determined by

(⋀rR)(x1∧⋯∧xr)=Rx1∧⋯∧Rxr.(\bigwedge^rR)(x_1\wedge\cdots\wedge x_r)=Rx_1\wedge\cdots\wedge Rx_r.

In degree zero, ⋀0R\bigwedge^0R is the identity on R\mathbb{R}. We write u∧ku^{\wedge k} for the kk-fold exterior product of uu, with u∧0=1u^{\wedge0}=1. In degree two, let u↦u^u\mapsto\widehat{u} denote the isometric alternating embedding ⋀2E↪E⊗E\bigwedge^2E\hookrightarrow E\otimes E determined by

x∧y^=12(x⊗y−y⊗x).\widehat{x\wedge y}=\frac{1}{\sqrt{2}}(x\otimes y-y\otimes x).

The assignment x∧y↦x⊗y−y⊗xx\wedge y\mapsto x\otimes y-y\otimes x extends linearly and continuously to an isomorphism from ⋀2E\bigwedge^2E onto the skew-adjoint Hilbert–Schmidt operators on EE. For such an operator TT, let ωT\omega_T denote the corresponding two-form. Then ∥T∥22=2∥ωT∥2\lVert T\rVert_2^2=2\lVert\omega_T\rVert^2. If the nonzero eigenvalue pairs of TT are ±iσj\pm i\sigma_j, with σj>0\sigma_j>0, define its determinant function, with scalar parameter λ\lambda, by

ZT(λ):=det⁡2(I+λT)=∑k≥0∥ωT∧kk!∥2λ2k=∏j(1+λ2σj2),λ≥0.(1)Z_T(\lambda):=\det_2(I+\lambda T)=\sum_{k\ge0}\left\lVert\frac{\omega_T^{\wedge k}}{k!}\right\rVert^2\lambda^{2k}=\prod_j(1+\lambda^2\sigma_j^2),\qquad\lambda\ge0. \tag*{(1)}

Indeed, for a suitable orthonormal family (ej)(e_j), the orthogonal block representation is ωT=∑jσje2j−1∧e2j\omega_T=\sum_j \sigma_j e_{2j-1}\wedge e_{2j}, and each exterior coefficient is the corresponding elementary symmetric function of (σj2)(\sigma_j^2). The series and product define the same entire function, also denoted by ZTZ_T, of the complex parameter.

Pathwise dilation for Gaussian exterior forms

Let (Ω,F,P)(\Omega,\mathcal{F},\mathbb{P}) be a probability space. Let H,KH,K be Hilbert spaces, and let Q≥0Q\ge0 be trace class on HH. Let G:K→L2(Ω;H)G:K\to L^2(\Omega;H) be linear, with the scalar family {⟨Gf,h⟩:f∈K, h∈H}\{\langle Gf,h\rangle:f\in K,\ h\in H\} centered and jointly Gaussian. Assume

E[Gf⊗Gg]=⟨f,g⟩Q,f,g∈K.(2)\mathbb{E}[Gf\otimes Gg]=\langle f,g\rangle Q,\qquad f,g\in K. \tag*{(2)}

Gaussian fourth moments imply Gf∧Gg∈L2(Ω;Λ2H)Gf\wedge Gg\in L^2(\Omega;\Lambda^2H) for f,g∈Kf,g\in K. The alternating bilinear map (f,g)↦Gf∧Gg(f,g)\mapsto Gf\wedge Gg therefore induces a linear map from the algebraic exterior square of KK into L2(Ω;Λ2H)L^2(\Omega;\Lambda^2H), which we call Gaussian evaluation. No samplewise bounded operator G:K→HG:K\to H is assumed. Write (qi)(q_i) for the finite or countable sequence of positive eigenvalues of QQ, counted with multiplicity. For integers r≥1r\ge1, set

S=tr⁡Q=∑iqi,er(q)=∑i1<⋯<irqi1⋯qir=tr⁡(ΛrQ),e0(q)=1.S=\operatorname{tr}Q=\sum_i q_i,\qquad e_r(q)=\sum_{i_1<\cdots<i_r}q_{i_1}\cdots q_{i_r}=\operatorname{tr}(\Lambda^rQ),\qquad e_0(q)=1.

An empty sum is zero. These quantities are finite and satisfy r!er(q)≤Srr!e_r(q)\le S^r.

Theorem 1 (Pathwise dilation). Gaussian evaluation has a unique continuous linear extension from Λ2K\Lambda^2K to L2(Ω;Λ2H)L^2(\Omega;\Lambda^2H). For each fixed η∈Λ2K\eta\in\Lambda^2K, let ωA\omega_A be its image and let AA be the associated skew-adjoint operator. Then A∈L2(Ω;S2)A\in L^2(\Omega;\mathcal{S}_2), and

P(λ):=EZA(λ)=∑k≥0(2k)!e2k(q)∥η∧kk!∥2λ2k,λ≥0.(3)P(\lambda):=\mathbb{E}Z_A(\lambda)=\sum_{k\ge0}(2k)!e_{2k}(q)\left\|\frac{\eta^{\wedge k}}{k!}\right\|^2\lambda^{2k},\qquad\lambda\ge0. \tag*{(3)}

The series on the right defines an entire function. With inf⁡∅=∞\inf\varnothing=\infty, define the least pathwise dilation by

L∗:=inf⁡{ℓ≥1:P(λ/ℓ)≤ZA(λ)≤P(ℓλ) for every λ≥0}.(4)L_*:=\inf\{\ell\ge1:P(\lambda/\ell)\le Z_A(\lambda)\le P(\ell\lambda)\text{ for every }\lambda\ge0\}. \tag*{(4)}

Then L∗L_* is a measurable [1,∞][1,\infty]-valued random variable and satisfies

P(L∗>ℓ)≤9ℓ,ℓ≥1,(5)\mathbb{P}(L_*>\ell)\le\frac{9}{\ell},\qquad\ell\ge1, \tag*{(5)}

with a constant independent of QQ and η\eta. For 0<p<20<p<2, define the deterministic quantity Np∈[0,∞]N_p\in[0,\infty] by

Npp=2psin⁡(πp/2)π∫0∞log⁡P(λ)λp+1 dλ.N_p^p=\frac{2p\sin(\pi p/2)}{\pi}\int_0^\infty\frac{\log P(\lambda)}{\lambda^{p+1}}\,d\lambda.

There is an event Ω0∈F\Omega_0\in\mathcal{F} with P(Ω0)=1\mathbb{P}(\Omega_0)=1 such that, on Ω0\Omega_0, 1≤L∗<∞1\le L_*<\infty and both inequalities in (4) hold with ℓ=L∗\ell=L_*. On the same event,

L∗−1Np≤∥A∥p≤L∗Npfor every 0<p<2,(6)L_*^{-1}N_p\le\|A\|_p\le L_*N_p\qquad\text{for every }0<p<2, \tag*{(6)}

with infinite values allowed, and A∈SpA\in\mathcal{S}_p for every $p\ge 2.

Proof. Step 1. Gaussian evaluation and its mean determinant. Let HQ=ran⁡Q‾H_Q=\overline{\operatorname{ran}Q}. Equation (2) implies that Gf∈L2(Ω;HQ)Gf\in L^2(\Omega;H_Q) and E∥Gf∥2=S∥f∥2\mathbb{E}\|Gf\|^2=S\|f\|^2. Choose an orthonormal eigenbasis (ei)(e_i) of HQH_Q with Qei=qieiQe_i=q_ie_i, and put ξi(f)=qi−1/2⟨Gf,ei⟩\xi_i(f)=q_i^{-1/2}\langle Gf,e_i\rangle. With δij\delta_{ij} denoting the Kronecker delta, we have

E[ξi(f)ξj(g)]=δij⟨f,g⟩.\mathbb{E}[\xi_i(f)\xi_j(g)]=\delta_{ij}\langle f,g\rangle.

Joint Gaussianity implies that the ξi\xi_i are independent isonormal Gaussian processes over KK: each is a centered Gaussian linear map from KK to L2(Ω)L^2(\Omega) with covariance ⟨f,g⟩\langle f,g\rangle. Moreover, Gf=∑iqi ξi(f)eiGf=\sum_i\sqrt{q_i}\,\xi_i(f)e_i in L2(Ω;H)L^2(\Omega;H). For background on isonormal processes, see [10].

Let ξ\xi be an isonormal Gaussian process over KK. For an integer m≥1m\ge1 and f∈Kf\in K, let ιf:ΛmK→Λm−1K\iota_f:\Lambda^mK\to\Lambda^{m-1}K be the adjoint of left exterior multiplication w↦f∧ww\mapsto f\wedge w. For v∈ΛmKv\in\Lambda^mK and an orthonormal basis (fa)(f_a) of KK, define

ιξv=∑aξ(fa)ιfavin L2(Ω;Λm−1K),E∥ιξv∥2=m∥v∥2.(7)\iota_\xi v=\sum_a\xi(f_a)\iota_{f_a}v\quad\text{in }L^2(\Omega;\Lambda^{m-1}K),\qquad\mathbb{E}\|\iota_\xi v\|^2=m\|v\|^2. \tag*{(7)}

To check convergence, expand vv in an orthonormal wedge basis. Each squared coefficient occurs exactly mm times in ∑a∥ιfav∥2\sum_a\|\iota_{f_a}v\|^2. The series is basis independent by the isonormal isometry applied to the Hilbert--Schmidt map f↦f∧vf\mapsto f\wedge v. If W∈L2(Ω;ΛmK)W\in L^2(\Omega;\Lambda^mK) is measurable with respect to a σ\sigma-field G⊂F\mathcal{G}\subset\mathcal{F} independent of ξ\xi, the same partial sums define ιξW\iota_\xi W in L2(Ω;Λm−1K)L^2(\Omega;\Lambda^{m-1}K) and satisfy

E[∥ιξW∥2∣G]=m∥W∥2.\mathbb{E}[\|\iota_\xi W\|^2\mid\mathcal{G}]=m\|W\|^2.

Conditionally on G\mathcal{G}, ιξW\iota_\xi W is centered Gaussian on Λm−1K\Lambda^{m-1}K with covariance ∑a(ιfaW)⊗(ιfaW)\sum_a(\iota_{f_a}W)\otimes(\iota_{f_a}W). This series converges in trace norm almost surely and has trace m∥W∥2m\|W\|^2. For an integer r≥1r\ge1, distinct indices $i_1,\ldots,i_r from the chosen covariance eigenbasis, and u∈ΛrKu \in\Lambda^r K, define the successive contractions by this conditional construction. At step 1≤j≤r1 \le j \le r, the input has degree r−j+1r-j+1 and is measurable with respect to the processes used in the preceding steps, hence independent of ξij\xi_{i_j}. Iterating the conditional second-moment identity gives

E∣ιξir⋯ιξi1u∣2=r!∥u∥2.(8)\mathbb{E}\left|\iota_{\xi_{i_r}}\cdots\iota_{\xi_{i_1}}u\right|^2=r!\lVert u\rVert^2. \tag*{(8)}

Define the Gaussian evaluation of η\eta by

ωA:=∑i<jqiqj (ιξjιξiη)ei∧ejin L2(Ω;Λ2HQ).\omega_A:=\sum_{i<j}\sqrt{q_iq_j}\,(\iota_{\xi_j}\iota_{\xi_i}\eta)e_i\wedge e_j\quad\text{in }L^2(\Omega;\Lambda^2H_Q).

The convergence follows from (8) with r=2r=2 and the finiteness of e2(q)e_2(q). Regard ωA\omega_A as an element of L2(Ω;Λ2H)L^2(\Omega;\Lambda^2H), and let AA be its associated skew-adjoint operator. The same identity gives

E∥ωA∥2=2e2(q)∥η∥2,E∥A∥22=4e2(q)∥η∥2.(9)\mathbb{E}\lVert\omega_A\rVert^2=2e_2(q)\lVert\eta\rVert^2,\qquad\mathbb{E}\lVert A\rVert_2^2=4e_2(q)\lVert\eta\rVert^2. \tag*{(9)}

Each coordinate depends linearly on η\eta, so (9) also applies to differences of inputs. On the algebraic exterior square, expansion of the coordinates identifies this map with the substitution f∧g↦Gf∧Ggf\wedge g\mapsto Gf\wedge Gg. Density therefore proves that the construction is the unique continuous extension of Gaussian evaluation. In particular, it is independent of the chosen eigenbasis.

For each integer k≥0k\ge0, set uk=η∧k/k!u_k=\eta^{\wedge k}/k!. For k≥1k\ge1 and an increasing multi-index I=(i1<⋯<i2k)I=(i_1<\cdots<i_{2k}), write eI=ei1∧⋯∧ei2ke_I=e_{i_1}\wedge\cdots\wedge e_{i_{2k}} and qI=∏i∈Iqiq_I=\prod_{i\in I}q_i. We claim that

⟨ωA∧kk!,eI⟩=qI ιξi2k⋯ιξi1uk(10)\left\langle\frac{\omega_A^{\wedge k}}{k!},e_I\right\rangle=\sqrt{q_I}\,\iota_{\xi_{i_{2k}}}\cdots\iota_{\xi_{i_1}}u_k \tag*{(10)}

almost surely, simultaneously for all such pairs (k,I)(k,I). For k=1k=1, this is the definition of ωA\omega_A. If η\eta is supported on a finite-dimensional subspace E⊂KE\subset K, choose representatives of the images under GG of a basis of EE. Outside a single null set, these vectors define a linear map E→HE\to H. Its induced maps on exterior powers preserve exterior products, and determinant expansion gives (10) for every k≥1k\ge1 and every increasing multi-index II of length 2k2k. If 2k>dim⁡E2k>\dim E, both sides vanish.

For general η\eta, choose finite-rank orthogonal projections PNP_N increasing strongly to IKI_K. For every fixed integer r≥1r\ge1, convergence on simple wedges and the bound ∥ΛrPN∥op≤1\lVert\Lambda^rP_N\rVert_{\mathrm{op}}\le1 imply, by density, that ΛrPN\Lambda^rP_N converges strongly to the identity on ΛrK\Lambda^rK. Set ηN=(Λ2PN)η\eta_N=(\Lambda^2P_N)\eta and denote its Gaussian image by ωAN\omega_{A_N}. Equation (9) gives

E∥ωAN−ωA∥2=2e2(q)∥ηN−η∥2⟶0.\mathbb{E}\lVert\omega_{A_N}-\omega_A\rVert^2=2e_2(q)\lVert\eta_N-\eta\rVert^2\longrightarrow0.

For each fixed kk, continuity of exterior multiplication gives ηN∧k/k!→uk\eta_N^{\wedge k}/k!\to u_k in Λ2kK\Lambda^{2k}K. Thus the right-hand side of (10), with η\eta replaced by ηN\eta_N, converges in L2L^2 by (8). For fixed II, the left-hand side is a polynomial in the finitely many coordinates of ωAN\omega_{A_N} indexed by pairs of elements of II. It therefore converges in probability to the corresponding coordinate of ωA∧k/k!\omega_A^{\wedge k}/k!. Uniqueness of limits in probability proves (10) almost surely for this pair (k,I)(k,I). Since there are only countably many such pairs, the identities hold on a single event of probability one for the fixed input η\eta.

Squaring and summing (10), then applying Tonelli, yields

E∥ωA∧kk!∥2=(2k)!e2k(q)∥uk∥2.\mathbb{E}\left\lVert\frac{\omega_A^{\wedge k}}{k!}\right\rVert^2=(2k)!e_{2k}(q)\lVert u_k\rVert^2.

Equation (1) and another application of Tonelli prove (3). Since (2k)!e2k(q)≤S2k(2k)!e_{2k}(q)\le S^{2k}, the product representation in (1), applied to the skew-adjoint operator associated with η\eta, gives, for z∈Cz\in\mathbb{C},

∑k≥0(2k)!e2k(q)∥uk∥2∣z∣2k≤∑k≥0∥uk∥2(S∣z∣)2k≤exp⁡(S2∣z∣2∥η∥2).\sum_{k\ge0}(2k)!e_{2k}(q)\lVert u_k\rVert^2|z|^{2k}\le\sum_{k\ge0}\lVert u_k\rVert^2(S|z|)^{2k}\le\exp(S^2|z|^2\lVert\eta\rVert^2).

Here the last inequality uses log⁡(1+x)≤x\log(1+x)\le x. The series therefore converges uniformly on compact subsets of C\mathbb{C} and defines an entire function.

Step 2. A small-ball estimate for the exterior coefficients. Fix 0≠v∈ΛmK0\ne v\in\Lambda^mK, where m≥2m\ge2, and put V=ιξvV=\iota_\xi v. The operator Tv:K→Λm−1KT_v:K\to\Lambda^{m-1}K defined by Tvf=ιfvT_vf=\iota_fv satisfies

∥Tv∥op≤∥v∥,∥Tv∥22=m∥v∥2.\lVert T_v\rVert_{\mathrm{op}}\le\lVert v\rVert,\qquad\lVert T_v\rVert_2^2=m\lVert v\rVert^2.

For f≠0f\ne0, complete f/∥f∥f/\lVert f\rVert to an orthonormal basis of KK. Contraction removes this basis vector from each wedge in which it occurs, giving ∥ιfv∥≤∥f∥∥v∥\lVert\iota_fv\rVert\le\lVert f\rVert\lVert v\rVert. The Hilbert--Schmidt identity follows from (7). The covariance of VV is C=TvTv∗C=T_vT_v^*, so

tr⁡C=m∥v∥2,0≤C≤∥v∥2I.(11)\operatorname{tr}C=m\lVert v\rVert^2,\qquad0\le C\le\lVert v\rVert^2I. \tag*{(11)}

Thus the exterior degree bounds from below the ratio of the covariance trace to its operator norm. This relation gives the inverse-norm estimate below.

First suppose that KK is finite dimensional. For ε>0\varepsilon> 0, set Rε=(∥V∥2+ε2)1/2R_\varepsilon= (\lVert V\rVert^2+\varepsilon^2)^{1/2}. Gaussian integration by parts, applied to the smooth vector field x/(∥x∥2+ε2)1/2x/(\lVert x\rVert^2+\varepsilon^2)^{1/2}, gives

E∥V∥2Rε=E[tr⁡CRε−⟨V,CV⟩Rε3]≥(m−1)∥v∥2ERε−1.\mathbb{E}\frac{\lVert V\rVert^2}{R_\varepsilon} = \mathbb{E}\left[\frac{\operatorname{tr} C}{R_\varepsilon}-\frac{\langle V,CV\rangle}{R_\varepsilon^3}\right] \ge(m-1)\lVert v\rVert^2\mathbb{E}R_\varepsilon^{-1}.

Here we used (11) and ∥V∥2≤Rε2\lVert V\rVert^2\le R_\varepsilon^2. The left-hand side is at most E∥V∥≤m ∥v∥\mathbb{E}\lVert V\rVert\le\sqrt{m}\,\lVert v\rVert. Monotone convergence as ε↓0\varepsilon\downarrow0 therefore proves

E∥ιξv∥−1≤m(m−1)∥v∥.(12)\mathbb{E}\lVert\iota_\xi v\rVert^{-1}\le\frac{\sqrt{m}}{(m-1)\lVert v\rVert}. \tag*{(12)}

We interpret the reciprocal of zero as +∞+\infty. The calculation also applies to singular covariance, since integration by parts may be carried out in the underlying independent Gaussian coordinates.

For general KK, let vN=(ΛmPN)vv_N=(\Lambda^mP_N)v, with the finite-rank projections used above. Then vN→vv_N\to v in ΛmK\Lambda^mK, and (7) gives ιξvN→ιξv\iota_\xi v_N\to\iota_\xi v in L2L^2. For all sufficiently large NN, vN≠0v_N\ne0. The finite-dimensional estimate applies to vNv_N and the restriction of ξ\xi to PNKP_NK. Along an almost surely convergent subsequence, Fatou’s lemma and ∥vN∥→∥v∥\lVert v_N\rVert\to\lVert v\rVert give (12). In particular, ιξv≠0\iota_\xi v\ne0 almost surely.

Fix an integer r≥2r\ge2 and a nonzero vector u∈ΛrKu\in\Lambda^rK. On an auxiliary probability space, let ζ1,…,ζr\zeta_1,\ldots,\zeta_r be independent isonormal Gaussian processes over KK, and put

V0=u,Vj=ιζjVj−1(1≤j≤r).V_0=u,\qquad V_j=\iota_{\zeta_j}V_{j-1}\quad(1\le j\le r).

Let F0\mathcal{F}_0 be the trivial σ\sigma-field and let Fj\mathcal{F}_j be generated by ζ1,…,ζj\zeta_1,\ldots,\zeta_j. Each VjV_j is Fj\mathcal{F}_j-measurable and square integrable. We prove inductively that E∥Vj∥−1<∞\mathbb{E}\lVert V_j\rVert^{-1}<\infty for 0≤j≤r−10\le j\le r-1. This is immediate for V0=uV_0=u. Suppose that 1≤j≤r−11\le j\le r-1 and E∥Vj−1∥−1<∞\mathbb{E}\lVert V_{j-1}\rVert^{-1}<\infty. Then Vj−1≠0V_{j-1}\ne0 almost surely and has exterior degree r−j+1≥2r-j+1\ge2. Conditional on Fj−1\mathcal{F}_{j-1}, we may therefore apply (12) with m=r−j+1m=r-j+1, since ζj\zeta_j is independent of Fj−1\mathcal{F}_{j-1}. This gives

E[∥Vj∥−1∣Fj−1]≤r−j+1r−j∥Vj−1∥−1,1≤j≤r−1.\mathbb{E}\left[\lVert V_j\rVert^{-1}\mid\mathcal{F}_{j-1}\right]\le\frac{\sqrt{r-j+1}}{r-j}\lVert V_{j-1}\rVert^{-1},\qquad1\le j\le r-1.

Taking expectations proves the induction step. Iterating the estimates yields

E∥Vr−1∥−1≤1∥u∥∏m=2rmm−1=r!(r−1)!∥u∥.\mathbb{E}\lVert V_{r-1}\rVert^{-1}\le\frac{1}{\lVert u\rVert}\prod_{m=2}^{r}\frac{\sqrt{m}}{m-1}=\frac{\sqrt{r!}}{(r-1)!\lVert u\rVert}.

In particular, Vr−1≠0V_{r-1}\ne0 almost surely. For the final contraction, set F=VrF=V_r. Conditional on Fr−1\mathcal{F}_{r-1}, FF is a centered real Gaussian random variable with variance ∥Vr−1∥2>0\lVert V_{r-1}\rVert^2>0. Its conditional density gives

P{∣F∣≤a∣Fr−1}≤2πa∥Vr−1∥,a≥0.\mathbb{P}\{|F|\le a\mid\mathcal{F}_{r-1}\}\le\sqrt{\frac{2}{\pi}}\frac{a}{\lVert V_{r-1}\rVert},\qquad a\ge0.

Repeated application of (7) gives E∣F∣2=r!∥u∥2\mathbb{E}|F|^2=r!\lVert u\rVert^2. Combining the conditional density bound with the preceding inverse-moment estimate and setting U=∣F∣2/(r!∥u∥2)U=|F|^2/(r!\lVert u\rVert^2), we obtain

EU=1,P(U≤x)≤2π x,x≥0.(13)\mathbb{E}U=1,\qquad\mathbb{P}(U\le x)\le\sqrt{\frac{2}{\pi}}\,\sqrt{x},\qquad x\ge0. \tag*{(13)}

We next apply (13) to the coefficients of the determinant. For a fixed k≥1k\ge1 with positive mean coefficient in (3), both e2k(q)e_{2k}(q) and ∥uk∥\lVert u_k\rVert are positive. Define

UI=∣ιξi2k⋯ιξi1uk∣2(2k)!∥uk∥2,wI=qIe2k(q).U_I=\frac{\left|\iota_{\xi_{i_{2k}}}\cdots\iota_{\xi_{i_1}}u_k\right|^2}{(2k)!\lVert u_k\rVert^2},\qquad w_I=\frac{q_I}{e_{2k}(q)}.

For each II, (13) applies to UIU_I with r=2kr=2k and u=uku=u_k, since ξi1,…,ξi2k\xi_{i_1},\ldots,\xi_{i_{2k}} are independent isonormal processes. The weights are deterministic, nonnegative, and sum to one. By (10), the normalized coefficient is

Yk:=∥ωAk/k!∥2(2k)!e2k(q)∥uk∥2=∑IwIUI.Y_k:=\frac{\lVert\omega_A^{k}/k!\rVert^2}{(2k)!e_{2k}(q)\lVert u_k\rVert^2}=\sum_I w_IU_I.

For x>0x>0, the event Yk≤xY_k\le x implies that the indices with UI≤2xU_I\le2x carry at least half of the total weight. Hence

1{Yk≤x}≤2∑IwI1{UI≤2x}.\mathbf{1}_{\{Y_k\le x\}}\le2\sum_Iw_I\mathbf{1}_{\{U_I\le2x\}}.

Taking expectations in the preceding inequality and using Tonelli gives

P(Yk≤x)≤2∑IwIP(UI≤2x).\mathbb{P}(Y_k\le x)\le2\sum_Iw_I\mathbb{P}(U_I\le2x).

Apply (13) with r=2kr = 2k and use ∑IwI=1\sum_I w_I = 1. Also, EYk=∑IwIEUI=1\mathbb{E}Y_k = \sum_I w_I\mathbb{E}U_I = 1. Consequently,

EYk=1,P(Yk≤x)≤8kπx,x≥0.(14)\mathbb{E}Y_k = 1,\qquad\mathbb{P}(Y_k \le x) \le\frac{8k}{\sqrt{\pi}}\sqrt{x},\qquad x \ge0. \tag*{(14)}

The bound at x=0x = 0 follows by continuity of probability from above. No independence between the variables UIU_I is required. For every ℓ>1\ell> 1, the bounds in (14) at the thresholds x=ℓ−2kx = \ell^{-2k} are summable over kk. This is the lower-tail estimate used to compare all determinant coefficients by one dilation.

Step 3. The least admissible dilation. Choose a measurable representative of AA that is skew-adjoint and Hilbert–Schmidt at every outcome, setting it equal to zero on a null exceptional set. For each realization, let A\mathcal{A} be the set of admissible dilations in (4). Since PP is nondecreasing, A\mathcal{A} is upward closed. Continuity of PP and ZAZ_A implies that A\mathcal{A} is closed in [1,∞)[1,\infty) and that admissibility can be tested at nonnegative rational values of λ\lambda. Consequently, for every finite ℓ≥1\ell\ge1,

{L∗≤ℓ}=⋂λ∈Q≥0{P(λ/ℓ)≤ZA(λ)≤P(ℓλ)}.\{L_* \le\ell\} = \bigcap_{\lambda\in\mathbb{Q}_{\ge0}}\{P(\lambda/\ell) \le Z_A(\lambda) \le P(\ell\lambda)\}.

For each fixed λ≥0\lambda\ge0, the function T↦det⁡2(I+λT)T \mapsto\det{}_2(I+\lambda T) is continuous on S2\mathcal{S}_2. Every event in this countable intersection is therefore measurable, which proves the measurability of L∗L_*. If A\mathcal{A} is nonempty, its infimum belongs to A\mathcal{A}; hence L∗L_* is admissible whenever it is finite.

Let J\mathcal{J} be the set of indices k≥1k \ge1 for which the coefficient of λ2k\lambda^{2k} in PP is positive. By nonnegativity and countability, all coefficients with zero mean vanish on a common event of probability one. On that event, the bounds

ℓ−2k≤Yk≤ℓ2k,k∈J,\ell^{-2k} \le Y_k \le\ell^{2k},\qquad k\in\mathcal{J},

imply P(λ/ℓ)≤ZA(λ)≤P(ℓλ)P(\lambda/\ell) \le Z_A(\lambda) \le P(\ell\lambda) for every λ≥0\lambda\ge0, by summing the nonnegative power series term by term. Thus, for ℓ>1\ell> 1, Markov's inequality and (14) yield

P(L∗>ℓ)≤∑k∈J[P(Yk<ℓ−2k)+P(Yk>ℓ2k)]≤∑k≥1(ℓ−2k+8kπℓ−k)=1ℓ2−1+8π(ℓ−1)2.(15)\begin{aligned} \mathbb{P}(L_* > \ell) \le\sum_{k\in\mathcal{J}}\left[\mathbb{P}(Y_k < \ell^{-2k})+\mathbb{P}(Y_k > \ell^{2k})\right] \\ &\le\sum_{k\ge1}\left(\ell^{-2k}+\frac{8k}{\sqrt{\pi}}\ell^{-k}\right) \\ &= \frac{1}{\ell^2-1}+\frac{8}{\sqrt{\pi}(\ell-1)^2}. \tag*{(15)} \end{aligned}

For ℓ≥4\ell\ge4, the right side multiplied by ℓ\ell is at most 4/15+128/(9π)<94/15+128/(9\sqrt{\pi})<9. For 1≤ℓ<41\le\ell<4, use probability at most one. This proves (5), including almost sure finiteness. If J\mathcal{J} is empty, then A=0A=0 and P=ZA=1P=Z_A=1 almost surely, so L∗=1L_*=1.

Step 4. Simultaneous Schatten comparison. For a skew-adjoint T∈S2T\in\mathcal{S}_2, we first verify the integral representation

∥T∥pp=2psin⁡(πp/2)π∫0∞log⁡ZT(λ) dλλp+1,0<p<2.(16)\|T\|_p^p=\frac{2p\sin(\pi p/2)}{\pi}\int_0^\infty\log Z_T(\lambda)\,\frac{\mathrm{d}\lambda}{\lambda^{p+1}},\qquad0<p<2. \tag*{(16)}

Let {±iσj}\{\pm i\sigma_j\} be the nonzero eigenvalue pairs of TT. For λ≥0\lambda\ge0,

log⁡ZT(λ)=∑jlog⁡(1+λ2σj2),0≤log⁡ZT(λ)≤λ2∑jσj2=λ22∥T∥22.\log Z_T(\lambda)=\sum_j\log(1+\lambda^2\sigma_j^2),\qquad0\le\log Z_T(\lambda)\le\lambda^2\sum_j\sigma_j^2=\frac{\lambda^2}{2}\|T\|_2^2.

All summands are nonnegative. For s>0s>0 and 0<p<20<p<2, integration by parts after the substitution u=λsu=\lambda s gives

∫0∞log⁡(1+λ2s2) dλλp+1=2spp∫0∞u1−p1+u2 du=πsppsin⁡(πp/2).\int_0^\infty\log(1+\lambda^2s^2)\,\frac{\mathrm{d}\lambda}{\lambda^{p+1}} =\frac{2s^p}{p}\int_0^\infty\frac{u^{1-p}}{1+u^2}\,\mathrm{d}u =\frac{\pi s^p}{p\sin(\pi p/2)}.

The boundary terms vanish because p<2p<2 at zero and p>0p>0 at infinity; for s=0s=0, both sides are zero. Tonelli's theorem and ∥T∥pp=2∑jσjp\|T\|_p^p=2\sum_j\sigma_j^p now prove (16), with both sides allowed to be infinite.

Let Ω0\Omega_0 be the intersection of the probability-one events obtained in Steps 1 and 3. On this event, A∈S2A\in\mathcal{S}_2, L∗<∞L_*<\infty, and the inequalities in (4) hold with ℓ=L∗\ell=L_* for every λ≥0\lambda\ge0. Fix a realization in Ω0\Omega_0. Then

log⁡P(λ/L∗)≤log⁡ZA(λ)≤log⁡P(L∗λ)(λ≥0).\log P(\lambda/L_*)\le\log Z_A(\lambda)\le\log P(L_*\lambda)\qquad(\lambda\ge0).

For any p∈(0,2)p\in(0,2), integrate these nonnegative functions against λ−p−1 dλ\lambda^{-p-1}\,\mathrm{d}\lambda. The changes of variables λ=L∗u\lambda=L_*u in the lower integral and λ=u/L∗\lambda=u/L_* in the upper integral, together with (16), give

L∗−pNpp≤∥A∥pp≤L∗pNpp.L_*^{-p}N_p^p\le\|A\|_p^p\le L_*^pN_p^p.

Taking pp-th roots proves (6). These operations remain valid for infinite integrals because 1≤L∗<∞1\le L_*<\infty on the fixed realization. Since pp was arbitrary and the event was chosen independently of pp, the comparison holds simultaneously for every 0<p<20<p<2. On the same event, A∈SpA\in\mathcal{S}_p for every p≥2p\ge2 by S2⊂Sp\mathcal{S}_2\subset\mathcal{S}_p.

□

Remark (Sharpness of the tail exponent). The power ℓ−1\ell^{-1} in (5) is optimal. Take H=K=R2H=K=\mathbb{R}^{2}, let f1,f2f_{1},f_{2} be the standard orthonormal basis of KK, and set Q=IQ=I and η=f1∧f2\eta=f_{1}\wedge f_{2}. Let Δ\Delta be the determinant of a 2×22\times2 matrix with independent standard normal entries. Then

ZA(λ)=1+Δ2λ2,P(λ)=1+2λ2,L∗=max⁡{∣Δ∣/2,2/∣Δ∣}.Z_{A}(\lambda)=1+\Delta^{2}\lambda^{2},\qquad P(\lambda)=1+2\lambda^{2},\qquad L_{*}=\max\{|\Delta|/\sqrt{2},\sqrt{2}/|\Delta|\}.

An orthogonal change of the four Gaussian coordinates expresses Δ\Delta as

Δ=g12+g222−g32+g422.\Delta=\frac{g_{1}^{2}+g_{2}^{2}}{2}-\frac{g_{3}^{2}+g_{4}^{2}}{2}.

where g1,…,g4g_{1},\ldots,g_{4} are independent standard Gaussian variables. The two terms are independent exponential variables with mean one, so Δ\Delta has density e−∣x∣/2e^{-|x|/2} on R\mathbb{R}. For ℓ≥1\ell\geq1, the events ∣Δ∣<2/ℓ|\Delta|<\sqrt{2}/\ell and ∣Δ∣>2ℓ|\Delta|>\sqrt{2}\ell are disjoint. Consequently,

P(L∗>ℓ)=1−e−2/ℓ+e−2ℓ,ℓ≥1.\mathbb{P}(L_{*}>\ell)=1-e^{-\sqrt{2}/\ell}+e^{-\sqrt{2}\ell},\qquad\ell\geq1.

In particular, lim⁡ℓ→∞ℓP(L∗>ℓ)=2\lim_{\ell\to\infty}\ell\mathbb{P}(L_{*}>\ell)=\sqrt{2} and EL∗=∞\mathbb{E}L_{*}=\infty. The tail exponent in (5) is therefore optimal; the numerical constant 9 is not claimed to be optimal.

Remark (Comparison at large parameters). Suppose that PP has infinitely many nonzero coefficients. Then every coefficient of even degree is positive: vanishing of e2k(q)e_{2k}(q) or η∧k\eta^{\wedge k} forces all subsequent coefficients to vanish. Thus YkY_{k} is defined for every k≥1k\geq1. For every ℓ>1\ell>1, Markov’s inequality and (14) give

∑k≥1P{Yk1/(2k)∉[ℓ−1,ℓ]}<∞.\sum_{k\geq1}\mathbb{P}\left\{Y_{k}^{1/(2k)}\notin[\ell^{-1},\ell]\right\}<\infty.

Apply the first Borel–Cantelli lemma to a countable sequence ℓ↓1\ell\downarrow1. On one event of probability one,

Yk1/(2k)⟶1.Y_{k}^{1/(2k)}\longrightarrow1.

Intersect this event with the event of Theorem 1. For each resulting realization, fix any ℓ>1\ell>1 and choose 1<ℓ0<ℓ1<\ell_{0}<\ell. There is an integer k0≥1k_{0}\geq1 such that ℓ0−2k≤Yk≤ℓ2k\ell_{0}^{-2k}\leq Y_{k}\leq\ell^{2k} for every k≥k0k\geq k_{0}. Consequently, each of

P(ℓλ)−ZA(λ),ZA(λ)−P(λ/ℓ)P(\ell\lambda)-Z_{A}(\lambda),\qquad Z_{A}(\lambda)-P(\lambda/\ell)

has nonnegative coefficients in degrees at least 2k02k_{0}, and its coefficient of degree 2k02k_{0} is strictly positive. Its remaining terms form a polynomial of degree less than 2k02k_{0}. Thus both differences are nonnegative for all sufficiently large λ\lambda.

The conclusion holds on the same event for every real ℓ>1\ell>1; only the threshold in λ\lambda depends on the realization and on ℓ\ell. The infimum of the eventually admissible dilations is therefore one. The value one is eventually admissible if and only if ZA=PZ_{A}=P identically: equality on an interval extends to the whole complex plane by analyticity. The least global dilation also depends on the behavior at bounded parameter values.

Wiener area and covariance entropy

Let W=(Wt)0≤t≤1W=(W_{t})_{0\leq t\leq1} be an HH-valued Wiener process with W0=0W_{0}=0 and covariance Q≥0Q\geq0, Q∈S1Q\in\mathcal{S}_{1}, on a filtered probability space (Ω,F,(Ft)0≤t≤1,P)(\Omega,\mathcal{F},(\mathcal{F}_{t})_{0\leq t\leq1},\mathbb{P}) satisfying the usual conditions. Write (qi)(q_{i}) for the positive eigenvalues of QQ and S=tr⁡Q=∑iqiS=\operatorname{tr}Q=\sum_{i}q_{i}. Its skew area on [0,1][0,1] is

A=12∫01(Ws⊗dWs−dWs⊗Ws).(17)A=\frac{1}{2}\int_{0}^{1}(W_{s}\otimes dW_{s}-dW_{s}\otimes W_{s}). \tag*{(17)}

The first integral in (17) is an S2\mathcal{S}_{2}-valued Itô integral. Indeed, the Itô isometry gives

E∥∫01Ws⊗dWs∥22=∫01E∥Ws∥2tr⁡Q ds=(tr⁡Q)22<∞.\mathbb{E}\left\|\int_{0}^{1}W_{s}\otimes dW_{s}\right\|_{2}^{2}=\int_{0}^{1}\mathbb{E}\|W_{s}\|^{2}\operatorname{tr}Q\,ds=\frac{(\operatorname{tr}Q)^{2}}{2}<\infty.

The second integral is the adjoint of the first, since taking adjoints is a bounded real-linear isometry on S2\mathcal{S}_{2}. Thus AA is a skew-adjoint Hilbert–Schmidt operator. The Itô–Stratonovich correction is symmetric and cancels in the antisymmetric difference, so either convention gives the same area. Standard Hilbert-space stochastic integration is treated in [16].

Corollary 1 (Wiener area). For each fixed trace-class covariance QQ, there is a single event of probability one on which A∈SpA\in\mathcal{S}_{p} for every p>1p>1 and, simultaneously for every 0<p≤10<p\leq1,

A∈Sp⟺{∑iqip<∞,0<p<1,∑iqilog⁡(S/qi)<∞,p=1.(18)A\in\mathcal{S}_{p}\quad\Longleftrightarrow\quad \begin{cases} \sum_{i}q_{i}^{p}<\infty, & 0<p<1,\\[6pt] \sum_{i}q_{i}\log(S/q_{i})<\infty, & p=1. \end{cases} \tag*{(18)}

For Q=0Q=0, the entropy sum is defined to be zero. The mean determinant is

DQ(λ):=EZA(λ)=det⁡(I+λQ/2)+det⁡(I−λQ/2)2,λ≥0.(19)D_Q(\lambda):=\mathbb{E}Z_A(\lambda)=\frac{\det(I+\lambda Q/2)+\det(I-\lambda Q/2)}{2},\qquad\lambda\ge0. \tag*{(19)}

Choose an ordered orthonormal eigenbasis (ei)(e_i) of ran⁡Q\operatorname{ran}Q, with Qei=qieiQe_i=q_i e_i, in any ordering of the positive eigenvalues. Let sgn⁡\operatorname{sgn} denote the sign function, with sgn⁡(0)=0\operatorname{sgn}(0)=0. Define TQT_Q by its matrix entries (TQ)ij:=⟨TQej,ei⟩(T_Q)_{ij}:=\langle T_Qe_j,e_i\rangle:

(TQ)ij=12qiqjsgn⁡(j−i).(20)(T_Q)_{ij}=\frac{1}{2}\sqrt{q_iq_j}\operatorname{sgn}(j-i). \tag*{(20)}

Extend TQT_Q by zero on ker⁡Q\ker Q. This gives a skew-adjoint Hilbert–Schmidt operator. Then ZTQ=DQZ_{T_Q}=D_Q, and the nonzero singular values of TQT_Q, including multiplicities, are independent of the chosen ordered eigenbasis. The dilation factor L∗L_* defined by (4) with P=DQP=D_Q satisfies (5) and, on the same probability-one event,

L∗−1∥TQ∥p≤∥A∥p≤L∗∥TQ∥p,0<p<2.(21)L_*^{-1}\|T_Q\|_p\le\|A\|_p\le L_*\|T_Q\|_p,\qquad0<p<2. \tag*{(21)}

The deterministic operator also satisfies, with infinite values allowed,

max⁡{e2(q),1π∑iqilog⁡Sqi}≤∥TQ∥1≤e2(q)+1π∑iqilog⁡Sqi.(22)\max\left\{\sqrt{e_2(q)},\frac{1}{\pi}\sum_i q_i\log\frac{S}{q_i}\right\}\le\|T_Q\|_1\le\sqrt{e_2(q)}+\frac{1}{\pi}\sum_i q_i\log\frac{S}{q_i}. \tag*{(22)}

Proof. The case Q=0Q=0 is immediate; assume henceforth that S=tr⁡Q>0S=\operatorname{tr}Q>0.

Step 1. A rank-one determinant identity. Work on HQ=ran⁡QH_Q=\operatorname{ran}Q, equipped with the ordered eigenbasis fixed above. Both QQ and TQT_Q vanish on ker⁡Q\ker Q, so this restriction does not change their determinants. Set

B=2TQ,v=∑iqi ei,b=e2(q).B=2T_Q,\qquad v=\sum_i\sqrt{q_i}\,e_i,\qquad b=e_2(q).

Then

B∗=−B,∥v∥2=S,∥B∥22=∑i≠jqiqj=2b.B^*=-B,\qquad\|v\|^2=S,\qquad\|B\|_2^2=\sum_{i\ne j}q_iq_j=2b.

The rank-one perturbation B+v⊗vB+v\otimes v is upper triangular with diagonal entries qiq_i: adding v⊗vv\otimes v cancels each entry of BB below the diagonal. After complexification, i(B+v⊗v)i(B+v\otimes v) has the form of Livšic’s triangular model for operators with rank-one imaginary part; see [2 Section 2.3]. The determinant identity below will determine both the exterior coefficients of the time form and the trace-class criterion. For λ∈R\lambda\in\mathbb{R}, define

a(λ)=⟨v,(I−λ2B2)−1v⟩.a(\lambda)=\langle v,(I-\lambda^2B^2)^{-1}v\rangle.

This is well defined because I−λ2B2=I+λ2B∗B≥II-\lambda^2B^2=I+\lambda^2B^*B\ge I. We claim that

det⁡(I+λQ)=ZB(λ)[1+λa(λ)],λ∈R.(23)\det(I+\lambda Q)=Z_B(\lambda)[1+\lambda a(\lambda)],\qquad\lambda\in\mathbb{R}. \tag*{(23)}

Suppose first that QQ has finite rank. By the preceding triangularity and the matrix determinant lemma,

det⁡(I+λQ)=det⁡(I+λ(B+v⊗v))=det⁡(I+λB)[1+λ⟨v,(I+λB)−1v⟩].\begin{aligned} \det(I+\lambda Q)=\det(I+\lambda(B+v\otimes v)) \\ &=\det(I+\lambda B)\left[1+\lambda\langle v,(I+\lambda B)^{-1}v\rangle\right]. \end{aligned}

For real λ\lambda, the inverse exists because B∗=−BB^*=-B, and

(I+λB)−1=(I−λB)(I−λ2B2)−1.(I+\lambda B)^{-1}=(I-\lambda B)(I-\lambda^2B^2)^{-1}.

The resolvent (I−λ2B2)−1(I-\lambda^2B^2)^{-1} is self-adjoint and commutes with BB. Thus B(I−λ2B2)−1B(I-\lambda^2B^2)^{-1} is skew-adjoint and has zero quadratic form on real vectors. In this finite-dimensional calculation, tr⁡B=0\operatorname{tr}B=0, so det⁡(I+λB)=det⁡2(I+λB)\det(I+\lambda B)=\det_2(I+\lambda B). These identities prove (23) when QQ has finite rank.

Now suppose that QQ has infinite rank. Let ΠN\Pi_N be the orthogonal projection onto the first NN covariance coordinates and put QN=ΠNQΠNQ_N=\Pi_NQ\Pi_N, BN=ΠNBΠNB_N=\Pi_NB\Pi_N, and vN=ΠNvv_N=\Pi_Nv. Writing SN=∑i≤NqiS_N=\sum_{i\le N}q_i and bN=e2(q1,…,qN)b_N=e_2(q_1,\ldots,q_N), one has

∥Q−QN∥1=∥v−vN∥2=S−SN⟶0,∥B−BN∥22=2(b−bN)⟶0.\|Q-Q_N\|_1=\|v-v_N\|^2=S-S_N\longrightarrow0,\qquad\|B-B_N\|_2^2=2(b-b_N)\longrightarrow0.

In particular, BN2→B2B_N^2\to B^2 in operator norm. For real λ\lambda, both (I−λ2BN2)−1(I-\lambda^2B_N^2)^{-1} and (I−λ2B2)−1(I-\lambda^2B^2)^{-1} are positive contractions. For every fixed M>0M>0, their resolvent identity gives, uniformly for ∣λ∣≤M|\lambda|\le M,

∥(I−λ2BN2)−1−(I−λ2B2)−1∥op≤M2∥BN2−B2∥op⟶0.\left\|(I-\lambda^2B_N^2)^{-1}-(I-\lambda^2B^2)^{-1}\right\|_{\mathrm{op}}\le M^2\|B_N^2-B^2\|_{\mathrm{op}}\longrightarrow0.

The ordinary Fredholm determinants converge by trace-norm continuity, and the regularized determinants by Hilbert–Schmidt continuity [17]. Together with vN→vv_N\to v, these convergences allow passage to the limit in the finite-rank identity, proving (23) for every real λ\lambda.

Both ZBZ_B and aa are even, so taking even parts yields

ZB(λ)=det⁡(I+λQ)+det⁡(I−λQ)2.(24)Z_B(\lambda)=\frac{\det(I+\lambda Q)+\det(I-\lambda Q)}{2}. \tag*{(24)}

Both sides of (24) are entire, so the identity extends to every complex parameter.

Step 2. The exterior coefficients of the time form. Let K=L2(0,1)K=L^2(0,1) and define the integral operator JJ by

(Jf)(s)=∫01j0(s,t)f(t) dt,j0(s,t)=12sgn⁡(t−s),f∈K.(Jf)(s)=\int_0^1 j_0(s,t)f(t)\,\mathrm{d}t,\qquad j_0(s,t)=\frac{1}{2}\operatorname{sgn}(t-s),\qquad f\in K.

The kernel j0j_0 is square integrable and antisymmetric, so JJ is skew-adjoint and Hilbert–Schmidt. Let η0∈Λ2K\eta_0\in\Lambda^2K be the corresponding two-form. Under the identification (f⊗g)(s,t)=f(s)g(t)(f\otimes g)(s,t)=f(s)g(t) of K⊗KK\otimes K with L2((0,1)2)L^2((0,1)^2), the normalized alternating embedding satisfies η^0=j0/2\widehat{\eta}_0=j_0/\sqrt{2}. For each integer n≥1n\geq1 and 1≤i≤n1\leq i\leq n, set In,i=((i−1)/n,i/n]I_{n,i}=((i-1)/n,i/n]. Let PnP_n be the orthogonal projection of L2(0,1)L^2(0,1) onto the span of the normalized indicators fn,i=n1In,if_{n,i}=\sqrt{n}\mathbf{1}_{I_{n,i}}, 1≤i≤n1\leq i\leq n. Put Jn=PnJPnJ_n=P_nJP_n, regarded as an operator on L2(0,1)L^2(0,1) which vanishes on the orthogonal complement of this span. Its matrix is

⟨Jnfn,j,fn,i⟩=sgn⁡(j−i)2n.\langle J_nf_{n,j},f_{n,i}\rangle=\frac{\operatorname{sgn}(j-i)}{2n}.

The kernel of JnJ_n agrees with j0j_0 off the nn diagonal partition squares and is zero on those squares. Hence

∥J−Jn∥22=∑i=1n∫In,i×In,i∣j0(s,t)∣2 ds dt=14n.\lVert J-J_n\rVert_2^2=\sum_{i=1}^n\int_{I_{n,i}\times I_{n,i}}|j_0(s,t)|^2\,\mathrm{d}s\,\mathrm{d}t=\frac{1}{4n}.

Apply (24) in dimension nn with all covariance eigenvalues equal to 1/(2n)1/(2n). The matrix BB in that identity has entries sgn⁡(j−i)/(2n)\operatorname{sgn}(j-i)/(2n), so it is precisely the matrix of JnJ_n on ran⁡Pn\operatorname{ran}P_n. Hence

ZJn(z)=(1+z/(2n))n+(1−z/(2n))n2⟶cosh⁡(z/2)Z_{J_n}(z)=\frac{(1+z/(2n))^n+(1-z/(2n))^n}{2}\longrightarrow\cosh(z/2)

locally uniformly for z∈Cz\in\mathbb{C}. Hilbert–Schmidt convergence and continuity of the regularized determinant identify the limit as ZJ(z)Z_J(z). Comparing coefficients in (1) gives

(2k)!∥η0∧kk!∥2=4−k,k≥0.(25)(2k)!\left\lVert\frac{\eta_0^{\wedge k}}{k!}\right\rVert^2=4^{-k},\qquad k\geq0. \tag*{(25)}

Step 3. Identification with Wiener area. For each ii, define βi(t)=qi−1/2⟨Wt,ei⟩\beta_i(t)=q_i^{-1/2}\langle W_t,e_i\rangle, 0≤t≤10\leq t\leq1. These are independent standard Brownian motions, and

Wt=∑iqi βi(t)eiin L2(Ω;H),0≤t≤1.W_t=\sum_i\sqrt{q_i}\,\beta_i(t)e_i\quad\text{in }L^2(\Omega;H),\qquad0\leq t\leq1.

For f∈K=L2(0,1)f\in K=L^2(0,1), define

ξi(f)=∫01f(s) dβi(s),Gf=∑iqi ξi(f)ei.\xi_i(f)=\int_0^1 f(s)\,\mathrm{d}\beta_i(s),\qquad Gf=\sum_i\sqrt{q_i}\,\xi_i(f)e_i.

The series defining GfGf converges in L2(Ω;H)L^2(\Omega;H), since its second moment is S∥f∥2S\lVert f\rVert^2. The map GG satisfies (2), and Wt=G1[0,t]W_t=G\mathbf{1}_{[0,t]} in L2(Ω;H)L^2(\Omega;H) for every t∈[0,1]t\in[0,1].

For i≠ji\ne j, define the mixed Gaussian integral Iij\mathcal{I}_{ij} first on simple tensors f⊗gf\otimes g, with f,g∈Kf,g\in K, by

Iij(f⊗g)=ξi(f)ξj(g).\mathcal{I}_{ij}(f\otimes g)=\xi_i(f)\xi_j(g).

Independence and the isonormal isometry imply that Iij\mathcal{I}_{ij} extends uniquely to a linear isometry

Iij:L2((0,1)2)⟶L2(Ω).\mathcal{I}_{ij}:L^2((0,1)^2)\longrightarrow L^2(\Omega).

For every u∈Λ2Ku\in\Lambda^2K,

ιξjιξiu=2 Iij(u^).\iota_{\xi_j}\iota_{\xi_i}u=\sqrt{2}\,\mathcal{I}_{ij}(\widehat{u}).

This identity follows first for simple wedges and then by continuity. Under the tensor convention fixed above, the off-diagonal entries of (17) are

⟨Aej,ei⟩=qiqj2(∫01βi(t) dβj(t)−∫01βj(t) dβi(t))=qiqj Iij(1{s<t}−1{t<s}2)=qiqj Iij(j0).\begin{aligned} \langle Ae_j,e_i\rangle=\frac{\sqrt{q_iq_j}}{2}\left(\int_0^1\beta_i(t)\,\mathrm{d}\beta_j(t)-\int_0^1\beta_j(t)\,\mathrm{d}\beta_i(t)\right) \\ &=\sqrt{q_iq_j}\,\mathcal{I}_{ij}\left(\frac{\mathbf{1}_{\{s<t\}}-\mathbf{1}_{\{t<s\}}}{2}\right)=\sqrt{q_iq_j}\,\mathcal{I}_{ij}(j_0). \end{aligned}

The equality follows first for step-process approximations and then by the Itô isometry. Since 2η^0=j0\sqrt{2}\widehat{\eta}_0=j_0, these entries are precisely those of the Gaussian evaluation of η0\eta_0 in (10). The diagonal entries vanish, and both operators vanish on ker⁡Q\ker Q. Moreover,

∑i≠jE∣⟨Aej,ei⟩∣2=14∑i≠jqiqj=b2<∞.\sum_{i\ne j}\mathbb{E}|\langle Ae_j,e_i\rangle|^2=\frac{1}{4}\sum_{i\ne j}q_iq_j=\frac{b}{2}<\infty.

It follows that the Gaussian evaluation of η0\eta_0 under GG agrees with the Wiener area AA in L2(Ω;S2)L^2(\Omega;\mathcal{S}_2). For the double Wiener–Itô approach to Gaussian area, see also [6].

Equations (3), (25), and the even Fredholm expansion now give

EZA(λ)=∑k≥04−ke2k(q)λ2k=ZB(λ/2).\mathbb{E}Z_A(\lambda)=\sum_{k\geq0}4^{-k}e_{2k}(q)\lambda^{2k}=Z_B(\lambda/2).

This proves (19) and ZTQ=DQZ_{T_Q}=D_Q. By (1), the zeros of this determinant determine the nonzero singular values with their multiplicities. Since DQD_Q depends only on the unordered covariance spectrum, these singular values are independent of the chosen ordered eigenbasis. By (16), the identity ZTQ=DQZ_{T_Q}=D_Q gives Np=∥TQ∥pN_p=\lVert T_Q\rVert_p for every 0<p<20<p<2, with infinite values allowed. Intersect the event on which the Gaussian evaluation and the Wiener area agree with the event in Theorem 1. On this common event, (6) gives (21) simultaneously for every 0<p<20<p<2. The same dilation factor satisfies the tail bound (5). We work on this event for the remainder of the proof.

Step 4. Schatten membership and the trace-class endpoint. It remains to determine the Schatten classes containing TQT_Q and to prove (22). The nonzero singular values of BB occur in equal pairs, so ∥B∥op2≤∥B∥22/2=b\lVert B\rVert_{\mathrm{op}}^2\leq\lVert B\rVert_2^2/2=b. Consequently,

S1+bλ2≤a(λ)≤S,λ≥0.\frac{S}{1+b\lambda^2}\leq a(\lambda)\leq S,\qquad\lambda\geq0.

For λ≥0\lambda\geq0, define

RQ(λ)=log⁡det⁡(I+λQ)−log⁡(1+λS).R_Q(\lambda)=\log\det(I+\lambda Q)-\log(1+\lambda S).

The Fredholm product formula shows that both logarithms have derivative SS at zero. Equation (23) gives

log⁡ZB(λ)−RQ(λ)=log⁡1+λS1+λa(λ).\log Z_B(\lambda)-R_Q(\lambda)=\log\frac{1+\lambda S}{1+\lambda a(\lambda)}.

The bounds on a(λ)a(\lambda) imply

1≤1+λS1+λa(λ)≤(1+bλ2)(1+λS)1+bλ2+λS≤1+bλ2.1\leq\frac{1+\lambda S}{1+\lambda a(\lambda)} \leq\frac{(1+b\lambda^2)(1+\lambda S)}{1+b\lambda^2+\lambda S} \leq1+b\lambda^2.

Moreover, ZB(λ)≥1+bλ2Z_B(\lambda)\geq1+b\lambda^2, since its coefficients are nonnegative and the coefficient of λ2\lambda^2 is bb. Therefore

max⁡{RQ(λ),log⁡(1+bλ2)}≤log⁡DQ(2λ)≤RQ(λ)+log⁡(1+bλ2).(26)\max\{R_Q(\lambda),\log(1+b\lambda^2)\}\leq\log D_Q(2\lambda)\leq R_Q(\lambda)+\log(1+b\lambda^2). \tag*{(26)}

The function log⁡(1+bλ2)\log(1+b\lambda^2) is integrable against λ−p−1 dλ\lambda^{-p-1}\,d\lambda for every 0<p<20<p<2.

For each ii and λ≥0\lambda\geq0, set

hi(λ)=log⁡(1+λqi)−qiSlog⁡(1+λS).h_i(\lambda)=\log(1+\lambda q_i)-\frac{q_i}{S}\log(1+\lambda S).

The Fredholm product formula and ∑iqi=S\sum_i q_i=S give RQ(λ)=∑ihi(λ)R_Q(\lambda)=\sum_i h_i(\lambda). Each hih_i is nonnegative, since hi(0)=0h_i(0)=0 and

hi′(λ)=qi(S−qi)λ(1+λqi)(1+λS)≥0.h_i'(\lambda)=\frac{q_i(S-q_i)\lambda}{(1+\lambda q_i)(1+\lambda S)}\geq0.

Fix 0<p<20<p<2, and let Γ\Gamma be Euler’s gamma function, Γ(x)=∫0∞ux−1e−u du\Gamma(x)=\int_0^\infty u^{x-1}e^{-u}\,du for x>0x>0. If qi=Sq_i=S, then hi=0h_i=0. Otherwise 0<qi<S0<q_i<S. Since hi(λ)=∫0λhi′(t) dth_i(\lambda)=\int_0^\lambda h_i'(t)\,dt, Tonelli’s theorem and the identity

S−qi(1+qiλ)(1+Sλ)=∫qiSdu(1+uλ)2\frac{S-q_i}{(1+q_i\lambda)(1+S\lambda)}=\int_{q_i}^{S}\frac{du}{(1+u\lambda)^2}

give

∫0∞hi(λ)dλλp+1=1p∫0∞hi′(λ)λ−p dλ=qip∫qiS∫0∞λ1−p(1+uλ)2 dλ du=qiΓ(p)Γ(2−p)p∫qiSup−2 du.\begin{aligned} \int_0^\infty h_i(\lambda)\frac{d\lambda}{\lambda^{p+1}} =\frac{1}{p}\int_0^\infty h_i'(\lambda)\lambda^{-p}\,d\lambda\\ &=\frac{q_i}{p}\int_{q_i}^{S}\int_0^\infty\frac{\lambda^{1-p}}{(1+u\lambda)^2}\,d\lambda\,du \\ &=\frac{q_i\Gamma(p)\Gamma(2-p)}{p}\int_{q_i}^{S}u^{p-2}\,du. \end{aligned}

The inner integral equals up−2Γ(2−p)Γ(p)u^{p-2}\Gamma(2-p)\Gamma(p) by Euler’s beta integral. For p≠1p\ne1, use Γ(p)Γ(2−p)=(1−p)π/sin⁡(πp)\Gamma(p)\Gamma(2-p)=(1-p)\pi/\sin(\pi p); at p=1p=1, the last integral equals log⁡(S/qi)\log(S/q_i). We obtain

∫0∞hi(λ)dλλp+1={πpsin⁡(πp)(qip−qiSp−1),p≠1,qilog⁡(S/qi),p=1.(27)\int_0^\infty h_i(\lambda)\frac{d\lambda}{\lambda^{p+1}} = \begin{cases} \frac{\pi}{p\sin(\pi p)}\left(q_i^p-q_iS^{p-1}\right), & p\ne1,\\ q_i\log(S/q_i), & p=1. \end{cases} \tag*{(27)}

Since hi≥0h_i \ge0, Tonelli permits summation without any prior finiteness assumption:

∫0∞RQ(λ) dλλp+1={πpsin⁡(πp)(∑iqip−Sp),p≠1,∑iqilog⁡(S/qi),p=1,0<p<2.(28)\int_0^\infty R_Q(\lambda)\,\frac{d\lambda}{\lambda^{p+1}} = \begin{cases} \frac{\pi}{p\sin(\pi p)}\left(\sum_i q_i^p-S^p\right), & p\ne1,\\ \sum_i q_i\log(S/q_i), & p=1, \end{cases} \qquad0<p<2. \tag*{(28)}

For 0<p<10<p<1, this integral is finite exactly when ∑iqip<∞\sum_i q_i^p<\infty. At p=1p=1, it is the covariance entropy. For 1<p<21<p<2, it is always finite, since ∑iqip≤Sp\sum_i q_i^p\le S^p. Equations (26) and (16) therefore give the corresponding membership criteria for TQT_Q. The comparison (21) proves (18) and A∈SpA\in S_p for 1<p<21<p<2 on the same event. Hilbert–Schmidt membership gives A∈SpA\in S_p for every p≥2p\ge2.

At p=1p=1, (16) and a change of variables give

∥TQ∥1=1π∫0∞log⁡DQ(2λ) dλλ2,∫0∞log⁡(1+bλ2) dλλ2=πb.\lVert T_Q\rVert_1=\frac{1}{\pi}\int_0^\infty\log D_Q(2\lambda)\,\frac{d\lambda}{\lambda^2},\qquad \int_0^\infty\log(1+b\lambda^2)\,\frac{d\lambda}{\lambda^2}=\pi\sqrt{b}.

Integrating (26) and using (28) proves (22). □

Remark (Schatten orders greater than one). The conclusion for p>1p>1 also follows from the exactness criterion for tensor norms of Ledoux, Lyons and Qian [13 Definition 1 and Theorem 3]. Let HH be a Hilbert space, let Q≥0Q\ge0 be trace class on HH, and let Gk,Gk′G_k,G'_k, k≥1k\ge1, be mutually independent centered HH-valued Gaussian random variables with covariance QQ. For every integer N≥1N\ge1,

E∥∑k=1NGk⊗Gk′∥p≤Nmax⁡{1/p,1/2}tr⁡Q,1<p<∞.\mathbb{E}\left\lVert\sum_{k=1}^{N}G_k\otimes G'_k\right\rVert_p\le N^{\max\{1/p,1/2\}}\operatorname{tr}Q,\qquad1<p<\infty.

For 1<p<21<p<2, this follows by interpolation between the trace-norm bound Ntr⁡QN\operatorname{tr}Q and the Hilbert–Schmidt bound N1/2tr⁡QN^{1/2}\operatorname{tr}Q. For p≥2p\ge2, the Hilbert–Schmidt bound suffices. In every case the exponent of NN is strictly less than one, as required by the exactness criterion.

For S>0S>0, set pi=qi/Sp_i=q_i/S. The entropy in (18) then equals S∑ipilog⁡(1/pi)S\sum_i p_i\log(1/p_i). Thus, for each fixed covariance, the area is trace class almost surely if and only if the normalized covariance spectrum has finite entropy. The Hilbert–Schmidt norm satisfies

∥TQ∥22=E∥A∥2=12e2(q),e2(q)=2 ∥TQ∥2.\lVert T_Q\rVert_2^2=\mathbb{E}\lVert A\rVert^2=\frac{1}{2}e_2(q),\qquad \sqrt{e_2(q)}=\sqrt{2}\,\lVert T_Q\rVert_2.

If QQ has rank one, then e2(q)=0e_2(q)=0 and A=0A=0. If QQ has rank two, then ∥TQ∥1=q1q2\lVert T_Q\rVert_1=\sqrt{q_1q_2}.

Remark (Covariance spectra). For positive sequences (an)(a_n) and (bn)(b_n), write an≍bna_n\asymp b_n if there are constants 0<c≤C<∞0<c\le C<\infty such that cbn≤an≤Cbncb_n\le a_n\le Cb_n for all sufficiently large nn. Write the positive covariance eigenvalues as (qn)(q_n). For each fixed covariance considered below, all membership statements hold on the single probability-one event of Corollary 1.

(i) Suppose that

qn≍n−βfor some β>1.q_n\asymp n^{-\beta}\qquad\text{for some }\beta>1.

Then qnlog⁡(S/qn)≍n−βlog⁡nq_n\log(S/q_n)\asymp n^{-\beta}\log n, so the covariance entropy is finite. Moreover, ∑nqnp<∞\sum_n q_n^p<\infty exactly when βp>1\beta p>1. Consequently,

A∈Sp⟺p>1β,p>0.A\in S_p\quad\Longleftrightarrow\quad p>\frac{1}{\beta},\qquad p>0.

In particular, A∉S1/βA\notin S_{1/\beta}.

(ii) Suppose that

qn≍1n(log⁡n)α,n≥2,α>1.q_n\asymp\frac{1}{n(\log n)^\alpha},\qquad n\ge2,\quad\alpha>1.

The covariance is trace class, while qnlog⁡(S/qn)≍[n(log⁡n)α−1]−1q_n\log(S/q_n)\asymp[n(\log n)^{\alpha-1}]^{-1}. Thus

A∈S1⟺α>2.A\in S_1\quad\Longleftrightarrow\quad\alpha>2.

For every α>1\alpha>1, the area belongs to SpS_p for all p>1p>1 and to none of the classes SpS_p with 0<p<10<p<1, since ∑nqnp\sum_n q_n^p diverges for every 0<p<10<p<1.

Declaration of AI use

The author used ChatGPT (GPT-6.0; OpenAI) interactively in developing proofs and refining the presentation. The author has independently verified all proofs and takes full responsibility for the content of the manuscript.

References

School of Mathematical Sciences, University of Science and Technology of China, Hefei, Anhui 230026, China

Email address: [email protected]

References

  1. [1]Aleksandrov, A. B. and Peller, V. V. (2022). Triangular projection on S_p, 0 < p < 1, and related inequalities. Preprint. arXiv:2207.02975v1.arxiv.org/abs/2207.02975
  2. [2]Arlinskiĭ, Y. and Tsekanovskiĭ, E. (2006). Non-self-adjoint Jacobi matrices with a rank-one imaginary part. J. Funct. Anal. 241, no. 2, 383–438. doi:10.1016/j.jfa.2006.05.002. Author preprint: arXiv:math/0602033v1.DOI
  3. [3]Barvinok, A. (1999). Polynomial time algorithms to approximate permanents and mixed discriminants within a simply exponential factor. Random Structures Algorithms 14, no. 1, 29–61. doi:10.1002/(SICI)1098-2418(1999010)14:1<29::AID-RSA2>3.0.CO;2-X.doi.org/10.1002/(sici)1098-2418(1999010)14:1<29::aid-rsa2>3.0.co;2-x
  4. [4]Dela Cruz, D. J. and Oberhauser, H. (2026). The singular values of Lévy’s area matrix. Preprint. arXiv:2606.08928v1.arxiv.org/abs/2606.08928
  5. [5]Dereich, S. (2010). Rough paths analysis of general Banach space-valued Wiener processes. J. Funct. Anal. 258, no. 9, 2910–2936. doi:10.1016/j.jfa.2010.01.018. Author manuscript.DOI
  6. [6]Ferreiro-Castilla, A. and Utzet, F. (2011). Lévy area for Gaussian processes: A double Wiener–Itô integral approach. Statist. Probab. Lett. 81, no. 9, 1380–1391. doi:10.1016/j.spl.2011.04.015. Author preprint: arXiv:1007.2516v1 (2010).
  7. [7]Friz, P. K. and Hairer, M. (2020). A Course on Rough Paths: With an Introduction to Regularity Structures, 2nd ed., Universitext, Springer, Cham. doi:10.1007/978-3-030-41556-3.DOI
  8. [8]Grong, E., Nilssen, T. and Schmеding, A. (2022). Geometric rough paths on infinite dimensional spaces. J. Differential Equations 340, 151–178. doi:10.1016/j.jde.2022.08.034. Author version: arXiv:2006.06362v4 (2021).DOI
  9. [9]Helmes, K. and Schwane, A. (1983). Lévy’s stochastic area formula in higher dimensions. J. Funct. Anal. 54, no. 2, 177–192. doi:10.1016/0022-1236(83)90053-8.DOI
  10. [10]Janson, S. (1997). Gaussian Hilbert Spaces, Cambridge Tracts in Mathematics, vol. 129, Cambridge University Press, Cambridge. doi:10.1017/CBO9780511526169.DOI
  11. [11]Krée, P. (1986). A remark on Paul Lévy’s stochastic area formula. In Aspects of Mathematics and its Applications (Barroso, J. A., ed.), North-Holland Mathematical Library, vol. 34, North-Holland, Amsterdam, pp. 533–550. doi:10.1016/S0924-6509(09)70279-X.DOI
  12. [12]Kwapień, S. and Pełczyński, A. (1970). The main triangle projection in matrix spaces and its applications. Studia Math. 34, no. 1, 43–67. doi:10.4064/sm-34-1-43-67.DOI
  13. [13]Ledoux, M., Lyons, T. and Qian, Z. (2002). Lévy area of Wiener processes in Banach spaces. Ann. Probab. 30, no. 2, 546–578. doi:10.1214/aop/1023481002.DOI
  14. [14]Lévy, P. (1951). Wiener’s random function, and other Laplacian random functions. In Proceedings of the Second Berkeley Symposium on Mathematical Statistics and Probability (Neyman, J., ed.), University of California Press, Berkeley, pp. 171–187. Project Euclid.DOI
  15. [15]Lyons, T. J. (1998). Differential equations driven by rough signals. Rev. Mat. Iberoam. 14, no. 2, 215–310. doi:10.4171/RMI/240.DOI
  16. [16]Métivier, M. (1982). Semimartingales: A Course on Stochastic Processes, De Gruyter Studies in Mathematics, vol. 2, Walter de Gruyter, Berlin–New York. doi:10.1515/9783110845563.DOI
  17. [17]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.
  18. [18]Zhao, G. (2026). Sharp weak-trace regularity of stochastic cross integrals. Preprint, version 6, 14 October 2026. arXiv:2608.08756v6.arxiv.org/abs/2608.08756

Paper details

vlnc:W_spd3h3yx
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 2

downloads · 2

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

  • Mathematics

Not recorded for this paper

View statement
Looking for related papers…
Contents