betaFeedback

WorksMathematics

Gelfand numbers of Schatten class embeddings in the quasi-Banach range

Abstract

We prove Conjecture 1.4 of Prochno and Strzelecki [J. Approx. Theory 277 (2022), 105736] on Gelfand numbers of embeddings between finite-dimensional Schatten classes. For 0<p≤10<p\le1 and indices bounded away from the ambient dimension by a fixed proportion, we obtain the conjectured lower bound, uniformly in the target exponent 2≤q≤∞2\le q\le\infty. Together with the known upper estimate obtained by Hinrichs, Prochno and Vybíral [Math. Ann., 380(3-4):1563-1593, 2021], this gives the exact asymptotic order in the previously unresolved range. As a consequence, we obtain a complete all-index description by combining our result with the estimates of Hinrichs, Prochno and Vybíral and of Prochno and Strzelecki. The proof is deterministic and is based on Hilbert–Schmidt projections on rectangular matrix spaces.

Introduction & Main Results

The Schatten classes form one of the classical families of operator ideals in functional analysis. For a Hilbert space HH and 0<p<∞0 < p < \infty, the Schatten pp-class Sp\mathcal{S}_p consists of all compact operators on HH whose sequence of singular values belongs to ℓp\ell_p. We write S∞\mathcal{S}_\infty for the compact operators equipped with the operator norm. The cases p=1p = 1 and p=2p = 2 give the trace-class, or nuclear, operators and the Hilbert–Schmidt operators, respectively. The origins of the theory lie in Schatten’s work on cross-spaces of linear transformations [26] and his subsequent papers with von Neumann [28, 29]. These investigations formed part of the development of nuclear operators and tensor products, with related contributions beyond the Hilbert space setting by Ruston [25] and Grothendieck [9]. Schatten’s monograph [27] subsequently provided a systematic account of norm ideals in terms of symmetric gauge functions and unitarily invariant crossnorms. As he observed there, the classical study of individual compact operators had only comparatively recently been complemented by an interest in the structure of spaces of such operators.

This shift from operators to operator spaces naturally leads to questions about their geometry. For finite pp, the Schatten class Sp\mathcal{S}_p is a non-commutative counterpart of ℓp\ell_p, sharing its nested scale Sp⊂Sq\mathcal{S}_p \subset\mathcal{S}_q for p<qp < q, uniform convexity for 1<p<∞1 < p < \infty, and, in this range, duality with the conjugate exponent under the trace pairing; see [27, 31]. Nevertheless, the geometry of Schatten classes is not determined by the corresponding sequence-space results. Building on the results of Kwapień and Pełczyński [21], Gordon and Lewis [8] proved that, on an infinite-dimensional Hilbert space, the Banach spaces Sp\mathcal{S}_p, 1≤p≤∞1 \le p \le\infty, fail to have local unconditional structure whenever p≠2p \ne2. In finite dimensions, the geometry of Schatten unit balls has been investigated through isotropic constants. [18], concentration and thin-shell phenomena [11, 23], asymptotic volumes and volume ratios [15], and volumes of intersections [16]. A complementary probabilistic perspective is provided by Sanov-type large deviation principles for empirical eigenvalue and singular-value distributions of random matrices [17].

For approximation theory, a particularly relevant aspect of this geometry is the behaviour of the identity mappings SpN↪SqN\mathcal{S}_p^N \hookrightarrow\mathcal{S}_q^N on subspaces of prescribed codimension. This is quantified by their Gelfand numbers. Geometrically, these numbers measure how small a section of the domain unit ball can be made in the target norm. Their determination therefore requires understanding the singular-value structure of matrices in subspaces of prescribed codimension and how it governs the comparison of different Schatten quasi-norms. This brings the distinction between commutative and non-commutative geometry directly into the problem: sharp estimates for the embeddings ℓpN↪ℓqN\ell_p^N \hookrightarrow\ell_q^N do not, by themselves, settle the corresponding questions for matrices.

The geometry of low-codimensional matrix subspaces that underlies the study of Gelfand numbers also plays a central role in low-rank matrix recovery and completion, where Schatten norms and quasi-norms play a role analogous to that of sequence norms and quasi-norms in sparse recovery; see [1, 3, 5, 20, 24]. In particular, the range 0<p<10 < p < 1 is relevant to non-convex recovery by Schatten quasi-norm minimization. The connection with Gelfand numbers is more than a formal analogy: Chávez-Domínguez and Kutzarova [3] characterized the stability of such recovery procedures in terms of Gelfand widths of suitable Schatten class embeddings. Sharp estimates for these quantities thus describe both the geometry of matrix subspaces and limitations on stable recovery from linear measurements.

The study of general approximation quantities for finite-dimensional sequence spaces has a long history. For Gelfand numbers, fundamental contributions are due to Gluskin [7] and Garnaev and Gluskin [6], whereas the entropy numbers of embeddings between finite-dimensional ℓp\ell_p-spaces were studied by Schütt [30] and subsequently in [4, 32, 10, 19]. In the non-commutative setting, related approximation quantities for embeddings between Schatten classes had already been investigated by Carl and Defant [2] and Hinrichs and Michels [12], in addition to the work of Chávez-Domínguez and Kutzarova [3] discussed above. Hinrichs, Prochno and Vybíral [13] subsequently determined the asymptotic behaviour of the entropy numbers of SpN↪SqN\mathcal{S}_p^N \hookrightarrow\mathcal{S}_q^N for all 0<p,q≤∞0 < p,q \leq\infty, up to constants depending only on pp and qq. However, the corresponding description of Gelfand numbers proved less complete. In [14], the authors obtained asymptotically sharp estimates in a wide range of parameters, extending classical sequence-space results and complementing the earlier bounds. Prochno and Strzelecki [22] subsequently established matching bounds for Gelfand numbers in the intermediate codimension range for 0<p≤2≤q≤∞0 < p \leq2 \leq q \leq\infty and obtained new estimates for approximation and Kolmogorov numbers of Schatten class embeddings. Still, two gaps remained in the description of Gelfand numbers in [22]. One concerns embeddings from SpN\mathcal{S}_p^N, with 0<p<10 < p < 1, into SqN\mathcal{S}_q^N, with 2<q≤∞2 < q \leq\infty: sharp upper bounds were available, but a matching lower estimate throughout the range of indices bounded away from N2N^2 by a fixed proportion had not been established. The purpose of the present paper is to prove the lower estimate proposed in [22 Conjecture 1.4]. Together with the previously known bounds, including the boundary cases p=1p = 1 and q=2q = 2, this completes the asymptotic description of the Gelfand numbers for all indices in the range 0<p≤10 < p \leq1 and 2≤q≤∞2 \leq q \leq\infty.

Presentation of the main results

For 0<p≤∞0 < p \le\infty and N∈NN \in\mathbb{N}, let SpN\mathcal{S}_p^N denote the space of all real N×NN \times N matrices equipped with the Schatten pp-norm (or quasi-norm when 0<p<10 < p < 1), defined by

∥A∥p:={(∑j=1Nσj(A)p)1/p,0<p<∞,σ1(A),p=∞.\lVert A\rVert_p := \begin{cases} \left(\sum_{j=1}^{N}\sigma_j(A)^p\right)^{1/p}, & 0 < p < \infty,\\ \sigma_1(A), & p = \infty. \end{cases}

where σ1(A)≥⋯≥σN(A)≥0\sigma_1(A) \ge\cdots\ge\sigma_N(A) \ge0 are the singular values of AA, that is, the eigenvalues of (ATA)1/2(A^{\mathsf{T}}A)^{1/2}, counted with multiplicity. In particular, the Schatten ∞\infty-norm coincides with the operator norm of A:ℓ2N→ℓ2NA:\ell_2^N \to\ell_2^N. We write SpN↪SqN\mathcal{S}_p^N \hookrightarrow\mathcal{S}_q^N for the natural identity mapping between these spaces. Given a bounded linear operator T:X→YT:X \to Y between quasi-Banach spaces, its nn-th Gelfand number is defined by

cn(T)=inf⁡{∥T∣F∥:F⊂X closed and linear, codim⁡F<n}.(1)c_n(T)=\inf\{\lVert T|_F\rVert:F\subset X\text{ closed and linear},\ \operatorname{codim}F<n\}. \tag*{(1)}

Equivalently, the Gelfand numbers are given by (here and below, all infima over FF are taken over linear subspaces, which are automatically closed in the finite-dimensional setting)

cn(SpN↪SqN)=inf⁡F⊂RN×Ncodim⁡F<n sup⁡A∈F∖{0}∥A∥q∥A∥p,1≤n≤N2.c_n(\mathcal{S}_p^N\hookrightarrow\mathcal{S}_q^N)=\inf_{\substack{F\subset\mathbb{R}^{N\times N}\\\operatorname{codim}F<n}}\ \sup_{A\in F\setminus\{0\}}\frac{\lVert A\rVert_q}{\lVert A\rVert_p},\qquad1\le n\le N^2.

Throughout, 1/∞:=01/\infty:=0 and constants implicit in ≲\lesssim and ≍\asymp are independent of NN and nn. Dependence on fixed parameters is indicated by subscripts.

As already mentioned, the purpose of this paper is to determine the Gelfand numbers of SpN↪SqN\mathcal{S}_p^N\hookrightarrow\mathcal{S}_q^N in the range 0<p≤10<p\le1 and 2≤q≤∞2\le q\le\infty. Hinrichs, Prochno and Vybíral [14 Proposition 4.5] obtained upper estimates for all indices in this range. Prochno and Strzelecki subsequently proved matching lower estimates in the intermediate range of codimensions [22 Proposition 3.1]. For small and moderately large codimensions, the missing estimate was formulated in [22 Conjecture 1.4], which asserts that

cn(SpN↪SqN)≳p,qmin⁡{1,Nn}1/p−1/2,1≤n≤(1−c0)N2;(2)c_n(\mathcal{S}_p^N\hookrightarrow\mathcal{S}_q^N)\gtrsim_{p,q}\min\left\{1,\frac{N}{n}\right\}^{1/p-1/2},\qquad1\le n\le(1-c_0)N^2; \tag*{(2)}

here 0<p≤10<p\le1, 2≤q≤∞2\le q\le\infty, and c0∈(0,1)c_0\in(0,1) is the fixed universal constant denoted by cc in [22]. The endpoint p=1p=1 was already covered by [22 Proposition 3.2], and the case q=2q=2 was also known; see [22 Proposition 3.3] and the references therein. Thus, the missing part of (2) concerns 0<p<10<p<1 and q>2q>2.

Our main result establishes the conjectured lower estimate for any fixed proportion of the ambient dimension. We include the known matching upper estimate to emphasise that both bounds can be chosen uniformly in the target space index qq.

Theorem A. Let 0<p≤10<p\le1 and η∈(0,1)\eta\in(0,1). There exist constants cp,η>0c_{p,\eta}>0 and Cp>0C_p>0 such that, for every 2≤q≤∞2\le q\le\infty, every N∈NN\in\mathbb{N}, and every integer 1≤n≤(1−η)N21\le n\le(1-\eta)N^2,

cp,ηmin⁡{1,Nn}1/p−1/2≤cn(SpN↪SqN)≤Cpmin⁡{1,Nn}1/p−1/2.(3)c_{p,\eta}\min\left\{1,\frac{N}{n}\right\}^{1/p-1/2}\le c_n(\mathcal{S}_p^N\hookrightarrow\mathcal{S}_q^N)\le C_p\min\left\{1,\frac{N}{n}\right\}^{1/p-1/2}. \tag*{(3)}

The lower estimate in Theorem A is the contribution of this paper. Taking η=c0\eta=c_0 proves [22 Conjecture 1.4]. The upper estimate is known and follows from the corresponding bound with target space S2N\mathcal{S}_2^N (see [22 Proposition 3.3, with $q=2$]). In particular, its constant depends only on pp, whereas the lower-bound constant depends only on pp and η\eta. An explicit choice of cp,ηc_{p,\eta} is given in (10) below.

Remark 1.1 (Complete asymptotic description). Combining Theorem A with [14 Proposition 4.5] and [22 Proposition 3.1] yields a complete description of the Gelfand numbers for 0<p≤10 < p \leq1 and 2≤q≤∞2 \leq q \leq\infty. More precisely, with the universal constant c0∈(0,1)c_0 \in(0,1) introduced above, we have

cn(SpN↪SqN)≍p,q{min⁡{1,(Nn)1/p−1/2}:1≤n≤(1−c0)N2,max⁡{N1/q−1/p,N−1/p−1/2N2−n+1}:(1−c0)N2<n≤N2.c_n(\mathcal{S}_p^N \hookrightarrow\mathcal{S}_q^N) \asymp_{p,q} \begin{cases} \min\left\{1,\left(\frac{N}{n}\right)^{1/p-1/2}\right\} & : 1 \leq n \leq(1-c_0)N^2,\\ \max\left\{N^{1/q-1/p},N^{-1/p-1/2}\sqrt{N^2-n+1}\right\} & : (1-c_0)N^2 < n \leq N^2. \end{cases}

The estimates in the second range were already sharp. The contribution of Theorem A is the missing matching lower bound in the first range for 0<p<10 < p < 1 and q>2q > 2. Moreover, in this first range, both comparison constants can be chosen independently of qq.

The remaining gap for Gelfand numbers identified in [22] remains open. In the notation used here, [22 Conjecture 1.3] predicts that, for 2<p<q≤∞2 < p < q \leq\infty and

(1−c0)N2≤n≤N2−c0N1+2/q+1,(1-c_0)N^2 \leq n \leq N^2-c_0N^{1+2/q}+1,

one has

cn(SpN↪SqN)≲p,q(N2−n+1N2)1/p−1/q1/2−1/q.c_n(\mathcal{S}_p^N \hookrightarrow\mathcal{S}_q^N) \lesssim_{p,q} \left(\frac{N^2-n+1}{N^2}\right)^{\frac{1/p-1/q}{1/2-1/q}}.

Note that the original conjecture in [22] is slightly stronger: it asserts the same upper bound for the corresponding approximation numbers.

Idea of the Proof of Theorem A

The starting point is the rank-one case. If n≤Nn \leq N, an arbitrary subspace of RN×N\mathbb{R}^{N\times N} of codimension less than nn must intersect nontrivially the NN-dimensional space of matrices supported on a single column. Since every nonzero matrix in this space has rank one, all Schatten quasi-norms coincide there. For larger codimensions, this space is too small. It is therefore natural to replace it by the space of matrices supported on rr columns, which we identify with RN×r\mathbb{R}^{N \times r}. This space has dimension NrNr, while all its elements have rank at most rr. Thus, in order to force a nontrivial intersection with a subspace of codimension less than nn, one is led to choose NrNr of order nn, or equivalently rr of order n/Nn/N. This already explains the scale entering the desired estimate.

A rank bound alone, however, would lose a factor of order r\sqrt{r}. The key observation is that the Hilbert–Schmidt geometry of the intersection produces a particularly well-positioned matrix. Indeed, projecting the standard basis matrices of RN×r\mathbb{R}^{N \times r} onto a dd-dimensional subspace and averaging their squared Hilbert–Schmidt norms yields one projected matrix AA for which

∥A∥∞≥∥A∥22.\lVert A\rVert_{\infty} \ge\lVert A\rVert_{2}^{2}.

Since every matrix in RN×r\mathbb{R}^{N \times r} has rank at most rr, we have ∥A∥p≤r1/p−1/2∥A∥2\lVert A\rVert_{p} \le r^{1/p-1/2}\lVert A\rVert_{2}. Together with ∥A∥∞≥∥A∥22\lVert A\rVert_{\infty} \ge\lVert A\rVert_{2}^{2}, this yields a lower bound of order r1/2−1/pr^{1/2-1/p} whenever the subspace occupies a fixed proportion of RN×r\mathbb{R}^{N \times r}. Choosing r≍n/Nr \asymp n/N then produces precisely the factor

(Nn)1/p−1/2.\left(\frac{N}{n}\right)^{1/p-1/2}.

Finally, the passage from the operator norm to all target exponents q≥2q \ge2 uses ∥A∥q≥∥A∥∞\lVert A\rVert_{q} \ge\lVert A\rVert_{\infty}.

Organization of the paper

The paper is organized as follows. Section 2 contains the notation and elementary facts about Schatten norms and Gelfand numbers. In Section 3, we establish the rectangular projection estimate and prove Theorem A.

Notation and preliminaries

We present some notation and preliminary facts that are used in the paper.

Matrix spaces and orthogonal projections. For positive integers a,ba,b, write Ma,b:=Ra×b\mathbb{M}_{a,b}:=\mathbb{R}^{a \times b} and MN:=MN,N\mathbb{M}_{N}:=\mathbb{M}_{N,N}, and Tr⁡\operatorname{Tr} for the trace of a linear operator on a finite-dimensional vector space. We equip Ma,b\mathbb{M}_{a,b} with the Hilbert–Schmidt inner product

⟨A,B⟩HS:=Tr⁡(ATB),A,B∈Ma,b.\langle A,B\rangle_{\mathrm{HS}}:=\operatorname{Tr}(A^{\mathsf{T}}B), \qquad A,B\in\mathbb{M}_{a,b}.

Let e1,…,eae_1,\ldots,e_a and f1,…,fbf_1,\ldots,f_b denote the standard basis vectors of Ra\mathbb{R}^{a} and Rb\mathbb{R}^{b}, respectively, regarded as column vectors. The matrices

Eij:=eifjT,1≤i≤a,1≤j≤b,E_{ij}:=e_i f_j^{\mathsf{T}}, \qquad1\le i\le a,\quad1\le j\le b,

each having a single entry equal to one in position (i,j)(i,j) and zeros elsewhere, form an orthonormal basis of Ma,b\mathbb{M}_{a,b}. For a linear subspace F⊂Ma,bF\subset\mathbb{M}_{a,b}, we write PFP_F for the orthogonal projection onto FF with respect to this inner product. Viewed as a linear operator on Ma,b\mathbb{M}_{a,b}, it has trace dim⁡F\dim F.

Schatten quasi-norms and rank estimates. For A∈Ma,bA\in\mathbb{M}_{a,b}, let σ1(A)≥⋯≥σmin⁡{a,b}(A)≥0\sigma_1(A)\ge\cdots\ge\sigma_{\min\{a,b\}}(A)\ge0 denote its singular values. We extend the notation ∥A∥p\lVert A\rVert_p from square to rectangular matrices by taking the ℓp\ell_p-norm, or quasi-norm when 0<p<10<p<1, of this sequence. In particular, ∥A∥∞=σ1(A)\lVert A\rVert_{\infty}=\sigma_1(A) is the operator norm of A:ℓ2b→ℓ2aA:\ell_2^b\to\ell_2^a, and ∥A∥22=⟨A,A⟩HS\lVert A\rVert_2^2=\langle A,A\rangle_{\mathrm{HS}}.

If rank⁡A≤r\operatorname{rank} A \le r, where r≥1r \ge1, then for 0<p≤q≤∞0 < p \le q \le\infty

∥A∥q≤∥A∥p≤r1/p−1/q∥A∥q.(4)\lVert A\rVert_q \le\lVert A\rVert_p \le r^{1/p-1/q}\lVert A\rVert_q. \tag*{(4)}

The first inequality follows from the monotonicity of sequence quasi-norms. For the second, note that there are at most rr nonzero singular values. If 0<p<q<∞0 < p < q < \infty, Hölder’s inequality with exponents q/pq/p and q/(q−p)q/(q-p) gives, after padding the nonzero singular-value sequence with zeros,

∑j=1rσj(A)p≤(∑j=1rσj(A)q)p/qr1−p/q.\sum_{j=1}^{r}\sigma_j(A)^p \le\left(\sum_{j=1}^{r}\sigma_j(A)^q\right)^{p/q}r^{1-p/q}.

Taking the pp-th root proves the assertion. The cases p=qp=q and q=∞q=\infty are immediate. Under the same rank assumption, we shall in particular use that for all 0<p≤20 < p \le2

∥A∥p≤r1/p−1/2∥A∥2(5)\lVert A\rVert_p \le r^{1/p-1/2}\lVert A\rVert_2 \tag*{(5)}

Rectangular matrices as subspaces. For 1≤r≤N1 \le r \le N, we identify MN,r\mathcal{M}_{N,r} with the subspace of MN\mathcal{M}_N consisting of matrices whose last N−rN-r columns vanish, via A↦A~:=(A 0)A \mapsto\widetilde{A}:=(A\ 0). Since A~A~T=AAT\widetilde{A}\widetilde{A}^{\mathsf{T}}=AA^{\mathsf{T}}, this identification leaves all nonzero singular values unchanged. Consequently, ∥A~∥p=∥A∥p\lVert\widetilde{A}\rVert_p=\lVert A\rVert_p, 0<p≤∞0 < p \le\infty. We shall therefore pass between these rectangular and square matrix representations without changing the notation.

Gelfand numbers. For 1≤n≤N21 \le n \le N^2, the condition codim⁡F<n\operatorname{codim} F<n in (1) is equivalent to dim⁡F≥N2−n+1\dim F\ge N^2-n+1. Restricting to a smaller subspace cannot increase the restriction norm. Hence the infimum may be taken over subspaces of dimension exactly N2−n+1N^2-n+1, and

cn(SpN↪SqN)=inf⁡F⊂MNdim⁡F=N2−n+1sup⁡A∈F∖{0}∥A∥q∥A∥p.(6)c_n(\mathcal{S}_p^N\hookrightarrow\mathcal{S}_q^N) = \inf_{\substack{F\subset\mathcal{M}_N\\ \dim F=N^2-n+1}} \sup_{A\in F\setminus\{0\}}\frac{\lVert A\rVert_q}{\lVert A\rVert_p}. \tag*{(6)}

For 0<p≤q≤∞0 < p \le q \le\infty, the full embedding has norm one: the upper bound follows from (4), and equality is attained on any nonzero rank-one matrix. Moreover, applying (4) with r=Nr=N gives

cn(SpN↪SqN)≥N1/q−1/p,1≤n≤N2.(7)c_n(\mathcal{S}_p^N\hookrightarrow\mathcal{S}_q^N)\ge N^{1/q-1/p},\qquad1\le n\le N^2. \tag*{(7)}

This lower bound is attained at the last index. Indeed, let IN=diag⁡(1,…,1)I_N=\operatorname{diag}(1,\ldots,1) denote the N×NN\times N identity matrix and choose F=span⁡{IN}F=\operatorname{span}\{I_N\} in (6) for n=N2n=N^2. All singular values of tINtI_N equal ∣t∣\lvert t\rvert, so for every t≠0t\ne0,

∥tIN∥q∥tIN∥p=N1/q−1/p.\frac{\lVert tI_N\rVert_q}{\lVert tI_N\rVert_p}=N^{1/q-1/p}.

Together with (7), for 0<p≤q≤∞0 < p \le q \le\infty this yields

cN2(SpN↪SqN)=N1/q−1/p.c_{N^2}(\mathcal{S}_p^N\hookrightarrow\mathcal{S}_q^N)=N^{1/q-1/p}.

Finally, the inequality ∥A∥q≥∥A∥∞\lVert A\rVert_q\ge\lVert A\rVert_\infty implies

cn(SpN↪SqN)≥cn(SpN↪S∞N),0<p,q≤∞,1≤n≤N2.c_n(\mathcal{S}_p^N\hookrightarrow\mathcal{S}_q^N)\ge c_n(\mathcal{S}_p^N\hookrightarrow\mathcal{S}_\infty^N),\qquad0<p,q\le\infty,\quad1\le n\le N^2.

Thus, to prove the lower bound in Theorem A, it suffices to consider the target space S∞N\mathcal{S}_\infty^N.

A rectangular projection argument

We first establish the subspace estimate that is the main ingredient of the proof. Its formulation includes the range 0<p≤20 < p \le2, although only 0<p≤10 < p \le1 is needed for Theorem A.

Lemma 3.1. Let 1≤r≤N1 \le r \le N, and let E⊂MN,rE \subset\mathbb{M}_{N,r} be a subspace of dimension d≥1d \ge1. Then, for 0<p≤20 < p \le2,

sup⁡A∈E∖{0}∥A∥∞∥A∥p≥dNr1/p.\sup_{A\in E\setminus\{0\}} \frac{\lVert A\rVert_{\infty}}{\lVert A\rVert_p} \ge\frac{\sqrt{d}}{\sqrt{N}r^{1/p}}.

Proof. Let PE:MN,r→MN,rP_E:\mathbb{M}_{N,r}\to\mathbb{M}_{N,r} be the Hilbert–Schmidt-orthogonal projection onto the subspace EE. Since PE=PE∗=PE2P_E=P_E^*=P_E^2, we have

∥PEEij∥22=⟨PEEij,Eij⟩HS.\lVert P_E E_{ij}\rVert_2^2=\langle P_E E_{ij},E_{ij}\rangle_{\mathrm{HS}}.

As the trace of a finite-dimensional linear operator is independent of the choice of basis and the matrices EijE_{ij} form an orthonormal basis of MN,r\mathbb{M}_{N,r}, we have

Tr⁡(PE)=∑i=1N∑j=1r⟨PEEij,Eij⟩HS.\operatorname{Tr}(P_E)=\sum_{i=1}^{N}\sum_{j=1}^{r}\langle P_E E_{ij},E_{ij}\rangle_{\mathrm{HS}}.

But since PEP_E is the orthogonal projection onto the dd-dimensional space EE, we have Tr⁡(PE)=d\operatorname{Tr}(P_E)=d and, consequently,

∑i=1N∑j=1r∥PEEij∥22=d.\sum_{i=1}^{N}\sum_{j=1}^{r}\lVert P_E E_{ij}\rVert_2^2=d.

Hence, by averaging, there must exist indices i,ji,j such that

t:=∥PEEij∥22≥dNr.t:=\lVert P_E E_{ij}\rVert_2^2\ge\frac{d}{Nr}.

Now put A:=PEEij=(akℓ)k,ℓA:=P_E E_{ij}=(a_{k\ell})_{k,\ell}. Since PEP_E is the orthogonal projection onto EE, we have

A=PEEij∈EandEij−A=(IMN,r−PE)Eij=PE⊥Eij∈E⊥,A=P_E E_{ij}\in E \qquad\text{and}\qquad E_{ij}-A=(I_{\mathbb{M}_{N,r}}-P_E)E_{ij}=P_{E^\perp}E_{ij}\in E^\perp,

and so ⟨A,Eij−A⟩HS=0\langle A,E_{ij}-A\rangle_{\mathrm{HS}}=0. Therefore

aij=⟨A,Eij⟩HS=⟨A,A⟩HS=∥A∥22=t.a_{ij}=\langle A,E_{ij}\rangle_{\mathrm{HS}}=\langle A,A\rangle_{\mathrm{HS}}=\lVert A\rVert_2^2=t.

Since the absolute value of every matrix entry is bounded by the operator norm,

∥A∥∞≥∣aij∣=t.\lVert A\rVert_\infty\ge|a_{ij}|=t.

On the other hand, A∈MN,rA\in\mathbb{M}_{N,r} and hence rank⁡A≤r\operatorname{rank}A\le r. By (5),

∥A∥p≤r1/p−1/2∥A∥2=r1/p−1/2t.\lVert A\rVert_p\le r^{1/p-1/2}\lVert A\rVert_2=r^{1/p-1/2}\sqrt{t}.

It follows that

∥A∥∞∥A∥p≥r1/2−1/pt≥r1/2−1/pdNr=dNr1/p,\frac{\lVert A\rVert_\infty}{\lVert A\rVert_p}\ge r^{1/2-1/p}\sqrt{t}\ge r^{1/2-1/p}\sqrt{\frac{d}{Nr}}=\frac{\sqrt{d}}{\sqrt{N}r^{1/p}},

which proves the result.

The parameter rr determines the dimension NrNr of the rectangular matrix space. Intersecting it with a subspace of codimension less than nn leaves a subspace of dimension at least Nr−n+1Nr-n+1. We keep rr free because the resulting lower bound depends on rr, and a suitable choice of rr later on will balance the intersection dimension against the rank-dependent norm estimate.

Proposition 3.2. Let 0<p≤2≤q≤∞0 < p \le2 \le q \le\infty, N∈NN \in\mathbb{N}, and 1≤n≤N21 \le n \le N^2. Then

cn(SpN↪SqN)≥max⁡1≤r≤NNr≥nNr−n+1Nr1/p.(8)c_n(\mathcal{S}_p^N \hookrightarrow\mathcal{S}_q^N) \ge\max_{\substack{1 \le r \le N \\ Nr \ge n}} \frac{\sqrt{Nr-n+1}}{\sqrt{N}r^{1/p}}. \tag*{(8)}

Moreover, cn(SpN↪SqN)=1c_n(\mathcal{S}_p^N \hookrightarrow\mathcal{S}_q^N)=1 whenever 1≤n≤N1 \le n \le N.

Proof. Fix a subspace F⊂MNF \subset\mathbb{M}_N with codim⁡F<n\operatorname{codim} F < n. For 1≤r≤N1 \le r \le N, let Hr⊂MNH_r \subset\mathbb{M}_N denote the subspace of matrices whose last N−rN-r columns vanish. By deleting these zero columns, we identify HrH_r with MN,r\mathbb{M}_{N,r}. As observed above, this identification preserves all Schatten norms and quasi-norms, and dim⁡Hr=Nr\dim H_r=Nr. Now consider the canonical quotient map

Q:MN⟶MN/F,Q(A)=A+F.Q:\mathbb{M}_N \longrightarrow\mathbb{M}_N/F,\qquad Q(A)=A+F.

Its restriction Q∣Hr:Hr→MN/FQ|_{H_r}:H_r\to\mathbb{M}_N/F has kernel ker⁡(Q∣Hr)=F∩Hr\ker(Q|_{H_r})=F\cap H_r. Moreover, rank⁡(Q∣Hr)≤dim⁡(MN/F)=codim⁡F\operatorname{rank}(Q|_{H_r})\le\dim(\mathbb{M}_N/F)=\operatorname{codim}F and hence, by the rank–nullity theorem,

dim⁡(F∩Hr)=dim⁡Hr−rank⁡(Q∣Hr)≥Nr−codim⁡F≥Nr−n+1.(9)\dim(F\cap H_r)=\dim H_r-\operatorname{rank}(Q|_{H_r})\ge Nr-\operatorname{codim}F\ge Nr-n+1. \tag*{(9)}

Suppose now that Nr≥nNr\ge n. Then F∩Hr≠{0}F\cap H_r\ne\{0\}. Put d:=dim⁡(F∩Hr)d:=\dim(F\cap H_r). Applying Lemma 3.1 to the subspace F∩Hr⊂Hr≅MN,rF\cap H_r\subset H_r\cong\mathbb{M}_{N,r}, and using d≥Nr−n+1d\ge Nr-n+1, we obtain

sup⁡A∈(F∩Hr)∖{0}∥A∥∞∥A∥p≥dNr1/p≥Nr−n+1Nr1/p.\sup_{A\in(F\cap H_r)\setminus\{0\}}\frac{\lVert A\rVert_\infty}{\lVert A\rVert_p}\ge\frac{\sqrt{d}}{\sqrt{N}r^{1/p}}\ge\frac{\sqrt{Nr-n+1}}{\sqrt{N}r^{1/p}}.

Since F∩Hr⊂FF\cap H_r\subset F and ∥A∥q≥∥A∥∞\lVert A\rVert_q\ge\lVert A\rVert_\infty for every q>0q>0, it follows that

sup⁡A∈F∖{0}∥A∥q∥A∥p≥Nr−n+1Nr1/p.\sup_{A\in F\setminus\{0\}}\frac{\lVert A\rVert_q}{\lVert A\rVert_p}\ge\frac{\sqrt{Nr-n+1}}{\sqrt{N}r^{1/p}}.

This estimate holds for every integer rr satisfying 1≤r≤N1\le r\le N and Nr≥nNr\ge n. We may therefore maximize over all such rr. Since FF was arbitrary, taking the infimum over all subspaces F⊂MNF\subset\mathbb{M}_N with codim⁡F<n\operatorname{codim}F<n yields

cn(SpN↪SqN)≥max⁡1≤r≤NNr≥nNr−n+1Nr1/p,c_n(\mathcal{S}_p^N \hookrightarrow\mathcal{S}_q^N)\ge\max_{\substack{1\le r\le N\\Nr\ge n}}\frac{\sqrt{Nr-n+1}}{\sqrt{N}r^{1/p}},

which proves (8).

Finally, suppose that 1≤n≤N1\le n\le N. Taking r=1r=1 in (9) gives dim⁡(F∩H1)≥N−n+1≥1\dim(F\cap H_1)\ge N-n+1\ge1, and hence F∩H1≠{0}F\cap H_1\ne\{0\}. Every nonzero matrix in H1H_1 has rank one, so all of its Schatten norms and quasi-norms coincide. Therefore

sup⁡A∈F∖{0}∥A∥q∥A∥p≥1.\sup_{A\in F\setminus\{0\}}\frac{\lVert A\rVert_q}{\lVert A\rVert_p}\ge1.

Since p≤qp\le q in the parameter range under consideration, the full embedding SpN↪SqN\mathcal{S}_p^N\hookrightarrow\mathcal{S}_q^N has norm one. Consequently, whenever 1≤n≤N1\le n\le N,

cn(SpN↪SqN)=1,c_n(\mathcal{S}_p^N\hookrightarrow\mathcal{S}_q^N)=1,

which completes the proof.

We now choose the rectangular dimension in Proposition 3.2 and prove the main result.

Proof of Theorem A. Let us define α:=1/p−1/2\alpha:= 1/p - 1/2. We first prove the lower estimate. The proof uses only 0<p≤20 < p \le2, and we distinguish three ranges of indices.

Case 1: 1≤n≤N1 \le n \le N. Proposition 3.2 gives cn(SpN↪SqN)=1c_n(\mathcal{S}_p^N \hookrightarrow\mathcal{S}_q^N) = 1. Since min⁡{1,N/n}=1\min\{1,N/n\} = 1, the lower estimate holds with constant one.

Case 2: N<n≤N2/2N < n \le N^2/2. Here we choose r:=⌈2nN⌉r := \left\lceil\frac{2n}{N} \right\rceil. Since n≤N2/2n \le N^2/2, we have r≤Nr \le N. Moreover, Nr≥2nNr \ge2n, and therefore

Nr−n+1≥Nr2.Nr-n+1 \ge\frac{Nr}{2}.

On the other hand,

r≤2nN+1≤3nN,r \le\frac{2n}{N}+1 \le\frac{3n}{N},

where the last inequality uses n>Nn > N. With our specific choice of rr, we can apply the lower estimate in (8) combined with the previous estimates to obtain

cn(SpN↪SqN)≥Nr/2Nr1/p=12r−α≥3−α2(Nn)α.c_n(\mathcal{S}_p^N \hookrightarrow\mathcal{S}_q^N) \ge\frac{\sqrt{Nr/2}}{\sqrt{N}r^{1/p}} = \frac{1}{\sqrt{2}}r^{-\alpha} \ge\frac{3^{-\alpha}}{\sqrt{2}}\left(\frac{N}{n}\right)^\alpha.

In the last estimate we used r≤3nNr \le\frac{3n}{N}.

Case 3: N2/2<n≤(1−η)N2N^2/2 < n \le(1-\eta)N^2. If this range is non-empty, take r=Nr=N in (8). Since in our range of nn, N2−n+1≥ηN2N^2-n+1 \ge\eta N^2, we obtain

cn(SpN↪SqN)≥N−1/2−1/pN2−n+1≥ηN−α.c_n(\mathcal{S}_p^N \hookrightarrow\mathcal{S}_q^N) \ge N^{-1/2-1/p}\sqrt{N^2-n+1} \ge\sqrt{\eta}N^{-\alpha}.

The inequality n>N2/2n > N^2/2 implies

min⁡{1,Nn}α≤(2N)α.\min\left\{1,\frac{N}{n}\right\}^{\alpha} \le\left(\frac{2}{N}\right)^{\alpha}.

Consequently,

cn(SpN↪SqN)≥η2−αmin⁡{1,Nn}α;c_n(\mathcal{S}_p^N \hookrightarrow\mathcal{S}_q^N) \ge\sqrt{\eta}2^{-\alpha}\min\left\{1,\frac{N}{n}\right\}^{\alpha};

empty ranges are omitted.

Combining the three cases above proves the lower estimate in (3); one may take

cp,η:=min⁡{1,3−α2,η2−α}.(10)c_{p,\eta}:=\min\left\{1,\frac{3^{-\alpha}}{\sqrt{2}},\sqrt{\eta}2^{-\alpha}\right\}. \tag*{(10)}

In particular, this constant is independent of qq.

For the upper estimate, we use the known bound

cn(SpN↪S2N)≤Cpmin⁡{1,Nn}α,1≤n≤N2;(11)c_n(\mathcal{S}_p^N \hookrightarrow\mathcal{S}_2^N) \le C_p\min\left\{1,\frac{N}{n}\right\}^{\alpha}, \qquad1 \le n \le N^2; \tag*{(11)}

see [22 Proposition 3.3, with $q=2$]; compare also [14 Proposition 4.5, Step 1]. Since q≥2q \ge2, the inequality ∥A∥q≤∥A∥2\lVert A\rVert_q \le\lVert A\rVert_2 holds for every matrix AA. Taking the restriction norm on each admissible subspace and then the infimum gives

cn(SpN↪SqN)≤cn(SpN↪S2N).c_n(\mathcal{S}_p^N \hookrightarrow\mathcal{S}_q^N) \le c_n(\mathcal{S}_p^N \hookrightarrow\mathcal{S}_2^N).

Together with (11), this proves the upper estimate in (3) with a constant independent of both qq and η\eta.

Acknowledgement

The author was supported by the German Research Foundation under DFG project 516672205 Limit theorems for the volume of random projections of ℓp\ell_p-balls. Most of this work was carried out during a research semester at Graz University of Technology. The author gratefully acknowledges the excellent working conditions and the hospitality provided during his stay.

ChatGPT (GPT-5.6 Luna, OpenAI) was used to improve the wording and for grammar and spelling checks, and to assist in the production of Figure 1.

Schematic overview of the Gelfand numbers of $\mathcal{S}_p^N \hookrightarrow \mathcal{S}_q^N$

Figure 1. Schematic overview of the Gelfand numbers of SpN↪SqN\mathcal{S}_p^N \hookrightarrow\mathcal{S}_q^N; compare [22 Figure 2(a)]. Light grey indicates parameter regions with previously known sharp asymptotics. Dark grey highlights 0<p<10 < p < 1 and 2<q≤∞2 < q \leq\infty, where Theorem A closes the remaining gap and Remark 1.1 gives the resulting all-index asymptotics. The hatched triangle corresponds to 2<p<q≤∞2 < p < q \leq\infty, where matching bounds are still missing for intermediate indices nn. The boundary of the two highlighted regions is not part of the open cases: sharp asymptotic estimates were already known for p=1p = 1 with 2<q≤∞2 < q \leq\infty, for q=2q = 2 with 0<p<10 < p < 1, for p=2p = 2 with 2<q≤∞2 < q \leq\infty, and along the diagonal p=q>2p = q > 2.

References

  1. [1]E. J. Candès and B. Recht. Exact matrix completion via convex optimization. Found. Comput. Math., 9(6):717–772, 2009.
  2. [2]B. Carl and A. Defant. Asymptotic estimates for approximation quantities of tensor product identities. J. Approx. Theory, 88(2):228–256, 1997.DOI
  3. [3]J. A. Chávez-Domínguez and D. Kutzarova. Stability of low-rank matrix recovery and its connections to Banach space geometry. J. Math. Anal. Appl., 427(1):320–335, 2015.
  4. [4]D. E. Edmunds and H. Triebel. Function spaces, entropy numbers, differential operators, volume 120 of Cambridge Tracts in Mathematics. Cambridge University Press, Cambridge, 1996.
  5. [5]S. Foucart and H. Rauhut. A mathematical introduction to compressive sensing. Applied and Numerical Harmonic Analysis. Birkhäuser/Springer, New York, 2013.DOI
  6. [6]A. Y. Garnaev and E. D. Gluskin. The widths of a Euclidean ball. Soviet Math. Dokl., 30:200–204, 1984.
  7. [7]E. D. Gluskin. Norms of random matrices and diameters of finite-dimensional sets. Mat. Sb. (N.S.), 120(162)(2):180–189, 286, 1983.DOI
  8. [8]Y. Gordon and D. R. Lewis. Absolutely summing operators and local unconditional structures. Acta Math., 133:27–48, 1974.DOI
  9. [9]A. Grothendieck. Sur une notion de produit tensoriel topologique d’espaces vectoriels topologiques, et une classe remarquable d’espaces vectoriels liée à cette notion. C. R. Acad. Sci. Paris, 233:1556–1558, 1951.
  10. [10]O. Guédon and A. E. Litvak. Euclidean projections of a p-convex body. In Geometric aspects of functional analysis, volume 1745 of Lecture Notes in Math., pages 95–108. Springer, Berlin, 2000.
  11. [11]O. Guédon and G. Paouris. Concentration of mass on the Schatten classes. Ann. Inst. H. Poincaré Probab. Statist., 43(1):87–99, 2007.DOI
  12. [12]A. Hinrichs and C. Michels. Approximation numbers of inclusions between Schatten classes. Rend. Circ. Mat. Palermo (2) Suppl., (76):395–411, 2005.
  13. [13]A. Hinrichs, I. Prochno, and J. Vybiral. Entropy numbers of embeddings of Schatten classes. J. Funct. Anal., 273(10):3241–3261, 2017.arxiv.org/abs/1612.08105
  14. [14]A. Hinrichs, J. Prochno, and J. Vybiral. Gelfand numbers of embeddings of Schatten classes. Math. Ann., 380(3–4):1563–1593, 2021.arxiv.org/abs/2011.06554
  15. [15]Z. Kabluchko, J. Prochno, and C. Thäle. Exact asymptotic volume and volume ratio of Schatten unit balls. J. Approx. Theory, 257:105457, 2020.arxiv.org/abs/1804.03467
  16. [16]Z. Kabluchko, J. Prochno, and C. Thäle. Intersection of unit balls in classical matrix ensembles. Israel J. Math., 239(1):129–172, 2020.arxiv.org/abs/1804.03466
  17. [17]Z. Kabluchko, J. Prochno, and C. Thäle. Sanov-type large deviations in Schatten classes. Ann. Inst. Henri Poincaré Probab. Stat., 56(2):928–953, 2020.arxiv.org/abs/1808.04862
  18. [18]H. König, M. Meyer, and A. Pajor. The isotropy constants of the Schatten classes are bounded. Math. Ann., 312(4):773–783, 1998.DOI
  19. [19]T. Kühn. A lower estimate for entropy numbers. J. Approx. Theory, 110(1):120–124, 2001.DOI
  20. [20]C. Kümmerle and J. Sigl. Harmonic mean iteratively reweighted least squares for low-rank matrix recovery. J. Mach. Learn. Res., 19:Paper no. 47, 49, 2018.arxiv.org/abs/1703.05038
  21. [21]S. Kwapień and A. Pełczyński. The main triangle projection in matrix spaces and its applications. Studia Math., 34:43–68, 1970.DOI
  22. [22]J. Prochno and M. Strzelecki. Approximation, Gelfand, and Kolmogorov numbers of Schatten class embeddings. J. Approx. Theory, 277:105736, 2022.DOI
  23. [23]J. Radke and B.-H. Vritsiou. On the thin-shell conjecture for the Schatten classes. Ann. Inst. H. Poincaré Probab. Statist., 56(1):87–119, 2020.arxiv.org/abs/1602.06934
  24. [24]A. Rohde and A. B. Tsybakov. Estimation of high-dimensional low-rank matrices. Ann. Statist., 39(2):887–930, 2011.DOI
  25. [25]A. F. Ruston. On the Fredholm theory of integral equations for operators belonging to the trace class of a general Banach space. Proc. London Math. Soc. (2), 53:109–124, 1951.DOI
  26. [26]R. Schatten. The cross-space of linear transformations. Ann. of Math. (2), 47:73–84, 1946.DOI
  27. [27]R. Schatten. Norm ideals of completely continuous operators. Ergebnisse der Mathematik und ihrer Grenzgebiete. N. F., Heft 27. Springer-Verlag, Berlin-Göttingen-Heidelberg, 1960.
  28. [28]R. Schatten and J. von Neumann. The cross-space of linear transformations. II. Ann. of Math. (2), 47:608–630, 1946.DOI
  29. [29]R. Schatten and J. von Neumann. The cross-space of linear transformations. III. Ann. of Math. (2), 49:557–582, 1948.DOI
  30. [30]C. Schütt. Entropy numbers of diagonal operators between symmetric Banach spaces. J. Approx. Theory, 40(2):121–128, 1984.DOI
  31. [31]N. Tomczak-Jaegermann. The moduli of smoothness and convexity and the Rademacher averages of trace classes Sₚ (1 ≤ p < ∞). Studia Math., 50:163–182, 1974.
  32. [32]H. Triebel. Fractals and spectra, volume 91 of Monographs in Mathematics. Birkhäuser Verlag, Basel, 1997. Related to Fourier analysis and function spaces.

Paper details

vlnc:W_rgky2q7b
Version
v2 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 16

downloads · 1

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

  • Mathematics

GPT-5.6 Luna (bare)

View statement
Looking for related papers…
Contents