betaFeedback

WorksMathematics

Cyclic group tensors and the volume ratio of projective tensor products

Abstract

We revisit the problem of estimating the norm of the natural embedding from the Euclidean space ℓ2n3\ell_2^{n^3} into the projective tensor product ℓpn⊗πℓqn⊗πℓrn\ell_p^n\otimes_π\ell_q^n\otimes_π\ell_r^n in the range 1≤p≤q≤2≤r≤∞1\le p\le q\le 2\le r\le\infty, studied by Giladi, Prochno, Schütt, Tomczak-Jaegermann and Werner [J. Funct. Anal. 273 (2017), 471-495]. Using convolution on finite cyclic groups, we establish new lower bounds that improve their estimates and determine the exact asymptotic order of the embedding norm whenever 1/p+1/q+1/r≥21/p+1/q+1/r\ge 2. We then apply an estimate of Boedihardjo for the injective norm of random tensors to a randomly signed cyclic convolution tensor and use multilinear interpolation to determine the optimal exponent throughout the entire parameter range, up to logarithmic factors. As consequences, we obtain corresponding improvements and asymptotically sharp estimates, up to logarithmic factors, for the volume ratio of these projective tensor products.

Introduction & Main Results

In the geometry of Banach spaces, the volume ratio vr⁡(X)\operatorname{vr}(X) of a finite-dimensional normed space XX compares the volume of its unit ball with that of its John ellipsoid, the unique ellipsoid of maximal volume contained in the unit ball. It is a fundamental concept in the local theory of Banach spaces and asymptotic geometric analysis with important applications in approximation theory (see, e.g., [17]). The notion goes back to the seminal works of Szarek [21] and Szarek and Tomczak-Jaegermann [20], which were inspired by Kashin’s celebrated result on nearly Euclidean decompositions of ℓ22n\ell_2^{2n} [15]; nearly Euclidean is to be understood in the sense of Banach–Mazur distance. To be precise, an important connection between what is now known as the volume ratio and the existence of nearly Euclidean decompositions was already implicit in the work of Szarek [21], namely that every 2n2n-dimensional Banach space XX admits a decomposition into two nn-dimensional subspaces whose Banach–Mazur distances to ℓ2n\ell_2^n are bounded by a universal constant times vr⁡(X)2\operatorname{vr}(X)^2. Building on this observation, Szarek and Tomczak-Jaegermann [20] formally introduced the notion of volume ratio and investigated its relationship with the cotype 2 constant of Banach spaces. They also established Kashin-type decompositions for several classes of Banach spaces, including the projective tensor products ℓpn⊗πℓ2n\ell_p^n \otimes_{\pi} \ell_2^n for 1≤p≤21 \le p \le2.

The foundations of the modern theory of tensor products of Banach spaces were laid by Grothendieck in his seminal 1953 Résumé [14] and had a huge impact on Banach space theory. This impact and the success of the concept of tensor products is to a large extent due to the work [16] of J. Lindenstrauss and A. Pełczyński in the late sixties who reformulated Grothendieck’s ideas in the context of operator ideals and made this theory accessible to a broader audience. Modern applications include quantum computing and quantum information theory [3, 4, 5, 22] or the theory of locally decodable codes [11].

Main Results & Previous Work

In this paper, we study the volume ratio of three-fold projective tensor products of finite-dimensional ℓp\ell_p spaces. Formally, the volume ratio is defined for an nn-dimensional normed space XX by

vr⁡(X):=(vol⁡n(BX)vol⁡n(EX))1/n,\operatorname{vr}(X) := \left(\frac{\operatorname{vol}_n(B_X)}{\operatorname{vol}_n(\mathcal{E}_X)}\right)^{1/n},

where EX\mathcal{E}_X is the ellipsoid of maximal volume contained in the unit ball BXB_X of XX. In the case of projective tensor products of finite-dimensional sequence spaces, the determination of this parameter can be reduced to norm estimates for a natural Euclidean embedding. More precisely, Giladi, Prochno, Schütt, Tomczak-Jaegermann and Werner [13] studied the volume ratio of the three-fold projective tensor products

Xp,q,rn:=ℓpn⊗πℓqn⊗πℓrn,1≤p≤q≤r≤∞.X_{p,q,r}^n := \ell_p^n \otimes_\pi\ell_q^n \otimes_\pi\ell_r^n,\qquad1 \le p \le q \le r \le\infty.

They obtained sharp estimates in several parameter regimes and applications to nearly Euclidean decompositions and cotype. In particular, they established asymptotically sharp estimates for the volume radii of the unit balls of Xp,q,rnX_{p,q,r}^n throughout the range $1 \le p \le q \le r \le \infty$; see [13 Theorem B]. Consequently, determining the volume ratio was reduced to identifying the John ellipsoids. Since the spaces Xp,q,rnX_{p,q,r}^n have enough symmetries (see [19, 20]), their John ellipsoids are Euclidean balls of the form

EXp,q,rn=∥id⁡:ℓ2n3→Xp,q,rn∥−1B2n3.\mathcal{E}_{X_{p,q,r}^n}=\left\| \operatorname{id}:\ell_2^{n^3}\to X_{p,q,r}^n\right\|^{-1}B_2^{n^3}.

Thus, the remaining task was to estimate the operator norm of the identity mapping id⁡:ℓ2n3→Xp,q,rn\operatorname{id}:\ell_2^{n^3}\to X_{p,q,r}^n; see [13 Theorem C and Corollary 5.7]. Their results leave open the precise order of the volume ratio in most parts of the range $1 \le p \le q \le 2 \le r \le \infty$; some boundary cases are covered in [13 Remark 2.1 and Corollary 5.7].

The purpose of the present paper is to sharpen the existing estimates in the range $1 \le p \le q \le 2 \le r \le \infty$ and, in particular, to determine the optimal asymptotic order throughout this parameter range, up to logarithmic factors. We identify the underlying vector space of Xp,q,rnX_{p,q,r}^n with Rn3\mathbb{R}^{n^3} through the canonical tensor basis (ei⊗ej⊗ek)1≤i,j,k≤n(e_i\otimes e_j\otimes e_k)_{1\le i,j,k\le n} and consider the identity operator

id⁡:ℓ2n3⟶Xp,q,rn.\operatorname{id}:\ell_2^{n^3}\longrightarrow X_{p,q,r}^n.

In what follows, constants implicit in ≲\lesssim and ≍\asymp are independent of nn and may depend on the fixed parameters p,q,rp,q,r. Whenever such a dependence is present, we indicate it by subscripts. We start by recalling the relevant estimates and results from [13]. By Lemmas 5.5 and 5.6 of [13],

∥id⁡:ℓ2n3⟶Xp,q,rn∥≲p,q,rnmin⁡{1/q+1/r,1}(1)\left\|\operatorname{id}:\ell_2^{n^3}\longrightarrow X_{p,q,r}^n\right\|\lesssim_{p,q,r} n^{\min\{1/q+1/r,1\}} \tag*{(1)}

and

∥id⁡:ℓ2n3⟶Xp,q,rn∥≳p,q,rmax⁡{n1/2+1/r,nmin⁡{1/q+1/r,1}+1/p−1,n1/q}.\left\|\operatorname{id}:\ell_2^{n^3}\longrightarrow X_{p,q,r}^n\right\|\gtrsim_{p,q,r}\max\left\{n^{1/2+1/r},n^{\min\{1/q+1/r,1\}+1/p-1},n^{1/q}\right\}.

In particular, the bounds were already sharp when p=1p=1 or q=2q=2, as well as in the boundary cases r=2r=2 and r=∞r=\infty, as follows from Lemmas 5.5 and 5.6 of [13]. Our aim is to improve the lower bounds in the remaining parameter range.

Our first result is based on an explicit (deterministic) construction involving a tensor supported on the relation j=i+kj=i+k in the finite cyclic group Zn\mathbb{Z}_n. The associated trilinear form can be analyzed using Young’s convolution inequality, leading to the first main result.

Theorem A. Let $1 \le p \le q \le 2 \le r \le \infty$. Then, for every n∈Nn \in\mathbb{N},

∥id⁡:ℓ2n3⟶Xp,q,rn∥≥nmin⁡{1,1/p+1/q+1/r−1}.(2)\left\lVert\operatorname{id}: \ell_2^{n^3} \longrightarrow X_{p,q,r}^n \right\rVert\ge n^{\min\{1,1/p+1/q+1/r-1\}}. \tag*{(2)}

In particular, if 1/p+1/q+1/r≥21/p+1/q+1/r \ge2, then

∥id⁡:ℓ2n3⟶Xp,q,rn∥≍p,q,rn.(3)\left\lVert\operatorname{id}: \ell_2^{n^3} \longrightarrow X_{p,q,r}^n \right\rVert\asymp_{p,q,r} n. \tag*{(3)}

The estimate (2) is proved in § 4; the upper estimate needed to obtain (3) is the previously known bound (1). Thus Theorem A extends the region in which matching bounds are available, rather than replacing the upper estimate of [13]. The deterministic construction unfortunately does not by itself settle the entire parameter range. Our second main result does so with a logarithmic loss.

Theorem B. For all $1 \le p \le q \le 2 \le r \le \infty$,

nmin⁡{1/q+1/r,1}(log⁡(en))4(1−1/q)≲∥id⁡:ℓ2n3⟶Xp,q,rn∥≲p,q,rnmin⁡{1/q+1/r,1}.\frac{n^{\min\{1/q+1/r,1\}}}{(\log(en))^{4(1-1/q)}} \lesssim\left\lVert\operatorname{id}: \ell_2^{n^3} \longrightarrow X_{p,q,r}^n \right\rVert\lesssim_{p,q,r} n^{\min\{1/q+1/r,1\}}.

The upper bound in Theorem B is again due to [13 Lemma 5.5]; the lower bound is the new assertion. For its proof, we place independent random signs on the cyclic support. The essential probabilistic input is the estimate of Boedihardjo [8 Corollary 1.4] for the injective norm of random tensors with independent centred bounded entries; we note that this bound is similar to the bound by Bandeira and van Handel for the expected spectral norm of a random matrix with independent entries [6]. In our setting, every one-dimensional coordinate fibre of the support contains exactly one nonzero coefficient. Boedihardjo’s result therefore yields a bound of order (log⁡(en))2(\log(en))^2 for the expected Hilbertian injective norm. We then combine a suitable choice of signs with elementary endpoint estimates and multilinear interpolation (see [7 Section 4.4]). The random-tensor inequality itself is not new; its use for this cyclic construction and the resulting lower embedding estimate are the ingredients of the present argument. In the region 1/p+1/q+1/r≥21/p+1/q+1/r \ge2, Theorem A removes the logarithmic loss. The corresponding consequences for volume ratio follow from the volume-radius estimate in [13 Theorem B] and the relation between volume ratio and the Euclidean embedding norm for spaces with enough symmetries, as used in [13 (2.3)].

Corollary 1.1. For $1 \le p \le q \le 2 \le r \le \infty$,

nmin⁡{1/q−1/2,1/2−1/r}(log⁡(en))4(1−1/q)≲vr⁡(Xp,q,rn)≲p,q,rnmin⁡{1/q−1/2,1/2−1/r}.\frac{n^{\min\{1/q-1/2,1/2-1/r\}}}{(\log(en))^{4(1-1/q)}} \lesssim\operatorname{vr}(X_{p,q,r}^n) \lesssim_{p,q,r} n^{\min\{1/q-1/2,1/2-1/r\}}.

If, in addition, 1/p+1/q+1/r≥21/p+1/q+1/r \ge2, then

vr⁡(Xp,q,rn)≍p,q,rn1/2−1/r.\operatorname{vr}(X_{p,q,r}^n) \asymp_{p,q,r} n^{1/2-1/r}.

Here the volume-radius asymptotics and the reduction to the embedding norm are inherited from [13]; the improvement comes from the new embedding lower bounds.

The paper is organized as follows. In § 3, we recall some basic facts about tensor products of Banach spaces. § 4 is devoted to the deterministic cyclic construction. In § 5, we investigate its randomly signed counterpart and derive the logarithmic estimates via interpolation.

Some background on tensor products of Banach spaces

Let (X,∥⋅∥X)(X,\|\cdot\|_X) and (Y,∥⋅∥Y)(Y,\|\cdot\|_Y) be two Banach spaces. The algebraic tensor product X⊗YX \otimes Y can be realized as a space of linear functionals on the vector space of all bilinear forms on X×YX \times Y. For x∈Xx \in X, y∈Yy \in Y, and a bilinear form BB on X×YX \times Y, we define

(x⊗y)(B):=B(x,y).(x \otimes y)(B) := B(x,y).

The projective tensor product X⊗πYX \otimes_{\pi} Y is the algebraic tensor product X⊗YX \otimes Y equipped with the projective tensor norm

∥A∥X⊗πY:=inf⁡{∑i=1m∥xi∥X∥yi∥Y:A=∑i=1mxi⊗yi, xi∈X, yi∈Y}.\|A\|_{X\otimes_{\pi}Y}:=\inf\left\{\sum_{i=1}^{m}\|x_i\|_X\|y_i\|_Y:A=\sum_{i=1}^{m}x_i\otimes y_i,\ x_i\in X,\ y_i\in Y\right\}.

Similarly, the injective tensor product X⊗εYX \otimes_{\varepsilon} Y is the algebraic tensor product X⊗YX \otimes Y equipped with the injective tensor norm

∥A∥X⊗εY:=sup⁡{∣∑i=1mφ(xi)ψ(yi)∣:φ∈BX∗, ψ∈BY∗},\|A\|_{X\otimes_{\varepsilon}Y}:=\sup\left\{\left|\sum_{i=1}^{m}\varphi(x_i)\psi(y_i)\right|:\varphi\in B_{X^*},\ \psi\in B_{Y^*}\right\},

where A=∑i=1mxi⊗yiA=\sum_{i=1}^{m}x_i\otimes y_i is any representation of AA; we may also use the short notation ∥⋅∥inj\|\cdot\|_{\mathrm{inj}} if the underlying space is clear from the context. Throughout this paper, we restrict ourselves to finite-dimensional Banach spaces, so that the algebraic tensor products equipped with either of these norms are complete. In this setting, the projective tensor product X⊗πYX\otimes_{\pi}Y can be identified isometrically with the space N(X∗,Y)\mathcal{N}(X^*,Y) of nuclear operators from X∗X^* into YY. Moreover, its dual can be identified isometrically with the space L(X,Y∗)\mathcal{L}(X,Y^*) of bounded linear operators from XX into Y∗Y^*, equipped with the operator norm.

The injective tensor product X⊗εYX\otimes_{\varepsilon}Y can be identified isometrically with L(X∗,Y)\mathcal{L}(X^*,Y). In particular, for A∈X⊗εYA\in X\otimes_{\varepsilon}Y, we have

∥A∥X⊗εY=sup⁡x∗∈BX∗∥Ax∗∥Y.\|A\|_{X\otimes_{\varepsilon}Y}=\sup_{x^*\in B_{X^*}}\|Ax^*\|_Y.

We refer the reader to the wonderful books of [10] and [18] for more information about tensor products.

A Lower Bound for the Identity Operator via Cyclic Convolution

Let n∈Nn\in\mathbb{N} and consider Xp,q,rn=ℓpn⊗πℓqn⊗πℓrnX^n_{p,q,r}=\ell_p^n\otimes_{\pi}\ell_q^n\otimes_{\pi}\ell_r^n with 1≤p≤q≤2≤r≤∞1\leq p\leq q\leq2\leq r\leq\infty. We show that the upper bound for the norm of the identity obtained in [13] is sharp in the region 1p+1q+1r≥2\frac{1}{p}+\frac{1}{q}+\frac{1}{r}\geq2. The argument is based on convolution on the finite cyclic group Zn:=Z/nZ\mathbb{Z}_n:=\mathbb{Z}/n\mathbb{Z} of nn elements, which we identify with {0,…,n−1}\{0,\ldots,n-1\}, with addition understood modulo nn. We use the additive structure of Zn\mathbb{Z}_n only to perform index arithmetic modulo nn and to define cyclic convolution, which allows us to apply Young’s convolution inequality; no further group-theoretic properties are needed. Throughout, vectors indexed by Zn\mathbb{Z}_n are identified with vectors in Rn\mathbb{R}^n, and all ℓs\ell_s-norms are taken with respect to counting measure, that is,

∥x∥s:=(∑i∈Zn∣xi∣s)1/s,1≤s<∞,\|x\|_s:=\left(\sum_{i\in\mathbb{Z}_n}|x_i|^s\right)^{1/s},\qquad1\leq s<\infty,

and

∥x∥∞:=max⁡i∈Zn∣xi∣.\|x\|_{\infty}:=\max_{i\in\mathbb{Z}_n}|x_i|.

We recall Young’s convolution inequality on the finite group Zn\mathbb{Z}_n. If 1≤a,b,c≤∞1\leq a,b,c\leq\infty satisfy

1a+1b=1+1c,\frac{1}{a}+\frac{1}{b}=1+\frac{1}{c},

then

∥x∗y∥c≤∥x∥a∥y∥b,(4)\lVert x * y\rVert_c \le\lVert x\rVert_a\lVert y\rVert_b, \tag*{(4)}

where

(x∗y)k:=∑i∈Znxiyk−i(5)(x * y)_k := \sum_{i\in\mathbb{Z}_n} x_i y_{k-i} \tag*{(5)}

is the discrete convolution of xx and yy. The constant in (4) is one (see, e.g., [12]).

Recall that, identifying the standard basis of Rn\mathbb{R}^n with (ei)i∈Zn(e_i)_{i\in\mathbb{Z}_n}, we may associate a trilinear form with every tensor in (Rn)⊗3(\mathbb{R}^n)^{\otimes3}. Indeed, using the standard Euclidean inner product on Rn\mathbb{R}^n, an elementary tensor ei⊗ej⊗eke_i\otimes e_j\otimes e_k corresponds to the trilinear form

(x,y,z)↦⟨ei,x⟩⟨ej,y⟩⟨ek,z⟩=xiyjzk.(x,y,z)\mapsto\langle e_i,x\rangle\langle e_j,y\rangle\langle e_k,z\rangle=x_i y_j z_k.

By linearity, a general tensor

A=∑i,j,k∈Znaijkei⊗ej⊗ekA=\sum_{i,j,k\in\mathbb{Z}_n} a_{ijk}e_i\otimes e_j\otimes e_k

is therefore identified with the trilinear form

TA(x,y,z)=∑i,j,k∈Znaijkxiyjzk,x,y,z∈Rn.T_A(x,y,z)=\sum_{i,j,k\in\mathbb{Z}_n}a_{ijk}x_i y_j z_k,\qquad x,y,z\in\mathbb{R}^n.

In what follows, we shall also use the notation x+:=max⁡{x,0}x_+:=\max\{x,0\}, x∈Rx\in\mathbb{R}. Our construction below is inspired by the classical structure tensor of a finite group algebra and, more specifically, by cyclic group tensors (see, e.g., Cohn and Umans [9], Alexeev, Forbes, and Tsimerman [1], and Alman [2]). For the cyclic group Zn\mathbb{Z}_n, the tensor AA defined below coincides, up to a permutation of its tensor factors, with the structure tensor of the group algebra R[Zn]\mathbb{R}[\mathbb{Z}_n], whose multiplication corresponds to cyclic convolution.

Proposition 3.1. Let 1≤p≤q≤2≤r≤∞1\le p\le q\le2\le r\le\infty and put S:=1p+1q+1rS:=\frac{1}{p}+\frac{1}{q}+\frac{1}{r}. For the tensor A:=∑i,k∈Znei⊗ei+k⊗ekA:=\sum_{i,k\in\mathbb{Z}_n}e_i\otimes e_{i+k}\otimes e_k and its associated trilinear form TA:ℓpn×ℓqn×ℓrn→RT_A:\ell_p^n\times\ell_q^n\times\ell_r^n\to\mathbb{R} with TA(x,y,z):=∑i,k∈Znxiyi+kzkT_A(x,y,z):=\sum_{i,k\in\mathbb{Z}_n}x_i y_{i+k}z_k, we have

∥TA∥=n(2−S)+.(6)\lVert T_A\rVert=n^{(2-S)_+}. \tag*{(6)}

Moreover,

∥A∥ℓ2n23=n.\lVert A\rVert_{\ell_2^{n^2}}^3=n.

Proof. The tensor AA in its coefficient array form A=(aijk)i,j,k∈ZnA=(a_{ijk})_{i,j,k\in\mathbb{Z}_n} has coefficients

aijk={1:j=i+k(modn),0:otherwise.a_{ijk}= \begin{cases} 1 & : j=i+k\pmod n,\\ 0 & : \text{otherwise}. \end{cases}

Now for every pair (i,k)∈Zn2(i,k)\in\mathbb{Z}_n^2 there is precisely one j∈Znj\in\mathbb{Z}_n for which aijk=1a_{ijk}=1. Thus AA has exactly n2n^2 non-zero coefficients, all of modulus one, and therefore

∥A∥ℓ2n23=(∑i,j,k∈Zn∣aijk∣2)1/2=n.\lVert A\rVert_{\ell_2^{n^2}}^3=\left(\sum_{i,j,k\in\mathbb{Z}_n}|a_{ijk}|^2\right)^{1/2}=n.

This establishes the second assertion of the proposition.

It remains to compute the norm of TAT_A. The key observation is that the trilinear form is associated with cyclic convolution, i.e.,

TA(x,y,z)=∑i,k∈Znxiyi+kzk=∑j∈Znyj∑i∈Znxizj−i=∑j∈Znyj(x∗z)j=⟨x∗z,y⟩.(7)\begin{aligned} T_A(x,y,z) &= \sum_{i,k\in\mathbb{Z}_n} x_i y_{i+k}z_k \\ &= \sum_{j\in\mathbb{Z}_n} y_j \sum_{i\in\mathbb{Z}_n} x_i z_{j-i} \\ &= \sum_{j\in\mathbb{Z}_n} y_j(x*z)_j \\ &= \langle x*z,y\rangle. \tag*{(7)} \end{aligned}

Here, we have used the substitution j=i+kj=i+k in Zn\mathbb{Z}_n and ⟨⋅,⋅⟩\langle\cdot,\cdot\rangle denotes the standard Euclidean inner product on Rn\mathbb{R}^n. We first prove the upper bound. Equip Zn\mathbb{Z}_n with the normalized counting measure

μn:=1n∑j∈Znδj\mu_n := \frac{1}{n}\sum_{j\in\mathbb{Z}_n}\delta_j

and write ∥⋅∥Ls\|\cdot\|_{L_s} for the corresponding Ls(μn)L_s(\mu_n)-norm. Recall that the relationship between unnormalized (ℓs\ell_s) and normalized (LsL_s) norms is

∥x∥Ls=n−1/s∥x∥s.(8)\|x\|_{L_s}=n^{-1/s}\|x\|_s. \tag*{(8)}

We also use the normalized convolution

(f∗μng)(k):=1n∑i∈Znfigk−i.(f*_{\mu_n}g)(k):=\frac{1}{n}\sum_{i\in\mathbb{Z}_n}f_i g_{k-i}.

In this setting, Young’s inequality (4) takes the form

∥f∗μng∥Lc≤∥f∥La∥g∥Lb(9)\|f*_{\mu_n}g\|_{L_c}\leq\|f\|_{L_a}\|g\|_{L_b} \tag*{(9)}

whenever

1a+1b=1+1c.\frac{1}{a}+\frac{1}{b}=1+\frac{1}{c}.

Suppose first that S≥2S\geq2. Then

1p+1q≥2−1r=1+1r∗,\frac{1}{p}+\frac{1}{q}\geq2-\frac{1}{r}=1+\frac{1}{r^*},

where r∗r^* denotes the Hölder conjugate of rr defined by the relation 1r+1r∗=1\frac{1}{r}+\frac{1}{r^*}=1. Hence, there exist a≥pa\geq p and b≥qb\geq q such that

1a+1b=2−1r=1+1r∗.(10)\frac{1}{a}+\frac{1}{b}=2-\frac{1}{r}=1+\frac{1}{r^*}. \tag*{(10)}

Since a≥pa\geq p and b≥qb\geq q, monotonicity of the unnormalized ℓs\ell_s-norms gives ∥x∥a≤∥x∥p\|x\|_a\leq\|x\|_p and ∥y∥b≤∥y∥q\|y\|_b\leq\|y\|_q. Now define the reflected vector x~\widetilde{x} by x~i:=x−i\widetilde{x}_i:=x_{-i}, i∈Zni\in\mathbb{Z}_n, and put

wk:=∑i∈Znxiyi+k,k∈Zn.w_k:=\sum_{i\in\mathbb{Z}_n}x_i y_{i+k},\qquad k\in\mathbb{Z}_n.

Then, by the definition of cyclic convolution (5),

(x~∗y)k=∑j∈Znx~jyk−j=∑j∈Znx−jyk−j=∑i∈Znxiyk+i=wk,(\widetilde{x}*y)_k=\sum_{j\in\mathbb{Z}_n}\widetilde{x}_j y_{k-j}=\sum_{j\in\mathbb{Z}_n}x_{-j}y_{k-j}=\sum_{i\in\mathbb{Z}_n}x_i y_{k+i}=w_k,

where we used the substitution i=−ji=-j in Zn\mathbb{Z}_n. Thus w=x~∗yw=\widetilde{x}*y. Moreover, we have ∥x~∥a=∥x∥a\|\widetilde{x}\|_a=\|x\|_a. Therefore, by Young’s inequality (4) and (9),

∥w∥r∗=∥x~∗y∥r∗≤∥x~∥a∥y∥b=∥x∥a∥y∥b≤∥x∥p∥y∥q.\|w\|_{r^*}=\|\widetilde{x}*y\|_{r^*}\leq\|\widetilde{x}\|_a\|y\|_b=\|x\|_a\|y\|_b\leq\|x\|_p\|y\|_q.

Consequently, since TA(x,y,z)=∑k∈ZnwkzkT_A(x,y,z)=\sum_{k\in\mathbb{Z}_n}w_kz_k, Hölder’s inequality gives

∣TA(x,y,z)∣=∣∑k∈Znwkzk∣≤∥w∥r∗∥z∥r≤∥x∥p∥y∥q∥z∥r.\begin{aligned} \lvert T_A(x,y,z)\rvert&=\left\lvert\sum_{k\in\mathbb{Z}_n}w_kz_k\right\rvert\\ &\le\lVert w\rVert_{r^*}\lVert z\rVert_r\\ &\le\lVert x\rVert_p\lVert y\rVert_q\lVert z\rVert_r. \end{aligned}

and so ∥TA∥≤1\lVert T_A\rVert\le1. Taking x=y=z=e0x=y=z=e_0 shows that ∥TA∥≥1\lVert T_A\rVert\ge1. Thus, whenever S≥2S\ge2,

∥TA∥=1.\lVert T_A\rVert=1.

Suppose now that S<2S<2. Then

1p+1q<2−1r.\frac{1}{p}+\frac{1}{q}<2-\frac{1}{r}.

Since 1≤p≤q≤21\le p\le q\le2, we have 1/p+1/q≥11/p+1/q\ge1, while 2−1/r≤22-1/r\le2. Hence, we may choose a≤pa\le p and b≤qb\le q such that

1a+1b=2−1r=1+1r∗.\frac{1}{a}+\frac{1}{b}=2-\frac{1}{r}=1+\frac{1}{r^*}.

Since (Zn,μn)(\mathbb{Z}_n,\mu_n) is a probability space, the monotonicity of LsL_s-norms for a≤pa\le p and b≤qb\le q yields

∥x∥La≤∥x∥Lpand∥y∥Lb≤∥y∥Lq.\lVert x\rVert_{L_a}\le\lVert x\rVert_{L_p} \qquad\text{and}\qquad \lVert y\rVert_{L_b}\le\lVert y\rVert_{L_q}.

Applying Young’s convolution inequality in the form (8) and then using (7), we obtain

n−1−1/r∗∥w∥r∗=∥x∗~μny∥Lr∗≤∥x∥La∥y∥Lb≤∥x∥Lp∥y∥Lq=n−1/p−1/q∥x∥p∥y∥q.\begin{aligned} n^{-1-1/r^*}\lVert w\rVert_{r^*} &=\lVert x\widetilde{*}_{\mu_n}y\rVert_{L_{r^*}}\\ &\le\lVert x\rVert_{L_a}\lVert y\rVert_{L_b}\\ &\le\lVert x\rVert_{L_p}\lVert y\rVert_{L_q}\\ &=n^{-1/p-1/q}\lVert x\rVert_p\lVert y\rVert_q. \end{aligned}

Since 1+1/r∗=2−1/r1+1/r^*=2-1/r, this yields

∥w∥r∗≤n2−S∥x∥p∥y∥q.(11)\lVert w\rVert_{r^*}\le n^{2-S}\lVert x\rVert_p\lVert y\rVert_q. \tag*{(11)}

Hölder’s inequality with conjugate exponents r∗r^* and rr, together with (10), therefore yields

∣TA(x,y,z)∣=∣∑k∈Znwkzk∣≤∥w∥r∗∥z∥r≤n2−S∥x∥p∥y∥q∥z∥r.\begin{aligned} \lvert T_A(x,y,z)\rvert&=\left\lvert\sum_{k\in\mathbb{Z}_n}w_kz_k\right\rvert\\ &\le\lVert w\rVert_{r^*}\lVert z\rVert_r\\ &\le n^{2-S}\lVert x\rVert_p\lVert y\rVert_q\lVert z\rVert_r. \end{aligned}

and hence

∥TA∥≤n2−S.\lVert T_A\rVert\le n^{2-S}.

For the reverse inequality, let

x=n−1/p(1,…,1),y=n−1/q(1,…,1),z=n−1/r(1,…,1).x=n^{-1/p}(1,\ldots,1),\qquad y=n^{-1/q}(1,\ldots,1),\qquad z=n^{-1/r}(1,\ldots,1).

Then ∥x∥p=∥y∥q=∥z∥r=1\lVert x\rVert_p=\lVert y\rVert_q=\lVert z\rVert_r=1, with the usual interpretation when r=∞r=\infty, and

TA(x,y,z)=∑i,k∈Znn−1/pn−1/qn−1/r=n2−S.T_A(x,y,z)=\sum_{i,k\in\mathbb{Z}_n}n^{-1/p}n^{-1/q}n^{-1/r}=n^{2-S}.

Thus ∥TA∥≥n2−S\lVert T_A\rVert\ge n^{2-S}. Together with the upper estimate this proves (6).

Corollary 3.2. Let $1 \le p \le q \le 2 \le r \le \infty$. Then

∥id⁡:ℓ2n3⟶Xp,q,rn∥≳p,q,rnγ(p,q,r),(12)\left\lVert\operatorname{id}: \ell_2^{n^3} \longrightarrow X_{p,q,r}^n \right\rVert\gtrsim_{p,q,r} n^{\gamma(p,q,r)}, \tag*{(12)}

where

γ(p,q,r):=max⁡{12+1r,1q,min⁡{1,1p+1q+1r−1}}.(13)\gamma(p,q,r) := \max\left\{\frac{1}{2}+\frac{1}{r},\frac{1}{q},\min\left\{1,\frac{1}{p}+\frac{1}{q}+\frac{1}{r}-1\right\}\right\}. \tag*{(13)}

The third term in (12) is furnished by the cyclic convolution tensor. More precisely,

∥id⁡:ℓ2n3⟶Xp,q,rn∥≥nmin⁡{1,1/p+1/q+1/r−1}.\left\lVert\operatorname{id}: \ell_2^{n^3} \longrightarrow X_{p,q,r}^n \right\rVert\ge n^{\min\{1,1/p+1/q+1/r-1\}}.

Proof. By duality, Proposition 3.1 gives

∥id⁡:ℓ2n3⟶Xp,q,rn∥=∥id⁡:(Xp,q,rn)∗⟶ℓ2n3∥\left\lVert\operatorname{id}: \ell_2^{n^3} \longrightarrow X_{p,q,r}^n \right\rVert= \left\lVert\operatorname{id}: (X_{p,q,r}^n)^* \longrightarrow\ell_2^{n^3} \right\rVert
=sup⁡0≠B∈(Xp,q,rn)∗∥B∥ℓ2n3∥B∥(Xp,q,rn)∗.= \sup_{0\ne B\in(X_{p,q,r}^n)^*} \frac{\lVert B\rVert_{\ell_2^{n^3}}}{\lVert B\rVert_{(X_{p,q,r}^n)^*}}.

Since the dual space (Xp,q,rn)∗(X_{p,q,r}^n)^* is canonically identified with the space of bounded trilinear forms on ℓpn×ℓqn×ℓrn\ell_p^n \times\ell_q^n \times\ell_r^n (see, e.g., [18 Section 2.2] in the two-fold case), the coefficient array AA represents an element of (Xp,q,rn)∗(X_{p,q,r}^n)^* with norm ∥TA∥\lVert T_A\rVert. Hence, evaluating the supremum at B=AB=A yields

∥id⁡:ℓ2n3⟶Xp,q,rn∥≥∥A∥ℓ2n3∥TA∥\left\lVert\operatorname{id}: \ell_2^{n^3} \longrightarrow X_{p,q,r}^n \right\rVert\ge\frac{\lVert A\rVert_{\ell_2^{n^3}}}{\lVert T_A\rVert}
=n1−(2−S)+= n^{1-(2-S)_+}
=nmin⁡{1,S−1},= n^{\min\{1,S-1\}},

where S=1/p+1/q+1/rS=1/p+1/q+1/r. Note that S≥1S\ge1 since p,q≤2p,q\le2, so the exponent is non-negative. Combining this estimate with the first and third lower bounds in Lemma 5.6 of [13] (which provide bounds of n1/2+1/rn^{1/2+1/r} and n1/qn^{1/q}, respectively) proves (11).

Remark 3.3. The cyclic term in (12) always dominates the second term in Lemma 5.6 of [13]. Indeed, writing S=1/p+1/q+1/rS=1/p+1/q+1/r, if 1/q+1/r≤11/q+1/r\le1, then S≤2S\le2 and

min⁡{1,S−1}=1p+1q+1r−1,\min\{1,S-1\}=\frac{1}{p}+\frac{1}{q}+\frac{1}{r}-1,

which coincides with the second exponent in Lemma 5.6. If instead 1/q+1/r>11/q+1/r>1, then

min⁡{1,S−1}≥1p,\min\{1,S-1\}\ge\frac{1}{p},

whereas the second exponent in Lemma 5.6 equals 1/p1/p.

To determine when the full lower bound is strictly improved, observe that for 1/q+1/r≤11/q+1/r\le1, the old and new lower exponents coincide. If 1/q+1/r>11/q+1/r>1, then, since p≤qp\le q, the lower exponent from Lemma 5.6 is

max⁡{12+1r,1p},\max\left\{\frac{1}{2}+\frac{1}{r},\frac{1}{p}\right\},

whereas (12) gives

max⁡{12+1r,min⁡{1,S−1}}.\max\left\{\frac{1}{2}+\frac{1}{r},\min\{1,S-1\}\right\}.

Consequently, the new lower exponent is strictly larger if and only if

min⁡{1,S−1}>max⁡{12+1r,1p}.\min\{1,S-1\}>\max\left\{\frac{1}{2}+\frac{1}{r},\frac{1}{p}\right\}.

This is equivalent to

p>1,r>2,1q+1r>1,1p+1q>32.p>1,\qquad r>2,\qquad\frac{1}{q}+\frac{1}{r}>1,\qquad\frac{1}{p}+\frac{1}{q}>\frac{3}{2}.

Thus our cyclic construction provides a genuine improvement of the lower bound in Lemma 5.6 of [13] precisely in this non-empty region.

Example 3.4. Take

p=65,q=54,r=3.p=\frac{6}{5},\qquad q=\frac{5}{4},\qquad r=3.

Then

S=56+45+13=5930<2,S=\frac{5}{6}+\frac{4}{5}+\frac{1}{3}=\frac{59}{30}<2,

and Corollary 3.2 gives

∥id⁡:ℓ2n3⟶ℓ6/5n⊗πℓ5/4n⊗πℓ3n∥≳n29/30.\left\|\operatorname{id}:\ell_{2}^{n^{3}}\longrightarrow\ell_{6/5}^{n}\otimes_{\pi}\ell_{5/4}^{n}\otimes_{\pi}\ell_{3}^{n}\right\|\gtrsim n^{29/30}.

By contrast, the three exponents in Lemma 5.6 of [13] are

12+13=56,1+56−1=56,45,\frac{1}{2}+\frac{1}{3}=\frac{5}{6},\qquad1+\frac{5}{6}-1=\frac{5}{6},\qquad\frac{4}{5},

so the previously known lower bound has exponent 5/65/6. Thus the cyclic convolution tensor improves the exponent from 5/65/6 to 29/3029/30.

Proposition 3.5. Let 1≤p≤q≤2≤r≤∞1\leq p\leq q\leq2\leq r\leq\infty and assume that 1p+1q+1r≥2\frac{1}{p}+\frac{1}{q}+\frac{1}{r}\geq2. Then

∥id⁡:ℓ2n3⟶Xp,q,rn∥≍p,q,rn.(14)\left\|\operatorname{id}:\ell_{2}^{n^{3}}\longrightarrow X_{p,q,r}^{n}\right\|\asymp_{p,q,r} n. \tag*{(14)}

In fact, the lower estimate in (13) holds with constant one.

Proof. If S=1/p+1/q+1/r≥2S=1/p+1/q+1/r\geq2, Proposition 3.1 (more precisely, (6)) gives ∥TA∥=1\|T_A\|=1, while ∥A∥ℓ2n3=n\|A\|_{\ell_{2}^{n^{3}}}=n. Hence, by duality,

∥id⁡:ℓ2n3⟶Xp,q,rn∥≥n.\left\|\operatorname{id}:\ell_{2}^{n^{3}}\longrightarrow X_{p,q,r}^{n}\right\|\geq n.

For the reverse estimate, Lemma 5.5 in [13] gives

∥id⁡:ℓ2n3⟶Xp,q,rn∥≲p,q,rnmin⁡{1/q+1/r,1}.(15)\left\|\operatorname{id}:\ell_{2}^{n^{3}}\longrightarrow X_{p,q,r}^{n}\right\|\lesssim_{p,q,r} n^{\min\{1/q+1/r,1\}}. \tag*{(15)}

Since S≥2S\geq2 and 1/p≤11/p\leq1, we have

1q+1r≥2−1p≥1.\frac{1}{q}+\frac{1}{r}\geq2-\frac{1}{p}\geq1.

Thus the right-hand side of (14) is of order nn, which proves the assertion. □\square

Corollary 3.6. Let 1≤p≤q≤2≤r≤∞1\leq p\leq q\leq2\leq r\leq\infty and suppose that 1p+1q+1r≥2\frac{1}{p}+\frac{1}{q}+\frac{1}{r}\geq2. Then

vr⁡(Xp,q,rn)≍p,q,rn1/2−1/r.(16)\operatorname{vr}(X_{p,q,r}^{n})\asymp_{p,q,r} n^{1/2-1/r}. \tag*{(16)}

Proof. Theorem B in [13] gives, in the regime $1 \le p \le q \le 2 \le r \le \infty$,

vol⁡n3(Bp,q,rXn)1/n3≍p,q,rn−2−1/r.(17)\operatorname{vol}_{n^3}\left(B_{p,q,r}^{X^n}\right)^{1/n^3} \asymp_{p,q,r} n^{-2-1/r}. \tag*{(17)}

Moreover, Xp,q,rnX_{p,q,r}^n has enough symmetries. Formula (2.3) in [13] therefore yields

vr⁡(Xp,q,rn)≍n3/2vol⁡n3(Bp,q,rXn)1/n3∥id⁡:ℓ2n3⟶Xp,q,rn∥.(18)\operatorname{vr}\left(X_{p,q,r}^n\right) \asymp n^{3/2}\operatorname{vol}_{n^3}\left(B_{p,q,r}^{X^n}\right)^{1/n^3}\left\|\operatorname{id}:\ell_2^{n^3}\longrightarrow X_{p,q,r}^n\right\|. \tag*{(18)}

Combining (17), (18) and Proposition 3.5, we obtain

vr⁡(Xp,q,rn)≍p,q,rn3/2n−2−1/rn=n1/2−1/r,\operatorname{vr}\left(X_{p,q,r}^n\right) \asymp_{p,q,r} n^{3/2}n^{-2-1/r}n=n^{1/2-1/r},

which proves (16).

Remark 3.7. Corollary 3.2 also quantifies the improvement outside the exact region of Proposition 3.5; see Remark 3.3 and Example 3.4.

Example 3.8. Take

p=2019,q=54,r=4.p=\frac{20}{19},\qquad q=\frac{5}{4},\qquad r=4.

Then 1<p<q<2<r1<p<q<2<r and

1p+1q+1r=1920+45+14=2.\frac{1}{p}+\frac{1}{q}+\frac{1}{r}=\frac{19}{20}+\frac{4}{5}+\frac{1}{4}=2.

Hence, Proposition 3.5 gives

∥id⁡:ℓ2n3⟶ℓ20/19n⊗πℓ5/4n⊗πℓ4n∥≍n.\left\|\operatorname{id}:\ell_2^{n^3}\longrightarrow\ell_{20/19}^n\otimes_\pi\ell_{5/4}^n\otimes_\pi\ell_4^n\right\|\asymp n.

The lower bound from Lemma 5.6 in [13] has exponent

max⁡{34,1920,45}=1920.\max\left\{\frac{3}{4},\frac{19}{20},\frac{4}{5}\right\}=\frac{19}{20}.

Thus, the cyclic convolution tensor improves the previously known lower-bound exponent from 19/2019/20 to 11 for this choice of parameters. Moreover, Corollary 3.6 gives

vr⁡(ℓ20/19n⊗πℓ5/4n⊗πℓ4n)≍n1/4.\operatorname{vr}\left(\ell_{20/19}^n\otimes_\pi\ell_{5/4}^n\otimes_\pi\ell_4^n\right)\asymp n^{1/4}.

A Lower Bound for the Identity Operator via Randomly Signed Cyclic Tensors

We now complement the deterministic cyclic construction by a randomized signing of the same support. The resulting estimate determines the optimal polynomial order of the norm of the identity throughout the full range 1≤p≤q≤2≤r≤∞1\le p\le q\le2\le r\le\infty, up to a power of a logarithm. In the region covered by Proposition 3.5, the deterministic result of the previous section is stronger and removes the logarithmic loss completely. For independent Rademacher random variables {εik:i,k∈Zn}\{\varepsilon_{ik}:i,k\in\mathbb{Z}_n\}, consider the random tensor

Aε:=∑i,k∈Znεikei⊗ei+k⊗ek(19)A_\varepsilon:=\sum_{i,k\in\mathbb{Z}_n}\varepsilon_{ik}e_i\otimes e_{i+k}\otimes e_k \tag*{(19)}

and the associated random trilinear form

Tε(x,y,z):=∑i,k∈Znεikxiyi+kzk.T_\varepsilon(x,y,z):=\sum_{i,k\in\mathbb{Z}_n}\varepsilon_{ik}x_i y_{i+k}z_k.

As before,

∥Aε∥ℓ2n3=n(20)\left\|A_\varepsilon\right\|_{\ell_2^{n^3}}=n \tag*{(20)}

for every realization of the signs. We first record the Hilbertian estimate that drives the argument.

Lemma 4.1. There is a constant C∈(0,∞)C \in(0,\infty) such that

Eε∥Tε∥ℓ2n×ℓ2n×ℓ2n≤C(log⁡(en))2.(21)\mathbb{E}_{\varepsilon}\|T_{\varepsilon}\|_{\ell_2^n \times\ell_2^n \times\ell_2^n} \le C(\log(en))^2. \tag*{(21)}

Consequently, for every n∈Nn \in\mathbb{N} there exists a deterministic choice of signs ε=(εik)\varepsilon= (\varepsilon_{ik}) for which

∥Tε∥ℓ2n×ℓ2n×ℓ2n≤C(log⁡(en))2.(22)\|T_{\varepsilon}\|_{\ell_2^n \times\ell_2^n \times\ell_2^n} \le C(\log(en))^2. \tag*{(22)}

Proof. We apply Corollary 1.4 of [8]. Define the random three-tensor X=(Xijk)X=(X_{ijk}) by

Xijk:={εik:j=i+k(modn),0:otherwise.X_{ijk}:= \begin{cases} \varepsilon_{ik} & : j=i+k \pmod n,\\ 0 & : \text{otherwise}. \end{cases}

Its entries are independent, centered, and bounded in absolute value by one; deterministic zero entries cause no difficulty. Moreover, every one-dimensional coordinate fiber of the support contains exactly one nonzero entry. Indeed, fixing any two of the indices i,j,ki,j,k, the relation j=i+kj=i+k (mod nn) uniquely determines the remaining index. Consequently,

max⁡j,k∈Zn(∑i∈ZnEXijk2)1/2=1,\max_{j,k\in\mathbb{Z}_n}\left(\sum_{i\in\mathbb{Z}_n}\mathbb{E}X_{ijk}^{2}\right)^{1/2}=1,
max⁡i,k∈Zn(∑j∈ZnEXijk2)1/2=1,\max_{i,k\in\mathbb{Z}_n}\left(\sum_{j\in\mathbb{Z}_n}\mathbb{E}X_{ijk}^{2}\right)^{1/2}=1,
max⁡i,j∈Zn(∑k∈ZnEXijk2)1/2=1.\max_{i,j\in\mathbb{Z}_n}\left(\sum_{k\in\mathbb{Z}_n}\mathbb{E}X_{ijk}^{2}\right)^{1/2}=1.

Corollary 1.4 of [8], specialized to tensors of order three and side length nn, therefore yields

Eε∥X∥inj≤123+C0(log⁡(en))2≲(log⁡(en))2.\mathbb{E}_{\varepsilon}\|X\|_{\mathrm{inj}} \le12\sqrt{3}+C_0(\log(en))^2 \lesssim(\log(en))^2.

Since the injective norm of XX equals the trilinear form norm of TεT_{\varepsilon} on ℓ2n×ℓ2n×ℓ2n\ell_2^n \times\ell_2^n \times\ell_2^n, this proves (21). The deterministic assertion follows by choosing a realization whose norm does not exceed the expectation.

Fix from now on a choice of signs satisfying (22), and put Ln:=C(log⁡(en))2L_n:=C(\log(en))^2. The same signed tensor has two further endpoint estimates. First, the one-point-fiber property of the cyclic support gives

∥Tε∥ℓ1n×ℓ1n×ℓ∞n=1.(23)\|T_{\varepsilon}\|_{\ell_1^n \times\ell_1^n \times\ell_\infty^n}=1. \tag*{(23)}

Indeed, for a trilinear form with coefficients (aijk)(a_{ijk}), we have

∥T∥ℓ1n×ℓ1n×ℓ∞n=max⁡i,j∑k∣aijk∣,\|T\|_{\ell_1^n \times\ell_1^n \times\ell_\infty^n}=\max_{i,j}\sum_k|a_{ijk}|,

and for the tensor in (18) the sum on the right-hand side equals one. Second, the Cauchy–Schwarz inequality in the kk-variable and (22) give

∥Tε∥ℓ2n×ℓ2n×ℓ∞n≤nLn.(24)\|T_{\varepsilon}\|_{\ell_2^n \times\ell_2^n \times\ell_\infty^n}\le\sqrt{n}L_n. \tag*{(24)}

Indeed, if ∥x∥2,∥y∥2≤1\|x\|_2,\|y\|_2\le1, then

sup⁡∥z∥∞≤1∣Tε(x,y,z)∣=∑k∣∑iεikxiyi+k∣≤n(∑k∣∑iεikxiyi+k∣2)1/2≤nLn.\sup_{\|z\|_\infty\le1}|T_{\varepsilon}(x,y,z)| =\sum_k\left|\sum_i\varepsilon_{ik}x_i y_{i+k}\right| \le\sqrt{n}\left(\sum_k\left|\sum_i\varepsilon_{ik}x_i y_{i+k}\right|^2\right)^{1/2} \le\sqrt{n}L_n.

Theorem 4.2. Let $1 \le p \le q \le 2 \le r \le \infty$. Then

∥id⁡:ℓ2n3⟶Xp,q,rn∥≳nmin⁡{1/q+1/r,1}(log⁡(en))4(1−1/q).(25)\left\lVert\operatorname{id}:\ell_2^{n^3}\longrightarrow X_{p,q,r}^n\right\rVert\gtrsim\frac{n^{\min\{1/q+1/r,1\}}}{(\log(en))^{4(1-1/q)}}. \tag*{(25)}

Consequently, together with the upper estimate from [13 Lemma 5.5, (5.8b)],

nmin⁡{1/q+1/r,1}(log⁡(en))4(1−1/q)≲∥id⁡:ℓ2n3⟶Xp,q,rn∥≲p,q,rnmin⁡{1/q+1/r,1}.(26)\frac{n^{\min\{1/q+1/r,1\}}}{(\log(en))^{4(1-1/q)}}\lesssim\left\lVert\operatorname{id}:\ell_2^{n^3}\longrightarrow X_{p,q,r}^n\right\rVert\lesssim_{p,q,r} n^{\min\{1/q+1/r,1\}}. \tag*{(26)}

In particular, the optimal polynomial growth exponent of the norm of the identity is determined throughout this parameter range up to logarithmic factors.

Proof. We first consider the region

1q+1r≤1.(27)\frac{1}{q}+\frac{1}{r}\le1. \tag*{(27)}

Set

a:=2q−1,b:=2r,c:=2−2q−2r.a:=\frac{2}{q}-1,\qquad b:=\frac{2}{r},\qquad c:=2-\frac{2}{q}-\frac{2}{r}.

Under (27), we have a,b,c≥0a,b,c\ge0 and a+b+c=1a+b+c=1. Moreover,

1q=a+b+c2,1r=b2.\frac{1}{q}=a+\frac{b+c}{2},\qquad\frac{1}{r}=\frac{b}{2}.

We now interpolate between the three endpoint spaces ℓ1n×ℓ1n×ℓ∞n\ell_1^n\times\ell_1^n\times\ell_\infty^n, ℓ2n×ℓ2n×ℓ∞n\ell_2^n\times\ell_2^n\times\ell_\infty^n, and ℓ2n×ℓ2n×ℓ2n\ell_2^n\times\ell_2^n\times\ell_2^n. By the multilinear Riesz–Thorin interpolation inequality for real-valued forms, obtained from the complex multilinear interpolation theorem (see, e.g., [7 Section 4.4]) with an absolute multiplicative constant at most 8, and using the endpoint estimates (22), (21), and (23), we obtain

∥Tε∥ℓqn×ℓqn×ℓrn≤8⋅1aLnb(nLn)c=8n1−1/q−1/rLn2−2/q\begin{aligned} \left\lVert T_\varepsilon\right\rVert_{\ell_q^n\times\ell_q^n\times\ell_r^n}\le8\cdot1^aL_n^b(\sqrt{n}L_n)^c \\ &=8n^{1-1/q-1/r}L_n^{2-2/q} \end{aligned}
≲n1−1/q−1/r(log⁡(en))4(1−1/q).(28)\lesssim n^{1-1/q-1/r}(\log(en))^{4(1-1/q)}. \tag*{(28)}

Since p≤qp\le q, we have Bℓpn⊆BℓqnB_{\ell_p^n}\subseteq B_{\ell_q^n}, and therefore the same upper estimate holds for ∥Tε∥ℓpn×ℓqn×ℓrn\left\lVert T_\varepsilon\right\rVert_{\ell_p^n\times\ell_q^n\times\ell_r^n}. By duality and (19),

∥id⁡:ℓ2n3⟶Xp,q,rn∥≳n∥Tε∥ℓpn×ℓqn×ℓrn≳n1/q+1/r(log⁡(en))4(1−1/q).\left\lVert\operatorname{id}:\ell_2^{n^3}\longrightarrow X_{p,q,r}^n\right\rVert\gtrsim\frac{n}{\left\lVert T_\varepsilon\right\rVert_{\ell_p^n\times\ell_q^n\times\ell_r^n}}\gtrsim\frac{n^{1/q+1/r}}{(\log(en))^{4(1-1/q)}}.

Since 1/q+1/r≤11/q+1/r\le1, this establishes (25) in the present case.

Suppose now that

1q+1r≥1.(29)\frac{1}{q}+\frac{1}{r}\ge1. \tag*{(29)}

Let r0:=q∗r_0:=q^*, so that 1q+1r0=1\frac{1}{q}+\frac{1}{r_0}=1. The boundary case of (28) gives

∥Tε∥ℓqn×ℓqn×ℓr0n≲(log⁡(en))4(1−1/q).\left\lVert T_\varepsilon\right\rVert_{\ell_q^n\times\ell_q^n\times\ell_{r_0}^n}\lesssim(\log(en))^{4(1-1/q)}.

Now, condition (29) is equivalent to r≤r0r\le r_0. Hence, we have the inclusion Bℓrn⊆Bℓr0nB_{\ell_r^n}\subseteq B_{\ell_{r_0}^n} and consequently

∥Tε∥ℓqn×ℓqn×ℓrn≲(log⁡(en))4(1−1/q).\left\lVert T_\varepsilon\right\rVert_{\ell_q^n\times\ell_q^n\times\ell_r^n}\lesssim(\log(en))^{4(1-1/q)}.

Again using p≤qp\le q, duality, and (19), we obtain

∥id⁡:ℓ2n3⟶Xp,q,rn∥≳n(log⁡(en))4(1−1/q).\left\lVert\operatorname{id}:\ell_2^{n^3}\longrightarrow X_{p,q,r}^n\right\rVert\gtrsim\frac{n}{(\log(en))^{4(1-1/q)}}.

Since 1/q+1/r≥11/q+1/r\ge1, this establishes (25) in the remaining case. Finally, the upper estimate in (26) follows from [13 Lemma 5.5, (5.8b)].

Remark 4.3. Theorem 4.2 and the deterministic results of § 4 complement each other. The former determines the optimal polynomial growth exponent of the norm of the identity throughout the range $1 \le p \le q \le 2 \le r \le \infty$, up to logarithmic factors. In the region 1p+1q+1r≥2\frac{1}{p}+\frac{1}{q}+\frac{1}{r} \ge2, Proposition 3.5 yields the stronger, log-free estimate

∥id⁡:ℓ2n3⟶Xp,q,rn∥≍p,q,rn.\left\lVert\operatorname{id}:\ell_2^{n^3}\longrightarrow X_{p,q,r}^n\right\rVert\asymp_{p,q,r} n.

Thus, the deterministic cyclic construction gives the exact asymptotic order in this region, while the randomized construction determines the optimal polynomial growth exponent throughout the full parameter range.

Corollary 4.4. Let $1 \le p \le q \le 2 \le r \le \infty$. Then

nmin⁡{1/q−1/2,1/2−1/r}(log⁡(en))4(1−1/q)≲p,q,rvr⁡(Xp,q,rn)≲p,q,rnmin⁡{1/q−1/2,1/2−1/r}.\frac{n^{\min\{1/q-1/2,1/2-1/r\}}}{(\log(en))^{4(1-1/q)}} \lesssim_{p,q,r} \operatorname{vr}(X_{p,q,r}^n) \lesssim_{p,q,r} n^{\min\{1/q-1/2,1/2-1/r\}}.

Proof. By (17) and (18),

vr⁡(Xp,q,rn)≍p,q,rn−1/2−1/r∥id⁡:ℓ2n3⟶Xp,q,rn∥.\operatorname{vr}(X_{p,q,r}^n) \asymp_{p,q,r} n^{-1/2-1/r}\left\lVert\operatorname{id}:\ell_2^{n^3}\longrightarrow X_{p,q,r}^n\right\rVert.

If 1/q+1/r≤11/q+1/r \le1, then

−12−1r+1q+1r=1q−12.-\frac{1}{2}-\frac{1}{r}+\frac{1}{q}+\frac{1}{r}=\frac{1}{q}-\frac{1}{2}.

whereas if 1/q+1/r≥11/q+1/r \ge1, then

−12−1r+1=12−1r.-\frac{1}{2}-\frac{1}{r}+1=\frac{1}{2}-\frac{1}{r}.

These two exponents are respectively the minimum of 1/q−1/21/q-1/2 and 1/2−1/r1/2-1/r in the corresponding regions. The assertion now follows from Theorem 4.2. □

The logarithmic loss in Theorem 4.2 arises from the estimate for the injective norm of the randomly signed cyclic tensor. It is natural to ask whether this loss is intrinsic to the embedding norm. We conjecture the following.

Conjecture 4.5. For every $1 \le p \le q \le 2 \le r \le \infty$,

∥id⁡:ℓ2n3⟶Xp,q,rn∥≍p,q,rnmin⁡{1/q+1/r,1}.\left\lVert\operatorname{id}:\ell_2^{n^3}\longrightarrow X_{p,q,r}^n\right\rVert\asymp_{p,q,r} n^{\min\{1/q+1/r,1\}}.

In particular, the conjecture would imply that

vr⁡(Xp,q,rn)≍p,q,rnmin⁡{1/q−1/2,1/2−1/r}\operatorname{vr}(X_{p,q,r}^n) \asymp_{p,q,r} n^{\min\{1/q-1/2,1/2-1/r\}}

throughout the full parameter range, without logarithmic corrections.

Acknowledgement

JP has been supported by the German Research Foundation under DFG project 516672205 *Limit theorems for the volume of random projections of ℓp\ell_p-balls$.

ChatGPT (GPT-6, OpenAI) was used for grammar and spelling checks and to assist in the construction of Examples 3.4 and 3.8.

References

  1. [1]B. Alexeev, M. A. Forbes, and J. Tsimerman. Tensor rank: Some lower and upper bounds. In 2011 IEEE 26th Annual Conference on Computational Complexity, pages 283–291, 2011.DOI
  2. [2]J. Alman. Limits on the universal method for matrix multiplication. In Proceedings of the 34th Computational Complexity Conference (CCC 2019), volume 137 of Leibniz International Proceedings in Informatics (LIPIcs), pages 12:1–12:24, Dagstuhl, Germany, 2019. Schloss Dagstuhl–Leibniz-Zentrum für Informatik.arxiv.org/abs/1812.08731
  3. [3]G. Aubrun and S. J. Szarek. Tensor products of convex sets and the volume of separable states on n qudits. Phys. Rev. A, 73:022109, Feb 2006.DOI
  4. [4]G. Aubrun, S. J. Szarek, and E. Werner. Nonadditivity of Rényi entropy and Dvoretzky’s theorem. J. Math. Phys., 51(2):022102, 7, 2010.DOI
  5. [5]G. Aubrun, S. J. Szarek, and E. Werner. Hastings’s additivity counterexample via Dvoretzky’s theorem. Comm. Math. Phys., 305(1):85–97, 2011.
  6. [6]A. S. Bandeira and R. van Handel. Sharp nonasymptotic bounds on the norm of random matrices with independent entries. Ann. Probab., 44(4):2479–2506, 2016.DOI
  7. [7]J. Bergh and J. Löfström. Interpolation spaces. An introduction, volume 223 of Grundlehren der mathematischen Wissenschaften. Springer-Verlag, Berlin-New York, 1976.
  8. [8]M. T. Boedihardjo. Injective norm of random tensors with independent entries, 2024. arXiv:2412.21193.arxiv.org/abs/2412.21193
  9. [9]H. Cohn and C. Umans. A group-theoretic approach to fast matrix multiplication. In 44th Annual IEEE Symposium on Foundations of Computer Science, 2003. Proceedings., pages 438–449, 2003.DOI
  10. [10]J. Diestel, J. H. Fourie, and J. Swart. The metric theory of tensor products. American Mathematical Society, Providence, RI, 2008. Grothendieck’s résumé revisited.DOI
  11. [11]K. Efremenko. 3-query locally decodable codes of subexponential length. In Proceedings of the Forty-First Annual ACM Symposium on Theory of Computing, STOC ’09, page 39–44, New York, NY, USA, 2009. Association for Computing Machinery.DOI
  12. [12]G. B. Folland. A course in abstract harmonic analysis, volume 29 of Textbooks in Mathematics. CRC Press, Boca Raton, FL, second edition, 2016.DOI
  13. [13]O. Giladi, J. Prochno, C. Schütt, N. Tomczak-Jaegermann, and E. Werner. On the geometry of projective tensor products. J. Funct. Anal., 273(2):471–495, 2017.arxiv.org/abs/1608.08436
  14. [14]A. Grothendieck. Résumé de la théorie métrique des produits tensoriels topologiques. Bol. Soc. Mat. São Paulo, 8:1–79, 1953.
  15. [15]B. S. Kashin. Diameters of some finite-dimensional sets and classes of smooth functions. Mathematics of the USSR-Izvestiya, 11(2):317–333, 1977.
  16. [16]J. Lindenstrauss and A. Pelczyński. Absolutely summing operators in L_p-spaces and their applications. Studia Math., 29:275–326, 1968.DOI
  17. [17]A. Pajor and N. Tomczak-Jaegermann. Volume ratio and other s-numbers of operators related to local properties of Banach spaces. J. Funct. Anal., 87(2):273–293, 1989.DOI
  18. [18]R. A. Ryan. Introduction to tensor products of Banach spaces. Springer Monographs in Mathematics. Springer-Verlag London, Ltd., London, 2002.DOI
  19. [19]C. Schütt. On the volume of unit balls in Banach spaces. Compositio Math., 47(3):393–407, 1982.
  20. [20]S. Szarek and N. Tomczak-Jaegermann. On nearly Euclidean decomposition for some classes of Banach spaces. Compositio Math., 40(3):367–385, 1980.
  21. [21]S. J. Szarek. On Kashin’s almost Euclidean orthogonal decomposition of lₙ¹. Bull. Acad. Polon. Sci. Sér. Sci. Math. Astronom. Phys., 26(8):691–694, 1978.
  22. [22]S. J. Szarek, E. Werner, and K. Życzkowski. How often is a random quantum state k-entangled? Journal of Physics A: Mathematical and Theoretical, 44(4):045303, dec 2010.DOI

Paper details

vlnc:W_vw4c4dy9
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

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

  • Mathematics

ChatGPT

View statement
Looking for related papers…
Contents