Arxiv:1902.02406V2 [Math.FA] 8 Sep 2020 Where Def 1 X Fb(A) = F(Δ)Wa(Δ) ∈ X

POLYNOMIAL INEQUALITIES ON THE HAMMING CUBE

ALEXANDROS ESKENAZIS AND PAATA IVANISVILI

Abstract. Let (X, k · kX ) be a Banach space. The purpose of this article is to systematically investigate dimension independent properties of vector valued functions f : {−1, 1}n → X on the Hamming cube whose spectrum is bounded above or below. Our proofs exploit contractivity properties of the heat ﬂow, induced by the geometry of the target space (X, k · kX ), combined with duality arguments and suitable tools from approximation theory and complex analysis. We obtain a series of improvements of various well-studied estimates for functions with bounded spectrum, including moment comparison results for low degree Walsh polynomials and Bernstein–Markov type inequalities, which constitute discrete vector valued analogues of Freud’s inequality in Gauss space (1971). Many of these inequalities are new even for scalar valued functions. Furthermore, we provide a short proof of Mendel and Naor’s heat smoothing theorem (2014) for functions in tail spaces with values in spaces of nontrivial type and we also prove a dual lower bound on the decay of the heat semigroup acting on functions with spectrum bounded from above. Finally, we improve the reverse Bernstein–Markov inequalities of Meyer (1984) and Mendel and Naor (2014) for functions with narrow enough spectrum and improve the bounds of Filmus, Hatami, Keller and 4 Lifshitz (2016) on the `p sums of inﬂuences of bounded functions for p ∈ 1, 3 .

2010 Mathematics Subject Classiﬁcation. Primary: 42C10; Secondary: 41A17, 41A63, 46B07. Key words. Hamming cube, heat semigroup, hypercontractivity, Bernstein–Markov inequality, moment comparison.

1. Introduction

Fix n ∈ N and let (X, k · kX ) be a Banach space. If p ∈ [1, ∞), the vector valued Lp norm of a function f : {−1, 1}n → X is deﬁned as 1/p def 1 X p kfk n = kf(ε)k . (1) Lp({−1,1} ;X) 2n X ε∈{−1,1}n

def n n As usual, we denote kfkL∞({−1,1} ;X) = maxε∈{−1,1} kf(ε)kX . For a subset A ⊆ {1, . . . , n} the n Q Walsh function wA : {−1, 1} → {−1, 1} is the Boolean function given by wA(ε) = i∈A εi, where n n ε = (ε1, . . . , εn) ∈ {−1, 1} . Every function f : {−1, 1} → X admits an expansion of the form X f = fb(A)wA, (2) A⊆{1,...,n} arXiv:1902.02406v2 [math.FA] 8 Sep 2020 where def 1 X fb(A) = f(δ)wA(δ) ∈ X. (3) 2n δ∈{−1,1}n For i ∈ {1, . . . , n} the i-th partial derivative of such a function f is given by

def f(ε) − f(ε1, . . . , εi−1, −εi, εi+1, . . . , εn) X ∂if(ε) = = fb(A)wA(ε) (4) 2 A⊆{1,...,n} i∈A

P. I. was partially supported by NSF DMS-1856486 and NSF CAREER-1945102. This work was carried out under the auspices of the Simons Algorithms and Geometry (A&G) Think Tank.

1 2 and satisﬁes ∂i f = ∂if. Therefore, the hypercube Laplacian of f is deﬁned as n def X X ∆f = ∂if = |A|fb(A)wA. (5) i=1 A⊆{1,...,n}

−t∆ Finally, the action of the discrete heat semigroup {e }t>0 on the function f is given by −t∆ def X −t|A| ∀ t > 0, e f = e fb(A)wA. (6) A⊆{1,...,n} A straightforward calculation shows that the heat semigroup can equivalently be expressed as 1 X X X X e−t∆f(ε) = e−t|B|(1 − e−t)n−|B|f ε e + δ e , (7) 2n j j j j δ∈{−1,1}n B⊆{1...,n} j∈B j∈{1,...,n}rB n where {e1, . . . , en} is the orthonormal basis of R , which by convexity implies that for every t > 0 −t∆ n n and p ∈ [1, ∞], ke fkLp({−1,1} ;X) 6 kfkLp({−1,1} ;X). Identity (7) also has a useful probabilistic n n interpretation. For a fixed ε ∈ {−1, 1} and t > 0, consider a random vector η ∈ {−1, 1} such that 1+e−t each coordinate ηi is chosen independently to coincide with εi with probability 2 and with −εi 1−e−t −t∆ with probability 2 . Then, the value e f(ε) is the expectation of the random vector f(η). The main purpose of this paper is to investigate finer contractivity properties of the heat semigroup under suitable assumptions on the spectrum of the function f and the geometry of the target space (X, k · kX ). The common feature in the proofs of most of the following results is a duality argument inspired by classical work of Figiel (see [MS86, Theorem 14.6]), which allows the self- improvement of contractivity properties of the heat semigroup relying on suitable inequalities from classical approximation theory and complex analysis. All Banach spaces in the ensuing discussion will be assumed to be over the field of complex numbers. In the rest of the introduction, we proceed to describe our results in decreasing order of generality. In Section 1.1, we present estimates for functions with spectrum bounded from above and values in a general Banach space. In Section 1.2, we improve the bounds of Section 1.1 under the additional assumption that the target space (X, k·kX ) is K-convex (we postpone the relevant definitions until Section 1.2). We also present a new proof of a theorem of Mendel and Naor [MN14] related to the heat smoothing conjecture. Finally, in Section 1.3 we present explicit estimates which hold true for scalar valued functions. These include various Bernstein–Markov type inequalities and their reverses, estimates on the influences of bounded functions and moment comparison results.

1.1. Estimates for a general Banach space. Fix n ∈ N and let (X, k · kX ) be a Banach space. If d ∈ {0, 1 . . . , n}, we say that a function f : {−1, 1}n → X has degree at most d if fb(A) = 0 when |A| > d and we say that f belongs in the d-th tail space if fb(A) = 0 when |A| < d. Finally, we say that f is d-homogeneous if fb(A) = 0 when |A|= 6 d.

1.1.1. A lower bound on the decay of the heat semigroup. The following theorem establishes a lower bound on the decay of the heat semigroup acting on functions of low degree.

Theorem 1. Fix n, d ∈ N with d ∈ {1, . . . , n} and let (X, k · kX ) be a Banach space. For every p ∈ [1, ∞] and every function f : {−1, 1}n → X of degree at most d, we have

−t∆ 1 ∀ t 0, ke fk n kfk n , (8) > Lp({−1,1} ;X) > t Lp({−1,1} ;X) Td(e ) where Td is the d-th Chebyshev polynomial of the ﬁrst kind.

2 We note that a weaker bound in the spirit of Theorem1, attributed partially to Oleszkiewicz, was established in [FHKL16, Lemma 5.4] (see Remark 18 below for a comparison with Theorem1). Using Theorem1 and the hypercontractivity of the discrete heat semigroup (see [Bon70]), we deduce the following moment comparison for functions of low degree.

Corollary 2. Fix n, d ∈ N with d ∈ {1, . . . , n} and let (X, k · kX ) be a Banach space. For every p > q > 1 and every function f : {−1, 1}n → X of degree at most d, we have rp − 1 kfk n T kfk n . (9) Lp({−1,1} ;X) 6 d q − 1 Lq({−1,1} ;X) To the extent of our knowledge, the best previously known moment comparison for general vector valued functions of low degree on the discrete cube can be extracted from an argument in the monograph [KW92] of Kwapieńand Woyczyński.In Remark 19 below, we quantify their argument and show that the bounds of [KW92, Proposition 6.5.1] are weaker than those of Corollary2. Another application of Theorem1 is a refinement of a celebrated inequality on the discrete cube due to Pisier [Pis86]. In connection with his work on nonlinear type, Pisier showed that for every n Banach space (X, k · kX ), every p ∈ [1, ∞] and every function f : {−1, 1} → X, we have n 1 X 1 X X p 1/p f − f(δ) (log n + 1) δi∂if . (10) n n 6 n n 2 Lp({−1,1} ;X) 2 Lp({−1,1} ;X) δ∈{−1,1}n δ∈{−1,1}n i=1

For X = R, the right hand side of (10) is equivalent to the Lp norm of the gradient of f due to Khintchine’s inequality [Khi23], thus (10) can be understood as a vector valued Poincar´einequality. The dependence on the dimension n in Pisier’s inequality (10) for various classes of spaces X is fundamental in investigations in the nonlinear geometry of Banach spaces and the Ribe program (see [Nao12] for a detailed discussion around this topic). For general Banach spaces, Talagrand [Tal93] has proven that the factor log n in (10) is asymptotically optimal for every p ∈ [1, ∞), whereas Wagner [Wag00] has shown that when p = ∞, Pisier’s inequality holds with an absolute constant. Here we show the following reﬁnement of Pisier’s inequality for functions of low degree.

Theorem 3. Fix n, d ∈ N with d ∈ {1, . . . , n} and let (X, k · kX ) be a Banach space. For every p ∈ [1, ∞) and every function f : {−1, 1}n → X of degree at most d, we have n 1 X 1 X X p 1/p f − f(δ) 3(log d + 1) δi∂if (11) n n 6 n n 2 Lp({−1,1} ;X) 2 Lp({−1,1} ;X) δ∈{−1,1}n δ∈{−1,1}n i=1 and the log d factor is asymptotically sharp for every p ∈ [1, ∞). 1.1.2. A Bernstein–Markov type inequality for ∆. A variant of the proof of Theorem1 implies the following Bernstein–Markov type inequality, which had previously appeared in [FHKL16].

Theorem 4. Fix n, d ∈ N with d ∈ {1, . . . , n} and let (X, k · kX ) be a Banach space. For every p ∈ [1, ∞] and every function f : {−1, 1}n → X of degree at most d, we have 2 n n k∆fkLp({−1,1} ;X) 6 d kfkLp({−1,1} ;X). (12) In [FHKL16, Lemma 5.4], the Bernstein–Markov type inequality (12) was stated only for real valued functions and p = 1. In the vector valued setting which is of interest here, inequality (12) is sharp for general Banach spaces. Recall that a Banach space (X, k · kX ) has cotype q ∈ [2, ∞] with constant C ∈ (0, ∞) if for every n ∈ N and vectors x1, . . . , xn ∈ X, we have n n 1 q 1/q 1 1/q X X X q εixi > kxikX . (13) 2n X C ε∈{−1,1}n i=1 i=1

3 The fact that every Banach space (X, k · kX ) has cotype q = ∞ with constant C = 1 follows from the triangle inequality, yet having ﬁnite cotype is a meaningful structural property of a given space. We refer to the survey [Mau03] for further information on the rich theory of type and cotype. We will show that any improvement of (12), forces the target space X to have ﬁnite cotype.

Theorem 5. Let (X, k · kX ) be a Banach space and p ∈ [1, ∞]. Suppose that there exist some n η ∈ (0, 1) and d ∈ N such that for every n ∈ N with d ∈ {1, . . . , n}, every function f : {−1, 1} → X of degree at most d satisﬁes the inequality 2 n n k∆fkLp({−1,1} ;X) 6 (1 − η)d kfkLp({−1,1} ;X). (14) Then X has ﬁnite cotype.

1.2. Estimates for K-convex Banach spaces. Let (X, k·kX ) be a Banach space. The X-valued n n Rademacher projection Rad : Lp({−1, 1} ; X) → Lp({−1, 1} ; X) is the operator given by n n def X ∀ ε ∈ {−1, 1} , Rad(f)(ε) = fb({i})εi. (15) i=1

We say that (X, k · k ) is K-convex (see also [Mau03]) if sup kRadk n n < X n∈N Lp({−1,1} ;X)→Lp({−1,1} ;X) ∞ for some (equivalently, for all) p ∈ (1, ∞). Pisier’s K-convexity theorem [Pis82] asserts that n ∞ a Banach space (X, k · kX ) is K-convex if and only if X does not contain copies of {`1 }n=1 with distortion arbitrarily close to 1. Moreover, both conditions are equivalent to X having nontrivial type, see [Pis73]. In the proofs of most results of this section, we will crucially use the deep fact that the heat semigroup with values in a K-convex space is a bounded analytic semigroup [Pis82].

1.2.1. The heat smoothing conjecture. In [MN14], Mendel and Naor asked whether for every K- convex Banach space (X, k · kX ) and p ∈ (1, ∞) there exist c(p, X),C(p, X) ∈ (0, ∞) such that for n every n ∈ N, every function f : {−1, 1} → X in the d-th tail space satisﬁes the estimate −t∆ −c(p,X)dt n n ∀ t > 0, ke fkLp({−1,1} ;X) 6 C(p, X)e kfkLp({−1,1} ;X). (16) In the direction of this question, currently known as the heat smoothing conjecture, they showed the following theorem, partially relying on ideas from [Pis07] (see also the work [HMO17] of Heilman, Mossel and Oleszkiewicz for an optimal result when d = 1 and X = C).

Theorem 6 (Mendel–Naor). Let (X, k · kX ) be a K-convex Banach space. For every p ∈ (1, ∞), there exist c(p, X),C(p, X) ∈ (0, ∞) and A(p, X) ∈ [1, ∞) such that for every n, d ∈ N with d ∈ {0, 1, . . . , n − 1} and every function f : {−1, 1}n → X in the d-th tail space, we have

−t∆ −c(p,X)d min{t,tA(p,X)} n n ∀ t > 0, ke fkLp({−1,1} ;X) 6 C(p, X)e kfkLp({−1,1} ;X). (17) In Section3 below, we present a simple proof of Theorem6 relying on a duality argument.

1.2.2. An improved lower bound on the decay of the heat semigroup. Under the assumption that the target space (X, k · kX ) is K-convex, we get the following improvement over Theorem1 (see also equation (45) in Remark 18 for comparison).

Theorem 7. Let (X, k · kX ) be a K-convex Banach space. For every p ∈ (1, ∞), there exist 1 c(p, X),C(p, X) ∈ (0, ∞) and η(p, X) ∈ 2 , 1 such that for every n, d ∈ N with d ∈ {1, . . . , n} and every function f : {−1, 1}n → X of degree at most d, we have

−t∆ −C(p,X)d max{t,tη(p,X)} n n ∀ t > 0, ke fkLp({−1,1} ;X) > c(p, X)e kfkLp({−1,1} ;X). (18)

4 Theorem7 should be understood as the dual of Mendel and Naor’s bound (17) for the heat smoothing conjecture. We conjecture that one can in fact take η(p, X) = 1 in Theorem7 for every p ∈ (1, ∞) and K-convex Banach space (X, k · kX ), but a proof of such a claim appears intractable with the technique presented here. A real valued version of this conjecture for p = 1 has previously appeared in [FHH+14, Section 5]. As in the case of Theorem1, Theorem7 also implies moment comparison for functions of low degree. We postpone the relevant result (Corollary 31) to Section3.

1.2.3. Bernstein–Markov type inequalities for the hypercube Laplacian and the gradient. Under the additional assumption that the target space (X, k · kX ) is K-convex, we can obtain the following asymptotic improvement of the bound of Theorem4.

Theorem 8. Let (X, k · kX ) be a K-convex Banach space. For every p ∈ (1, ∞), there exist α(p, X) ∈ [1, 2) and C(p, X) ∈ (0, ∞) such that for every n, d ∈ N with d ∈ {1, . . . , n} and every function f : {−1, 1}n → X of degree at most d, we have α(p,X) n n k∆fkLp({−1,1} ;X) 6 C(p, X)d kfkLp({−1,1} ;X). (19) We conjecture that the conclusion of Theorem8 holds true with α(p, X) = 1 for every p ∈ (1, ∞) and K-convex Banach space (X, k · kX ) (see also Remark 33 below). Since every K-convex Banach space has finite cotype, the conclusion of Theorem8 is consistent with Theorem5. However, there exist Banach spaces (e.g. X = `1) which have finite cotype but are not K-convex. It remains an interesting (and potentially challenging) open problem to understand whether the dependence on the degree in the vector valued Bernstein–Markov inequality (12) (even for, say, p = 2) can be 2 improved to o(d ) under the minimal assumption that (X, k · kX ) has finite cotype. We conclude this section by presenting a Bernstein–Markov type inequality where the hypercube Laplacian is replaced by the vector valued gradient appearing in Pisier’s inequality (10).

Theorem 9. Let (X, k · kX ) be a K-convex Banach space. For every p ∈ (1, ∞) there exist α(p, X) ∈ [1, 2) and C(p, X) ∈ (0, ∞) such that for every n, d ∈ N with d ∈ {1, . . . , n} and every function f : {−1, 1}n → X of degree at most d, we have n p 1/p 1 X X α(p,X) δi∂if C(p, X)d (log d + 1)kfk n . (20) n n 6 Lp({−1,1} ;X) 2 Lp({−1,1} ;X) δ∈{−1,1}n i=1 1.2.4. A reverse Bernstein–Markov inequality for ∆. In [MN14], Mendel and Naor asked if for every K-convex Banach space (X, k · kX ) and p ∈ (1, ∞) there exists c(p, X) ∈ (0, ∞) such that for every n n ∈ N and d ∈ {0, 1, . . . , n − 1}, every function f : {−1, 1} → X in the d-th tail space satisﬁes

n n k∆fkLp({−1,1} ;X) > c(p, X)dkfkLp({−1,1} ;X). (21) Similar estimates for scalar valued functions had also been obtained by Meyer in [Mey84]. The following theorem, whose proof relies on a recent inequality of Erd´elyi,contains a result in this direction under the additional assumption that the spectrum of the function f is also bounded above.

Theorem 10. Let (X, k · kX ) be a K-convex Banach space. For every p ∈ (1, ∞) there exists n c(p, X) ∈ (0, ∞) such that for every n, d, m ∈ N with d+m 6 n and every function f : {−1, 1} → X of degree at most d + m which is also in the d-th tail space, we have d k∆fk n c(p, X) kfk n . (22) Lp({−1,1} ;X) > m Lp({−1,1} ;X) In particular, Theorem 10 provides a positive answer to the question of [MN14] in the special case when m = O(1) and improves upon the previously known bounds of Meyer [Mey84] and Mendel and Naor [MN14] when m is a small enough power of d.

5 1.3. Estimates for scalar valued functions. In this section we will present explicit estimates which hold true for scalar valued functions. Even though several of the following results are special cases of theorems from the previous section, we present the full statements for the convenience of the reader not interested in the general vector valued setting. 1.3.1. The decay of the heat semigroup. The optimal bounds that can be derived from our approach for the action of the heat semigroup on functions with bounded spectrum are the following. √ 2 p−1 Theorem 11. For p ∈ [1, ∞], let θp = 2 arcsin p . Then, for every p ∈ (1, ∞), n, d ∈ N with n d ∈ {0, 1, . . . , n} and every function f : {−1, 1} → C of degree at most d, we have π π !d t 2π−θp t 2π−θp −t∆ (e + 1) − (e − 1) n n ∀ t > 0, ke fkLp({−1,1} ;C) > π π kfkLp({−1,1} ;C). (23) (et + 1) 2π−θp + (et − 1) 2π−θp n Moreover, for every function f : {−1, 1} → C in the d-th tail space, we have π π !d −t θp −t θp −t∆ (1 + e ) − (1 − e ) n n ∀ t > 0, ke fkLp({−1,1} ;C) 6 π π kfkLp({−1,1} ;C). (24) (1 + e−t) θp + (1 − e−t) θp The above lower bound (23) on the decay of the heat semigroup combined with classical hy- percontractivite estimates [Bon70] implies the following improved moment comparison result for functions of low degree. √ 2 p−1 Corollary 12. For p ∈ [1, ∞], let θp = 2 arcsin p . Then, for every p > q > 1, every n, d ∈ N n with d ∈ {1, . . . , n} and every function f : {−1, 1} → C of degree at most d, we have √ √ π √ √ π !d ( p − 1 + q − 1) 2π−θp + ( p − 1 − q − 1) 2π−θp n n kfkLp({−1,1} ;C) 6 √ √ π √ √ π kfkLq({−1,1} ;C). (25) ( p − 1 + q − 1) 2π−θp − ( p − 1 − q − 1) 2π−θp Even though Bonami’s hypercontractive estimates [Bon70] break down at the endpoint p = 1, a n well known trick (see [O’D14, Theorem 9.22]) implies that if f : {−1, 1} → C has degree at most d, d n n kfkL2({−1,1} ;C) 6 e kfkL1({−1,1} ;C). (26) Relying on works of Beckner [Bec75] and Weissler [Wei79], we prove the following improved bound. n Theorem 13. For every n, d ∈ N with d ∈ {1, . . . , n} and every function f : {−1, 1} → C of degree at most d, we have d n n kfkL2({−1,1} ;C) 6 (2.69076) kfkL1({−1,1} ;C). (27) 1.3.2. Bernstein–Markov type inequalities. The bounds (19) in the Bernstein–Markov inequality for the hypercube Laplacian take the following explicit form for scalar valued functions. √ 2 p−1 Theorem 14. For p ∈ [1, ∞], let θp = 2 arcsin p . Then, for every n, d ∈ N with d ∈ n {1, . . . , n} and every function f : {−1, 1} → C of degree at most d, we have θ 2− p n π n k∆fkLp({−1,1} ;C) 6 10d kfkLp({−1,1} ;C). (28) n For p ∈ [1, ∞] and a function f : {−1, 1} → C, denote by n 1/2 def X 2 k∇fkL ({−1,1}n; ) = (∂if) . (29) p C L ({−1,1}n; ) i=1 p C In contrast to the vector valued Theorem9, we can prove the following improved Bernstein–Markov type inequality for the gradient of scalar valued functions as a consequence of Theorem 14.

6 √ 2 p−1 Theorem 15. For p ∈ [1, ∞], let θp = 2 arcsin p . Then, for every p ∈ (1, ∞), there exists n Cp ∈ (0, ∞) such that for every n, d ∈ N with d ∈ {1, . . . , n} and every function f : {−1, 1} → C of degree at most d, we have

 θ 2 − p  p pπ n Cpd log(d + 1)kfkLp({−1,1} ;C), if p ∈ 1, 2 k∇fkL ({−1,1}n; ) θ . (30) p C 6 1− p 2π n Cpd kfkLp({−1,1} ;C), if p ∈ [2, ∞) The change in the exponent from the range p ∈ (1, 2) to the range p ∈ [2, ∞) and its relation to discrete Riesz transforms is further discussed in Section4. We also postpone until then the statement of some endpoint (p = ∞ and p = 1) Bernstein–Markov inequalities for the discrete gradient (see Proposition 46 and Remark 49 respectively). We ﬁnally turn to a problem studied by Filmus, Hatami, Keller and Lifshitz in [FHKL16] (see n also [BB14]). Following their notation, for p ∈ [1, ∞) and a function f : {−1, 1} → C, denote by n (p) def X p Inf f = k∂ifk n . (31) Lp({−1,1} ;C) i=1 Motivated by a question of Aaronson and Ambrainis [AA14], the authors were interested in obtain- ing bounds of the form (p) p Inf f fp(d)kfk n , (32) 6 L∞({−1,1} ;C) 3−p where fp(d) is a polynomial in d. In this direction, they proved (32) with fp(d) = d when p ∈ [1, 2] and fp(d) = d when p ∈ [2, ∞), the latter of which is sharp. Using Theorem 11, we deduce 4 the following improved bounds for p ∈ 1, 3 . 4 Corollary 16. For every p ∈ 1, 3 there exists a constant Kp ∈ (0, ∞) such that for every n, d ∈ N n with d ∈ {1, . . . , n} and every function f : {−1, 1} → C of degree at most d, we have √ 1 2 p−1 (p) 2− π arcsin p p Inf f Kpd kfk n . (33) 6 L∞({−1,1} ;C) 1.3.3. A reverse Bernstein–Markov type inequality for ∇. For the case of scalar valued functions we will also prove the following variant of Theorem 10 for the discrete gradient.

Theorem 17. For every p ∈ (1, ∞) there exists cp ∈ (0, ∞) such that for every n, d, m ∈ N with n d + m 6 n and every function f : {−1, 1} → C of degree at most d + m which is also in the d-th tail space, we have r d k∇fk n c kfk n . (34) Lp({−1,1} ;C) > p m Lp({−1,1} ;C) Acknowledgements. We are indebted to Assaf Naor for many helpful discussions. We are also very grateful to Tam´asErd´elyifor proving the main result of [Erd20] upon our request and to an anonymous referee for sharing with us an argument which improved Corollary 21. Finally, we would like to thank Franoise Lust-Piquard for valuable feedback.

2. Estimates for a general Banach space We first present the proof of Theorem1, the lower bound on the decay of the heat semigroup acting on functions with values in a general Banach space. Recall that the d-th Chebyshev polynomial of the first kind Td(x) is the unique polynomial of degree d such that Td(cos θ) = cos(dθ) for every θ ∈ R. Chebyshev’s inequality [BE95, p. 235] asserts that the d-th Chebyshev polynomial of the first kind is characterized by the extremal property ∀ x ∈ R r [−1, 1], |Td(x)| = max |p(x)| : deg(p) 6 d and kpkC([−1,1]) = 1 , (35)

7 where for a continuous function h : K → C on a compact space K, we set khkC(K) = maxx∈K |h(x)|. n Proof of Theorem1. Fix n ∈ N, d ∈ {1, . . . , n}, p ∈ [1, ∞] and let f : {−1, 1} → X be a function of degree at most d. The contractivity of the heat semigroup which we derived from (7), can be ∆ n n rewritten as kx fkLp({−1,1} ;X) 6 kfkLp({−1,1} ;X) for every x ∈ [0, 1]. However, for x ∈ R, ∆ X |A| X |A| ∆ (−x) f(ε) = (−x) fb(A)wA(ε) = x fb(A)wA(−ε) = (x f)(−ε), (36) A⊆{1,...,n} A⊆{1,...,n} ∆ ∆ n n thus k(−x) fkLp({−1,1} ;X) = kx fkLp({−1,1} ;X). Consequently, ∆ n n ∀ x ∈ [−1, 1], kx fkLp({−1,1} ;X) 6 kfkLp({−1,1} ;X). (37) Let µ be any complex measure on [−1, 1]. Then, averaging over (37), we get 1 1 ∆ ∆ x f dµ(x) kx fk n d|µ|(x) n 6 Lp({−1,1} ;X) ˆ Lp({−1,1} ;X) ˆ −1 −1 (38) (37) 1 n n 6 kfkLp({−1,1} ;X) d|µ|(x) = kµkM([−1,1])kfkLp({−1,1} ;X), ˆ−1 where for a complex measure µ on a compact space K, we denote by kµkM(K) the total variation d of µ. For t > 0, consider the linear functional ϕt : span{1, x, . . . , x }, k · kC([−1,1]) → C given by d d X k def X tk ϕt akx = ake , (39) k=0 k=0 t or ϕt(p) = p(e ) when p is a polynomial with deg(p) 6 d. Then, Chebyshev’s inequality (35) can be rewritten as d t ∀ p ∈ span{1, x, . . . , x }, |ϕt(p)| 6 Td(e )kpkC([−1,1]). (40) Therefore, by the Hahn–Banach theorem and the Riesz representation theorem, there exists a t complex measure µt on [−1, 1] such that kµtkM([−1,1]) 6 Td(e ) and for every polynomial p, we have 1 t deg(p) 6 d =⇒ p(x) dµt(x) = p(e ). (41) ˆ−1

Since fb(A) = 0 when |A| > d, applying (38) for the measure µt satisfying (41), we deduce that 1 (40) t∆ (41) ∆ t ∀ t 0, ke fkL ({−1,1}n;X) = x f dµt(x) Td(e )kfkL ({−1,1}n;X), (42) > p n 6 p ˆ−1 Lp({−1,1} ;X) which is equivalent to the desired inequality (8). Remark 18. A weaker lower bound for the decay of the heat semigroup, attributed partially to Oleszkiewicz, was established in [FHKL16, Lemma 5.4] and asserts that for every function f : {−1, 1}n → X of degree at most d, we have −t∆ −td2 n n ke fkLp({−1,1} ;X) > e kfkLp({−1,1} ;X). (43) Inequality (43) was stated in [FHKL16] only for X = R and p = 1, but the argument presented there works in greater generality. Checking that (43) is weaker than the estimate (8) amounts to d2 showing that for y > 1, we have Td(y) 6 y , which can be easily derived from the identity (y + py2 − 1)d + (y − py2 − 1)d ∀ |y| 1,T (y) = . (44) > d 2 Furthermore, using (44), it is elementary to check that √ t 3 max{t, t}d ∀ t > 0,Td(e ) 6 e , (45)

8 therefore inequality (18) is also an improvement over (8) for t ≈ 0.

n Proof of Corollary2. Fix n ∈ N, d ∈ {1, . . . , n}, p > q > 1 and a function f : {−1, 1} → X of degree at most d. Bonami’s hypercontractive inequality [Bon70] (the straightforward vector valued extension of which is due to Borell, see [Bor79]) asserts that

1 p − 1 −t∆ ∀ t log , ke fk n kfk n . (46) > 2 q − 1 Lp({−1,1} ;X) 6 Lq({−1,1} ;X) 1 p−1 Combining (8) and (46), we deduce that for t > 2 log q−1 ,

(8) (46) t −t∆ t n n n kfkLp({−1,1} ;X) 6 Td(e )ke fkLp({−1,1} ;X) 6 Td(e )kfkLq({−1,1} ;X). (47) 1 p−1 Plugging t = 2 log q−1 in (47), we deduce the moment comparison (9). Remark 19. Inequality (9) for d = 1 with some constant C(p, q) originated in the work [Kah64] of q p−1 Kahane and the implicit constant was later improved to q−1 by Kwapie´nin [Kwa76]. The use of hypercontractivity for moment comparison of vector valued Walsh polynomials was initiated by Borell in [Bor79] (see the exposition [Pis78]), who later showed in [Bor84] that inequality (9) holds true with some constant depending only on p, q and d (this had previously been proven for scalar valued functions by Bourgain in [Bou80] via a square function approach). To the extent of our knowledge, the best known dependence in (9) before the present work could be extracted from an argument in the monograph [KW92, Proposition 6.5.1] which goes as follows. For k ∈ {0, 1, . . . , n} n n let Radk : Lp({−1, 1} ; X) → Lp({−1, 1} ; X) be the Rademacher projection on level k given by

n def X ∀ ε ∈ {−1, 1} , Radk(f)(ε) = fb(A)wA(ε). (48) A⊆{1,...,n} |A|=k

d If d ∈ {0, 1, . . . , n}, consider the linear functional σk : span{1, x, . . . , x }, k·kC([−1,1]) → C given by p(k)(0) σk(p) = k! . Then, for every k ∈ {1, . . . , d}, there exists some constant c(d, k) ∈ (0, ∞) such that

deg(p) 6 d =⇒ |σk(p)| 6 c(d, k)kpkC([−1,1]). (49)

(k) |Td (0)| The optimal value of c(d, k) is known, see [BE95, p. 248], but clearly c(d, k) > k! . Repeating the duality argument used in the proof of Theorem1, we deduce that for every function f : {−1, 1}n → X of degree at most d, q ∈ [1, ∞] and k ∈ {0, . . . , d},

Radkf n c(d, k)kfk n . (50) Lq({−1,1} ;X) 6 Lq({−1,1} ;X) Furthermore, by the moment comparison of homogeneous Walsh polynomials [Bor79], for p > q > 1, p − 1k/2 Radkf n Radkf n . (51) Lp({−1,1} ;X) 6 q − 1 Lq({−1,1} ;X) Therefore, combining (50) and (51), we get

d d X (51) X p − 1k/2 kfkL ({−1,1}n;X) Radkf n Radkf n p 6 Lp({−1,1} ;X) 6 q − 1 Lq({−1,1} ;X) k=0 k=0 (52) d (50) X p − 1k/2 c(d, k) kfk n . 6 q − 1 Lq({−1,1} ;X) k=0

9 However, the constant obtained by this argument is d d (k) d (k) r X p − 1k/2 X |T (0)|p − 1k/2 X T (0)p − 1k/2 p − 1 c(d, k) d d = T , (53) q − 1 > k! q − 1 > k! q − 1 d q − 1 k=0 k=0 k=0 and therefore Corollary2 improves over the result of [KW92]. Combining Corollary2 and (45), it follows that for p > q > 1 and every function f : {−1, 1}n → X of degree at most d, we have 1/2 p−1 p−1 3d max 1 log , √1 log n 2 q−1 2 q−1 n kfkLp({−1,1} ;X) 6 e kfkLq({−1,1} ;X). (54) √ In particular, if p − 1 > (1 + ε)(q − 1) for some ε > 0, then there exists Cε = O(1/ ε) > 0 such that every function f : {−1, 1}n → X of degree at most d satisfies p − 1Cεd kfk n kfk n , (55) Lp({−1,1} ;X) 6 q − 1 Lq({−1,1} ;X) which should be understood as an extension of Borell’s moment comparison (51) for vector valued functions of low degree instead of homogeneous. The validity of Theorem7 with η(p, X) = 1 would imply the estimate (55) with Cε replaced by a universal contant C ∈ (0, ∞). Remark 20. Some similar moment comparison estimates can be obtained for functions on the α- biased hypercube using the biased hypercontractivity of Oleszkiewicz [Ole03] and Wolff [Wol07]. Indeed, it follows from the main results of [Ole03, Wol07] and convexity considerations that for every p > q > 1 and α ∈ (0, 1), there exists a closed interval Ip,q(α) ⊆ [−1, 1] with nonempty interior containing 0 such that for every n ∈ N and every coefficients {aS}S⊆{1,...,n} ⊆ X, X |S| X sup x aSwS aSwS , (56) −1 n n 6 −1 n n Lp({−γ ,γ} ,µ ;X) Lq({−γ ,γ} ,µ ;X) x∈Ip,q(α) S⊆{1,...,n} α S⊆{1,...,n} α

q 1 where µα = αδ−γ−1 + (1 − α)δγ and γ = α − 1. Running the same duality argument as in the proof of Theorem1 combined with an application of Chebyshev’s inequality (35) for a symmetric −1 n n interval Jp,q(α) ⊆ Ip,q(α), proves an analogue of (9) on {−γ , γ} , µα . Finally we state one “endpoint” version1 of our moment comparison (9). Recall that for a function h : {−1, 1}n → [0, ∞), we denote its entropy by def 1 X 1 X 1 X Ent(h) = h(ε) log h(ε) − h(ε) · log h(ε) . (57) 2n 2n 2n ε∈{−1,1}n ε∈{−1,1}n ε∈{−1,1}n

Corollary 21. Fix n, d ∈ N with d ∈ {1, . . . , n} and let (X, k · kX ) be a Banach space. For every n q ∈ (0, ∞) there exists a constant Cq ∈ (0, ∞) such that every function f : {−1, 1} → X of degree at most d satisﬁes q q Ent(kfk ) C dkfk n . (58) X 6 q Lq({−1,1} ;X) Proof. By H¨older’sinequality, for every α, β ∈ (0, ∞) and λ ∈ (0, 1), λ 1−λ kfkL ({−1,1}n;X) kfk n kfk n . (59) ( λ + 1−λ )−1 6 Lα({−1,1} ;X) Lβ({−1,1} ;X) α β In other words, the function φ : (0, ∞) → R given by def ∀ r ∈ (0, ∞), φ(r) = log kfk n (60) L1/r({−1,1} ;X)

1 2 The bound (58) of Corollary 21 was pointed out to us by an anonymous referee, who improved a suboptimal Oq(d ) estimate appearing in an earlier version of this manuscript. We are grateful to them for sharing their improvement with us and for their helpful comments.

10 is convex. Hence, the function (0, ∞) 3 r 7−→ φ0(1/r) is nonincreasing, which implies that 0 1 1 1 1 1 ∀ p < q, φ q · p − q > φ p − φ q . (61) Writing 1 1 X α kfk n = exp log kf(ε)k , (62) Lα({−1,1} ;X) α 2n X ε∈{−1,1}n one deduces that α d EntkfkX kfk n = kfk n · , (63) Lα({−1,1} ;X) Lα({−1,1} ;X) 2 α dα α kfk n Lα({−1,1} ;X) which implies that q 01 EntkfkX φ = − q . (64) q kfk n Lq({−1,1} ;X) By the monotonicity of r 7→ φ0(1/r) and (64), in order to prove (58) for every q ∈ (0, ∞) it suﬃces to consider the case q > 4. By (64), (61) can be rewritten as q ! kfk n EntkfkX pq Lq({−1,1} ;X) q 6 log kfk n q − p kfkL ({−1,1}n;X) Lq({−1,1} ;X) p (65) (9) pq rq − 1 pqd 4(q − 1) log T log , 6 q − p d p − 1 6 2(q − p) p − 1

d where in the last inequality we used the estimate Td(y) 6 (2y) for y > 1 which follows from (44). Choosing p = q/2 completes the proof with Cq = O(q) as q → ∞. Remark 22. The application (59) of H¨older’sinequality can be rewritten as ! ! n n kfkLp({−1,1} ;X) (p − q)r kfkLr({−1,1} ;X) ∀ 0 < q < p < r, log 6 · log . (66) n n kfkLq({−1,1} ;X) (r − q)p kfkLq({−1,1} ;X) It is classical that using (66), one can improve moment comparison bounds for low degree functions (see, e.g., [IT19] for a recent application of this idea). The same applies to the moment comparison inequalities obtained in this paper. For instance, combining (9) with (66), we deduce that for every function f : {−1, 1}n → X of degree at most d and every p > q > 1,

r (p−q)r r − 1 (r−q)p kfkL ({−1,1}n;X) inf Td · kfkL ({−1,1}n;X). (67) p 6 r>p q − 1 q Moreover it is well known that the same trick relying on (66) can be used to prove moment comparison estimates beyond the range p > q > 1 (see, e.g., [O’D14, Theorem 9.22] or Theorem 13 below). Since all such applications of (66) are automatic, in the sequel we will omit stating improvements such as (67) (for instance, of Corollary 12 or Corollary 31). We proceed by proving Pisier’s inequality (11) for functions of low degree. The proof is an almost mechanical adaptation of Pisier’s argument from [Pis86] (see also the exposition in [Nao12]) with the exception of suitably using the lower bound (8) instead of the trivial lower bound

−t∆ −nt n n ∀ t > 0, ke fkLp({−1,1} ;X) > e kfkLp({−1,1} ;X), (68) which holds true for every function f : {−1, 1}n → X. For completeness, we present the full proof.

11 n P Proof of Theorem3. Fix p ∈ [1, ∞) and let f : {−1, 1} → X be such that δ∈{−1.1}n f(δ) = 0. ∗ −s∆ n Fix s > 0 and consider a normalizing functional gs of e f in Lp({−1, 1} ; X), that is, a function ∗ n ∗ ∗ 1 1 n ∗ gs : {−1, 1} → X such that kgs kLq({−1,1} ;X ) = 1, where p + q = 1, and

−s∆ 1 X ∗ −s∆ 1 X −s∆ ∗ ke fk n = hg (ε), e f(ε)i = he g (ε), f(ε)i, (69) Lp({−1,1} ;X) 2n s 2n s ε∈{−1,1}n ε∈{−1,1}n where in the last equality we used the fact that e−s∆ is self-adjoint. However, for ε ∈ {−1, 1}n, n n 1 X X X he−s∆g∗(ε), f(ε)i = he−s∆∆−1 δ ∂ g∗(ε), δ ∂ f(ε)i s 2n i i s i i δ∈{−1,1}n i=1 i=1 n n (70) 1 X ∞ X X = h δ ∂ e−t∆g∗(ε), δ ∂ f(ε)i dt. 2n ˆ i i s i i δ∈{−1,1}n s i=1 i=1 ∗ n n ∗ For s, t > 0, consider the fucntion gs,t : {−1, 1} × {−1, 1} → X given by

∗ def X ∗ Y −t −t gs,t(ε, δ) = gbs (A) e εi + (1 − e )δi A⊆{1,...,n} i∈A n (71) −t∆ ∗ t X −t∆ ∗ ∗ = e gs (ε) + (e − 1) εiδi∂ie gs (ε) + Φs,t(ε, δ), i=1 P ∗ n where δ∈{−1,1}n δiΦs,t(ε, δ) = 0 for every ε ∈ {−1, 1} and i ∈ {1, . . . , n}. Furthermore, for every s, t > 0, expanding (71) in the Walsh basis, we see that ∗ X −t|B| −t n−|B| ∗ X X gs,t(ε, δ) = e (1 − e ) gs εiei + δiei , (72) B⊆{1,...,n} i∈B i∈{1,...,n}rB n n where {ei}i=1 is the standard basis of R . Therefore, for every s, t > 0, we have n ∗ X n −tk −t n−k ∗ kg k n n ∗ e (1 − e ) kg k n ∗ = 1. (73) s,t Lq({−1,1} ×{−1,1} ;X ) 6 k s Lq({−1,1} ;X ) k=0 Combining (69), (70), (71) and (73) with H¨older’sinequality, we deduce that ∞ n −s∆ (70)∧(71) 1 X 1 ∗ X ke fk n = hg (ε, δ), δ ∂ f(ε)i dt Lp({−1,1} ;X) 4n ˆ et − 1 s,t i i ε,δ∈{−1,1}n s i=1 n (73) ∞ 1 1 X X p 1/p dt δi∂if 6 t n n (74) ˆ e − 1 2 Lp({−1,1} ;X) s δ∈{−1,1}n i=1 n es 1 X X p 1/p = log δi∂if . s n n e − 1 2 Lp({−1,1} ;X) δ∈{−1,1}n i=1 Furthermore, since f has degree at most d, Theorem1 implies that for s > 0, 1 −s∆ kfk n ke fk n . (75) s Lp({−1,1} ;X) 6 Lp({−1,1} ;X) Td(e ) Combining (74) and (75), we conclude that for every s > 0, n s p 1/p s e 1 X X kfk n Td(e ) log δi∂if . (76) Lp({−1,1} ;X) 6 s n n e − 1 2 Lp({−1,1} ;X) δ∈{−1,1}n i=1

12 s d2s As in Remark 18, using that Td(e ) 6 e for s > 0 and a straightforward optimization we get s s s e d2s e min Td(e ) log s 6 min e log s 6 3(log d + 1). (77) s>0 e − 1 s>0 e − 1 Plugging (77) in (76), we ﬁnally deduce Pisier’s inequality (11). We now prove the Bernstein–Markov type inequality (12) for the hypercube Laplacian. In the proof, we will need Markov’s inequality (see [BE95, Theorem 5.1.8]) which asserts that 0 2 max{|p (1)| : deg(p) 6 d and kpkC([−1,1]) = 1 = d , (78) where the equality is achieved for the d-th Chebyshev polynomial of the ﬁrst kind Td(x). Our proof is similar to the one presented in [FHKL16], which also crucially relied on Markov’s inequality (78). n Proof of Theorem4. Fix n ∈ N, d ∈ {1, . . . , n}, p ∈ [1, ∞] and let f : {−1, 1} → X be a function of d degree at most d. Consider the linear functional ψ : span{1, x, . . . , x }, k · kC([−1,1]) → C given by d d X k def X ψ akx = kak, (79) k=0 k=0 0 or ψ(p) = p (1) when p is a polynomial with deg(p) 6 d. Then, Markov’s inequality (78) can be rewritten as d 2 ∀ p ∈ span{1, x, . . . , x }, |ψ(p)| 6 d kpkC([−1,1]). (80) Therefore, by the Hahn–Banach theorem and the Riesz representation theorem, there exists a 2 complex measure ν on [−1, 1] such that kνkM([−1,1]) 6 d and for every polynomial p, we have 1 0 deg(p) 6 d =⇒ p(x) dν(x) = p (1). (81) ˆ−1 Since fb(A) = 0 when |A| > d, we get that 1 1 (81) ∆ ∆ k∆fk n = x f dν(x) kx fk n d|ν|(x) Lp({−1,1} ;X) n 6 Lp({−1,1} ;X) ˆ Lp({−1,1} ;X) ˆ −1 −1 (82) (37) 1 (80) 2 n n n 6 kfkLp({−1,1} ;X) d|ν|(x) = kνkM([−1,1])kfkLp({−1,1} ;X) 6 d kfkLp({−1,1} ;X), ˆ−1 which concludes the proof of the theorem.

Remark 23. One can generalize Theorem4 by showing that for every k ∈ N, we have d2d2 − 12 ··· d2 − (k − 1)2 k∆(∆ − 1) ··· (∆ − k + 1)fk n kfk n . (83) Lp({−1,1} ;X) 6 1 · 3 ··· (2k − 1) Lp({−1,1} ;X) Theorem1 corresponds to the case k = 1. To derive (83), one can use the same duality argument (k) used in the proofs of Theorems1 and4 for the linear functional ψk(p) = p (1) along with Markov’s inequality for higher derivatives (see [BE95, Theorem 5.2.1]), which asserts that d2d2 − 12 ··· d2 − (k − 1)2 max{|p(k)(1)| : deg(p) d and kpk = 1 = T (k)(1) = . (84) 6 C([−1,1]) d 1 · 3 ··· (2k − 1) We conclude this section by proving Theorem5.

Proof of Theorem5. Assume that (X, k · kX ) does not have ﬁnite cotype. By the Maurey–Pisier m theorem [MP76], for every θ > 0 and m ∈ N there exists a linear operator Sm : `∞ → X such that m ∀ y ∈ `∞, kyk∞ 6 kSmykX 6 (1 + θ)kyk∞. (85)

13 n n Let n ∈ N with d ∈ {1, . . . , n}. Define the function f : {−1, 1} → L∞({−1, 1} ; R) by ε δ + ··· ε δ f(ε)(δ) def= T 1 1 n n , (86) d n where Td(x) is the d-th Chebyshev polynomial of the first kind and consider the composition n n 2n F : {−1, 1} → X given by F = S2n ◦ f, where L∞({−1, 1} ; R) is naturally identified with `∞ . Since Td(x) is a polynomial of degree d, the functions f and F are also of degree at most d and (85) ε1δ1 + ··· εnδn n kF (ε)kX 6 (1 + θ)kf(ε)kL∞({−1,1} ; ) = (1 + θ) max Td = 1 + θ, (87) R δ∈{−1,1}n n n for every ε ∈ {−1, 1} , because kTdkC([−1,1]) = Td(1) = 1. Therefore, we also have 1/p 1 X p kF k n = kF (ε)k 1 + θ. (88) Lp({−1,1} ;X) 2n X 6 ε∈{−1,1}n Furthermore, for every ε, δ ∈ {−1, 1}n, 1 ε δ + ··· + ε δ ε δ + ··· + ε δ 2ε δ ∂ f(ε)(δ) = T 1 1 n n − T 1 1 n n − i i , (89) i 2 d n d n n which implies the identity n 1 ε1δ1 + ··· + εnδn X ε1δ1 + ··· + εnδn 2εiδi ∆f(ε)(δ) = nT − T − . (90) 2 d n d n n i=1 Therefore, taking ε = δ, we deduce that n 2 n 2 k∆f(ε)k n T (1) − T 1 − = 1 − T 1 − , (91) L∞({−1,1} ;R) > 2 d d n 2 d n from which we derive the estimate (85) (91) n 2 k∆F k n k∆fk n n 1 − T 1 − . (92) Lp({−1,1} ;X) > Lp({−1,1} ;L∞({−1,1} ;R)) > 2 d n Applying (14) for the function F : {−1, 1}n → X and using (88) and (92), we finally get n 2 (92) n 1 − Td 1 − 6 k∆F kLp({−1,1} ;X) 2 n (93) (14) (88) 2 2 n 6 (1 − η)d kF kLp({−1,1} ;X) 6 (1 − η)(1 + θ)d . Letting n → ∞, we get 2 0 2 d = Td(1) 6 (1 − η)(1 + θ)d . (94) + 2 2 Finally, letting θ → 0 ,(94) becomes d 6 (1 − η)d which is a contradiction.

n ε1+···+εn Remark 24. Considering the function f : {−1, 1} → C given by f(ε) = Td n for n → ∞ shows that Theorem4 is also sharp when X = C and p = ∞. 3. Estimates for K-convex Banach spaces The main vector valued Fourier analytic tool which we will exploit in this section is the following fact which lies at the heart of the proof of Pisier’s K-convexity theorem [Pis82].

Theorem 25 (Pisier). A Banach space (X, k · kX ) is K-convex if and only if for every p ∈ (1, ∞) π there exist θ = θ(p, X) ∈ 0, 2 and M = M(p, X) ∈ [1, ∞) such that for every n ∈ N −z∆ n n | arg z| 6 θ =⇒ ke kLp({−1,1} ;X)→Lp({−1,1} ;X) 6 M, (95) n n n where kT kLp({−1,1} ;X)→Lp({−1,1} ;X) is the operator norm of T from Lp({−1, 1} ; X) to itself.

14 π The fact that a Banach space satisfying (95) for some θ ∈ 0, 2 and M ∈ [1, ∞) is K-convex is π simple. Indeed, let a = tan θ so that every point z in the interval with endpoints a + iπ, a − iπ has | arg z| 6 θ. Then, for k ∈ {0, 1, . . . , n}, n 1 π 1 π X eikte−(a+it)∆ dt = eikt−(a+it)jRad dt = e−kaRad , (96) 2π ˆ 2π ˆ j k −π −π j=0 which implies that π ka 1 ikt −(a+it)∆ Radk L ({−1,1}n;X)→L ({−1,1}n;X) = e e e dt p p n n 2π ˆ−π Lp({−1,1} ;X)→Lp({−1,1} ;X) (97) ka π (95) e −(a+it)∆ ka n n 6 ke kLp({−1,1} ;X)→Lp({−1,1} ;X) dt 6 Me . 2π ˆ−π a In particular, sup kRadk n n Me < ∞, i.e. (X, k · k ) is K-convex. n∈N Lp({−1,1} ;X)→Lp({−1,1} ;X) 6 X Here and throughout, we will denote by D = {z ∈ C : |z| < 1} the open unit disc. For r ∈ [1, ∞) consider the lens domain def p p Ω(r) = z ∈ C : max z − i r2 − 1 , z + i r2 − 1 < r ⊆ D (98) and notice that ∂Ω(r) is a Jordan curve with an interior angle θ(r) = 2 arcsin(1/r) at z = ±1. We will need the following simple consequence of Theorem 25.

Corollary 26. Let (X, k · kX ) be a K-convex Banach space and p ∈ (1, ∞). Then there exist some r = r(p, X) ∈ [1, ∞) and K = K(p, X) ∈ [1, ∞) such that ∆ n n sup sup kw kLp({−1,1} ;X)→Lp({−1,1} ;X) 6 K. (99) w∈Ω(r) n∈N Proof. By Theorem 25 and (36), for every w ∈ {e−z : |argz| < θ}, we have ∆ ∆ n n n n kw kLp({−1,1} ;X)→Lp({−1,1} ;X) = k(−w) kLp({−1,1} ;X)→Lp({−1,1} ;X) 6 M. (100) Moreover, by (97), we get ∆ n n kw kLp({−1,1} ;X)→Lp({−1,1} ;X) n (97) n X k X ak (101) |w| Radk n n M |w|e , 6 Lp({−1,1} ;X)→Lp({−1,1} ;X) 6 k=0 k=0 which implies that for |w| < e−2a, we also have the estimate

∆ M kw k n n . (102) Lp({−1,1} ;X)→Lp({−1,1} ;X) 6 1 − e−a Combining the domains for which (100) and (102) hold, one can easily deduce that there exists r ∈ [1, ∞) such that −z −z −2a Ω(r) ⊆ {e : |argz| < θ} ∪ {−e : |argz| < θ} ∪ {w ∈ D : |w| < e } (103) M and (99) follows with K = 1−e−a . We record for ease of future reference the following calculation of conformal mappings. α Lemma 27. For α ∈ (0, ∞), let φα : C r (−∞, 0] → C be a holomorphic branch of z 7→ z . Also, 1−z z+1 consider the M¨obiustransformations ψ1(z) = 1+z and ψ2(z) = z−1 . Then, for every r ∈ [1, ∞), the map ψ1 ◦φπ/θ(r) ◦ψ1 is a conformal equivalence between Ω(r) and the unit disc D. Furthermore, c the map ψ2 ◦ φπ/(2π−θ(r)) ◦ ψ2 is a conformal equivalence between the complement Ω(r) of Ω(r) and c the complement D of the closed unit disc.

15 Proof. Since ψ1(−1) = 0, ψ1(1) = ∞ and M¨obiustransformations preserve angles, we have n θ(r)o ψ Ω(r) = z ∈ : | arg z| < . (104) 1 C 2 Therefore, composing with φπ/θ(r), we get φπ/θ(r) ◦ ψ1 Ω(r) = z ∈ C : Rez > 0 , (105) which immediately implies that ψ1 ◦ φπ/θ(r) ◦ ψ1 Ω(r) = D. Since all the functions involved are conformal equivalences at their domains of deﬁnition the proof is complete. The proof of the claim c for the complement Ω(r) of Ω(r) is identical. We can now proceed with the proof of Theorem6. The crucial proposition is the following.

Proposition 28. Fix a K-convex Banach space (X, k·kX ), p ∈ (1, ∞) and K ∈ (0, ∞). Let Ω ⊆ D be a simply connected domain bounded by a Jordan curve γ such that (−1, 1) ⊆ Ω and suppose that ∆ n n sup sup kw kLp({−1,1} ;X)→Lp({−1,1} ;X) 6 K. (106) w∈Ω n∈N Also, let φΩ :Ω → D be a conformal mapping of Ω onto D with φΩ(0) = 0 which extends continuously as a homeomorphism between Ω ∪ γ and D. Then, for every n, d ∈ N with d ∈ {0, 1, . . . , n − 1} and every function f : {−1, 1}n → X in the d-th tail space, we have −t∆ −t d n n ∀ t > 0, ke fkLp({−1,1} ;X) 6 K φΩ e kfkLp({−1,1} ;X). (107) n Proof. Fix n ∈ N, d ∈ {0, 1, . . . , n−1} and let f : {−1, 1} → X be a function in the d-th tail space. d d+1 n For t > 0, consider the linear functional ζt : span{w , w , . . . , w }, k · kC(Ω) → C given by n n X k def X −tk ζt akw = ake , (108) k=d k=d −t d or ζt(p) = p(e ) if p(w) is a polynomial of degree n which is a multiple of w . If p is such a polynomial, consider the function h :Ω → C given by def p(w) ∀ w ∈ Ω, h(w) = d . (109) φΩ(w) Notice that φΩ does not vanish on Ω r {0} and furthermore it has a single root at 0, therefore the multiplicity of the root 0 in the numerator is at least the multiplicity in the denominator. Thus h is a holomorphic function on Ω. Furthermore, for w ∈ γ, we have |φΩ(w)| = 1 since φΩ is a homeomorphism between γ and ∂D (such a conformal map φΩ always exists by Caratheodory’s theorem, see [Kra06, Theorem 5.1.1]). Therefore, by the maximum principle, |p(w)| sup |h(w)| = max |h(w)| = max d = kpkC(γ) = kpkC(Ω), (110) w∈Ω w∈γ w∈γ |φΩ(w)| which implies that for w ∈ Ω, we have d |p(w)| 6 |φΩ(w)| kpkC(Ω). (111) −t Applying (111) for w = e ∈ Ω, we deduce that the linear functional ζt satisﬁes d d+1 n −t d ∀ p ∈ span{w , w , . . . , w }, |ζt(p)| 6 φΩ e kpkC(Ω). (112) Therefore, by the Hahn–Banach theorem and the Riesz representation theorem, there exists a −t d complex measure τt on Ω such that kτtkM(Ω) 6 |φΩ(e )| and

d d+1 n −t p ∈ span{w , w , . . . , w } =⇒ p(w) dτt(w) = p e . (113) ˆΩ

16 Since fb(A) = 0 when |A| < d, we get that

−t∆ (113) ∆ ∆ ke fkL ({−1,1}n;X) = w f dτt(w) kw fkL ({−1,1}n;X) d|τt|(w) p n 6 p ˆΩ Lp({−1,1} ;X) ˆΩ (106) n n (114) 6 K kfkLp({−1,1} ;X) d|τt|(w) = KkτtkM(Ω)kfkLp({−1,1} ;X) ˆΩ (112) −t d n 6 K φΩ e kfkLp({−1,1} ;X), which concludes the proof. n Proof of Theorem6. Fix n ∈ N, d ∈ {0, 1, . . . , n − 1}, p ∈ (1, ∞) and let f : {−1, 1} → X be a function in the d-th tail space. Using the notation introduced after Theorem 25, for every t > 2a, n (97) X 1 −t∆ −tk − 2 td ke fk n e Radkf n Ke kfk n , (115) Lp({−1,1} ;X) 6 Lp({−1,1} ;X) 6 Lp({−1,1} ;X) k=d M where K = 1−e−a . We will now treat the range t 6 2a. By Corollary 26, there exists r = r(p, X) ∈ [1, ∞) such that ∆ n n sup sup kw kLp({−1,1} ;X)→Lp({−1,1} ;X) 6 K, (116) w∈Ω(r) n∈N where Ω(r) is the domain (98). Therefore, by Proposition 28, we have −t∆ −t d n n ∀ t > 0, ke fkLp({−1,1} ;X) 6 K φΩ(r) e kfkLp({−1,1} ;X). (117) To conclude the proof, we will show that there exists c = c(p, X) such that −t −ctπ/θ(r) ∀ 0 6 t 6 2a, φΩ(r) e 6 e , (118) where θ(r) = 2 arcsin(1/r) ∈ (0, π] which would then imply the conclusion of Theorem6 with π A(p, X) = θ(r(p,X)) ∈ [1, ∞). By Lemma 27, if θ = θ(r), we have (1 + e−t)π/θ − (1 − e−t)π/θ φ e−t = . (119) Ω(r) (1 + e−t)π/θ + (1 − e−t)π/θ Therefore, −t π/θ ! −t π/θ −t 2(1 − e ) 2(1 − e ) log φΩ(r) e = log 1 − − (120) (1 + e−t)π/θ + (1 − e−t)π/θ 6 (1 + e−t)π/θ + (1 − e−t)π/θ and for t 6 2a, −t π/θ 1− π −2a π/θ 2(1 − e ) 1− π −t π/θ −2 θ (1 − e ) π/θ − −2 θ (1 − e ) t , (121) (1 + e−t)π/θ + (1 − e−t)π/θ 6 6 (2a)π/θ π 1− −2a π/θ which implies (118) with c = 2 θ (1−e ) . (2a)π/θ We now proceed to prove the dual of Theorem6, namely the improved lower bound on the decay of the heat semigroup, Theorem7. Even though Theorems6 and7 are not formally dual to one another, there is a continuous analogy between the techniques used in their proofs. We will need the following elementary result from complex analysis. Lemma 29. Let Ω ⊆ be a bounded simply connected domain and consider a conformal mapping c c C φΩc : Ω → D of the complement of Ω onto the complement of the closed unit disc with φΩc (∞) = ∞. Then, φΩc (z) lim = β ∈ C r {0}. (122) z→∞ z

17 def 1 c Proof. Without loss of generality assume that 0 ∈ Ω and let V = z ∈ C : z ∈ Ω . Then V ∪ {0} is a domain and the function F : V → C given by def 1 ∀ z ∈ V,F (z) = (123) φΩc 1/z is holomorphic, injective and satisﬁes |F (z)| 6 1 for every z ∈ V . Therefore, by Riemann’s theorem on removable singularities, F can be holomorphically extended at 0. Furthermore, F (0) = 0, since 0 otherwise limz→∞ φΩc (z) ∈ C. Since F is injective, we also have F (0) 6= 0. Thus, φΩc (z) w 1 lim = lim wφΩc (1/w) = lim = ∈ C r {0}, (124) z→∞ z w→0 w→0 F (w) F 0(0) which concludes the proof.

Proposition 30. Fix a K-convex Banach space (X, k·kX ), p ∈ (1, ∞) and K ∈ (0, ∞). Let Ω ⊆ D be a simply connected domain bounded by a Jordan curve γ and suppose that ∆ n n sup sup kw kLp({−1,1} ;X)→Lp({−1,1} ;X) 6 K. (125) w∈Ω n∈N c c Also, let φΩc : Ω → D be a conformal mapping of the complement of Ω onto the complement of the c c closed unit disc which extends continuously as a homeomorphism between Ω and D . Then, for ev- n ery n, d ∈ N with d ∈ {1, . . . , n} and every function f : {−1, 1} → X of degree at most d, we have −t∆ 1 ∀ t 0, ke fkL ({−1,1}n;X) kfkL ({−1,1}n;X). (126) > p > t d p K φΩc e n Proof. Fix n ∈ N, d ∈ {1, . . . , n} and let f : {−1, 1} → X be a function of degree at most d. For d t > 0, consider the linear functional ϕt : span{1, w, . . . , w }, k · kC(Ω) → C given by d d X k def X tk ϕt akw = ake , (127) k=0 k=0 t or ϕt(p) = p(e ) when p is a polynomial with deg(p) d. If p is such a polynomial, consider the c 6 function h : Ω → C given by c def p(z) ∀ z ∈ Ω , h(z) = d . (128) φΩc (z) c c Notice that, since the conformal map φΩc extends as a homeomorphism between Ω and D , it also satisﬁes φΩc (∞) = ∞. Thus, (122) implies that limz→∞ h(z) ∈ C and therefore h is a holomorphic c function on Ω ∪ {∞}. Furthermore, for z ∈ γ, we have |φΩc (z)| = 1 since φΩc is a homeomorphism between γ and ∂D. Therefore, by the maximum principle, |p(z)| sup |h(z)| = max |h(z)| = max d = kpkC(γ) = kpkC(Ω), (129) z∈Ωc z∈γ z∈γ |φΩc (z)| which implies that for z ∈ Ωc, we have d |p(z)| 6 |φΩc (z)| kpkC(Ω). (130) t c Applying (130) for z = e ∈ Ω , we deduce that the linear functional ϕt satisﬁes d t d ∀ p ∈ span{1, w, . . . , w }, |ϕt(p)| 6 φΩc (e ) kpkC(Ω). (131) Therefore, by the Hahn–Banach theorem and the Riesz representation theorem, there exists a t d complex measure ρt on Ω such that kρtkM(Ω) 6 |φΩc (e )| and

d t p ∈ span{1, w, . . . , w } =⇒ p(w) dρt(w) = p e . (132) ˆΩ

18 Since fb(A) = 0 when |A| > d, we get that

t∆ (132) ∆ ∆ ke fkL ({−1,1}n;X) = w f dρt(w) kw fkL ({−1,1}n;X) d|ρt|(w) p n 6 p ˆΩ Lp({−1,1} ;X) ˆΩ (125) n n (133) 6 K kfkLp({−1,1} ;X) d|ρt|(w) = KkρtkM(Ω)kfkLp({−1,1} ;X) ˆΩ (131) t d c n 6 K φΩ e kfkLp({−1,1} ;X), which is equivalent to (126). n Proof of Theorem7. Fix n ∈ N, d ∈ {1, . . . , n}, p ∈ (1, ∞) and let f : {−1, 1} → X be a function of degree at most d. By Theorem1, for every t > 0,

−t∆ 1 ke fk n kfk n (134) Lp({−1,1} ;X) > t Lp({−1,1} ;X) Td(e ) and, by (44), for t > 1, (44) t t d 2td Td(e ) 6 (2e ) 6 e . (135) We will now treat the range t 6 1. By Corollary 26, there exists r = r(p, X) ∈ [1, ∞) such that ∆ n n sup sup kw kLp({−1,1} ;X)→Lp({−1,1} ;X) 6 K, (136) w∈Ω(r) n∈N where Ω(r) is the domain (98). Therefore, by Proposition 30, we have

−t∆ 1 ∀ t 0, ke fkL ({−1,1}n;X) kfkL ({−1,1}n;X). (137) > p > t d p K φΩ(r)c e To conclude the proof, we will show that there exists C = C(p, X) such that

t Ctπ/(2π−θ(r)) ∀ 0 6 t 6 1, φΩ(r)c e 6 e , (138) where θ(r) = 2 arcsin(1/r) ∈ (0, π] which would then imply the conclusion of Theorem7 with π 1 η(p, X) = 2π−θ(r(p,X)) ∈ 2 , 1 . By Lemma 27, if θ = θ(r), we have

π π t 2π−θ t 2π−θ t (e + 1) + (e − 1) φΩ(r)c e = π π . (139) (et + 1) 2π−θ − (et − 1) 2π−θ Therefore, π ! π t 2π−θ t 2π−θ t 2(e − 1) 2(e − 1) log φΩ(r)c e = log 1 + π π 6 π π (140) (et + 1) 2π−θ − (et − 1) 2π−θ (et + 1) 2π−θ − (et − 1) 2π−θ and for t 6 1, t π t π π 2(e − 1) 2π−θ 2(e − 1) 2π−θ 2(e − 1) 2π−θ π 2π−θ π π 6 π π 6 π π t , (141) (et + 1) 2π−θ − (et − 1) 2π−θ (e + 1) 2π−θ − (e − 1) 2π−θ (e + 1) 2π−θ − (e − 1) 2π−θ

π 2(e−1) 2π−θ which implies (138) with C = π π . (e+1) 2π−θ −(e−1) 2π−θ An argument identical to the one used in the proof of Corollary2 implies the following improved moment comparison of low degree functions with values in a K-convex space (X, k · kX ).

19 Corollary 31. Let (X, k · kX ) be a K-convex Banach space. For every p > q > 1, there exist 1 C = C(p, X) ∈ (0, ∞) and η = η(p, X) ∈ 2 , 1 such that for every n, d ∈ N with d ∈ {1, . . . , n} and every function f : {−1, 1}n → X of degree at most d, we have η Cd max log p−1 ,log p−1 n q−1 q−1 n kfkLp({−1,1} ;X) 6 Ce kfkLq({−1,1} ;X). (142) We now proceed to prove Theorem8, the improved Bernstein–Markov inequality for the Lapla- cian. For this, we will use the following classical approximation theoretic result of Szeg¨o[Sze25].

Theorem 32 (Szegö). Let Ω ⊆ C be a domain bounded by a Jordan curve γ and fix w ∈ γ. Denote by θ(Ω, w) ∈ [0, 2π] the exterior angle of γ at w. Then, there exists a constant K(Ω, w) ∈ (0, ∞) such that for every d ∈ N and every polynomial p of degree at most d, we have 0 θ(Ω,w)/π |p (w)| 6 K(Ω, w)d kpkC(Ω). (143) A simple proof of Szegö’stheorem specifically for lens domains of the form (98) was given in [EI18]. n Proof of Theorem8. Fix n ∈ N, d ∈ {1, . . . , n}, p ∈ (1, ∞) and let f : {−1, 1} → X be a function π of degree at most d. Also let θ = θ(p, X) ∈ 0, 2 and M = M(p, X) ∈ [1, ∞) be given by Theorem −z 25. Consider the domain V = {e : | arg z| < θ} ⊆ D. Then, by Theorem 25, we have ∆ n n ∀ w ∈ V, kw kLp({−1,1} ;X)→Lp({−1,1} ;X) 6 M. (144) d As in the proof of Theorem4, consider the linear functional ψ : span{1, w, . . . , w }, k · kC(V ) → C given by d d X k def X ψ akw = kak, (145) k=0 k=0 0 or ψ(p) = p (1) when p is a polynomial with deg(p) 6 d. Notice that V is a domain bounded by a Jordan curve γ with 1 ∈ γ which forms an exterior angle 2π − 2θ at 1. Then, Szegö’sinequality (143) implies that there exists some K = K(p, X) ∈ (0, ∞) such that

2θ d 2− π ∀ p ∈ span{1, w, . . . , w }, |ψ(p)| 6 Kd kpkC(V ). (146) Therefore, by the Hahn–Banach theorem and the Riesz representation theorem, there exists a 2θ 2− π complex measure ν on V such that kνkM(V ) 6 Kd such that for every polynomial p, we have

0 deg(p) 6 d =⇒ p(w) dν(w) = p (1). (147) ˆV Since fb(A) = 0 when |A| > d, we get that

(147) ∆ ∆ k∆fk n = w f dν(w) kw fk n d|ν|(w) Lp({−1,1} ;X) n 6 Lp({−1,1} ;X) ˆ Lp({−1,1} ;X) ˆ V V (148) (95) (146) 2− 2θ n π n 6 M kfkLp({−1,1} ;X) d|ν|(w) 6 KMd kfkLp({−1,1} ;X), ˆV 2θ(p,X) which concludes the proof of the theorem with α(p, X) = 2 − π ∈ [1, 2). Remark 33. It is straightfoward to see that if the Bernstein–Markov inequality (12) holds true with linear dependence on the degree for a given Banach space, then the asymptotically optimal lower bound for the action of the heat semigroup conjectured after Theorem7 follows. Indeed, assume that for a Banach space (X, k · kX ) and p ∈ [1, ∞) there exists C ∈ (0, ∞) such that for every n n, d ∈ N with d ∈ {1, . . . , n} and every function f : {−1, 1} → X of degree at most d, we have

n n k∆fkLp({−1,1} ;X) 6 CdkfkLp({−1,1} ;X). (149)

20 Then, for every t > 0, we get ∞ k ∞ n (149) k t∆ X k(t∆) fkLp({−1,1} ;X) X (Ctd) ke fk n kfk n Lp({−1,1} ;X) 6 k! 6 k! Lp({−1,1} ;X) k=0 k=0 (150) Ctd n = e kfkLp({−1,1} ;X), which, applied to e−t∆f, is equivalent to the conjectured optimal version of Theorem7. The Bernstein–Markov inequality (20) for the vector valued gradient is an immediate consequence of Theorem8 combined with the following dual to Pisier’s inequality (11) for low degree functions.

Proposition 34. Let (X, k · kX ) be a K-convex Banach space. For every p ∈ (1, ∞) there exists n B(p, X) ∈ (0, ∞) such that for every n, d ∈ N with d ∈ {1, . . . , n}, every function g : {−1, 1} → X of degree at most d satisﬁes n 1 X X p 1/p δi∂ig B(p, X)(log d + 1)k∆gk n . (151) n n 6 Lp({−1,1} ;X) 2 Lp({−1,1} ;X) δ∈{−1,1}n i=1

Proof. Following the notation (71) of the proof of Theorem3, for t > 0, consider the function n n gt : {−1, 1} × {−1, 1} → X given by def X Y −t −t gt(ε, δ) = gb(A) e εi + (1 − e )δi A⊆{1,...,n} i∈A n (152) −t∆ t X −t∆ = e g(ε) + (e − 1) εiδi∂ie g(ε) + Φt(ε, δ), i=1 P n where δ∈{−1,1}n δiΦt(ε, δ) = 0 for every ε ∈ {−1, 1} and i ∈ {1, . . . , n}. Therefore, for every t > 0, we have n p 1/p 1 X X −t∆ (152) 1 δi∂ie g = Radδgt n n n n t Lp({−1,1} ×{−1,1} ;X) 2 Lp({−1,1} ;X) e − 1 δ∈{−1,1}n i=1 (153) K (73) K kg k n n kgk n , 6 et − 1 t Lp({−1,1} ×{−1,1} ;X) 6 et − 1 Lp({−1,1} ;X) n where Radδgt is the Rademacher projection of gt with respect to the variable δ ∈ {−1, 1} , K = K(p, X) = sup kRadk n n < ∞ and the proof of the last inequality is n∈N Lp({−1,1} ;X)→Lp({−1,1} ;X) identical to the proof of (73) via (72). Integrating the above inequality, we deduce that for s > 0, n p 1/p 1 X X −1 −s∆ δi∂i∆ e g n n 2 Lp({−1,1} ;X) δ∈{−1,1}n i=1 n ∞ p 1/p 1 X X −t∆ = δi∂ie g dt (154) n n 2 ˆ Lp({−1,1} ;X) δ∈{−1,1}n s i=1 (153) ∞ 1 es K dt kgk n = K log kgk n . 6 t Lp({−1,1} ;X) s Lp({−1,1} ;X) ˆs e − 1 e − 1 s d2s By Theorem1 and the elementary inequality Td(e ) 6 e , where s > 0, we conclude that for every δ ∈ {−1, 1}n, n n X −1 −s∆ −d2s X −1 δi∂i∆ e g > e δi∂i∆ g . (155) L ({−1,1}n;X) L ({−1,1}n;X) i=1 p i=1 p

21 Therefore, combining (154) and (155), n p 1/p s 1 X X −1 d2s e δi∂i∆ g K mine log kgkL ({−1,1}n;X) n n 6 s p 2 Lp({−1,1} ;X) s>0 e − 1 δ∈{−1,1}n i=1 (156)

n 6 3K(log d + 1)kgkLp({−1,1} ;X), which is equivalent to the desired inequality (151) with B(p, X) = 3K(p, X). Remark 35. It has been shown in [NS02] that the validity of (151) with the factor C(log d + 1) replaced by a constant C(p, X) depending only on p and the Banach space (X, k · kX ), where p ∈ (1, ∞), implies that X is K-convex. Nevertheless, (151) is the best known bound to date for general K-convex spaces, even when d = n. Under additional assumptions (e.g. when X is a UMD+ space or when X is a K-convex Banach lattice), inequality (151) is known to hold true with a constant C(p, X) independent of the dimension n for functions of arbitrary degree d, see [HN13].

n Proof of Theorem9. Fix n ∈ N, d ∈ {1, . . . , n} and let f : {−1, 1} → X be a function of degree at most d. Then, we have n 1/p (151) 1 X X p δi∂if n B(p, X)(log d + 1)k∆fkL ({−1,1}n;X) 2n Lp({−1,1} ;X) 6 p δ∈{−1,1}n i=1 (157) (19) α(p,X) n 6 B(p, X)C(p, X)d (log d + 1)kfkLp({−1,1} ;X), which completes the proof. Finally, we will prove the reverse Bernstein–Markov inequality of Theorem 10. Recall that for γ ∈ (0, ∞) and a function f : {−1, 1}n → X, we denote by

γ def X γ ∆ f = |A| fb(A)wA (158) A⊆{1,...,n} the action of a fractional power of the hypercube Laplacian ∆ on f. We will prove the following statement for ∆1/2 of which Theorem 10 is an immediate consequence.

Theorem 36. Let (X, k · kX ) be a K-convex Banach space. For every p ∈ (1, ∞) there exists n c(p, X) ∈ (0, ∞) such that for every n, d, m ∈ N with d+m 6 n and every function f : {−1, 1} → X of degree at most d + m which is also in the d-th tail space, we have r 1/2 d k∆ fk n c(p, X) kfk n . (159) Lp({−1,1} ;X) > m Lp({−1,1} ;X) We start by proving the following lemma.

Lemma 37. Let (X, k · kX ) be a K-convex Banach space. For every p ∈ (1, ∞) there exists n K = K(p, X) ∈ (0, ∞) such that for every n ∈ N and every function f : {−1, 1} → X, we have

1/2 −t∆ K n n ∀ t > 0, k∆ e fkLp({−1,1} ;X) 6 √ kfkLp({−1,1} ;X). (160) e2t − 1 Proof. By (97), there exist M = M(p, X) ∈ (0, ∞) and a = a(p, X) ∈ (0, ∞) such that for every n n ∈ N, k ∈ {1, . . . , n} and every function f : {−1, 1} → X, we have ka Radkf n Me kfk n . (161) Lp({−1,1} ;X) 6 Lp({−1,1} ;X)

22 Therefore, for t > 2a, n √ 1/2 −t∆ X −tk k∆ e fk n ke Radkf n Lp({−1,1} ;X) 6 Lp({−1,1} ;X) k=1 (162) (161) n √ ∞ √ X −(t−a)k a X −ak −t n n 6 M ke kfkLp({−1,1} ;X) 6 Me ke e kfkLp({−1,1} ;X), k=1 k=1 which implies that for t > 2a,

1/2 −t∆ K n n k∆ e fkLp({−1,1} ;X) 6 √ kfkLp({−1,1} ;X), (163) e2t − 1 √ a P∞ −ak where K = Me k=1 ke . To prove (160) for t ∈ (0, 2a), recall that by Pisier’s K-convexity π theorem, there exists θ = θ(p, X) ∈ 0, 2 such that −z∆ n n | arg z| 6 θ =⇒ ke kLp({−1,1} ;X)→Lp({−1,1} ;X) 6 M. (164) Let r = t sin θ and notice that the closed disc D(t, r) of radius r centered at t is contained in {z ∈ C : | arg z| 6 θ}. By the Cauchy integral formula for the derivative, −ζ∆ −t∆ 1 e − ∆e = 2 dζ, (165) 2πi ˆ∂D(t,r) (ζ − t) hence (165) (164) −t∆ 1 −ζ∆ M n n n k∆e fkLp({−1,1} ;X) 6 sup ke fkLp({−1,1} ;X) 6 kfkLp({−1,1} ;X). (166) r ζ∈∂D(t,r) t sin θ Furthermore, by an inequality of Naor and Schechtman [BELP08, Lemma 5.6], for every β ∈ (0, 1) and function g : {−1, 1}n → X, we have

β β 1−β k∆ gk n 4k∆gk n kgk n . (167) Lp({−1,1} ;X) 6 Lp({−1,1} ;X) Lp({−1,1} ;X) −t∆ 1 Therefore, combining (167) for g = e f and β = 2 with (166) and the contractivity of the heat semigroup, we deduce that √ 1/2 −t∆ 4 M n n k∆ e fkLp({−1,1} ;X) 6 √ √ kfkLp({−1,1} ;X), (168) t sin θ

√which completes√ the proof of (160), since for every a ∈ (0, ∞), there exists ca ∈ (0, ∞) such that 2t e − 1 6 ca t for t ∈ (0, 2a). In the proof of Theorem 36, we will use a reverse Bernstein inequality for incomplete polynomials, proven recently by Erd´elyi[Erd20].

Theorem 38 (Erd´elyi). Fix d, m ∈ N and let P (x) be a polynomial of the form d d+1 d+m P (x) = adx + ad+1x + ··· + ad+mx . (169) Then, rm p |P (1)| 6 k 1 − x2P 0(x)k . (170) 6 d C([0,1]) Furthermore, the estimate is sharp up to the value of the universal constant.

23 n Proof of Theorem 36. Fix n, d, m ∈ N with d+m 6 n and a function f : {−1, 1} → X of degree at most d + m which is also in the d-th tail space. Writing x = e−t ∈ [0, 1], (160) can be rewritten as

p ∆−1 1/2 2 n n ∀ x ∈ [0, 1], k 1 − x ∆x fkLp({−1,1} ;X) 6 Kk∆ fkLp({−1,1} ;X). (171) √ √ 2 d−1 2 d+m−1 Consider the linear functional ξ : span 1 − x x ,..., 1 − x x , k · kC([0,1]) given by

d+m d+m p 2 X k−1 def X ξ 1 − x kakx = ak. (172) k=d k=d Then, Erd´elyi’sinequality (170) can be rewritten as p p rm ∀ p ∈ span 1 − x2xd−1,..., 1 − x2xd+m−1 , |ξ(p)| 6 kpk . (173) 6 d C([0,1]) Therefore, by the Hahn–Banach theorem and the Riesz representation theorem, there exists a p m complex measure σ on [0, 1] such that kσkM([0,1]) 6 6 d and p p 1 ∀ p ∈ span 1 − x2xd−1,..., 1 − x2xd+m−1 , p(x) dσ(x) = ξ(p). (174) ˆ0 Since fb(A) = 0 when |A| ∈/ {d, . . . , d + m}, we deduce that 1 (174) p 2 ∆−1 kfkL ({−1,1}n;X) = 1 − x ∆x f dσ(x) p n ˆ0 Lp({−1,1} ;X) 1 (171) 1 p 2 ∆−1 1/2 1 − x ∆x f d|σ|(x) Kk∆ fk n d|σ|(x) (175) 6 L ({−1,1}n;X) 6 Lp({−1,1} ;X) ˆ0 p ˆ0 (173) r 1/2 m 1/2 = Kkσk k∆ fk n 6K k∆ fk n , M([0,1]) Lp({−1,1} ;X) 6 d Lp({−1,1} ;X) which is equivalent to (159). n Proof of Theorem 10. Fix n, d, m ∈ N with d + m 6 n and a function f : {−1, 1} → X of degree at most d + m which is also in the d-th tail space. Then, Theorem 36 implies that 1/2 1/2 n n k∆fkLp({−1,1} ;X) = k∆ ∆ fkLp({−1,1} ;X) (159) r (159) (176) d 1/2 2 d c(p, X) k∆ fk n c(p, X) kfk n , > m Lp({−1,1} ;X) > m Lp({−1,1} ;X) which is the desired inequality.

4. Estimates for scalar valued functions We noticed in Corollary 26 that Pisier’s K-convexity theorem easily implies that for every K- convex Banach space (X, k·kX ) and p ∈ (1, ∞) there exist r = r(p, X) ∈ [1, ∞) and K = K(p, X) ∈ [1, ∞) such that ∆ n n sup sup kw kLp({−1,1} ;X)→Lp({−1,1} ;X) 6 K, (177) w∈Ω(r) n∈N where the domain Ω(r) is given by (98). The scalar valued version of this result was ﬁrst studied in classical work of Weissler [Wei79] who found the exact domain Ωp ⊆ D for which the operator ∆ n n w : Lp({−1, 1} ; C) → Lp({−1, 1} ; C) is uniformly bounded (equivalently, a contraction) for 3 p ∈ (1, ∞) r 2 , 2 ∪ (2, 3) and w ∈ Ωp. Finally, the domain Ωp for p in the remaining range 3 2 , 2 ∪ (2, 3) was recently identiﬁed by the second named author and Nazarov in [IN19].

24 √p Theorem 39 (Weissler, Ivanisvili–Nazarov). For p ∈ (1, ∞), let rp = 2 p−1 . Then, for every p ∈ (1, ∞) and w ∈ C, we have ∆ n n w ∈ Ω(rp) ⇐⇒ sup kw kLp({−1,1} ;C)→Lp({−1,1} ;C) = 1. (178) n∈N Furthermore, both conditions are equivalent to ∆ n n sup kw kLp({−1,1} ;C)→Lp({−1,1} ;C) < ∞. (179) n∈N The decay properties of Theorem 11 are straightforward consequences of Theorem 39 and the results of Section3. Proof of Theorem 11. Inequality (23) for functions of low degree is a straightforward consequence of Theorem 39 combined with Proposition 30 and (139). Similarly, inequality (24) for functions in the tail space follows from Theorem 39 combined with Proposition 28 and (119). Proof of Corollary 12. The deduction of Corollary 12 from Theorem 11 is identical to the deduction of Corollary2 from Theorem1; it follows by concatenating (23) with Bonami’s hypercontractive 1 p−1 inequality (46) and choosing t = 2 log q−1 .

We now proceed with the proof of Theorem 13, the improved L1 −L2 moment comparison for low degree functions. We use the following important result of Beckner [Bec75] and Weissler [Wei79]. Theorem 40 (Beckner, Weissler). Let p ∈ (2, ∞) and p∗ ∈ (1, 2) its conjugate exponent, that is ∗ p p = p−1 . Then, a complex number w ∈ D satisfies n p − 2 p − 2 o p o max w − , w + (180) 2(p − 1) 2(p − 1) 6 2(p − 1) if and only if ∆ sup kw k n n = 1. (181) Lp({−1,1} ;C)→Lp∗ ({−1,1} ;C) n∈N Furthermore, both conditions are equivalent to ∆ sup kw k n n < ∞. (182) Lp({−1,1} ;C)→Lp∗ ({−1,1} ;C) n∈N √ In [Bec75], Beckner proved that w = ±i p∗ − 1 satisfies (181) and then Weissler [Wei79] modified his argument to obtain (181) for w in the full domain(180).

Proof of Theorem 13. Fix n ∈ N and d ∈ {1, . . . , n}. For p ∈ (2, ∞), denote by Vp ⊆ D the set of all w ∈ D that satisfy (180). Then, a straightforward computation shows that p ipp − 1V = Ω √ , (183) p 2 p − 1 c c where Ω(r) is deﬁned by (98). Therefore, by Lemma 27, the function ϕp : Vp → D √ π √ π def (i p − 1w − 1) 2π−θp + (i p − 1w + 1) 2π−θp ∀ w ∈ C r Vp, ϕp(w) = √ π √ π , (184) (i p − 1w − 1) 2π−θp − (i p − 1w + 1) 2π−θp √ 2 p−1 where θp = 2 arcsin p , is a conformal equivalence between the complement of Vp and the complement of the closed unit disc D. A duality argument identical to that of Proposition 30 n combined with (181) now implies that for every function f : {−1, 1} → C of degree at most d,

−t∆ 1 ∀ t 0, ke fkL ∗ ({−1,1}n; ) kfkL ({−1,1}n; ). (185) > p C > t d p C ϕp e

25 In particular, for t = 0, we have d kfk n |ϕ (1)| kfk n . (186) Lp({−1,1} ;C) 6 p Lp∗ ({−1,1} ;C) Now, by Hölder’sinequality, we get p−2 p 2(p−1) 2(p−1) kfkL ({−1,1}n; ) kfk n kfk n (187) 2 C 6 L1({−1,1} ;C) Lp({−1,1} ;C) and p−2 1 p−1 p−1 kfkL ∗ ({−1,1}n; ) kfk n kfk n . (188) p C 6 L1({−1,1} ;C) Lp({−1,1} ;C) Combining the two, we deduce that ! n p kfkLp({−1,1} ;C) n n 2(p−2) kfkL2({−1,1} ;C) 6 kfkL1({−1,1} ;C) kfk n Lp∗ ({−1,1} ;C) (189) (186) pd 2(p−2) n 6 |ϕp(1)| kfkL1({−1,1} ;C). Consequently, d n n kfkL2({−1,1} ;C) 6 C kfkL1({−1,1} ;C), (190) where p π π √ √ 2(p−2) (i p − 1 − 1) 2π−θp + (i p − 1 + 1) 2π−θp def C = inf √ π √ π < 2.69076. (191) p>2 (i p − 1 − 1) 2π−θp − (i p − 1 + 1) 2π−θp The last inequality can be checked numerically. d/2 n n We note in passing that the bound kfkL2({−1,1} ;C) 6 e kfkL1({−1,1} ;C), which improves upon n Theorem 13, was obtained in the recent work [IT19] for d-homogeneous functions f : {−1, 1} → C. Remark 41. Following Beckner’s pioneering work [Bec75], significant efforts were devoted in iden- tifying the complex domains consisting of those w ∈ D for which ∆ n n sup kw kLp({−1,1} ;C)→Lq({−1,1} ;C) < ∞ (192) n∈N for general p > q > 1. In [Wei79], Weissler managed to precisely characterize these complex 3 numbers w for all p > q > 1 apart from the cases 2 < q 6 p < 2 and 2 < q 6 p < 3 and posed a conjecture for p, q in the remaining ranges. The case p = q of his conjecture was recently settled by the second named author and Nazarov [IN19]. In contrast to this long standing problem, Epperson [Epp89] (see also [Jan97]) has characterized those w ∈ D for which L n n sup kw kLp((R ,γn);C)→Lq((R ,γn);C) < ∞, (193) n∈N n where γn is the standard Gaussian measure on R and L = ∆ − hx, ∇i is the generator of the Ornstein–Uhlenbeck semigroup, for every p > q > 1. For p and q not belonging in the missing ranges mentioned earlier, the domains of complex hypercontractivity for the Hamming cube and 3 the Gauss space coincide and it is natural to believe that this is also the case when 2 < q < p < 2 and 2 < q < p < 3. It is evident from the proof above that the constant 2.69076 appearing in Theorem 13 is not optimal. In fact, one can run a similar argument starting with the inequality pq ! 2(p−q) n kfkLp({−1,1} ;C) n n kfkL2({−1,1} ;C) 6 kfkL1({−1,1} ;C) , (194) n kfkLq({−1,1} ;C) n which is valid for any p > 2 > q and any function f : {−1, 1} → C. Then, to obtain an Lq − Lp moment comparison as in the proof of Theorem 13, one should use the general Lq − Lp

26 complex hypercontractivity of [Wei79] and explicitly compute the conformal map of the domain provided by Weissler’s theorem. In fact, such a computation could also provide an improvement of Corollary 12. We did not attempt to optimize any of these computations as the domains of Lq − Lp hypercontractivity for q∈ / {p, p∗} tend to be quite complicated. We also note that for p > 2, the n least constant Cp for which every function f : {−1, 1} → C of degree at most d satisﬁes d n n kfkLp({−1,1} ;C) 6 Cp kfkL2({−1,1} ;C) (195) √ is known to be Cp = p −√1 (see [IT19]), yet the sharp constant in (27) is still unknown. In- equality (195) with Cp = p − 1 is usually proven via an orthogonality argument (see [O’D14, Theorem 9.21]), but a duality based proof can be given using the result of Weissler [Wei79] who showed that for p > 2

1 ∆ √ n n |w| 6 ⇐⇒ sup kw kLp({−1,1} ;C)→L2({−1,1} ;C) = 1. (196) p − 1 n∈N A straightfoward adaptation of the proof of Proposition 30 then implies that for every function n f : {−1, 1} → C of degree at most d, we have −t d −t∆ e ∀ t 0, ke fk n √ kfk n , (197) > L2({−1,1} ;C) > p − 1 Lp({−1,1} ;C) √ which for t = 0 coincides with (195) with Cp = p − 1. Proof of Theorem 14. The proof is a mechanical adaptation of the proof of Theorem8, where (144) is replaced by the characterization (178) of the complex domain where the heat ﬂow is a contraction. The fact that the underlying constant K = K(Ω, w) in Szeg¨o’stheorem can be taken to be the absolute constant 10 if Ω is a lens domain of the form (98) and w = 1 was shown in [EI18, Proposition 15]. This proves the Bernstein–Markov inequality (28). To derive the Bernstein–Markov inequalities for the discrete gradient presented in Theorem 15, we will need to make use of Lust-Piquard’s Riesz transform inequalities [LP98] (see also [BELP08] where the implicit dependence in p was improved).

Theorem 42 (Lust-Piquard). For every p ∈ [2, ∞), there exist cp,Cp ∈ (0, ∞) such that for every n n ∈ N, every function f : {−1, 1} → C satisﬁes 1/2 1/2 n n n cpk∆ fkLp({−1,1} ;C) 6 k∇fkLp({−1,1} ;C) 6 Cpk∆ fkLp({−1,1} ;C). (198)

n Proof of Theorem 15 for p > 2. Fix n ∈ N, d ∈ {1, . . . , n} and let f : {−1, 1} → C be a function of degree at most d. Then, by Lust-Piquard’s inequality (198), we have

1/2 n n k∇fkLp({−1,1} ;C) 6 Cpk∆ fkLp({−1,1} ;C). (199) 1 Combining (199), Naor and Schechtman’s inequality (167) for β = 2 and Theorem 14, we derive the Bernstein–Markov inequalities of Theorem 15 for p > 2. Even though the one-sided Riesz transform inequality 1/2 n n ∀ p ∈ (1, ∞), cpk∆ fkLp({−1,1} ;C) 6 k∇fkLp({−1,1} ;C) (200) is true for p ∈ (1, 2) (see [LP98]), its reverse is known to be false in this range. In fact, it has been shown by Naor and Schechtman (see [BELP08, Lemma 5.5]) that, if p ∈ (1, 2), a dimension independent inequality of the form β n n k∇fkLp({−1,1} ;C) 6 Ck∆ fkLp({−1,1} ;C) (201)

27 1 implies that β > p . We will now show that this is almost optimal. The following proposition is due to A. Naor, to whom we are grateful for allowing us to include it here. The proof presented here is diﬀerent than Naor’s original proof, which will appear elsewhere. 1 Proposition 43 (Naor). For every p ∈ (1, 2) and every ε ∈ 0, 2 , there exists Cp ∈ (0, ∞) such n that for every n ∈ N, every function f : {−1, 1} → C satisﬁes

Cp 1/p+ε k∇fk n k∆ fk n . (202) Lp({−1,1} ;C) 6 ε Lp({−1,1} ;C) For the proof of Proposition 43 we will need the following lemma.

Lemma 44. For every p ∈ (1, 2], there exists Ap ∈ (0, ∞) such that for every n ∈ N, every function n f : {−1, 1} → C satisﬁes

−t∆ Ap ∀ t > 0, k∇e fk n · kfk n . (203) Lp({−1,1} ;C) 6 (ept − 1)1/p Lp({−1,1} ;C) n n n Proof. For t > 0, consider the operator St : Lp({−1, 1} ; C) → Lp({−1, 1} × {−1, 1} ; C) given by def X Y −t −t −t∆ St(f)(ε, δ) = fb(A) e εi + (1 − e )δi − e f(ε), (204) A⊆{1,...,n} i∈A where (ε, δ) ∈ {−1, 1}n × {−1, 1}n. Notice that, because of (72), (73) and the contractivity of the heat semigroup, we have

n n n kSt(f)kL1({−1,1} ×{−1,1} ;C) 6 2kfkL1({−1,1} ;C). (205)

Furthermore, if f = wA for a subset A ⊆ {1, . . . , n},

X −t|B| −t |ArB| St(wA)(ε, δ) = e (1 − e ) wB(ε)wArB(δ), (206) B(A which by orthogonality implies that |A|−1 2 X |A| −2tk −t 2(|A|−k) kSt(wA)k n n = e (1 − e ) L2({−1,1} ×{−1,1} ;C) k k=0 (207) −2t −t 2|A| −2t|A| −t = e + (1 − e ) − e 6 1 − e , where the last inequality is elementary. Therefore,

kStkL ({−1,1}n; )→L ({−1,1}n×{−1,1}n; ) = max kSt(wA)kL ({−1,1}n×{−1,1}n; ) 2 C 2 C A⊆{1,...,n} 2 C (208) p 6 1 − e−t. Using the Riesz–Thorin interpolation theorem, we conclude that for every p ∈ [1, 2],

(205)∧(208) 2 −1 −t 1− 1 −t 1− 1 n n n p p p kStkLp({−1,1} ;C)→Lp({−1,1} ×{−1,1} ;C) 6 2 (1 − e ) 6 2(1 − e ) (209) and, by (71), we get n p 1/p 1 X X −t∆ (71) 1 δi∂ie f = RadδSt(f) n n 2n L ({−1,1}n; ) et − 1 Lp({−1,1} ×{−1,1} ;C) n i=1 p C δ∈{−1,1} (210) (209) n n Kp 2KpkfkLp({−1,1} ;C) 4KpkfkLp({−1,1} ;C) kSt(f)k n n , 6 t Lp({−1,1} ×{−1,1} ;C) 6 t(1− 1 ) 1 6 1 e − 1 e p (et − 1) p (ept − 1) p

28 where K = sup kRadk n n . Finally, by Khintchine’s inequality, p n∈N Lp({−1,1} ;C)→Lp({−1,1} ;C) n p 1/p −t∆ 1 X X −t∆ k∇e fkL ({−1,1}n; ) Bp δi∂ie f (211) p C 6 n n 2 Lp({−1,1} ;C) δ∈{−1,1}n i=1 √ for some Bp ∈ [1, 2], which combined with (210) completes the proof. 1 n Proof of Proposition 43. Fix ε ∈ 0, 2 , n ∈ N, d ∈ {1, . . . , n} and let f : {−1, 1} → C. A change of variables gives the integral representation ∞ 1 1 +ε−1 ∆−1/p−εf = t p e−t∆f dt. (212) 1 ˆ Γ p + ε 0 Therefore, we can write ∞ −1/p−ε (212) 1 1 +ε−1 −t∆ k∇∆ fkL ({−1,1}n; ) = ∇ t p e f dt p C 1 ˆ L ({−1,1}n; ) Γ p + ε 0 p C ∞ 1 1 +ε−1 −t∆ t p k∇e fk n dt 6 1 ˆ Lp({−1,1} ;C) (213) Γ p + ε 0 1 (203) ∞ +ε−1 Ap t p dt kfk n . 6 pt 1/p Lp({−1,1} ;C) Γ(1/2) ˆ0 (e − 1) To conclude the proof, notice that 1 +ε−1 1 +ε−1 1 +ε−1 ∞ t p 1 t p ∞ t p dt = dt + dt = O(1/ε) + O(1) = O(1/ε) (214) pt 1/p pt 1/p pt 1/p ˆ0 (e − 1) ˆ0 (e − 1) ˆ1 (e − 1) and thus (213) becomes

−1/p−ε Cp k∇∆ fk n kfk n , (215) Lp({−1,1} ;C) 6 ε Lp({−1,1} ;C) which is equivalent to (202).

n n Remark 45. Combining (202) with (167) and the estimate k∆fkLp({−1,1} ;C) 6 nkfkLp({−1,1} ;C) n which holds true for every function f : {−1, 1} → C, we get ε 4Cpn 1/p k∇fk n k∆ fk n , (216) Lp({−1,1} ;C) 6 ε Lp({−1,1} ;C) 1 which for ε = log(n+1) becomes 1/p n n k∇fkLp({−1,1} ;C) 6 4eCp log(n + 1)k∆ fkLp({−1,1} ;C). (217) In the upcoming manuscript [EN20], the ﬁrst named author and Naor use a new Littlewood–Paley– Stein inequality [Ste70] on the discrete hypercube, which allows to improve the logarithmic term in cp (217) to (log n) for some cp ∈ (0, 1), where p ∈ (1, 2). We conjecture that (217) holds true with a dimension independent constant. n Proof of Theorem 15 for p ∈ (1, 2). Fix ε ∈ (0, 1), n ∈ N, d ∈ {1, . . . , n} and let f : {−1, 1} → C 1 be a function of degree at most d. Combining (202), (167) for β = p + ε and (28), we deduce that θ 40Cp (2− p )( 1 +ε) k∇fk n d π p kfk n . (218) Lp({−1,1} ;C) 6 ε Lp({−1,1} ;C) 1 Choosing ε = log(d+1) gives the conclusion. The next proposition is an asymptotically sharp endpoint Bernstein–Markov inequality for the discrete gradient and p = ∞.

29 n Proposition 46. For every n, d ∈ N with d ∈ {1, . . . , n} and every function f : {−1, 1} → C,

n n k∇fkL∞({−1,1} ;C) 6 2dkfkL∞({−1,1} ;C). (219) Furthermore, the factor d is asymptotically optimal. We will need the following lemma, which is a special case of a more general semigroup statement from [BGL14, Proposition 8.6.1] and is the L∞ analogue of Lemma 44. n Lemma 47. For every n ∈ N and every function f : {−1, 1} → C, we have

−t∆ 1 n n ∀ t > 0, k∇e fkL∞({−1,1} ; ) 6 √ kfkL∞({−1,1} ; ). (220) C e2t − 1 C n Proof. Fix n ∈ N and consider a function f : {−1, 1} → C. Then, we have the pointwise identity n n X t X (e2t − 1) (∂ e−t∆f)2 = 2 e2s (∂ e−t∆f)2 ds. (221) i ˆ i i=1 0 i=1 −(t−s)∆ Denote by gi(ε) = εi∂ie f(ε) and notice that n s −t∆ −s∆ ∀ ε ∈ {−1, 1} , εie ∂ie f(ε) = e gi(ε). (222) Therefore, (221) implies that n n X t X (e2t − 1) (∂ e−t∆f)2 = 2 (e−s∆g )2 ds i ˆ i i=1 0 i=1 n n (223) (†) t X t X 2 e−s∆ g2 ds = 2 e−s∆ (∂ e−(t−s)∆f)2 ds, 6 ˆ i ˆ i 0 i=1 0 i=1 where inequality (†) follows from Jensen’s inequality since e−s∆ is an averaging operator. Using n the deﬁnition of the Laplacian, one can check that for every function h : {−1, 1} → C the identity n n 2 X 2 ∀ ε ∈ {−1, 1} , 2h(ε) · ∆h(ε) − ∆h (ε) = 2 (∂ih(ε)) (224) i=1 holds true. Combining (223) and (224), we conclude that n X t (e2t − 1) (∂ e−t∆f)2 e−s∆2e−(t−s)∆f · ∆e−(t−s)∆f − ∆(e−(t−s)∆f)2 ds. (225) i 6 ˆ i=1 0 However, we also have d e−s∆e−(t−s)∆f2 = e−s∆2e−(t−s)∆f · ∆e−(t−s)∆f − ∆(e−(t−s)∆f)2, (226) ds which implies that (225) can be rewritten as

n t X d 2 (e2t − 1) (∂ e−t∆f)2 e−s∆e−(t−s)∆f ds = e−t∆f 2 − (e−t∆f)2. (227) i 6 ˆ ds i=1 0 Therefore,

p 2t −t∆ −t∆ 2 1/2 e − 1k∇e fk n ke f k n kfk n , (228) L∞({−1,1} ;C) 6 L∞({−1,1} ;C) 6 L∞({−1,1} ;C) which completes the proof of the lemma.

30 n Remark 48. In fact, it follows from (227) that for every scalar valued function f : {−1, 1} → C and p ∈ [2, ∞], we have the inequality

−t∆ 1 n n ∀ t > 0, k∇e fkLp({−1,1} ; ) 6 √ kfkLp({−1,1} ; ). (229) C e2t − 1 C Combined with Lemma 44, these estimates provide the sharp dimension free decay of the operator −t∆ n norm of ∇e from Lp({−1, 1} ; C) to itself for every p ∈ (1, ∞].

n Proof of Proposition 46. Fix n ∈ N, d ∈ {1, . . . , n} and let f : {−1, 1} → C be a function of degree t td2 at most d. Then, by Lemma 47, Theorem1 and the estimate Td(e ) 6 e , we have td2 1 t∆ e n n n k∇fkL∞({−1,1} ; ) 6 √ ke fkL∞({−1,1} ; ) 6 √ kfkL∞({−1,1} ; ). (230) C e2t − 1 C e2t − 1 C √ √ 1 2t Plugging t = d2 and using that e − 1 > 2t, we thus get ed n n n k∇fkL∞({−1,1} ; ) 6 √ kfkL∞({−1,1} ; ) < 2dkfkL∞({−1,1} ; ). (231) C 2 C C n To prove that a O(d) factor is necessary, consider the function f : {−1, 1} → C given by ε + ··· + ε ∀ ε = (ε , . . . , ε ) ∈ {−1, 1}n, f(ε) def= T 1 n , (232) 1 n d n where Td(x) is the d-th Chebyshev polynomial of the ﬁrst kind. Clearly f has degree at most d and furthermore ε1 + ··· + εn kfkL ({−1,1}n; ) = max Td = Td(1) = 1. (233) ∞ C ε∈{−1,1}n n On the other hand, n 1/2 X 2 n k∇fkL∞({−1,1} ;C) > (∂if(1,..., 1)) (234) i=1 and for every i ∈ {1, . . . , n}, 1 2 ∂ f(1,..., 1) = 1 − T 1 − . (235) i 2 d n Therefore, √ n 2 k∇fk n 1 − T 1 − , (236) L∞({−1,1} ;C) > 2 d n which for n = d2 becomes d 2 k∇fk n 1 − T 1 − . (237) L∞({−1,1} ;C) > 2 d d2 Finally notice that 2 2 1 − T 1 − = 1 − cos d arccos 1 − 1 (238) d d2 d2 > for large enough values of d, since 2 lim 1 − cos d arccos 1 − = 1 − cos(2). (239) d→∞ d2 The asymptotic optimality of (219) follows from (237) and (238).

31 Remark 49. Using Theorem 13, one can also derive an endpoint Bernstein–Markov inequality for p = 1. Indeed, using Jensen’s inequality and orthogonality, for every n, d ∈ N with d ∈ {1, . . . , n}, n every function f : {−1, 1} → C satisﬁes √ n n n k∇fkL1({−1,1} ;C) 6 k∇fkL2({−1,1} ;C) 6 dkfkL2({−1,1} ;C) (240) and invoking Theorem 13 we conclude that √ d n n k∇fkL1({−1,1} ;C) 6 (2.69076) dkfkL1({−1,1} ;C). (241) It would be interesting if one could obtain a polynomial bound in this inequality. Remark 50. In our recent paper [EI18], we investigated Bernstein–Markov inequalities in the spirit n of Theorem 15 for polynomials on R equipped with the Gaussian measure. The main tools we used were Epperson’s Gaussian complex hypercontractivity [Epp89] (see also [Jan97]) and Meyer’s Gaussian Riesz transform inequalities [Mey84]. Even though in [EI18] we did not consider vector valued inequalities, we note that all the results of Section3 also hold true in the Gaussian setting by the holomorphicity of the Ornstein–Uhlenbeck semigroup e−tL on K-convex targets which follows from Pisier’s K-convexity theorem [Pis82]. Moreover, it has been shown by Pisier [Pis88] that for every UMD Banach space (X, k · kX ) and p ∈ (1, ∞) there exist c(p, X),C(p, X) ∈ (0, ∞) such that n for every n ∈ N and every function f : R → X, we have n 1/p 1/2 X n n c(p, X)kL fkLp((R ,γn);X) 6 yi∂if Lp((R ,γn);X) dγn(y) ˆ n R i=1 (242) 1/2 n 6 C(p, X)kL fkLp((R ,γn);X). Therefore, since UMD spaces have non-trivial type (e.g. by [Pis73]), we then conclude that for every p ∈ (1, ∞) there exist β = β(p, X) ∈ [0, 1) and K = K(p, X) ∈ (0, ∞) such that if P is a n polynomial on R of degree at most d, then n 1/p def X β n n k∇P kLp((R ,γn);X) = yi∂if dγn(y) 6 Kd kP kLp((R ,γn);X). (243) ˆ n L (( n,γ );X) R i=1 p R n Such a result for UMD-valued functions on the Hamming cube is unknown. Inequality (243) is a high dimensional vector valued analogue of Freud’s inequality [Fre71], who showed (243) for X = C 1 and n = 1 with the optimal exponent β = 2 for every p ∈ [1, ∞). To the best of our knowledge no vector valued versions of Freud’s inequality (even for n = 1) were available in the literature. We will now prove Corollary 16, the improved estimate on the inﬂuences of bounded functions. We will need the following lemma, whose proof relies on an inequality of Sarantopoulos [Sar91]. n Lemma 51. Let n ∈ N, d ∈ {1, . . . , n} and p ∈ [1, ∞). Then, for every function f : {−1, 1} → C of degree at most d, we have p ! kgk n (p) Lp({−1,1} ;C) d p Inf f inf sup √ kfk n . (244) 6 −t∆ p 2t L∞({−1,1} ;C) t>0 ke gk n e − 1 deg(g)6d Lp({−1,1} ;C) Proof. For t > 0, denote by p kgk n def Lp({−1,1} ;C) Cd(t) = sup −t∆ p (245) ke gk n deg(g)6d Lp({−1,1} ;C) and notice that n n (p) X p X −t∆ p Inf f = k∂ifk n Cd(t) k∂ie fk n . (246) Lp({−1,1} ;C) 6 Lp({−1,1} ;C) i=1 i=1

32 Moreover, we have n n X −t∆ p X −t∆ −t∆ p−1 k∂ie fk n k∂ie fk n ke fk n Lp({−1,1} ;C) 6 L1({−1,1} ;C) L∞({−1,1} ;C) i=1 i=1 (247) p−1 (1) −t∆ kfk n Inf (e f), 6 L∞({−1,1} ;C) where the last inequality follows from the contractivity of the heat semigroup. Therefore, it suﬃces to show that (1) −t∆ d n Inf (e f) 6 √ kfkL∞({−1,1} ; ). (248) e2t − 1 C n Let F (x1, . . . , xn) be the multilinear polynomial on R given by n def X Y ∀ (x1, . . . , xn) ∈ R ,F (x1, . . . , xn) = fb(A) xi. (249) A⊆{1,...,n} i∈A n −t∆ −t −t∆ Notice that for ε ∈ {−1, 1} , we have e f(ε) = F (e ε) and for i ∈ {1, . . . , n}, |∂ie f(ε)| = −t e |∂iF (ε)|, where ∂iF is the usual partial derivative of F . Consequently, n n (1) −t∆ 1 X X −t∆ −t X −t Inf (e f) = |∂ie f(ε)| 6 e max |∂iF (e ε)| 2n ε∈{−1,1}n ε∈{−1,1}n i=1 i=1 (250) n n −t X −t −t X = e max δi∂iF (e ε) = e max yi∂iF (x), ε,δ∈{−1,1}n kxk e−t,kyk 1 i=1 ∞6 ∞6 i=1 Pn where in the last equality we used the multilinearity of the polynomial i=1 yi∂iF (x) in the vari- 2n ables (x, y) ∈ R . By Sarantopoulos’ vector valued Bernstein inequality from [Sar91], we deduce n X d d n max yi∂iF (x) 6 √ max |F (x)| 6 √ kfkL∞({−1,1} ; ), (251) kxk e−t,kyk 1 −2t kxk 1 −2t C ∞6 ∞6 i=1 1 − e ∞6 1 − e again by the multilinearity of F . Therefore, (250) and (251) imply (248), which combined with (246) and (247) completes the proof of the lemma. Equipped with Lemma 51, we can prove Corollary 16. 4 n Proof of Corollary 16. Fix n ∈ N, d ∈ {1, . . . , n}, p ∈ 1, 3 and let f : {−1, 1} → C be a function of degree at most d. Denote by def π ηp = √ . (252) 2 p−1 2π − 2 arcsin p

It follows from (23) and (138) that there exists a constant Cp ∈ (0, ∞) such that for every function n g : {−1, 1} → C be a function of degree at most d, ηp −t∆ −Cpt d n n ∀ 0 6 t 6 1, ke gkLp({−1,1} ;C) > e kgkLp({−1,1} ;C). (253) Therefore, by Lemma 51, we conclude that ηp pCpt d (p) e d p Inf f 6 inf √ · kfk n . (254) t>0 e2t − 1 L∞({−1,1} ;C) √ √ −1/η Choosing t = d p and using that e2t − 1 > 2t, we deduce that there exists a constant Kp ∈ (0, ∞) such that √ 1+ 1 1 2 p−1 (p) 2ηp p 2− π arcsin p p Inf f Kpd kfk n = Kpd kfk n , (255) 6 L∞({−1,1} ;C) L∞({−1,1} ;C) which concludes the proof.

33 Finally, we present the proof of Theorem 17. n Proof of Theorem 17. Fix n, d, m ∈ N with d + m 6 n and let f : {−1, 1} → C be a function of degree at most d + m which is also in the d-th tail space. Then, by Lust-Piquard’s lower Riesz transform inequalities [LP98], for every p ∈ (1, ∞) there exists cp ∈ (0, ∞) such that (200) 1/2 n n k∇fkLp({−1,1} ;C) > cpk∆ fkLp({−1,1} ;C). (256) The conclusion of the theorem now follows from (159).

References [AA14] S. Aaronson and A. Ambainis. The need for structure in quantum speedups. Theory Comput., 10:133–166, 2014. [BB14] A. Baˇckursand M. Bavarian. On the sum of L1 influences. In IEEE 29th Conference on Computational Complexity—CCC 2014, pages 132–143. IEEE Computer Soc., Los Alamitos, CA, 2014. [BE95] P. Borwein and T. Erdélyi. Polynomials and polynomial inequalities, volume 161 of Graduate Texts in Mathematics. Springer-Verlag, New York, 1995. [Bec75] W. Beckner. Inequalities in Fourier analysis. Ann. of Math. (2), 102(1):159–182, 1975. [BELP08] L. Ben Efraim and F. Lust-Piquard. Poincarétype inequalities on the discrete cube and in the CAR algebra. Probab. Theory Related Fields, 141(3-4):569–602, 2008. [BGL14] D. Bakry, I. Gentil, and M. Ledoux. Analysis and geometry of Markov diffusion operators, volume 348 of Grundlehren der Mathematischen Wissenschaften. Springer, Cham, 2014. [Bon70] A. Bonami. Etude´ des coefficients de Fourier des fonctions de Lp(G). Ann. Inst. Fourier (Grenoble), 20(fasc. 2):335–402 (1971), 1970. [Bor79] C. Borell. On the integrability of Banach space valued Walsh polynomials. In Séminaire de Probabilités, XIII, volume 721 of Lecture Notes in Math., pages 1–3. Springer, Berlin, 1979. [Bor84] C. Borell. On polynomial chaos and integrability. Probab. Math. Statist., 3(2):191–203, 1984. [Bou80] J. Bourgain. Walsh subspaces of Lp-product spaces. In Seminar on Functional Analysis, 1979–1980 (French), pages Exp. No. 4A, 9. Ecole´ Polytech., Palaiseau, 1980. [EI18] A. Eskenazis and P. Ivanisvili. Dimension independent Bernstein–Markov inequalities in Gauss space. To appear in J. Approx. Theory. Preprint available at https://arxiv.org/abs/1808.01273, 2018. [EN20] A. Eskenazis and A. Naor. Discrete Littlewood–Paley–Stein theory and Pisier’s inequality for superreflexive targets. Preprint, 2020. [Epp89] J. B. Epperson. The hypercontractive approach to exactly bounding an operator with complex Gaussian kernel. J. Funct. Anal., 87(1):1–30, 1989. [Erd20] T. Erdélyi.Reverse Markov- and Bernstein-type inequalities for incomplete polynomials. J. Approx. The- ory, 251:105341, 2020. [FHH+14] Y. Filmus, H. Hatami, S. Heilman, E. Mossel, R. O’Donnell, S. Sachdeva, A. Wan, and K. Wimmer. Real Analysis in Computer Science: a collection of open problems. Preprint available at https://simons. berkeley.edu/sites/default/files/openprobsmerged.pdf, 2014. [FHKL16] Y. Filmus, H. Hatami, N. Keller, and N. Lifshitz. On the sum of the L1 influences of bounded functions. Israel J. Math., 214(1):167–192, 2016. [Fre71] G. Freud. A certain inequality of Markov type. Dokl. Akad. Nauk SSSR, 197:790–793, 1971. [HMO17] S. Heilman, E. Mossel, and K. Oleszkiewicz. Strong contraction and influences in tail spaces. Trans. Amer. Math. Soc., 369(7):4843–4863, 2017. [HN13] T. Hytönenand A. Naor. Pisier’s inequality revisited. Studia Math., 215(3):221–235, 2013. [IN19] P. Ivanisvili and F. Nazarov. On Weissler’s conjecture on the Hamming cube I. Preprint available at https://arxiv.org/abs/1907.11359, 2019. [IT19] P. Ivanisvili and T. Tkocz. Comparison of moments of Rademacher chaoses. Ark. Mat., 57(1):121–128, 2019. [Jan97] S. Janson. On complex hypercontractivity. J. Funct. Anal., 151(1):270–280, 1997. P [Kah64] J.-P. Kahane. Sur les sommes vectorielles ±un. C. R. Acad. Sci. Paris, 259:2577–2580, 1964. [Khi23] A. Khintchine. Uber¨ dyadische Brüche. Math. Z., 18(1):109–116, 1923. [Kra06] S. G. Krantz. Geometric function theory. Cornerstones. BirkhäuserBoston, Inc., Boston, MA, 2006. Ex- plorations in complex analysis. [KW92] S. Kwapieńand W. A. Woyczyński. Random series and stochastic integrals: single and multiple. Probability and its Applications. BirkhäuserBoston, Inc., Boston, MA, 1992.

34 [Kwa76] S. Kwapień.A theorem on the Rademacher series with vector valued coefficients. In Probability in Banach spaces, pages 157–158. Lecture Notes in Math., Vol. 526. Springer, Berlin, 1976. [LP98] F. Lust-Piquard. Riesz transforms associated with the number operator on the Walsh system and the fermions. J. Funct. Anal., 155(1):263–285, 1998. [Mau03] B. Maurey. Type, cotype and K-convexity. In Handbook of the geometry of Banach spaces, Vol. 2, pages 1299–1332. North-Holland, Amsterdam, 2003. [Mey84] P.-A. Meyer. Transformations de Riesz pour les lois gaussiennes. In Seminar on probability, XVIII, volume 1059 of Lecture Notes in Math., pages 179–193. Springer, Berlin, 1984. [MN14] M. Mendel and A. Naor. Nonlinear spectral calculus and super-expanders. Publ. Math. Inst. Hautes Etudes´ Sci., 119:1–95, 2014. [MP76] B. Maurey and G. Pisier. Séries de variables aléatoires vectorielles indépendantes et propriétés géométriquesdes espaces de Banach. Studia Math., 58(1):45–90, 1976. [MS86] V. D. Milman and G. Schechtman. Asymptotic theory of finite-dimensional normed spaces, volume 1200 of Lecture Notes in Mathematics. Springer-Verlag, Berlin, 1986. With an appendix by M. Gromov. [Nao12] A. Naor. An introduction to the Ribe program. Jpn. J. Math., 7(2):167–233, 2012. [NS02] A. Naor and G. Schechtman. Remarks on non linear type and Pisier’s inequality. J. Reine Angew. Math., 552:213–236, 2002. [O’D14] R. O’Donnell. Analysis of Boolean functions. Cambridge University Press, New York, 2014. [Ole03] K. Oleszkiewicz. On a nonsymmetric version of the Khinchine-Kahane inequality. In Stochastic inequalities and applications, volume 56 of Progr. Probab., pages 157–168. Birkhäuser,Basel, 2003. 1 [Pis73] G. Pisier. Sur les espaces de Banach qui ne contiennent pas uniformément de ln. C. R. Acad. Sci. Paris Sér.A-B, 277:A991–A994, 1973. [Pis78] G. Pisier. Les inégalitésde Khintchine-Kahane, d’aprèsC. Borell. In Séminaire sur la Géométriedes Espaces de Banach (1977–1978), pages Exp. No. 7, 14. Ecole´ Polytech., Palaiseau, 1978. [Pis82] G. Pisier. Holomorphic semigroups and the geometry of Banach spaces. Ann. of Math. (2), 115(2):375–392, 1982. [Pis86] G. Pisier. Probabilistic methods in the geometry of Banach spaces. In Probability and analysis (Varenna, 1985), volume 1206 of Lecture Notes in Math., pages 167–241. Springer, Berlin, 1986. [Pis88] G. Pisier. Riesz transforms: a simpler analytic proof of P.-A. Meyer’s inequality. In Séminaire de Proba- bilités,XXII, volume 1321 of Lecture Notes in Math., pages 485–501. Springer, Berlin, 1988. [Pis07] G. Pisier. A remark on hypercontractive semigroups and operator ideals. Preprint available at https: //arxiv.org/abs/0708.3423, 2007. [Sar91] Y. Sarantopoulos. Bounds on the derivatives of polynomials on Banach spaces. Math. Proc. Cambridge Philos. Soc., 110(2):307–312, 1991. [Ste70] E. M. Stein. Topics in harmonic analysis related to the Littlewood-Paley theory. Annals of Mathematics Studies, No. 63. Princeton University Press, Princeton, N.J.; University of Tokyo Press, Tokyo, 1970. [Sze25] G. Szegö. Uber¨ einen Satz von A. Markoff. Math. Z., 23(1):45–61, 1925. [Tal93] M. Talagrand. Isoperimetry, logarithmic Sobolev inequalities on the discrete cube, and Margulis’ graph connectivity theorem. Geom. Funct. Anal., 3(3):295–314, 1993. [Wag00] R. Wagner. Notes on an inequality by Pisier for functions on the discrete cube. In Geometric aspects of functional analysis, volume 1745 of Lecture Notes in Math., pages 263–268. Springer, Berlin, 2000. [Wei79] F. B. Weissler. Two-point inequalities, the Hermite semigroup, and the Gauss-Weierstrass semigroup. J. Funct. Anal., 32(1):102–121, 1979. [Wol07] P. Wolff. Hypercontractivity of simple random variables. Studia Math., 180(3):219–236, 2007.

(A. E.) Department of Mathematics, Princeton University, Princeton, NJ 08544-1000, USA E-mail address: [email protected]

(P. I.) Department of Mathematics, University of California, Irvine, CA 92617, USA E-mail address: [email protected]