5
$\begingroup$

[Edit: added details on the Reed-Muller codes]

Are there explicit (non random) constructions of probability measures on $D_N = \{0,1\}^N$ with support of size $O(N)$ and with all nontrivial Fourier coefficients less than $\frac 1 2$ is absolute value ?

I would expect that such questions are very well understood, but being from a different background I probably lack the common knowledge. Pointers to the relevent litterature are very much welcome.

Below are some explanations on the terminology and on the $O(N)$. I am happy with any answer with $\frac 1 2$ replaced by $c<1$ (but independent from $N$).


When I write that $\mu$ has all nontrivial Fourier coefficients less than $\frac 1 2$ in absolute value, I mean that $|\hat \mu(A)| \leq \frac 1 2$ for every nonempty subset $A$ of $\{1,\dots,N\}$, where $\hat \mu(A)$ is the Fourier (or Fourier-Walsh) coefficient: $$ \hat \mu(A) = \int (-1)^{\sum_{i \in A} \varepsilon_i} d\mu(\varepsilon).$$ The trivial Fourier coefficient, corresponding to $A=\emptyset$, is always $1$. The uniform probability measure on $D_N$ has all non-trivial Fourier coefficients equal to $0$, but has full support of size $2^N$. So I am looking for a probability measure that looks like the uniform probability on the Fourier side, but is very far on the other side, having very small support.


Why the $O(N)$? Because it is the optimal rate for the mere existence of $\mu$.

Indeed, seeing $D_N$ as a vector space of dimension $N$ over the field with two elements, the question can be translated in linear algebra terms. In particular, we see that for all nontrivial Fourier coefficients of $\mu$ to be $\neq 1$, we need that the support of $\mu$ contains at least a basis of $D_N$ and therefore has to have cardinality at least $N$.

Conversely, taking for $\mu$ the uniform measure on a subset of cardinality $K N$ chosen uniformly at random, a basic application of union bound and large deviation (Hoeffding's inequality) arguments gives that $\mu$ works with high probability, provided that $K$ is large enough. What I am looking for is explicit non-random constructions.


Using Reed-Muller codes, a back-of-the enveloppe-computation seems to provide an explicit $\mu$ with support $N^{K \log \log N}$.

[Added after request in comments] Here is the construction. I do not know how standard it is ; I learnt it from the MIP*=RE paper. There is an integer parameter $a$, and we using the code coming from multilinear functions over the field $\mathbf{F}_q$ with $q=2^{a}$ elements in $m = 2^{a-1}$ variables.

Denote $\mathcal{P}(m)$ the set of all subsets of $\{1,\dots,m\}$, and define a map $\alpha:\mathbf{F}_q^{m+1} \to \mathbf{F}_q^{\mathcal{P}(m)}$ by setting $$\alpha(u_1,\dots,u_m,a) = (a \prod_{i \in A} u_i \prod_{i \notin A} (1-u_i))_{A \in \mathcal{P}(m)}.$$ We can identify $\mathbf{F}_q^{\mathcal{P}(m)}$ as an additive group with $D_N$ for $N=a 2^m = a 2^{2^{a-1}}$, and I claim that $\mu$, the image by $\alpha$ of the uniform measure on $\mathbf{F}_q^{m+1}$ does the job.

$\mu$ clearly has support of size $q^{m+1} = 2^{a +a2^{a-1}} = 2^{O(\log N \log\log N)}$, so we are fine as far as the support is concerned.

Let me prove that $|\hat\mu(\chi)|\leq \frac 1 2$ for every nontrivial character $\chi$. Let us fix a nonzero character $\eta$ of $\mathbf{F_q}$. Then by counting, every character of $\mathbf{F_q}^{\mathcal P(m)}$ is of the form $\eta \circ T$ for a $\mathbf{F}_q$-linear map $T:\mathbf{F}_q^{\mathcal{P}(m)} \to \mathbf{F}_q$. Moreover, for every $x \in \mathbf{F}_q$, we have $\mathbf{E}_{a \in \mathbf{F_q}} \eta(ax)=1_{x=0}$, so we can compute

\begin{align*}\hat \mu(\chi) &= \mathbf{E}_{u \in \mathbf{F}_q^m} \mathbf{E}_{a \in \mathbf{F}_q} \eta(a T(\alpha(u,1)))\\ &= \mathbf{E}_{u \in \mathbf{F}_q^m} 1_{T(\alpha(u,1))}=0\\ & = \mathbf{P}_{u \in\mathbf{F}_q^m}( T(\alpha(u,1))). \end{align*}

But if $T$ is a nonzero linear map, then $T(\alpha(u,1))$ is a nonzero polynomial in $m$ variables of individual degree $\leq 1$, so the last probability is $\leq\frac{m}{q} = \frac 1 2$. QED

$\endgroup$
5
  • $\begingroup$ What is the order of the Reed-Muller codes that you used for that estimate. Can you include details? $\endgroup$ Commented Jan 24, 2022 at 15:31
  • 1
    $\begingroup$ I think what you're looking for is an $\epsilon$-biased space (where in your case, $\epsilon=1/2$) (i.e. the expectation of any nontrivial parity function with respect to the distribution is at most $\epsilon$ in absolute value). To my knowledge, the best current construction is by Ta-Shma (Theorem 1 of eccc.weizmann.ac.il/report/2017/041) and gives a set of size $O(N/\epsilon^{2+o(1)})$ which is near-optimal, so should be what you need. The references there might give earlier and simpler constructions giving $O(N)$ for $\epsilon = 1/2$ from better codes. $\endgroup$ Commented Jan 24, 2022 at 19:54
  • 1
    $\begingroup$ @J.G you are correct, thanks a lot, this is precisely what I was looking for. If you make your comment into an answer, I will accept it. $\endgroup$ Commented Jan 24, 2022 at 22:22
  • $\begingroup$ @MikaeldelaSalle just made it into an answer. Very glad to have helped! $\endgroup$ Commented Jan 25, 2022 at 15:57
  • $\begingroup$ @J.G I would like to acknowledge your help in a paper that I am finishing to write. I can simply thank "MO user J.G", but if you agree to disclose your identity, I can thank you with your real-life identity. You can write me an email if you wish. $\endgroup$ Commented Mar 31, 2022 at 21:30

1 Answer 1

7
$\begingroup$

Just to close the lid on my comment: the desired construction is known as an $\epsilon$-biased space, where here $\epsilon = 1/2$. The best known construction seems to be by Ta-Shma from 2017 (https://eccc.weizmann.ac.il/report/2017/041/), which yields a subset with support size $O(N/\epsilon^{2+o(1)})$. This is near-optimal, as it is known that $\Omega(N/\epsilon^2\log(1/\epsilon))$ is necessary. However, for the special case of $\epsilon = 1/2$ where the precise $\epsilon$-dependence is not relevant, it appears that earlier constructions based on well-known codes suffice to obtain $O(N)$.

$\endgroup$

You must log in to answer this question.

Start asking to get answers

Find the answer to your question by asking.

Ask question

Explore related questions

See similar questions with these tags.