7

This question is inspired by the recent one about Gaussian measures over the reals:

What makes Gaussian distributions special?

I would be interested in a similar list of characterizations for the probability measure, on the field of $p$-adic numbers $\mathbb{Q}_p$, whose density with respect to the standard additive Haar measure is given by the indicator function of the ring of $p$-adic integers $\mathbb{Z}_p$.

  • 2
    Probably worth linking to this paper: https://arxiv.org/abs/math/9803046 – Sam Hopkins Feb 03 '21 at 04:10
  • Thanks! If you have time it would be good if you could extract one of the characterizations in the paper and post that as an answer. – Abdelmalek Abdesselam Feb 03 '21 at 20:27
  • Have you checked the refs in https://mathoverflow.net/questions/339214/applications-of-number-theory-in-dynamical-systems/339226#339226 ? – Tom Copeland Feb 28 '21 at 03:58
  • Another question: is an appropriate $p$-adic analogue of the Gaussian still real-valued, or should one look for a $p$-adic-valued analogue? – LSpice Mar 03 '21 at 04:07
  • 2
    I would say that it ought to be $\Bbb{C}$-valued owing to comments in my post below. – Jeanne Scott Mar 03 '21 at 04:19
  • @InesInstitoris, it's actually your post (which I enjoyed—thank you!) that made me think of it. Your post comes from the same point of view as my training, the complex represetation theory of finite and $p$-adic groups; but there is also a rich (if newer) theory of $p$-adic represetations of $p$-adic groups, and one might look there to find an appropriate description of $p$-adic valued special functions. (I don't know if that community already has such a theory, or even if they have a Weil representation.) – LSpice Mar 03 '21 at 04:25
  • @LSpice: I know that there are very interesting things going on when looking a p-adic valued stuff (e.g., yesterday's article by Fargues and Scholze which looks quite exciting to me). However, I would like this question to stay within the realm of Archimedian-valued things. What you said though could be the premise of another MO question which I think would be quite nice. I already took inspiration from an MO question to write this one. You could do the same with my question. – Abdelmalek Abdesselam Mar 03 '21 at 14:52
  • 1

1 Answers1

4

Perhaps one answer is the appearance of the Gaussian in the oscillator (a.k.a. Weil) representation --- which makes sense even in non-zero characteristic. You're probably aware of the construction of this infinite dimensional representation for the double cover of $\mathrm{SL}_2(\Bbb{R})$. The story for $\mathrm{SL}_2 \big( \Bbb{F}_q \big)$ where $q$ a power of a prime strongly parallels the continuous story except, of course, it's finite:

Choose any non-square $\delta \in \Bbb{F}_q$ and form the quadratic extension $\Bbb{F}_q(\delta) \cong \Bbb{F}_q^2$. As usual, we identify elements of $\Bbb{F}_q(\delta)$ as linear combinations $z = a + \sqrt{\delta} b$ with $a, b \in \Bbb{F}_q$ subject the usual formulae for addition and multiplication; conjugation and norm are expressed as $\bar{z} = a - \sqrt{\delta} b$ and $\mathrm{N}(z) = a^2 - \delta b^2$ respectively. We'll need to choose (any) non-trivial additive character $\psi: \Bbb{F}_q \longrightarrow \Bbb{C}^*$ together with any multiplicative character $\chi: \Bbb{F}_q(\delta)^* \longrightarrow \Bbb{C}^*$ which generates the character group of $\Bbb{F}_q(\delta)^*$. Morally $\psi$ plays the role of the exponential function $\exp: \Bbb{R} \longrightarrow \Bbb{R}^*$ in this finite context. Now set

\begin{equation} W_\chi := \ \left\{ \begin{array}{l} \displaystyle \text{all functions} \ f: \Bbb{F}_q(\delta) \longrightarrow \Bbb{C} \ \ \text{such that} \\ \displaystyle f(wz) = \overline{\chi(w)} \, f(z) \ \text{whenever $\mathrm{N}(w)=1$} \end{array} \right\} \end{equation}

which is $q-1$ dimensional. The oscillator representation $\varrho_\chi: \mathrm{SL}_2(\Bbb{F}_q) \longrightarrow \mathrm{GL}(W_\chi)$ is determined by the action of the follow elements, which generate $\mathrm{SL}_2(\Bbb{F}_q)$:

\begin{equation} \begin{array}{ll} \displaystyle \varrho_\chi { \scriptstyle \begin{pmatrix} x & 0 \\ 0 & x^{-1} \end{pmatrix}} f(z) &\displaystyle = \ f(xz) \\ \displaystyle \displaystyle \varrho_\chi { \scriptstyle \begin{pmatrix} 1 & \ \ y \\ 0 & \ \ 1 \end{pmatrix}} f(z) &\displaystyle = \ \underbrace{\psi \big( y \, \mathrm{N}(z) \big)}_{\text{the Gaussian $G_y(z)$}} \cdot f(z) \\ \displaystyle \displaystyle \varrho_\chi { \scriptstyle \begin{pmatrix} 0 & 1 \\ -1 & 0 \end{pmatrix}} f(z) &\displaystyle = \ -q \widehat{f}(-\bar{z}) \end{array} \end{equation}

where $\widehat{f}$ is the Fourier transform of $f$ with respect to the additive structure of $\Bbb{F}_q(\sqrt{\delta})$. Like the continuous Gaussian, $G_y(z)$ is an eigenfunction of this Fourier transform.

As far as I know, the construction that I've outlined (and which I learned from Amritanshu Prasad's online notes) can be carried out for $\mathrm{SL}_2\big( \Bbb{Z}/q\Bbb{Z} \big)$ where $q= p^N$ is still a power of a fixed prime $p$. Furthermore a coherent choice of additive and multiplicative characters $\psi_N$ and $\chi_N$ can be made for each $N \geq 1$ so that these oscillator representations agree with the inverse system

\begin{equation} 1 \longleftarrow \Bbb{Z}/p\Bbb{Z} \stackrel{\ \text{mod $p$}}{\longleftarrow} \Bbb{Z}/p^2\Bbb{Z} \stackrel{\ \text{mod $p^2$}}{\longleftarrow} \Bbb{Z}/p^3\Bbb{Z} \longleftarrow \, \cdots \end{equation}

thus allowing the oscillator representations $\varrho_{\chi_N}$ to be exported to the $p$-adic integers $\Bbb{Z}_p$, which will inherent some kind of Gaussian-like function $\varprojlim G_{\bf y}$ where ${\bf y}= (y_1,y_2,y_3, \dots)$ and $y_N = y_{N+1} \, \text{mod $p^N$}$ for each $N \geq 1$.

Jeanne Scott
  • 1,847
  • 1
    I would follow this up with the observation that Young's convolution inequality $| f * g|r \leq c{p,q} | f |_p , | g |_q$ with ${1 \over p} + {1 \over q} = 1 + {1 \over r}$ is maximized when $f$ and $g$ are Gaussians. One might ask whether or not the same characterization holds for the $\Bbb{F}_q$-Gaussians $G_y$. – Jeanne Scott Mar 03 '21 at 04:06
  • 1
    In a similar vein, the Babenko-Beckner inequality $| \widehat{f} |q \leq \tilde{c}{p,q} | f |_p$ with ${1 \over p} + {1 \over q} = 1$ and $1 < p \leq q$ is maximized by Gaussians. Does this characterize $\Bbb{F}_q$-Gaussians ? Is there some analogue at least ? – Jeanne Scott Mar 03 '21 at 04:33
  • 1
    Thank you for a very nice answer! I remember reading a long time ago about Segal-Shale-Weil and the metaplectic representation, Mehler's formula etc. Do you know if one gets the Fourier transform as a particular value of the corresponding unitary representation, like at $t=\frac{\pi}{4}$ if I remember correctly for the Oscillator? If so are the eigenfunctions determined, like Hermite functions in the Archimedean case? – Abdelmalek Abdesselam Mar 03 '21 at 14:58
  • 1
    I also hope someone will write an answer along the line of your comments, namely some charaterization of the indicator function of $\mathbb{Z}_p$ as some kind of maximizer/minimizer. – Abdelmalek Abdesselam Mar 03 '21 at 14:59
  • 1
    If $\delta = -1$ is a non-square in $\Bbb{F}_q$ then the Fourier transform is indeed a special value of the action the subgroup $\mathrm{SO}_2(\Bbb{F}_q)$ consisting of all $2 \times 2$ matrices of the form $\begin{pmatrix} a & b \ \delta b & a \end{pmatrix}$ with $a^2- \delta b^2 =1$. As far as Hermite functions are concerned, wouldn't you need a good candidate for a laplace operator $\Delta: \Bbb{C} \big[\Bbb{F}_q (\sqrt{ \delta }) \big] \longrightarrow \Bbb{C} \big[\Bbb{F}_q (\sqrt{ \delta }) \big]$ as well as an appropriate $\Bbb{C}$-valued potential function to play the role of $|z|^2$ ? – Jeanne Scott Mar 03 '21 at 16:33
  • Actually there is already a candidate for $\Delta$. The action of the associated heat-propagator $\exp(- t\Delta)$ on $\Bbb{C} \big[ \Bbb{F}q (\sqrt{\delta}) \big]$ should be implemented by convolution with the $\Bbb{F}_q$-Gaussian, which is how the operator $\varrho\chi \begin{pmatrix}1 & 0 \ t & 1 \end{pmatrix}$ acts on functions $f \in \Bbb{C} \big[ \Bbb{F}_q (\sqrt{\delta}) \big]$. – Jeanne Scott Mar 03 '21 at 18:42
  • ... or more simply put: The potential function ought to be something like $v(z) := \log G_{, ! 1}(z)$ while the Laplacian $\Delta$ should be defined by $\Delta f (z) := \sum_{w \in \Bbb{F}_q( \sqrt{\delta})} , \widehat{v}(z-w) f(w)$ for functions $f \in \Bbb{C}\big[ \Bbb{F}_q( \sqrt{\delta}) \big]$. Now we can try to identify Hermite functions by fishing for eigenfunctions of the $f \mapsto \Delta f - vf$ operator. – Jeanne Scott Mar 03 '21 at 19:49
  • I don't know if trying to mimic the real case too closely is good, namely, defining $H$ as some schrodinger operator before taking $e^{-tH}$ with $t$ real. I think time is p-adic here. Rather I would look for simultaneous eigenfunctions for the representation operators $U(M)$ where the $M$'s are some of your $2\times 2$ matrices. – Abdelmalek Abdesselam Mar 03 '21 at 20:09