10
$\begingroup$

Let $H = [-1, 1]^d$ be the $d$-dimensional hypercube, and let $\mu \in \text{int} H$.

Under these conditions, I can explicitly construct a tractable probability measure $P$, supported on on $H$, which has $\mu$ as its mean. For my purposes, tractability means:

(a) $P$ has a density, which can be written down and cheaply evaluated.

(b) $P$ can be exactly sampled from, efficiently.

This is a relatively simple task, since the problem basically decouples across dimensions; if you can find a probability measure $P_i$ on $[-1,1]$ with mean $\mu_i$, then setting $P = P_1 (x_1) \cdots P_d (x_d)$ suffices.

I've been using the family of random variables given by $P(x) \propto \exp( \langle \theta, x \rangle)$ to do this; the mean of each coordinate is then given by $m_i(\theta) = \coth \theta_i - \theta_i^{-1}$ which is surjective onto $(-1,1)$ and so one can always find a $\theta$ which gives you the mean you want.

Now, suppose I want to extend this, and construct a probability measure on $H$ which also has the covariance matrix I want, i.e. let $\Sigma$ be a general $d \times d$ positive semi-definite matrix. Then, I want all the above, plus:

\begin{align} \textbf{Var}_{X \sim P} [X] &= \Sigma \end{align}

This is more or less possible when $\Sigma$ is diagonal, but otherwise I haven't figured out how to do it. Of course, in the general case, I will need $\Sigma$ to be sufficiently small, as I can't have arbitrarily high variance when constrained to a cube,

I've considered sampling from a truncated Gaussian, i.e. sampling from $\mathcal{N}(\mu, \Sigma)$ until you land in $H$. However, this won't generally have a tractable density, due to the normalising constant. It also won't quite have the right moments, though I can basically forgive this; I ultimately don't mind too much if I only have

\begin{align} \mathbf{E}_{X \sim P} [X] &\approx \mu \\ \textbf{Var}_{X \sim P} [X] &\approx \Sigma \end{align}

with some control of the bias. I should perhaps highlight that it's not sufficient for my purposes to approximate $\Sigma$ by a diagonal matrix.

So, the concise version of my question is: is there a way of explicitly constructing probability measures on the hypercube which have the mean and covariance matrix which I want it to have?

There needn't be a 'canonical' / optimal' answer (whatever that might mean here); something functional will entirely suffice. It's also preferable to have a construction where it's easy to adapt things to sample from a range of different $(\mu, \Sigma)$.

$\endgroup$
1
  • 2
    $\begingroup$ Interesting question. I just want to mention that the truncated Gaussian (and truncated exponential distribution for matching means only) is in fact the usual canonical/optimal/theoretical answer: the maximal entropy distribution. $\endgroup$
    – Dap
    Mar 13, 2019 at 20:00

1 Answer 1

1
$\begingroup$

If only mean and covariance are important then using localized distributions is probably the easiest solution. By placing Dirac distributions of various weight on gridpoints of with coordinates in $\{-1/2,0,1/2\}$ you can obtain variety of covariance matricies still keeping mean equal to $0$. Then translate by a constant vector to get desired mean. Sadly this constrains the range of means that you can pick but maybe that's good enough.

If you replace Diracs by symmetric Gaussians with small enough sigma s.t. truncation error is small enough you can obtain cts pdf with approximate values of mean and covariance.

Finding weights on the Gaussians given covariance requires solving a system of equations size dependent on dimension but this has to be done once. Then sampling should be really quick (as it's almost sampling from a discrete distribution).

$\endgroup$
1
  • $\begingroup$ Hi - thank you for this. I agree that it's probably possible to construct such a distribution using a mixture of Dirac spikes / Gaussians; do you have a good sense of how one should decide where to locate them? (or e.g. a simple optimisation objective which accomplishes this) $\endgroup$
    – πr8
    Mar 21, 2019 at 8:29

You must log in to answer this question.

Not the answer you're looking for? Browse other questions tagged .