Mathematics Stack Exchange is a question and answer site for people studying math at any level and professionals in related fields. Join them; it only takes a minute:

Sign up
Here's how it works:
  1. Anybody can ask a question
  2. Anybody can answer
  3. The best answers are voted up and rise to the top

Suppose $X_1,\ldots,X_n$ are each unimodal random variables with mode around zero. In particular, for each $i$, $X_i=X_i'-X'$ where $X'_i$ and $X'$ are i.i.d. unimodal random ($X_i'$ and $X_j'$ are ind. as well) variables so that $X_i$ is indeed unimodal. Is it the case that joint distribution of $X_i$ has the form, say $F(\cdot)$ satisfies the following:

Given $x_2,\ldots,x_n$ $F(u,x_2,\ldots,x_n)$ is convex for $u\in (-\infty,a)$ and concave for $u \in (a,\infty)$. In particular, the conditional (on truncating the other variables) distribution is unimodal. If so, how would one go about proving such a claim?

Thanks in advance

share|cite|improve this question

I doubt it is unimodal about $0$. I would expect the conditional distribution to be unimodal about a value closer to the centre of the other values.

If $n$ is large and the $X'_i$ and $X'$ have unimodal symmetric distributions centred at $0$, I would expect the conditional distribution of $X_1$ given the values of $X_2,X_3,\ldots,X_n$ to be centred (and unimodal about) a value close to the centre of the values of $X_2,X_3,\ldots,X_n$ (which will be approximately the value of $-X'$).

share|cite|improve this answer
this was a reply to an earlier version of the question – Henry Jul 6 '12 at 20:40

Notation: we have iid variables $Y,Y_1,\dots,Y_n$ and we observe $X_i=Y_i-Y$.

In general this is not true. If a distribution has a sharp peak at 0 and a heavy tail, then the a posteriori distribution on $Y$ caused by the observation of $x_1,\dots,x_n$ will have discrete peaks near $0,-x_1,\dots,-x_n$, rather than a single large peak at e.g. their empirical average (if $Y$ has a Gaussian distribution) or e.g. their median (if $Y$ has a Laplace distribution). So the distribution of $Y$ won't be unimodal, in the sense that its CDF won't be quasiconcave.

One might wonder if the sum $X_1=Y_1-Y$ might "blur" these peaks, but precisely because $Y_1$ has a sharp peak at 0 this is not going to change anything.

A simple example is, for $\delta<\tfrac{1}{2n}$ and for small $\varepsilon>0$, to let $Y$ be distributed: uniformly in $[-\delta/2,\delta/2]$ with probability $1-\varepsilon$, and uniformly in $[-1/2,1/2]$ with probability $\varepsilon$.

The corresponding density is $$\psi=(1-\varepsilon)\chi_{[-\delta/2,\delta/2]}+\varepsilon\chi_{[-1/2,1/2]}$$

Defining $x_0=0$, we have the equalities given by leonbloy in his answer: $$\psi(y) f(x_1x_2\dots x_n\mid y)=\prod_{i=0}^n \psi(x_i+y)\\ F(x_1x_2\dots x_n)=\int_{\mathbb R} \psi(y) f(x_1x_2\dots x_n\mid y)\,dy$$

So we have that when all $x_i$ (including $x_0$) are a distance greater than $\delta$ apart of each other: $$\psi(y) f(x_1x_2\dots x_n\mid y)=O(\varepsilon^{n-1})$$ $$F(x_1x_2\dots x_n) = O(\varepsilon^{n-1})$$ But when we relax the condition to allow $x_1=x_k$ (for some $k=0,2,3,\dots,n$), we have $$\psi(y) f(x_1x_2\dots x_n\mid y)=\frac{\varepsilon^{n-2}}{\delta^2} \Big[|x_k+y|\le\delta/2\Big]^2 + O(\varepsilon^{n-1})$$ $$F(x_1x_2\dots x_n) = \varepsilon^{n-2}/\delta + O(\varepsilon^{n-1})$$

So, for small enough $\varepsilon$ the conditional density of $X_1$ has $n$ peaks (near $x_0,x_2,\dots,x_n$) with up to $n-1$ valleys (this bound is reached when the peaks are more than $2\delta$ apart), therefore it cannot be unimodal when $n\ge 2$.

This counterexample can also be made smooth and made to have a unique mode by convolving it with a normal distribution of small variance, so the distribution is as well-behaved as we could hope for: it is symmetric, smooth and unimodal.

However, a weaker form of your question is true: if $Y$ has a log-concave distribution, then the a posteriori distribution of $Y$ will be log-concave (since the product of log-concave functions is log-concave), and so will $X_1$ (integrating a multivariate log-concave function along an axis yields a log-concave function, e.g. see theorem 1.8.3 of this book). So if I'm not mistaken your statement is true when replacing "unimodal" with "log-concave". This explains why the Gaussian and Laplace examples made sense!

Edit: Henry was wondering about the large $n$ limit case, for a fixed distribution. It is an interesting question that is true for a larger class of distributions, but I don't think it is true in general. For example if $Y$ is absolutely continuous with a density $\psi$ that diverges in 0 so that $\psi(y)^2$ is not integrable, then the density will diverge in each of $x_2,\dots,x_n$, so the conditional distribution is never unimodal unless $x_2=\dots=x_n$. A formal statement of the large $n$ limit problem could be: "Does the probability that $x_2,\dots,x_n$ satisfy the question converge to 1 as $n\to\infty$?" (according to the measure on $X_2\times\dots\times X_n$)

share|cite|improve this answer
For $x_2=x_3=\cdots=x_n=0$, do you think the statement is true? – user25720 Jul 6 '12 at 20:53

Just some thougths. Let me change a little notation: we have $n+1$ iid variables $Y, Y_1, Y_2 \cdots Y_n$ with density $f_Y(y)=\psi(\cdot)$, and we are interested in the joint density of $X_i = Y_i - Y$.

We have $f_{X_1|Y}(x_1|y)=\psi(x_1+y)$ and they are jointly conditional independent. So

$$f(x_1 x_2 \cdots x_n |y)=\psi(x_1+y) \, \psi(x_2+y) \cdots \, \psi(x_n+y)=\frac{f(x_1 x_2 \cdots x_n )}{f_Y(y)}$$


$$ f_{X_1 X_2 \cdots X_n}(x_1 x_2 \cdots x_n ) = \int_{-\infty}^{+\infty} \psi(y) \,\psi(x_1+y) \, \psi(x_2+y) \, \cdots \, \psi(x_n+y) \, dy =\\ =\int_{-\infty}^{+\infty} \prod_{i=0}^n \psi(x_i+y) \, dy$$

where we've defined $x_0=0$. I'm not sure if much more can be said in general.

If $\psi(\cdot)$ is symmetric, then obviously this has the mode at $0$.

It's also true in general (but slightly trivial) that $E(X_1,X_2 \cdots X_n) = (0,0, \cdots 0)$. Further, it's also straighforward that $$E(X_1|X_2 \cdots X_n) = E(E( X_1 | X_2 \cdots X_n Y)) = \mu_y - E( Y | X_2 \cdots X_n) = \frac{\sum_{i=2}^n X_i}{n-1}$$

which agrees with Henry's observation.

share|cite|improve this answer

Your Answer


By posting your answer, you agree to the privacy policy and terms of service.

Not the answer you're looking for? Browse other questions tagged or ask your own question.