Suppose $U$ is a random vector satisfying $\mathbb E[U] = \mu$ and $\mathrm{var}(\|U\|_2) \le V$. Let $\bar{U} = U / \|U\|_2$ and $\bar\mu = \mu / \|\mu\|$. What is a lower bound on $\mathbb E[\bar U^\intercal\bar\mu]$ in terms of $V$?

(For what it's worth, I have in mind a setting where $V$ is quite small compared to $\|\mu\|_2$---seems like there should be some non-trivial bound in this case.)


The best possible general bound is the trivial one, $-1$. (This deserves to be called "trivial" because $\bar U^\prime \bar \mu \ge -1$ in any event.) To demonstrate this, I offer examples whose bounds come arbitrarily close to $-1$.

The intuition is this: the mean of the normalized vector $\bar U$ can be the opposite of the mean of $U$ itself because the normalization can greatly expand values close to zero. If we pick a vector $U$ that is far from zero in one direction a small fraction of the time and otherwise is very close to zero but in the opposite direction the rest of the time, then (a) the mean of $U$ can lie in this privileged direction but (b) the mean of $\bar U$ can be arbitrarily close to a unit vector in the opposite direction. This will cause $\bar U^\prime \bar \mu$ generally to be close to $-1$.

This construction works in any number of dimensions. I will provide an explicit example in one dimension.

Let $0\lt\epsilon \ll 1/3$. This makes $p=2\epsilon/(1+\epsilon)$ a probability because $0 \lt p \lt 1$. Let $U$ take on the value $1$ with probability $p$ and the value $-\epsilon$ with probability $1-p$. Straightforward computations yield

$$\mu=\mathbb{E}(U) = \epsilon \gt 0,$$

$$\bar\mu = 1,$$


$$V = \operatorname{Var}(||U||_2) = \frac{2\epsilon(1-3\epsilon+3\epsilon^2-\epsilon^3)}{(1+\epsilon)^2} \approx 2\mu.$$

The random variable $\bar U$ takes on the value $1$ with probability $p$ and $-1$ with probability $1-p$. Therefore

$$\mathbb{E}(\bar U^\prime \bar\mu)=\mathbb{E}(\bar U)=2p-1 = -1 + \frac{4\epsilon}{1+\epsilon} \lt 0.$$

This value remains unchanged even when we scale $U$ by $\sigma \gt 0$. That scaling changes $\mu$ to $\mu\sigma$ and $V$ to $V\sigma^2$, so by taking $\sigma^2$ small we may make $V\sigma^2$ as small as we like compared to $\mu\sigma$. Moreover, the greatest lower bound of $\mathbb{\bar U^\prime \bar \mu}$ is $-1$, QED.

  • $\begingroup$ Thank you for the thoughtful response! What if I added the condition that $V \ll \|\mu\|_2^2$ ? Is there a non-trivial bound then? $\endgroup$ – Jeff Jan 22 '17 at 20:27
  • $\begingroup$ Yes, that inequality makes more sense. When it holds, Chebyshev's Inequality implies there is a high chance that $U$ is close to $\mu$ and therefore $\bar U$ has to be nearly parallel to $\bar \mu$. $\endgroup$ – whuber Jan 22 '17 at 21:18
  • $\begingroup$ Whew! That's a relief. I'll use Chebyshev's Inequality then. Thanks for your help. $\endgroup$ – Jeff Jan 22 '17 at 21:36

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

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