I am wondering what is the $\mathbb{E}[(x\cdot v)^2]$ where $x$ is a random unit vector in $\mathbb{R}^n$ and $v$ is a given unit vector in $\mathbb{R}^n$. By $(x\cdot v)$ I mean the dot product between $x$ and $v$.

I read somewhere that $\mathbb{E}[(x \cdot v)^2] = \frac{1}{n}$. But I can not prove it. Does anyone have an idea on this?

  • 4
    $\begingroup$ This expectation is merely that of the squared component of $x$ in the direction $v$. It cannot be computed without additional assumptions on the distribution of $x$. What are those assumptions? (When that distribution is spherical, the distributions of all components of $x$ in any orthonormal frame are equal and sum of the squares of the components is constantly $1$, whence the result.) $\endgroup$ – whuber Oct 5 '14 at 15:58
  • $\begingroup$ No distribution is specified in the problem, but I think what you are saying makes sense. Can you elaborate your answer a bit more, please? (sorry I'm not strong in probability) why if x is from spherical distribution, the expectation goes through? $\endgroup$ – Learner Oct 6 '14 at 17:04
  • $\begingroup$ This is arguably a duplicate of Distribution of dot products between two random unit vectors in $\mathbb R^D$ $\endgroup$ – amoeba Oct 6 '14 at 20:46

Judging from the result, it appears the context implicitly supposes the distribution of $x$ is invariant under orthogonal transformations: I would call this a spherically-symmetric distribution.

(There are plenty of spherically-symmetric distributionns. Starting with any distribution $F$ in $\mathbb{R}^n$, define $\tilde F$ to be the values of $F$ averaged over the action of the orthogonal group $O(n)$ of rotations and reflections about the origin. The average exists because $O(n)$ is compact and acts continuously. It is immediate that $\tilde F$ is invariant under $O(n)$. In particular, a Normal distribution of mean $(0,0,\ldots,0)$, diagonal variance matrix, and equal variances is spherical.)

It is a (simple and geometrically obvious) algebraic result that any unit vector $v$ can be extended to an orthonormal frame $(v=v_1, v_2, \ldots, v_n)$. Because the distribution is spherically symmetric and any element $v_i$ can be rotated into any other element $v_j$, the coordinates $x\cdot v_i$ all have the same distribution. Let $\mu_2$ be the common expected value of all the $(x\cdot v_i)^2$.

Since $x$ is assumed to be a unit vector,

$$1 = 1^2 = x\cdot x = \sum_i (x\cdot v_i)^2.$$

Take expectations of both sides and use the linearity of expectation to compute

$$1 = \mathbb{E}(1) = \mathbb{E}(x\cdot x) = \mathbb{E}\left( \sum_{i=1}^n (x\cdot v_i)^2\right) = \sum_{i=1}^n \mathbb{E}((x\cdot v_i)^2) = \sum_{i=1}^n \mu_2 = n\mu_2,$$

implying $1/n = \mu_2 = \mathbb E ((x\cdot v)^2)$.

| cite | improve this answer | |

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.