I'm trying to calculate a good mean shrinkage parameter for a custom quadratic discriminant analysis (QDA), and I ran into a math problem. Suppose $X=(X_1, X_2, \ldots, X_k)^T\sim{\mathcal{N}(\textbf{0},\Sigma)}$, where $\Sigma_{ii} = 1$ for $i=1,2,\ldots, k$. Define $$ \text{mean}(X) = \frac{1}{k}\sum_{i = 1}^k X_i. $$ I would like to calculate $E[\text{mean}(X)|\ \text{mean}(X) > 0]$, if there's a closed form.

I suspect there isn't and curve fitting will be necessary. If that's the case, is there a programmatic way of fitting a curve to results obtained via Monte Carlo simulation?

I have been trying to fit a curve as a function of $k$ and $\mathbb{1}^T \tilde{\Sigma}\mathbb{1}$, where $\tilde{\Sigma} = \text{Var}(X|\ \text{mean}(X) > 0)$, since $\tilde{\Sigma}$ should be close to the covariance matrix I have for my labels in QDA. Unfortunately, I've had limited success. Neural networks and ML techniques are not an option due to speed concerns.


1 Answer 1


Since under the assumption $\bar{X} \sim N(0, \sigma^2)$ with $\sigma^2 = \frac{1}{k^2}1'\Sigma 1$, the problem can be reduced to finding $E[X|X > 0]$ given $X \sim N(0, \sigma^2)$, which by definition of conditional expectation (conditioning on an event) is \begin{align} E[X|X > 0] = \frac{E[XI_{[X > 0]}]}{P(X > 0)} = 2\int_0^\infty x\frac{1}{\sqrt{2\pi}\sigma}e^{-\frac{x^2}{2\sigma^2}}dx = \frac{\sqrt{2}\sigma}{\sqrt{\pi}}. \end{align}

  • $\begingroup$ Thank you for your help! I had considered the case where $X\sim{\mathcal{N}(\textbf{0}, I)}$, and I don't know why I didn't notice this. Your solution may be good enough. However, you need to assume that the off-diagonal elements are 0 to use the central limit theorem. Some components may be correlated here. $\endgroup$ May 19, 2023 at 22:51
  • 1
    $\begingroup$ @Charles0349 I didn't assume the off-diagonal elements to be $0$ (please read my first sentence more carefully) nor use CLT (there is no place in the answer invoked any asymptotic argument). The point is if $(X_1, \ldots, X_n) \sim N(0, \Sigma)$ (for any valid covariance matrix $\Sigma$), then the sample mean $\bar{X}$ is always normally distributed with mean $0$ and variance $1'\Sigma 1/n^2$. This is a consequence of affine transformation property of multivariate normal distribution. $\endgroup$
    – Zhanxiong
    May 19, 2023 at 23:46
  • $\begingroup$ You're right. I apologize. I guess there was less to this than I thought. At any rate, thank you for your help! $\endgroup$ May 20, 2023 at 2:13

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

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