Data

Prove MSE is non-convex for logistic regression

Asked at

Microsoft

1Give yourself 5 minutes

2Answer out loud, not in your head

3Then compare with the answer below

Reference answer

The logistic model is sigma(theta x). The MSE surrogate is J(theta) = 1/2n sum (sigma(theta x_i) - y_i)^2. J is non-convex because sigma is sigmoidal: composing a non-linear function with a squared error creates a landscape with multiple local minima at extreme theta (saturated predictions). Hessian is uninformative globally. This is exactly why cross-entropy is preferred for classification - the log sits inside the sum producing a convex objective. Equivalent sigma-in-MSE views go through X^T diag(sigma(1-sigma)) but with a (1 - 2y_i) factor that breaks sign-definiteness.

Then expect these follow-ups

What other surrogate losses are non-convex?
When might you still use MSE?

Hiring now

Get the Microsoft interview, not just the prep

landed applies to roles like these for you, on each company’s own site, with answers written for the role.

Get 5x interviews10 free auto-applies. No card.

Or practice this question first

Next question