What is Hessian-free optimization?

Updated May 16, 2026

Short answer

Hessian-free optimization approximates second-order methods without explicitly computing the Hessian.

Deep explanation

Instead of computing or storing the Hessian matrix, Hessian-free methods use iterative approximations like conjugate gradient to compute Hessian-vector products. This makes second-order optimization feasible for large neural networks.

Real-world example

Early deep learning research on recurrent neural networks.

Common mistakes

  • Thinking full Hessian is explicitly computed.

Follow-up questions

  • What is Hessian-vector product?
  • Why is it scalable?

More Gradient Descent interview questions

View all →