Learning rate η =
0.10
Condition number κ =
5
Steps/sec
10
step 0
▶ Play
↺ Reset
⚠ Diverging — η · max-eigenvalue exceeds 2. Lower the learning rate.
f(x,y) = ½(x² + κ·y²) ∇f = (x, κ·y) update:
θ ← θ − η ∇f(θ)
Stable iff η < 2/L where L = max eigenvalue of Hessian = κ. Optimal η
*
=
—
. Convergence rate ≈
—
per step.