White cross = global minimum (1, 1). The narrow valley (Rosenbrock's banana) challenges gradient-based methods.
Loss: (1−w₁)² + 100(w₂−w₁²)² + 0.1sin(10w₁) + 0.1cos(8w₂) · Adam adapts per-parameter learning rates; SGD follows raw gradients; RMSProp normalizes by a running mean of squared gradients.