How does learning rate affect bias-variance dynamics in gradient descent?

Learning rate influences convergence speed and stability, indirectly affecting bias (underfitting) and variance (overfitting).

Learning rate controls how aggressively a model updates its parameters during optimization. A very high learning rate can cause unstable training and prevent convergence, leading to high bias (underfitting). A very low learning rate may overfit noise if trained too long, increasing variance. Proper scheduling (decay, cosine annealing) helps balance convergence and generalization.

In deep learning, learning rate is often more critical than model architecture in controlling generalization behavior.