The Bellman operator shrinks distances
Recall the Bellman optimality backup: take a value estimate, apply one step of look-ahead, and you get a new estimate. The key fact is that this map is a γ-contraction in the max-norm: if two value functions differ by at most d somewhere, after one backup they differ by at most γ·d. Errors shrink by a factor of γ < 1 every iteration.
The Bellman optimality backup: one step of look-ahead. Because of the discount factor γ, this map is a γ-contraction.
A contraction has exactly one fixed point
The Banach fixed-point theorem says any contraction on a complete space has a unique fixed point, and iterating the map from any starting guess converges to it geometrically. For RL that fixed point is the optimal value function, and the satisfied equation is the Bellman optimality equation. This is the whole proof of value iteration in one breath: it is a contraction being iterated, so it must converge to the unique optimum.
Interactive Markov chain whose state distribution converges to a single stationary distribution regardless of the starting state.
# Value iteration: contraction => geometric convergence
V = zeros(num_states)
for k in range(max_iters):
V_new = bellman_optimality_backup(V) # one application T*
if max_norm(V_new - V) < (1 - gamma) * eps:
break # error now <= eps
V = V_new
# ||V_k - V*||_inf <= gamma**k * ||V_0 - V*||_infLearning, not planning: noisy averaging
Value iteration needs the full model. Tabular Q-learning instead updates from single sampled transitions, so each backup is a noisy estimate of the true Bellman backup. To analyse it we leave pure contraction behind and pick up stochastic approximation: the theory of solving a fixed-point equation when you only observe unbiased-but-noisy values of the operator. Q-learning is exactly stochastic approximation applied to the Bellman optimality operator.
Interactive gridworld where an agent learns Q-values cell by cell from sampled moves.
Step sizes that converge: Robbins–Monro
Noisy updates only settle if the step sizes are tuned just right. The Robbins–Monro conditions capture the balance: the step sizes must sum to infinity (so you can still move arbitrarily far and reach the answer) yet their squares must sum to a finite number (so the accumulated noise is eventually damped out). A schedule like αₜ = 1/t satisfies both; a constant step size satisfies the first but not the second, so it converges to a noisy neighbourhood rather than a point.
The Q-learning update: the step size αₜ weights the TD error. The Robbins–Monro conditions say exactly how αₜ must shrink for this to converge.
- Σ αₜ = ∞ — keep enough total motion to reach any target.
- Σ αₜ² < ∞ — make later steps shrink fast enough to average out noise.
- Every state-action pair must be visited infinitely often — exploration must not starve.
Putting it together — and where rates come from
With a contracting operator, Robbins–Monro step sizes, and infinitely-often visitation, Q-learning converges to the optimal action values with probability one. To go beyond 'it converges' to how fast, you bring in concentration bounds — inequalities (Hoeffding, Bernstein, Azuma) that say a sample average is close to its mean with high probability. They turn 'the noise washes out eventually' into an explicit sample complexity of the form Õ(SA·H³/ε²) per the dependence on states, actions, horizon, and accuracy.