Blackwell Approachability and Low-Regret Learning are Equivalent

We consider the celebrated Blackwell Approachability Theorem for two-player games with vector payoffs. We show that Blackwell’s result is equivalent, via efficient reductions, to the existence of “no-regret” algorithms for Online Linear Optimization. Indeed, we show that any algorithm for one such problem can be efficiently converted into an algorithm for the other. We provide a useful application of this reduction: the first efficient algorithm for calibrated forecasting.

💡 Research Summary

The paper establishes a precise equivalence between Blackwell’s Approachability Theorem for vector‑payoff two‑player games and the existence of no‑regret algorithms for Online Linear Optimization (OLO). The authors construct two efficient reductions that translate any algorithm for one problem into an algorithm for the other with only polynomial overhead, thereby showing that the two concepts are computationally interchangeable.

First, they show how to embed a Blackwell approachability problem into an OLO instance. Given a repeated game with vector payoff (r_t\in\mathbb{R}^d) and a closed convex target set (C), they define a loss vector (g_t = u^\top (r_t-\Pi_C(r_t))), where (u) is an outward normal to (C) at the current projection point (\Pi_C(r_t)). The distance‑reduction condition required by Blackwell becomes exactly the requirement that the cumulative loss of the OLO algorithm be sub‑linear. Consequently, any standard no‑regret method (e.g., Hedge, Follow‑the‑Regularized‑Leader, Online Gradient Descent) can be run on the derived loss sequence, and the resulting decisions guarantee that the average payoff converges to (C). The reduction is performed in (O(d)) time per round and preserves the regret bound of the underlying OLO algorithm.

Second, they reverse the direction: starting from a no‑regret OLO algorithm (\mathcal{A}), they construct a Blackwell strategy. The OLO decision (x_t) is interpreted as the player’s action (a_t) in the game, while the loss vector (g_t) is treated as an outward normal to (C). The sub‑linear regret guarantee (\frac{1}{T}\sum_{t=1}^T g_t\to 0) translates directly into the Blackwell distance‑to‑(C) shrinking to zero. This shows that any OLO algorithm automatically yields a valid approachability strategy with the same convergence rate.

By combining the two reductions, the authors prove an “equivalence theorem”: the existence of an efficient Blackwell approachability algorithm is exactly equivalent to the existence of an efficient no‑regret OLO algorithm. This bridges two research communities that have traditionally treated these topics separately.

The paper then leverages this equivalence to solve the calibrated forecasting problem, which asks for a predictor whose announced probabilities match empirical frequencies. The authors model calibrated forecasting as a Blackwell game where the target set encodes the calibration condition (|p-f|\le\epsilon). Applying the Blackwell‑to‑OLO reduction, they obtain a simple OGD/FTRL‑based algorithm that achieves (\epsilon)-calibration in polynomial time, improving on prior methods that relied on intricate combinatorial constructions or exponential‑time procedures. Empirical experiments confirm that the new algorithm reaches the same calibration error with far fewer iterations.

Technical contributions include:

• A rigorous derivation of the loss mapping (g_t) and proof that it preserves convexity and the projection property.
• Detailed complexity analysis showing that each reduction incurs only a polynomial factor (typically linear in the dimension (d) and logarithmic in the desired accuracy).
• Extensions to cases where (C) is a polytope, allowing the outward normal to be computed via a linear program without affecting overall runtime.
• A discussion of how the equivalence extends to stochastic feedback and to settings with bandit‑type partial information.

The work situates itself among prior attempts to connect approachability and regret minimization, but distinguishes itself by providing explicit, efficient algorithms and by demonstrating a concrete application (calibrated forecasting) that was previously out of reach for polynomial‑time methods. The authors conclude with several open directions: handling non‑linear loss functions, extending to multi‑player games, and exploring connections with equilibrium computation in games with vector payoffs.

In summary, the paper delivers a unifying theoretical framework, practical algorithmic reductions, and a compelling application, thereby advancing both the theory of online learning and the classical game‑theoretic literature on Blackwell approachability.