---
title: Blackwell Approachability
url: https://www.emergentmind.com/topics/blackwell-approachability
type: topic
---

# Blackwell Approachability

Blackwell approachability is a framework for repeated vector-payoff games in which a decision maker seeks to force the running average payoff vector toward a target set against an adversary. In its classical convex form, a closed convex set is approachable exactly when every supporting halfspace containing the set can be forced, or equivalently when for every adversary mixed action there exists a player mixed action whose expected payoff lies in the set. The theory yields constructive projection-based strategies and canonical \(O(1/\sqrt{T})\) convergence under bounded payoffs, and it has become a unifying language for no-regret learning, calibration, partial monitoring, and a broad range of modern online decision problems [1301.2663][1503.00255].

## 1. Formal setup

In the standard model, two players repeatedly choose actions and observe a vector payoff \(g(a_t,b_t)\in\mathbb{R}^d\), or equivalently a vector loss \(\ell(x_t,y_t)\in\mathbb{R}^d\). The running average is
\[
\bar g_T=\frac{1}{T}\sum_{t=1}^T g(a_t,b_t)
\quad\text{or}\quad
\bar \ell_T=\frac{1}{T}\sum_{t=1}^T \ell(x_t,y_t).
\]
A closed target set \(C\subseteq \mathbb{R}^d\) is approachable if the player has a strategy such that \(\operatorname{dist}(\bar g_T,C)\to 0\) or \(\operatorname{dist}(\bar \ell_T,C)\to 0\) against any opponent strategy, with distance usually taken in Euclidean norm [1301.2663][2502.03919].

For convex compact action spaces, a common formulation uses a bilinear map \(\ell:X\times Y\to\mathbb{R}^d\), Euclidean unit ball \(B\), and
\[
\operatorname{dist}(v,S)=\inf_{s\in S}\|v-s\|, \qquad
h_S(w)=\sup_{s\in S}\langle w,s\rangle.
\]
In Euclidean norm,
\[
d(x,K)=\max_{w\in B}\{\langle w,x\rangle-h_K(w)\},
\]
so distance to a convex target can be written as a support-function maximization over directions \(w\) [1503.00255][2502.03919].

A weaker finite-horizon notion is weak approachability: for a fixed horizon \(n\), the player may use a horizon-dependent strategy \(\sigma_n\) and require only \(\mathbb{E}[d(\bar g_n,C)]\le \varepsilon\). For convex sets, weak approachability and approachability coincide in the classical model, but later extensions show that this equivalence can fail once absorbing or quitting dynamics are introduced [1301.2663][1609.08870].

## 2. Geometric characterization

For a closed convex target \(C\), Blackwell’s theorem admits several equivalent formulations. One classical statement is
\[
\forall\, y\in\Delta(B)\;\exists\, x\in\Delta(A)\ \text{such that}\ g(x,y)\in C.
\]
An equivalent support-function form is
\[
\forall u\in B_* \;\exists x\in\Delta(A)\ \text{such that}\ \langle u,g(x,y)\rangle \le h_C(u)\quad \forall y\in\Delta(B),
\]
or, in the loss convention,
\[
\forall w\in B\;\exists x\in X:\quad \max_{y\in Y}\langle w,\ell(x,y)\rangle \le h_S(w).
\]
These are the halfspace formulations: every supporting halfspace of the target set can be forced by the player [1503.00255][1011.1936].

The projection geometry is central. If \(x\notin C\) and \(c=\Pi_C(x)\), then \(u=x-c\) is an outward normal to a supporting hyperplane of \(C\) at \(c\). Blackwell’s condition requires an action whose expected next payoff lies on the \(C\)-side of that hyperplane. This is the geometric reason the running average is steered toward the target [2406.07585][2604.06621].

For arbitrary closed, not necessarily convex, sets, the exact criterion is more subtle. Blackwell’s sufficient condition is the notion of a \(B\)-set, and a closed set is approachable if and only if it contains a \(B\)-set. In proximal-normal form, the condition is
\[
\forall e\in E,\;\forall p\in NP_E(e),\qquad
\min_{x\in\Delta(A)}\max_{y\in\Delta(B)}\langle p,\;g(x,y)-e\rangle\le 0.
\]
This extends the convex normal-cone picture and gives a necessary-and-sufficient characterization beyond convexity [1301.2663][1110.1514].

## 3. Algorithms and convergence rates

Blackwell’s original algorithm is projection-based. At round \(t\), let \(z_t\) be the current average payoff and \(c_t=\Pi_C(z_t)\). If \(z_t\notin C\), define the steering direction
\[
u_t=\frac{z_t-c_t}{\|z_t-c_t\|}.
\]
The player then chooses a mixed action satisfying
\[
\langle g(x_t,y)-c_t,\; z_t-c_t\rangle \le 0 \qquad \forall y,
\]
or equivalently forces the supporting halfspace through \(c_t\) with normal \(u_t\). Under bounded payoffs, this yields the canonical \(O(1/\sqrt{T})\) rate; in classical Euclidean analyses one obtains \(\mathbb{E}[d(\bar g_T,C)]\le c\,G/\sqrt{T}\), and the survey literature also gives high-probability bounds [1301.2663][2406.07585].

A direct online-convex-optimization formulation replaces projection geometry by the support-function losses
\[
\ell_t(u)=\langle u,r_t\rangle-h_C(u), \qquad
f_t(u)=-\ell_t(u)=-\langle u,r_t\rangle+h_C(u).
\]
If an OCO algorithm over the dual unit ball guarantees regret \(\mathrm{Reg}_T\), then
\[
\operatorname{dist}(\bar r_T,C)\le \frac{\mathrm{Reg}_T}{T}.
\]
With standard OCO guarantees, this again gives \(O(1/\sqrt{T})\). In Euclidean geometry, unregularized FTL recovers Blackwell’s original steering direction, while under smoothness of \(\partial C\) and bounded curvature one obtains
\[
\operatorname{dist}(\bar r_T,C)\le \frac{C_0(1+\ln T)}{T}.
\]
Regularized FTL recovers the classical \(O(1/\sqrt{T})\) guarantee without the smoothness assumption [1503.00255].

Two important algorithmic refinements are predictive and response-based variants. Predictive Blackwell approachability assumes an estimate \(v^t\) of the next payoff vector and yields the bound
\[
\min_{s\in S}\left\|s-\frac{1}{T}\sum_{t=1}^T u(x^t,y^t)\right\|_2
\le
\frac{1}{\sqrt{T}\left(1+\frac{2}{T}\sum_{t=1}^T \|u(x^t,y^t)-v^t\|_2^2\right)},
\]
so convergence improves as the prediction error decreases [2007.14358]. A dual alternative replaces projection by computing responses to opponent actions; in that formulation one maintains target points \(r_t^\ast\in S\), steering vectors \(\lambda_t=\bar r_t^\ast-\bar r_t\), and obtains
\[
d(\bar r_n,S)\le \|\lambda_n\| \le \rho/\sqrt{n},
\]
avoiding direct projection onto \(S\) when response computation is simpler [1312.7658].

## 4. Equivalence to regret minimization and calibration

A central modern insight is that Blackwell approachability and no-regret online linear optimization are algorithmically equivalent. Efficient reductions map an approachability problem to an OLO problem by cone lifting, and conversely map an OLO problem to an approachability problem through halfspace oracles. In the reduction from OLO to approachability, regret is controlled by the distance of a lifted average vector payoff to a polar cone; in the reverse reduction, approachability distance is controlled by OLO regret [1011.1936].

This perspective subsumes standard regret notions. External regret arises by approaching the negative orthant of the instantaneous regret vector. Internal and swap regret are obtained by higher-dimensional regret matrices or \(\Phi\)-regret coordinates. Calibration can likewise be reduced to an approachability problem in an auxiliary game, and internally consistent strategies yield calibrated forecasting [1301.2663][1011.1936].

The equivalence to regret is not, however, automatically rate-preserving. It has been shown that the classical reduction of Abernethy–Bartlett–Hazan can map an approachability instance with optimal rate \(R_1\) to a no-regret instance with optimal regret-per-round \(R_2\) such that \(R_2/R_1\) is arbitrarily large, including cases with \(R_1=0\) and \(R_2>0\). A tight reduction is instead obtained through improper \(\phi\)-regret minimization, where
\[
\AppLoss=\Reg,\qquad
\Rate(\text{approachability})=\Rate(\text{improper }\phi\text{-regret}),
\]
and some improper \(\phi\)-regret instances are not linearly reducible to either external regret or proper \(\phi\)-regret while preserving rates [2406.07585].

## 5. Generalized information structures and dynamics

Approachability extends beyond full monitoring. In partial monitoring, the purely informative game replaces stage payoffs by the maximal information players obtain, represented as probability measures; objectives are then expressed as convergence of averages of these measures. This framework yields a unified characterization of approachable sets in games with or without signals, and for convex targets it recovers the familiar simple condition in the measure-valued space [1301.3609].

A different generalization appears in generalized quitting games, where either player may have quitting actions that absorb play. In that setting, three geometric conditions, denoted \((S)\), \((M)\), and \((N)\), govern weak approachability. In Big-Match Type I, where only the approaching player can quit, these conditions are equivalent and collapse to Blackwell’s classical condition. In Big-Match Type II, where only the opponent can quit, none of the three conditions is both sufficient and necessary for weak approachability, and weak approachability need not imply uniform approachability [1609.08870].

Dynamic extensions also include stochastic games and time-varying geometry. In Stackelberg stochastic games with vector costs, a Blackwell-style scalarization along the current normal direction yields a tractable leader strategy, and a reinforcement-learning algorithm learns the approachable strategy when the transition kernel is unknown [1411.0728]. In a separate time-dependent extension, both the outcome functions \(g_t\) and the inner products \(\langle\cdot,\cdot\rangle_t\) may vary with time; if the induced norms are nonincreasing, then
\[
\min_{r\in\mathcal{C}}\|\bar v_T-r\|_T
\le
\frac{1}{T}\sqrt{\sum_{t=1}^T \|r_t\|_t^2},
\]
which recovers the classical \(O(1/\sqrt{T})\) rate when the geometry is constant [2303.04956].

## 6. Computational relaxations and structured domains

Recent work has focused on computationally constrained approachability. When exact optimization over the player’s or adversary’s action set is intractable, approximation oracles can replace exact saddle-point computations. Under nonnegativity and monotone preferences, if \(S\) is approachable in the exact game and approximation ratios satisfy \(\alpha_X\ge 1\) and \(0<\alpha_Y\le 1\), then the downward closure of the scaled set
\[
(\alpha_X\alpha_Y^{-1}S)_{\downarrow}
\]
is efficiently approachable with rate \(O(1/\sqrt{T})\). If only one side has approximation access, the scaled targets simplify to \((\alpha_X S)_{\downarrow}\) or \((\alpha_Y^{-1}S)_{\downarrow}\) [2502.03919].

A different computational direction studies non-Euclidean objectives, especially \(\ell_\infty\)-style approachability that arises naturally in regret minimization. High-dimensional \(\ell_\infty\)-approachability can be rewritten as lower-dimensional pseudonorm approachability, and the resulting algorithmic theory gives rates independent of the original vector-payoff dimension. Under mild normalization assumptions, one obtains a dimension-independent \(O(1/\sqrt{T})\) guarantee, and a maximum-entropy FTRL construction yields an \(O(\sqrt{\log d/T})\) dependence on the original dimension [2302.01517].

Structured domains from game solving and optimization also admit specialized Blackwell algorithms. On treeplexes, which encode sequence-form strategies in extensive-form games, Predictive Treeplex Blackwell\(^+\) gives an \(O(1/\sqrt{T})\) convergence rate to Nash equilibrium in self-play, while a stabilized variant attains an \(O(1/T)\) rate [2403.04680]. In convex optimization, the Conic Blackwell Algorithm\(^+\) is a parameter- and scale-free regret minimizer for general convex sets with \(O(\sqrt{T})\) regret, and its saddle-point counterpart SP-CBA\(^+\) achieves an \(O(1/\sqrt{T})\) ergodic rate [2202.12277].

## 7. Contemporary applications and conceptual limits

Blackwell approachability has become a design principle for constrained and multi-objective online learning. In fair online learning with sensitive and non-sensitive contexts, an approachability formulation yields a necessary and sufficient condition for compatibility between learning objectives and fairness constraints, and it instantiates this condition for group-wise no-regret, group-wise calibration, and demographic parity; when compatibility fails, the same framework characterizes the optimal trade-off [2106.12242].

A more recent development is simultaneous Blackwell approachability, where multiple target sets must be approached through coupled actions. Individual approachability of each target does not suffice in general. The key sufficient condition is an \(\varepsilon\)-mixture linear optimization oracle that satisfies every convex mixture of one halfspace per target. This framework leads to multiclass omniprediction, where one seeks suboptimality bounds simultaneously for a family of losses and a family of comparator predictors, and it yields horizon or sample complexity approximately \(\varepsilon^{-(k+1)}\) for \(\varepsilon\)-omniprediction in a \(k\)-class problem [2602.17577].

The current limits are both conceptual and computational. Exact reductions to standard regret classes may lose optimal rates [2406.07585]. Exact target sets may have to be replaced by scaled downward closures when only approximation oracles are available [2502.03919]. Weak approachability need not imply uniform approachability once quitting dynamics are present [1609.08870]. At the same time, survey work now places Blackwell approachability at the center of modern AI topics including no-regret online learning, calibrated forecasting, fair online learning, and multi-objective RLHF, which suggests that its geometric control viewpoint remains technically live far beyond the repeated games setting in which it was introduced [2604.06621].

Source: https://www.emergentmind.com/topics/blackwell-approachability