Blackwell Approachability

• Blackwell approachability is a framework that allows a decision-maker in repeated vector-payoff games to steer average outcomes toward a target set by enforcing conditions on every supporting halfspace.
• It employs projection-based strategies that guarantee an O(1/√T) convergence rate under bounded payoffs, serving as a unifying method for no-regret learning and calibrated forecasting.
• The theory extends to complex settings including partial monitoring, generalized quitting games, and computational relaxations, with applications in game theory, optimization, and fair online learning.

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 (Perchet, 2013, Shimkin, 2015).

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 (Perchet, 2013, Garber et al., 6 Feb 2025).

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,

$g(a_t,b_t)\in\mathbb{R}^d$0

so distance to a convex target can be written as a support-function maximization over directions $g(a_t,b_t)\in\mathbb{R}^d$1 (Shimkin, 2015, Garber et al., 6 Feb 2025).

A weaker finite-horizon notion is weak approachability: for a fixed horizon $g(a_t,b_t)\in\mathbb{R}^d$2, the player may use a horizon-dependent strategy $g(a_t,b_t)\in\mathbb{R}^d$3 and require only $g(a_t,b_t)\in\mathbb{R}^d$4. 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 (Perchet, 2013, Flesch et al., 2016).

2. Geometric characterization

For a closed convex target $g(a_t,b_t)\in\mathbb{R}^d$5, Blackwell’s theorem admits several equivalent formulations. One classical statement is

$g(a_t,b_t)\in\mathbb{R}^d$6

An equivalent support-function form is

$g(a_t,b_t)\in\mathbb{R}^d$7

or, in the loss convention,

$g(a_t,b_t)\in\mathbb{R}^d$8

These are the halfspace formulations: every supporting halfspace of the target set can be forced by the player (Shimkin, 2015, Abernethy et al., 2010).

The projection geometry is central. If $g(a_t,b_t)\in\mathbb{R}^d$9 and $\ell(x_t,y_t)\in\mathbb{R}^d$0, then $\ell(x_t,y_t)\in\mathbb{R}^d$1 is an outward normal to a supporting hyperplane of $\ell(x_t,y_t)\in\mathbb{R}^d$2 at $\ell(x_t,y_t)\in\mathbb{R}^d$3. Blackwell’s condition requires an action whose expected next payoff lies on the $\ell(x_t,y_t)\in\mathbb{R}^d$4-side of that hyperplane. This is the geometric reason the running average is steered toward the target (Dann et al., 2024, Paxton, 8 Apr 2026).

For arbitrary closed, not necessarily convex, sets, the exact criterion is more subtle. Blackwell’s sufficient condition is the notion of a $\ell(x_t,y_t)\in\mathbb{R}^d$5-set, and a closed set is approachable if and only if it contains a $\ell(x_t,y_t)\in\mathbb{R}^d$6-set. In proximal-normal form, the condition is

$\ell(x_t,y_t)\in\mathbb{R}^d$7

This extends the convex normal-cone picture and gives a necessary-and-sufficient characterization beyond convexity (Perchet, 2013, Telgarsky, 2011).

3. Algorithms and convergence rates

Blackwell’s original algorithm is projection-based. At round $\ell(x_t,y_t)\in\mathbb{R}^d$8, let $\ell(x_t,y_t)\in\mathbb{R}^d$9 be the current average payoff and $$\bar g_T=\frac{1}{T}\sum_{t=1}^T g(a_t,b_t) \quad\text{or}\quad \bar \ell_T=\frac{1}{T}\sum_{t=1}^T \ell(x_t,y_t).$$0. If $$\bar g_T=\frac{1}{T}\sum_{t=1}^T g(a_t,b_t) \quad\text{or}\quad \bar \ell_T=\frac{1}{T}\sum_{t=1}^T \ell(x_t,y_t).$$1, define the steering direction

$$\bar g_T=\frac{1}{T}\sum_{t=1}^T g(a_t,b_t) \quad\text{or}\quad \bar \ell_T=\frac{1}{T}\sum_{t=1}^T \ell(x_t,y_t).$$2

The player then chooses a mixed action satisfying

$$\bar g_T=\frac{1}{T}\sum_{t=1}^T g(a_t,b_t) \quad\text{or}\quad \bar \ell_T=\frac{1}{T}\sum_{t=1}^T \ell(x_t,y_t).$$3

or equivalently forces the supporting halfspace through $$\bar g_T=\frac{1}{T}\sum_{t=1}^T g(a_t,b_t) \quad\text{or}\quad \bar \ell_T=\frac{1}{T}\sum_{t=1}^T \ell(x_t,y_t).$$4 with normal $$\bar g_T=\frac{1}{T}\sum_{t=1}^T g(a_t,b_t) \quad\text{or}\quad \bar \ell_T=\frac{1}{T}\sum_{t=1}^T \ell(x_t,y_t).$$5. Under bounded payoffs, this yields the canonical $$\bar g_T=\frac{1}{T}\sum_{t=1}^T g(a_t,b_t) \quad\text{or}\quad \bar \ell_T=\frac{1}{T}\sum_{t=1}^T \ell(x_t,y_t).$$6 rate; in classical Euclidean analyses one obtains $$\bar g_T=\frac{1}{T}\sum_{t=1}^T g(a_t,b_t) \quad\text{or}\quad \bar \ell_T=\frac{1}{T}\sum_{t=1}^T \ell(x_t,y_t).$$7, and the survey literature also gives high-probability bounds (Perchet, 2013, Dann et al., 2024).

A direct online-convex-optimization formulation replaces projection geometry by the support-function losses

$$\bar g_T=\frac{1}{T}\sum_{t=1}^T g(a_t,b_t) \quad\text{or}\quad \bar \ell_T=\frac{1}{T}\sum_{t=1}^T \ell(x_t,y_t).$$8

If an OCO algorithm over the dual unit ball guarantees regret $$\bar g_T=\frac{1}{T}\sum_{t=1}^T g(a_t,b_t) \quad\text{or}\quad \bar \ell_T=\frac{1}{T}\sum_{t=1}^T \ell(x_t,y_t).$$9, then

$C\subseteq \mathbb{R}^d$0

With standard OCO guarantees, this again gives $C\subseteq \mathbb{R}^d$1. In Euclidean geometry, unregularized FTL recovers Blackwell’s original steering direction, while under smoothness of $C\subseteq \mathbb{R}^d$2 and bounded curvature one obtains

$C\subseteq \mathbb{R}^d$3

Regularized FTL recovers the classical $C\subseteq \mathbb{R}^d$4 guarantee without the smoothness assumption (Shimkin, 2015).

Two important algorithmic refinements are predictive and response-based variants. Predictive Blackwell approachability assumes an estimate $C\subseteq \mathbb{R}^d$5 of the next payoff vector and yields the bound

$C\subseteq \mathbb{R}^d$6

so convergence improves as the prediction error decreases (Farina et al., 2020). A dual alternative replaces projection by computing responses to opponent actions; in that formulation one maintains target points $C\subseteq \mathbb{R}^d$7, steering vectors $C\subseteq \mathbb{R}^d$8, and obtains

$C\subseteq \mathbb{R}^d$9

avoiding direct projection onto $\operatorname{dist}(\bar g_T,C)\to 0$0 when response computation is simpler (Bernstein et al., 2013).

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 (Abernethy et al., 2010).

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 $\operatorname{dist}(\bar g_T,C)\to 0$1-regret coordinates. Calibration can likewise be reduced to an approachability problem in an auxiliary game, and internally consistent strategies yield calibrated forecasting (Perchet, 2013, Abernethy et al., 2010).

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 $\operatorname{dist}(\bar g_T,C)\to 0$2 to a no-regret instance with optimal regret-per-round $\operatorname{dist}(\bar g_T,C)\to 0$3 such that $\operatorname{dist}(\bar g_T,C)\to 0$4 is arbitrarily large, including cases with $\operatorname{dist}(\bar g_T,C)\to 0$5 and $\operatorname{dist}(\bar g_T,C)\to 0$6. A tight reduction is instead obtained through improper $\operatorname{dist}(\bar g_T,C)\to 0$7-regret minimization, where

$\operatorname{dist}(\bar g_T,C)\to 0$8

and some improper $\operatorname{dist}(\bar g_T,C)\to 0$9-regret instances are not linearly reducible to either external regret or proper $\operatorname{dist}(\bar \ell_T,C)\to 0$0-regret while preserving rates (Dann et al., 2024).

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 (Perchet et al., 2013).

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 $\operatorname{dist}(\bar \ell_T,C)\to 0$1, $\operatorname{dist}(\bar \ell_T,C)\to 0$2, and $\operatorname{dist}(\bar \ell_T,C)\to 0$3, 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 (Flesch et al., 2016).

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 (Kalathil et al., 2014). In a separate time-dependent extension, both the outcome functions $\operatorname{dist}(\bar \ell_T,C)\to 0$4 and the inner products $\operatorname{dist}(\bar \ell_T,C)\to 0$5 may vary with time; if the induced norms are nonincreasing, then

$\operatorname{dist}(\bar \ell_T,C)\to 0$6

which recovers the classical $\operatorname{dist}(\bar \ell_T,C)\to 0$7 rate when the geometry is constant (Kwon et al., 2023).

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 $\operatorname{dist}(\bar \ell_T,C)\to 0$8 is approachable in the exact game and approximation ratios satisfy $\operatorname{dist}(\bar \ell_T,C)\to 0$9 and $\ell:X\times Y\to\mathbb{R}^d$0, then the downward closure of the scaled set

$\ell:X\times Y\to\mathbb{R}^d$1

is efficiently approachable with rate $\ell:X\times Y\to\mathbb{R}^d$2. If only one side has approximation access, the scaled targets simplify to $\ell:X\times Y\to\mathbb{R}^d$3 or $\ell:X\times Y\to\mathbb{R}^d$4 (Garber et al., 6 Feb 2025).

A different computational direction studies non-Euclidean objectives, especially $\ell:X\times Y\to\mathbb{R}^d$5-style approachability that arises naturally in regret minimization. High-dimensional $\ell:X\times Y\to\mathbb{R}^d$6-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 $\ell:X\times Y\to\mathbb{R}^d$7 guarantee, and a maximum-entropy FTRL construction yields an $\ell:X\times Y\to\mathbb{R}^d$8 dependence on the original dimension (Dann et al., 2023).

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$\ell:X\times Y\to\mathbb{R}^d$9 gives an $B$0 convergence rate to Nash equilibrium in self-play, while a stabilized variant attains an $B$1 rate (Chakrabarti et al., 2024). In convex optimization, the Conic Blackwell Algorithm$B$2 is a parameter- and scale-free regret minimizer for general convex sets with $B$3 regret, and its saddle-point counterpart SP-CBA$B$4 achieves an $B$5 ergodic rate (Grand-Clément et al., 2022).

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 (Chzhen et al., 2021).

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 $B$6-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 $B$7 for $B$8-omniprediction in a $B$9-class problem (Hu et al., 19 Feb 2026).

The current limits are both conceptual and computational. Exact reductions to standard regret classes may lose optimal rates (Dann et al., 2024). Exact target sets may have to be replaced by scaled downward closures when only approximation oracles are available (Garber et al., 6 Feb 2025). Weak approachability need not imply uniform approachability once quitting dynamics are present (Flesch et al., 2016). 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 (Paxton, 8 Apr 2026).