Contribution
It brings performativity, “the prediction influences the target”, into risk minimization by a map D(·) from model parameters to the distribution of (features, outcome) that deploying the model produces. It defines performative optimality and performative stability, shows that they differ, and gives sufficient conditions for repeated retraining to converge to a stable point, each shown necessary by a counterexample. It also shows that stable points are near optimal. Strategic classification becomes a special case, with Stackelberg equilibria as performative optima.
Key insight
Retraining is “a natural equilibrating dynamic rather than a nuisance”, and what it equilibrates to is self-consistency, not truth about other policies. “Performative stability implies that the predictions are calibrated not against past outcomes, but against the future outcomes that manifest from acting on the prediction.” A stable model is optimal for the world it makes, and retraining only ever samples that world. It learns nothing about D(θ′) for policies never deployed.
Assumptions
- Θ ⊆ ℝ^d closed and convex; instances z = (x, y), x ∈ ℝ^(m−1), y ∈ ℝ.
- ε-sensitivity (Definition 3.1): W₁(D(θ), D(θ′)) ≤ ε‖θ − θ′‖₂.
- (A1) β-joint smoothness: ∇_θℓ(z; θ) is β-Lipschitz in θ and in z.
- (A2) γ-strong convexity of ℓ(z; ·) for every z.
- D(·) is fixed: “deploying the same model at two different points in time must induce the same distribution”. The paper reads D(θ) as the interventional distribution of a do-intervention on θ.
- Theorem 4.3: ℓ is L_z-Lipschitz in z. Theorem 3.10: a finite exponential moment ξ_{α,µ}, for Fournier–Guillin concentration.
Key results
- Definitions 2.1, 2.3. PR(θ) = E_{Z∼D(θ)} ℓ(Z; θ); θ_PO = argmin PR. θ_PS = argmin_θ E_{Z∼D(θ_PS)} ℓ(Z; θ). With DPR(θ, θ′) = E_{D(θ)} ℓ(Z; θ′), stability minimizes DPR(θ_PS, ·), optimality DPR(θ, θ).
- Example 2.2, re-derived. X uniform on {±1}, Y | X ~ Bernoulli(½ + µX + εθX), f_θ(x) = θx + ½, squared loss. PR(θ) = ¼ − 2µθ + (1 − 2ε)θ², so θ_PO = µ/(1 − 2ε). Retraining gives θ_{t+1} = µ + εθ_t, so θ_PS = µ/(1 − ε). They coincide iff ε = 0. At µ = 0.2, ε = 0.3: θ_PS = 0.286, θ_PO = 0.5, PR = 0.168 against 0.150, and the static answer θ = µ scores 0.186.
- Example 3.4. Point mass at 0 if θ ≥ ½ and at 1 otherwise: no stable point, and retraining alternates 1, 0, 1, …
- Theorem 3.5. Under (A1), (A2) and ε-sensitivity, ‖G(θ) − G(θ′)‖ ≤ ε(β/γ)‖θ − θ′‖, and for ε < γ/β retraining converges linearly to a unique stable point. Re-derived in THEORY-861.
- Proposition 3.6. Retraining can fail to converge with (a) convexity without strong convexity (ℓ = βyθ, Y = εθ: iterates …, 1, −1, 1, …), (b) strong convexity without joint smoothness (a regularized hinge), or (c) ε ≥ γ/β (ℓ = (y − θ)², Y = 1 + εθ: θ_t = Σ ε^k + θ₀ε^t diverges for ε > 1, and there is no stable point at ε = 1).
- Theorem 3.8. Repeated gradient descent with step η ≤ 2/(β + γ) contracts and converges for ε < γ/((β + γ)(1 + 1.5ηβ)).
- Theorem 3.10. With n_t = O(log(t/p)/(εδ)^m) samples per round, repeated empirical risk minimization reaches and stays within δ of θ_PS with probability 1 − p, for ε < γ/(2β).
- Proposition 4.1. A stable point exists for convex, jointly continuous ℓ on compact Θ (Kakutani). Proposition 4.2. PR can be concave in θ although ε < γ/β (Example 2.2 with ε ∈ [½, 1)).
- Theorem 4.3. For ℓ L_z-Lipschitz in z and γ-strongly convex, ‖θ_PO − θ_PS‖ ≤ 2L_zε/γ for every stable point and every optimum. Re-derived in THEORY-839.
- Corollary 5.1. In strategic classification retraining converges to a stable classifier within 2L_zε(L_θ + L_zε)/γ of the Stackelberg equilibrium in objective.
- Simulations (§5.2). A Kaggle credit dataset (18,357 points, three strategic features), best response x′_S = x_S − εθ_S, regularized logistic loss. Retraining converges in a few rounds for ε = 0.01 and diverges for large ε. Appendix B.2 notes that ε < γ/β holds only at ε = 0.01 and that repeated gradient descent never meets its theoretical cutoff, so most of the convergence seen is outside the theorems.
Claims
| id | claim | strength | support |
|---|---|---|---|
| C1 | Performative optimality and stability are distinct; neither implies the other | strong | Example 2.2, Remark 2.4; re-derived |
| C2 | Smooth, strongly convex loss and ε < γ/β make retraining contract to a unique stable point at linear rate | strong | Theorem 3.5, Appendix D.1; re-derived |
| C3 | Each of the three conditions is needed: dropping any one admits non-convergence | strong | Proposition 3.6, three counterexamples; (a) and (c) checked |
| C4 | Stable points and optima lie within 2L_zε/γ of each other | strong | Theorem 4.3; re-derived |
| C5 | Directly minimizing performative risk is in general harder than finding stable points | moderate | Proposition 4.2 shows non-convexity; no hardness result |
| C6 | Strategic classification is a special case, with Stackelberg equilibria as performative optima | strong | §5.1, by definition |
| C7 | Retraining converges quickly in a simulated credit setting when ε is small | moderate | §5.2, one dataset; mostly outside the theorems’ conditions (B.2) |
Concepts
- Distribution map D(·): model parameters to the distribution of instances that deploying the model produces.
- Performative risk PR(θ): E_{D(θ)} ℓ(Z; θ).
- Decoupled performative risk DPR(θ, θ′): loss of θ′ on the distribution θ induces.
- Performative optimum: argmin PR. Performatively stable: θ minimizes DPR(θ, ·), “fixed points of risk minimization”.
- Repeated risk minimization (RRM): θ_{t+1} = argmin DPR(θ_t, ·).
- ε-sensitivity: D(·) is ε-Lipschitz from Euclidean to W₁.
Connections
It casts concept drift (Gama et al. 2014; Žliobaitė’s application table) as performativity where it is, and strategic classification (Hardt et al. 2016; Brückner et al. 2012; Dalvi et al. 2004) as its special case. It relates D(θ) to a do-intervention (Pearl) and repeated risk minimization to off-policy learning in reinforcement learning. Its convergence proofs extend the classical contraction of gradient descent (Nesterov; Hardt, Recht and Singer 2016) to a moving distribution, through Kantorovich–Rubinstein duality. It cites performativity in the social sciences (Healy 2015 and others) without engaging that literature.
Bearing on the record
- CLAIM-402. A25’s warning, that a description “well calibrated against outcomes generated under its own use need not be an adequate description of the system under a different policy”, is the paper’s stability notion read negatively. Stability is optimality for D(θ_PS) alone, and retraining samples only the distributions it visits. A26 §2’s toy, which A26 called “a toy model rather than a claimed theorem from Perdomo et al.”, has two stable points. That is consistent with the paper. Its threshold rule is not strongly convex, and Theorem 3.5’s uniqueness needs strong convexity, joint smoothness and ε < γ/β (THEORY-861). The claim’s “Perdomo et al.’s performative prediction … not read here” is now out of date.
- CLAIM-407. Its point, that a description closed under ordinary dynamics need not preserve responses to intervention, has a learning analogue here: a stable model can be far from the optimum (Example 2.2: θ_PS = 0.286 against θ_PO = 0.5), because the optimum depends on how D(θ) varies away from the deployed θ, which stability never consults. The record’s join, not the paper’s.
- TERM-064. A predictor an institution deploys is a legibility regime that reshapes the population it reads. Theorem 3.5 says when such a loop settles, and where: at self-consistency, not at the best achievable legibility. The paper does not use the word.
It is a machine-learning paper and carries no instruction for practice.
Its LIT is flagged anthology-candidate.
Limitations
- Conditions are strong and seldom checkable. ε must be known relative to γ/β. In the paper’s own simulation the conditions hold only at the smallest ε (B.2).
- D(·) is static. A distribution that depends on the history of deployed models, not only the current one, is outside the framework.
- Uniqueness depends on strong convexity. Many models of interest, including any discrete classification rule, have no strong convexity, and then multiple stable points or none are possible (Example 3.4, Proposition 3.6).
- No method for the optimum. Finding θ_PO needs knowledge of D(·) beyond the deployed point, and the paper gives none here. Theorem 4.3 only bounds the distance.
- The social-science literature on performativity is cited, not used.
Open questions
- The authors’: is the gap between repeated gradient descent’s ε < γ/(β + γ) and retraining’s ε < γ/β fundamental or an artifact?
- How do stable points behave when D depends on the whole history of deployments, as reputations and rankings do?
- With multiple stable points, which one retraining reaches depends on the starting model. Is that path dependence measurable from deployment data alone?