# Anchored second-order minimax optimizer

- ID: 1996
- Canonical URL: https://synthcore.org/idea/1996/anchored-second-order-minimax-optimizer
- API JSON: https://synthcore.org/api/idea/1996.json
- API Markdown: https://synthcore.org/api/idea/1996.md
- Verification status: unverified
- Source: [arXiv:2608.08463](https://arxiv.org/abs/2608.08463)
- Category: optimization
- Solves: stability, speedup, accuracy
- ML areas: optimizer, gan, rlhf, training-dynamics
- Math tags: optimization, convex-analysis, dynamical-systems, numerical-analysis
- Ratings: usefulness 6/10; difficulty 6/10; novelty 6/10

## Idea description

Replace the ordinary update in a differentiable minimax game with a Halpern-anchored second-order operator step. The current game iterate is first corrected using the local Jacobian of the game gradient, and the corrected point is then contracted toward a fixed anchor with a decreasing Halpern weight. This is intended to reduce cycling in adversarial training while preserving the faster asymptotic behavior associated with second-order monotone-operator methods.

## Mathematical statement

The paper formulates a monotone variational inequality: find x* in a nonempty compact convex set X such that <F(x*), x-x*> >= 0 for every x in X, where F:X -> R^d is continuously differentiable and monotone, meaning <F(x)-F(y), x-y> >= 0 for all x,y in X. For a convex-concave loss L(u,v), use x=(u,v) and F(x)=(∇_u L(u,v), -∇_v L(u,v)); a saddle point is a zero of F plus the normal cone of X. The paper introduces an Anchored Tensor Method with base rate O(T^{-(p-1)}) and combines it with a large-step inexact Halpern iteration to obtain O~(T^{-p}); for p=2 this is O~(T^{-2}). The implementable adaptation uses the second-order Taylor operator F_2(y;z)=F(y)+J_F(y)(z-y), where J_F(y) is the Jacobian of F, followed by a projected correction z and Halpern mixing x_{t+1}=alpha_t x_0+(1-alpha_t)z_t. Here x_0 is the fixed anchor, alpha_t is a positive decreasing weight such as 2/(t+2), eta is the correction step size, and Pi_X is projection onto the feasible parameter box or trust region. The exact theorem's rate requires monotonicity, smoothness, and controlled inexactness; neural games generally violate global monotonicity, so the empirical test must measure whether local Jacobian regularization makes the method useful.

## Key formulas

- $$\langle F(x^*),x-x^*\rangle\geq 0\quad\forall x\in\mathcal X,$$
- $$\langle F(x)-F(y),x-y\rangle\geq 0\quad\forall x,y\in\mathcal X,$$
- $$F_2(y;z)=F(y)+J_F(y)(z-y),\qquad z_t=\Pi_{\mathcal X}\!\left(y_t-\eta F_2(y_t;z_t)\right),$$
- $$x_{t+1}=\alpha_t x_0+(1-\alpha_t)z_t,\qquad \alpha_t=\frac{2}{t+2},\qquad \text{target rate }\widetilde{\mathcal O}(T^{-2})\text{ for }p=2.$$

## Implementation notes

(1) Integrate this at the optimizer step of a two-player differentiable game, not inside the network forward pass. Flatten generator/discriminator or actor/critic parameters into x=(u,v), define F=(grad_u L, -grad_v L), and optionally constrain each parameter update to a radius-r trust region around the current parameters. Keep x0 as the parameters at the beginning of a training phase. (2) Pseudocode: compute gu=grad_u L(u,v) and gv=grad_v L(u,v), set Fy=concat(gu,-gv), and initialize z=y-eta*Fy. For K=1 to 3 fixed-point refinements, compute q=z-y, estimate J_F(y)q with one forward-over-reverse or reverse-over-forward autodiff Jacobian-vector product, set F2=Fy+J_F(y)q, and update z=project(y-eta*F2). Then set alpha=2/(t+2) and x_next=alpha*x0+(1-alpha)*z. Use separate parameter slices when restoring u and v. (3) Computed from the mathematical construction are the operator F, the Jacobian-vector product, the projection, and the Halpern convex combination. Estimate empirically the local monotonicity m_hat=<F(y+q)-F(y),q>/(||q||^2+epsilon), the Jacobian norm, and a stable eta ceiling; reduce eta when m_hat is negative or ||J_F(y)q|| is excessively large. The implicit fixed-point equation is intentionally solved approximately, matching the inexact-iteration setting. (4) First experiment: train a 2D bilinear game L(u,v)=u^T A v, a nonlinear matrix game, and a small MNIST GAN with identical base learning rates. Compare simultaneous gradient descent, extragradient, optimistic Adam, and this anchored second-order method at equal gradient/JVP budgets. Record distance to the known equilibrium for toy games, loss oscillation, Jacobian norm, FID, and validation performance. Success means fewer operator evaluations to reach a fixed equilibrium gap, visibly damped rotational trajectories, and improved GAN stability without a prohibitive wall-clock increase.

## Disclaimer

AI-generated research hypothesis, automatically tested. Not peer-reviewed.