Sequential Minimax Optimization (SMO) Overview

• Sequential Minimax Optimization (SMO) is a framework where the decision-maker (leader) optimizes against a sequentially adapting worst-case opponent, emphasizing the order of play.
• It spans diverse applications such as robust sequential testing, online prediction with minimax regret, and bilevel optimization, each addressing nonconvex-nonconcave challenges.
• Key algorithms like Follow-the-Ridge and Newton-type methods leverage SMO’s saddle-point structure to achieve local minimax solutions in complex decision settings.

Searching arXiv for recent and foundational papers on sequential minimax optimization and closely related terminology. Sequential Minimax Optimization (SMO) denotes a family of optimization and decision formulations in which a learner, controller, or outer variable is optimized against a sequentially responding adversary, follower, uncertainty set, or continuation policy. In current arXiv usage, the sequential structure appears in several technically distinct ways: leader–follower zero-sum games of the form $\min_x \max_y f(x,y)$, robust sequential tests in which the decision variable is a stopping rule, online prediction problems characterized by minimax regret, and bilevel methods that solve a sequence of minimax subproblems generated from an original constrained problem (Jin et al., 2019, Fiez et al., 2021, Liu et al., 2024, Lu et al., 10 Nov 2025). A recurring theme across these formulations is that exact global minimization against a globally optimal opponent is either structurally inappropriate or computationally unrealistic, so the central object becomes a local, adversarial, or worst-case sequential response.

1. Problem class and scope

A canonical continuous formulation of SMO is the sequential two-player zero-sum problem

$$\min_{x\in \mathbb{R}^n}\max_{y\in \mathbb{R}^m} f(x,y),$$

where $x$ is the leader or minimizer and $y$ is the follower or maximizer. The order of play is essential in the nonconvex-nonconcave regime because $\min_x \max_y f(x,y)$ need not equal $\max_y \min_x f(x,y)$, so simultaneous equilibrium notions are generally not interchangeable with sequential ones (Jin et al., 2019, Zhang et al., 2020). In this sense, SMO is not merely minimax optimization with time indices; it is optimization in which the second player’s response is part of the problem definition.

Other literatures instantiate the same minimax pattern with different decision variables. In continuous-time Poisson testing, the optimization variable is a sequential policy consisting of a stopping time $T$ and terminal decision $d\in\{0,1\}$, and the minimax criterion is

$\min_{T\in\mathcal S}\ \sup_{\psi>0} J(\psi;T),$

where the supremum is taken over all prior odds $\psi$ for the hidden intensity (Mei, 2023). In sequential probability assignment with contexts, the game is an adversarial $$\min_{x\in \mathbb{R}^n}\max_{y\in \mathbb{R}^m} f(x,y),$$0-round protocol and the central object is the minimax regret

$$\min_{x\in \mathbb{R}^n}\max_{y\in \mathbb{R}^m} f(x,y),$$1

defined relative to a class of sequential experts (Liu et al., 2024). In constrained bilevel optimization, SMO appears as an algorithmic scheme that repeatedly solves saddle subproblems of the form

$$\min_{x\in \mathbb{R}^n}\max_{y\in \mathbb{R}^m} f(x,y),$$2

generated by a modified augmented Lagrangian and penalty construction (Lu et al., 10 Nov 2025).

This variety suggests that “sequential” has several technical meanings in the literature: order of play, sequential observation, sequential regret accumulation, and sequential solution of surrogate minimax problems. What unifies these uses is the presence of an adaptive worst-case counterpart that is not collapsed into a static one-shot objective.

2. Local optimality and the sequential game viewpoint

For nonconvex-nonconcave problems, the basic local solution concept is local minimax rather than local Nash. The definition proposed for the sequential setting requires that $$\min_{x\in \mathbb{R}^n}\max_{y\in \mathbb{R}^m} f(x,y),$$3 satisfy

$$\min_{x\in \mathbb{R}^n}\max_{y\in \mathbb{R}^m} f(x,y),$$4

for nearby $$\min_{x\in \mathbb{R}^n}\max_{y\in \mathbb{R}^m} f(x,y),$$5, where $$\min_{x\in \mathbb{R}^n}\max_{y\in \mathbb{R}^m} f(x,y),$$6 as $$\min_{x\in \mathbb{R}^n}\max_{y\in \mathbb{R}^m} f(x,y),$$7 (Jin et al., 2019). This definition encodes the sequential order explicitly: the max-player is allowed a local response after the min-player moves. An equivalent characterization is that $$\min_{x\in \mathbb{R}^n}\max_{y\in \mathbb{R}^m} f(x,y),$$8 is a local maximum of $$\min_{x\in \mathbb{R}^n}\max_{y\in \mathbb{R}^m} f(x,y),$$9 and $x$0 is a local minimum of the locally maximized value function $x$1 for all small $x$2 (Jin et al., 2019).

The first-order condition at a local minimax point is standard stationarity,

$x$3

but the second-order structure is distinctive. If $x$4 is twice differentiable, then $x$5 is necessary, and when $x$6, the Schur-complement condition

$x$7

must hold (Jin et al., 2019). The sufficient strict version replaces the semidefinite inequalities by definite ones. The Newton-type literature uses the closely related notion of strict local minimax points as the local targets of second-order methods and as surrogates of global solutions (Zhang et al., 2020).

Two caveats are central. First, local minimax points need not exist. Second, a global minimax point can fail to be a local minimax point and may even fail to be stationary (Jin et al., 2019). These facts distinguish sequential local optimality from ordinary local minimization and explain why algorithms designed for simultaneous games or for unconstrained stationarity can converge to objects that are irrelevant for the sequential problem.

3. Continuous algorithms for nonconvex-nonconcave SMO

A major algorithmic line studies how to reach local minimax points in smooth leader–follower games. “Follow-the-Ridge” (FR) modifies gradient descent-ascent by adding a curvature-based correction term to the follower update: $x$8

$x$9

The term $y$0 approximates motion along the local response manifold $y$1 obtained from implicit differentiation of $y$2 (Wang et al., 2019). The method is designed so that, with a suitable learning rate, all strictly stable fixed points are local minimax, and all local minimax points are stable fixed points. The same work argues that FR addresses the rotational behavior of gradient dynamics and is compatible with preconditioning and positive momentum (Wang et al., 2019).

Newton-type methods extend this perspective by exploiting second-order structure more directly. Gradient-Descent-Newton (GDN) uses a Newton step in the follower block and a gradient step in the leader block, whereas Complete Newton (CN) applies Newton updates to both blocks via a Schur-complement elimination: $y$3

$y$4

These methods are implemented in Hessian-free form via Hessian-vector products and conjugate gradient, and they are proven to converge locally at strict local minimax points, with GDN achieving local linear convergence and CN local quadratic convergence (Zhang et al., 2020).

A different reformulation replaces exact inner maximization by an algorithmic adversary. Instead of assuming that the max-player can solve $y$5, one studies

$y$6

or, in randomized form,

$y$7

When the adversary algorithms $y$8 are smooth, the induced outer objective is weakly convex, which makes Moreau-envelope stationarity the appropriate optimality notion (Fiez et al., 2021). The framework yields monotonic progress and excludes limit cycles. Under the stated assumptions, stochastic subgradient descent reaches an $y$9-first-order stationary point in $\min_x \max_y f(x,y)$0 iterations, while a deterministic proximal outer method attains improved complexity, including $\min_x \max_y f(x,y)$1 and $\min_x \max_y f(x,y)$2 variants depending on the inner solver (Fiez et al., 2021). This suggests an important shift in SMO theory: the adversary may be a smooth algorithm rather than an exact optimizer.

4. Sequential statistical decision, least favorable priors, and least favorable distributions

In sequential testing, SMO becomes a robust stopping problem. For Poisson processes with two possible intensities,

$\min_x \max_y f(x,y)$3

the performance criterion is

$\min_x \max_y f(x,y)$4

and the minimax problem is

$\min_x \max_y f(x,y)$5

The solution has a saddle-point structure

$\min_x \max_y f(x,y)$6

where $\min_x \max_y f(x,y)$7 is an optimal stopping rule and $\min_x \max_y f(x,y)$8 a least favorable prior (Mei, 2023). For known prior, the optimal rule is of threshold form in the posterior-odds process $\min_x \max_y f(x,y)$9; in the minimax setting, the least favorable prior in the nontrivial regime is characterized by the root condition $\max_y \min_x f(x,y)$0, with existence guaranteed when the boundary functional satisfies $\max_y \min_x f(x,y)$1 (Mei, 2023). One consequence is that symmetry of error costs does not imply a symmetric least favorable prior.

The discrete-time Markov formulation generalizes this logic to multiple hypotheses and distributional uncertainty. A sequential test is a policy $\max_y \min_x f(x,y)$2 on a Markovian sufficient state $\max_y \min_x f(x,y)$3, and the weighted robust objective is

$\max_y \min_x f(x,y)$4

Under mild Markov assumptions, the optimal cost function satisfies a Bellman equation, and its generalized partial derivatives coincide with expected run length and weighted error probabilities of optimal tests (Fauss et al., 2018). Least favorable distributions are characterized implicitly, state by state, as maximizers of continuation-value functionals. The resulting minimax-optimal cost is identified as a generalized $\max_y \min_x f(x,y)$5-dissimilarity, and the least favorable distributions are those most similar with respect to that dissimilarity (Fauss et al., 2018).

A third line studies one-sided mixture-based stopping rules against composite alternatives and optimizes the mixing distribution in a minimax Kullback–Leibler sense. For a discrete alternative $\max_y \min_x f(x,y)$6, the nearly minimax choice is

$\max_y \min_x f(x,y)$7

which equalizes the worst-case $\max_y \min_x f(x,y)$8 up to an $\max_y \min_x f(x,y)$9 term as the false-alarm level $T$0 (Fellouris et al., 2011). For continuous exponential-family alternatives, the analogue is

$T$1

These results show that robust sequential design often reduces to identifying a least favorable prior, least favorable distribution, or least favorable mixture.

5. Minimax regret in sequential prediction

SMO also appears in online learning with logarithmic loss, where the objective is minimax regret rather than expected sample size or Bayes risk. In the contextual formulation, Nature reveals contexts $T$2, the learner outputs distributions $T$3, and the loss is

$T$4

The comparator class $T$5 may consist of fully sequential experts $T$6, not merely static mappings $T$7 (Liu et al., 2024). The exact minimax value is characterized by the contextual Shtarkov sum

$T$8

through the theorem

$T$9

This characterization yields the minimax-optimal prediction rule, contextual Normalized Maximum Likelihood (cNML): $d\in\{0,1\}$0 The proof requires a minimax swap argument under truncation because log loss is unbounded, using the smoothing map

$d\in\{0,1\}$1

to make Sion’s minimax theorem applicable (Liu et al., 2024). An important consequence is that sequential $d\in\{0,1\}$2 entropy does not characterize minimax regret in general, even though it still yields upper bounds. In this setting, SMO is less about saddle-point dynamics than about exact value characterization and optimal sequential probability assignment.

6. Structured reformulations: bilevel and mixed continuous-discrete SMO

In constrained bilevel optimization, SMO becomes an outer algorithm that repeatedly solves minimax surrogates generated from the lower-level problem. The model class considered is

$d\in\{0,1\}$3

with possibly nonconvex upper-level objective and convex or strongly convex lower-level objective (Lu et al., 10 Nov 2025). The method constructs the saddle subproblem

$d\in\{0,1\}$4

where

$d\in\{0,1\}$5

The sequential scheme uses decreasing tolerances $d\in\{0,1\}$6, parameters $d\in\{0,1\}$7 and $d\in\{0,1\}$8, a warm-started lower-level solve, an inner first-order solver for the minimax subproblem, and the multiplier update

$d\in\{0,1\}$9

The resulting operation complexity is $\min_{T\in\mathcal S}\ \sup_{\psi>0} J(\psi;T),$0 for merely convex lower-level objectives and $\min_{T\in\mathcal S}\ \sup_{\psi>0} J(\psi;T),$1 for strongly convex ones, with the latter improving the previous best-known bound by a factor of $\min_{T\in\mathcal S}\ \sup_{\psi>0} J(\psi;T),$2 (Lu et al., 10 Nov 2025).

A different structured setting mixes continuous minimization and discrete maximization. The convex-submodular problem

$\min_{T\in\mathcal S}\ \sup_{\psi>0} J(\psi;T),$3

assumes $\min_{T\in\mathcal S}\ \sup_{\psi>0} J(\psi;T),$4 convex and $\min_{T\in\mathcal S}\ \sup_{\psi>0} J(\psi;T),$5 submodular (Adibi et al., 2021). Exact saddle points are computationally intractable: finding an $\min_{T\in\mathcal S}\ \sup_{\psi>0} J(\psi;T),$6-saddle point is NP-hard for any $\min_{T\in\mathcal S}\ \sup_{\psi>0} J(\psi;T),$7, and polynomial-time approximation beyond $\min_{T\in\mathcal S}\ \sup_{\psi>0} J(\psi;T),$8 would imply $\min_{T\in\mathcal S}\ \sup_{\psi>0} J(\psi;T),$9 (Adibi et al., 2021). The paper therefore adopts the weaker minimax criterion

$\psi$0

Its alternating procedures—Gradient Greedy, Extra-Gradient Greedy, Gradient Replacement-Greedy, Extra-Gradient Replacement-Greedy, and Extra-Gradient on the continuous extension—combine projected gradient or extra-gradient updates in $\psi$1 with greedy or replacement-greedy updates in $\psi$2. Their guarantees range from $\psi$3-approximate minimax solutions in $\psi$4 iterations to $\psi$5-approximate minimax solutions in $\psi$6 iterations for the continuous-extension method (Adibi et al., 2021). This suggests that SMO is also a design pattern for alternating continuous-discrete robust optimization, not only a theory of smooth two-player games.

7. Terminological ambiguity and adjacent uses of “SMO”

The acronym “SMO” is overloaded in the arXiv literature, and distinguishing these meanings is necessary because they refer to different algorithmic objects.

Expansion of “SMO”	Technical role	Representative papers
Sequential Minimax Optimization	Leader–follower games, robust testing, bilevel minimax subproblems	(Jin et al., 2019, Lu et al., 10 Nov 2025)
Sequential Minimal Optimization	Two-variable decomposition methods for SVM-type quadratic programs; also coordinate-wise VQE updates	(Benavides-Herrera et al., 2 May 2026, Lange et al., 20 Jun 2026, Glasmachers, 2013, Torres-Barrán et al., 2020, Pedrielli et al., 15 May 2026)
Sequential Model-Based Optimization	Surrogate-guided configuration search in software effort estimation	(Xia et al., 2019)

Sequential Minimal Optimization in the SVM sense is a decomposition method that updates a very small working set, typically two dual variables at a time. Recent examples include an SMO algorithm for $\psi$7-SVR with MAPE loss and sample-dependent box constraints, where the only structural changes from standard SMO occur in feasibility sets and clipping bounds (Benavides-Herrera et al., 2 May 2026); an SMO algorithm for OC-SVM+ under the LUPI paradigm with finite-time convergence (Lange et al., 20 Jun 2026); “Planning-ahead SMO,” which modifies the current step size by anticipating the next working set while preserving convergence (Glasmachers, 2013); and Conjugate SMO, which replaces the plain SMO direction by a conjugate direction and proves linear convergence when the kernel matrix is positive definite (Torres-Barrán et al., 2020). In variational quantum eigensolvers, NFT or Rotosolve is explicitly described as Sequential Minimal Optimization because it performs analytic one-dimensional minimization of trigonometric coordinate subproblems and then studies bias accumulation and regularization in that coordinate-wise process (Pedrielli et al., 15 May 2026).

Sequential Model-Based Optimization is yet another distinct usage. In software process control, ROME wraps FLASH around a learner such as CART, maintains an archive of tried configurations and observed scores, trains a surrogate model on that archive, scores candidate settings, and selects the setting with the smallest predicted error (Xia et al., 2019). That procedure is an adaptive configuration search, not a minimax formulation.

The overlap of acronyms can obscure the literature. In current usage, Sequential Minimax Optimization concerns worst-case sequential response or saddle-point structure, whereas Sequential Minimal Optimization concerns decomposition-based quadratic programming and Sequential Model-Based Optimization concerns surrogate-guided search.