Dynamic Duality Gap (D-DGap) Overview

• Dynamic Duality Gap (D-DGap) is a performance measure that quantifies evolving primal–dual discrepancies in online optimization settings.
• It integrates adaptive learning rates, comparator sequence drift, and prediction error aggregation to benchmark deviations from equilibrium.
• D-DGap serves as both a convergence diagnostic and minimax-optimal objective across applications like zero-sum games, GANs, and variational inequalities.

Searching arXiv for recent and relevant sources on Dynamic Duality Gap (D-DGap). Dynamic Duality Gap (D-DGap) denotes a family of gap-based performance measures in which primal–dual discrepancy is evaluated in a dynamic, adaptive, or parameter-varying setting rather than only at a fixed point. In the most explicit formulation currently available for online convex-concave games, D-DGap is the cumulative comparator-dependent quantity

$$D\text{-}DGap\,(u_{1:T}, v_{1:T}) := \sum_{t=1}^T \left[ f_t(x_t, v_t) - f_t(u_t, y_t) \right], \qquad \forall (u_t, v_t) \in X \times Y,$$

where the comparator sequence itself may be non-stationary (Meng et al., 9 Sep 2025). More broadly, the term has been used, or invoked as a natural extension, in several adjacent literatures: descent on duality-gap objectives in zero-sum games (Fasoulakis et al., 31 Jan 2025), proximal duality gaps for GAN training with dynamic or adaptive regularization (Sidheekh et al., 2021), and D-gap merit functions for variational inequalities whose values act as dynamic error measures along an algorithmic trajectory (Li et al., 2022). Across these settings, the common role of D-DGap is to quantify deviation from equilibrium, optimality, or strong duality under temporal evolution, adaptive comparison, or perturbation.

1. Formalization in online convex-concave optimization

The clearest formal definition of Dynamic Duality Gap appears in Online Convex-Concave Optimization (OCCO), which extends Online Convex Optimization to two-player time-varying convex-concave games (Meng et al., 9 Sep 2025). At round $t$, the two players choose $x_t \in X$ and $y_t \in Y$, after which the environment reveals a continuous convex-concave payoff $f_t : X \times Y \to \mathbb{R}$, convex in $x$ and concave in $y$. The D-DGap against an arbitrary comparator sequence $(u_t, v_t)$ is

$$D\text{-}DGap\,(u_{1:T}, v_{1:T}) := \sum_{t=1}^T \left[ f_t(x_t, v_t) - f_t(u_t, y_t) \right].$$

This definition generalizes dynamic regret to a two-player setting (Meng et al., 9 Sep 2025). When the comparator is static, D-DGap reduces to a standard duality-gap-type benchmark; when the comparator is dynamic, it measures adaptation to a changing best-in-hindsight sequence. The paper’s assumptions are compact convex action sets $X, Y$, bounded payoffs, and bounded subgradients, together with conditions supporting mirror maps and adaptive learning-rate design (Meng et al., 9 Sep 2025).

A central structural quantity is the path-length of the comparator sequence,

$t$0

which appears in the minimax characterization of achievable D-DGap (Meng et al., 9 Sep 2025). This makes the metric explicitly dynamic: the comparator is not a single equilibrium, but an evolving reference trajectory.

2. Algorithmic minimization of D-DGap

For OCCO, a modular algorithm has been proposed to minimize D-DGap under both adversarial and favorable environments (Meng et al., 9 Sep 2025). It contains three components.

The Adaptive Module runs a pair of state-of-the-art OCO algorithms, such as ADER, for the two players and achieves

$t$1

where $t$2 is a data-dependent upper bound on $t$3 (Meng et al., 9 Sep 2025).

The Multi-Predictor Aggregator uses the clipped Hedge algorithm to combine $t$4 predictors into

$t$5

and yields the guarantee

$t$6

where $t$7 is the cumulative prediction error of predictor $t$8 (Meng et al., 9 Sep 2025).

The Integration Module combines the adaptive expert and the prediction-error expert. Its final decision uses convex combinations

$t$9

where the pair $x_t \in X$0 is prediction-driven and $x_t \in X$1 is adaptive (Meng et al., 9 Sep 2025). The joint update is posed as a coupled variational inequality; existence and uniqueness are guaranteed via monotone operator theory and Browder-Minty, and implementation is by a proximal-point style algorithm (Meng et al., 9 Sep 2025).

The resulting upper bound is

$x_t \in X$2

up to logarithmic factors (Meng et al., 9 Sep 2025). The same work also gives a minimax lower bound

$x_t \in X$3

for suitable path-length $x_t \in X$4, showing optimality up to logarithmic terms (Meng et al., 9 Sep 2025).

This suggests that, in the online game-theoretic setting, D-DGap is not merely a diagnostic metric but the primary regret-like objective around which algorithm design is organized.

3. Relation to standard duality gaps and zero-sum game descent

The static duality gap remains the immediate precursor to D-DGap. In two-player zero-sum games with payoff matrix $x_t \in X$5, mixed strategies $x_t \in X$6, and standard regrets $x_t \in X$7, the duality gap is

$x_t \in X$8

with $x_t \in X$9 if and only if $y_t \in Y$0 is a Nash equilibrium (Fasoulakis et al., 31 Jan 2025). The paper proves that $y_t \in Y$1 is convex in bilinear zero-sum games and studies steepest descent directly on this convex merit function (Fasoulakis et al., 31 Jan 2025).

The directional derivative along $y_t \in Y$2 is

$y_t \in Y$3

and the update takes the form

$y_t \in Y$4

Variants include a fixed-$y_t \in Y$5 method, a decaying-$y_t \in Y$6 schedule, and a decaying $y_t \in Y$7 schedule (Fasoulakis et al., 31 Jan 2025).

The main convergence results are geometric decrease of the duality gap and iteration bounds of

$y_t \in Y$8

for the adaptive choice $y_t \in Y$9 (Fasoulakis et al., 31 Jan 2025). The same work reports that a fixed-support variant can outperform or match OGDA in large games with thousands of strategies per player (Fasoulakis et al., 31 Jan 2025).

Although this is not the same object as OCCO’s comparator-sequence D-DGap, it establishes the central methodological template: a gap quantity can be treated as a directly optimizable, geometrically decreasing merit function rather than only as a post hoc certificate. A plausible implication is that online D-DGap can be viewed as the time-accumulated analogue of this static equilibrium residual.

4. Proximal and dynamic duality gaps in GANs

In GANs, the standard duality gap for

$f_t : X \times Y \to \mathbb{R}$0

is

$f_t : X \times Y \to \mathbb{R}$1

and vanishes at a Nash equilibrium (Sidheekh et al., 2021, Sidheekh et al., 2020). The limitation emphasized in later work is that standard DG assumes Nash equilibria exist and, under practical estimation procedures, may fail to distinguish Nash from non-Nash critical points (Sidheekh et al., 2020).

To address the first issue, the proximal duality gap $f_t : X \times Y \to \mathbb{R}$2 was introduced: $f_t : X \times Y \to \mathbb{R}$3

$f_t : X \times Y \to \mathbb{R}$4

Here, $f_t : X \times Y \to \mathbb{R}$5 interpolates between Stackelberg-like and Nash-like notions: as $f_t : X \times Y \to \mathbb{R}$6, one recovers Nash equilibrium; as $f_t : X \times Y \to \mathbb{R}$7, Stackelberg equilibrium (Sidheekh et al., 2021). The key theorem is that

$f_t : X \times Y \to \mathbb{R}$8

for classic GAN, WGAN, and F-GAN formulations, with $f_t : X \times Y \to \mathbb{R}$9 instantiated by the appropriate divergence and $x$0 accounting for realizability (Sidheekh et al., 2021). Thus, if $x$1, then the distributional divergence also tends to zero (Sidheekh et al., 2021).

Within that framework, the possibility of dynamically adapting $x$2 is explicitly discussed “in the spirit of a ‘Dynamic Duality Gap’” (Sidheekh et al., 2021). The paper primarily studies fixed $x$3, but notes that dynamic or adaptive $x$4 could tune how local or global the monitored equilibrium is (Sidheekh et al., 2021). This usage does not define a new canonical formula for D-DGap, but it places “dynamic duality gap” in the context of adaptive equilibrium monitoring rather than only online regret.

A second line of GAN work studies the estimation problem itself. Standard gradient-based DG estimation can return values near zero at non-Nash saddle points because auxiliary optimizers initialized at the current parameters fail to move when local gradients vanish (Sidheekh et al., 2020). The proposed remedy is a perturbation-based estimate: initialize the worst-case discriminator and generator at locally perturbed parameters,

$x$5

with small random noise, then optimize and evaluate

$x$6

Over training time, $x$7 with perturbations was presented as a dynamic monitoring signal (Sidheekh et al., 2020). Experiments on mixture-of-Gaussians tasks and image datasets showed that perturbed DG saturates near zero only in true convergence, while remaining positive under mode collapse or divergence (Sidheekh et al., 2020).

These two GAN strands illustrate two distinct meanings of “dynamic” in D-DGap: adaptive regularization of the equilibrium concept (Sidheekh et al., 2021) and time-indexed monitoring of training trajectories via a robust DG estimate (Sidheekh et al., 2020).

5. D-gap functions in variational inequalities and error bounds

In the variational inequality literature, the D-gap function is a difference of regularized gap functions rather than a comparator-sequence regret. For a closed convex set $x$8, a locally Lipschitz mapping $x$9, and parameters $y$0, the regularized gap function is

$y$1

and the D-gap function is

$y$2

It vanishes if and only if $y$3 solves the variational inequality problem (Li et al., 2022).

The paper derives exact formulas for the subderivative, regular subdifferential, limiting subdifferential, and Clarke subdifferential of $y$4, for example

$y$5

and

$y$6

with $y$7 (Li et al., 2022).

The paper then proves sufficient and necessary conditions for the Kurdyka-Łojasiewicz inequality and the error bound property. Under appropriate monotonicity-like conditions,

$y$8

The sequence produced by a derivative-free descent algorithm with inexact Armijo line search converges linearly; more specifically, merit values converge Q-linearly, iterates converge R-linearly, and the sequence has finite length (Li et al., 2022).

The paper states that “the dynamic duality gap (D-gap) value is proportional (in square root) to the distance to the solution set” under these conditions (Li et al., 2022). Here, “dynamic” refers to the evolving merit value along an algorithmic run. This suggests a broader unifying viewpoint: D-DGap can function as a time-indexed certificate whose square-root scale controls actual solution distance.

6. Broader meanings of “dynamic” in duality-gap research

Outside online games and GANs, “dynamic” is also used to describe parameter-varying or time-recursive duality-gap phenomena. In dynamic stochastic optimization, sufficient conditions for the absence of a duality gap are expressed through extended dynamic programming equations and a recession-linearity condition that generalizes no-arbitrage (Pennanen et al., 2011). The value function

$y$9

is closed and admits optimal solutions under these conditions, implying zero duality gap for the corresponding primal–dual pair (Pennanen et al., 2011). Although this is not the OCCO definition of D-DGap, it is a genuinely dynamic duality setting because the primal–dual structure is embedded in multistage stochastic recursion.

In conic linear programming, the duality gap function

$(u_t, v_t)$0

tracks how the primal–dual gap changes under perturbations of $(u_t, v_t)$1 and $(u_t, v_t)$2 (Zalinescu, 2022). In the perturbed Gale example,

$(u_t, v_t)$3

showing that the gap may remain strictly positive under perturbation even in structured settings (Zalinescu, 2022). Similarly, for semidefinite programs with nonzero finite duality gap, a limiting value function

$(u_t, v_t)$4

parameterizes the values attained along perturbation directions and, under singularity degree one for both primal and dual, continuously fills the interval between $(u_t, v_t)$5 and $(u_t, v_t)$6 (Tsuchiya et al., 2023).

These works do not formalize D-DGap as a standard term, but they show that “dynamic” may refer to three distinct mechanisms: temporal comparator drift (Meng et al., 9 Sep 2025), adaptive equilibrium regularization (Sidheekh et al., 2021), and perturbation- or recursion-induced evolution of primal–dual discrepancy [(Pennanen et al., 2011); (Zalinescu, 2022); (Tsuchiya et al., 2023)].

7. Conceptual significance and recurring misconceptions

A common misconception is that any duality gap is merely a scalar convergence diagnostic. The recent literature shows otherwise. In OCCO, D-DGap is the primary benchmark for online decision-making against arbitrary comparator sequences and admits minimax-optimal upper and lower bounds (Meng et al., 9 Sep 2025). In zero-sum games, direct descent on the duality gap yields algorithms with geometric decrease and competitive performance relative to OGDA (Fasoulakis et al., 31 Jan 2025). In GANs, the choice of equilibrium concept and the estimation procedure fundamentally affect whether a gap quantity is informative (Sidheekh et al., 2021, Sidheekh et al., 2020). In variational inequalities, D-gap functions act as merit functions with subdifferential structure, KL inequalities, and explicit error bounds (Li et al., 2022).

A second misconception is that “dynamic duality gap” has a unique universal definition. The evidence does not support that interpretation. The OCCO definition is exact and comparator-based (Meng et al., 9 Sep 2025). In GANs, “Dynamic Duality Gap” is used more loosely to describe adaptive $(u_t, v_t)$7 schemes or time-varying perturbed DG monitoring (Sidheekh et al., 2021, Sidheekh et al., 2020). In variational inequalities, dynamic language refers to the trajectory of D-gap values during descent (Li et al., 2022). The phrase therefore names a family of related constructions rather than a single invariant across all disciplines.

A plausible implication is that D-DGap is best regarded as a design pattern: one starts from a primal–dual or min–max inconsistency measure, then renders it dynamic by allowing comparator drift, parameter schedules, perturbation paths, or iterative evolution. What remains invariant across domains is the role of the quantity as a bridge between optimization dynamics and equilibrium or optimality certification.

Setting	D-DGap object	Role
OCCO	$(u_t, v_t)$8	Dynamic benchmark against arbitrary comparator sequences
Zero-sum games	$(u_t, v_t)$9 and its descent dynamics	Static gap used as directly minimized merit function
GANs	$$D\text{-}DGap\,(u_{1:T}, v_{1:T}) := \sum_{t=1}^T \left[ f_t(x_t, v_t) - f_t(u_t, y_t) \right].$$0 or perturbed $$D\text{-}DGap\,(u_{1:T}, v_{1:T}) := \sum_{t=1}^T \left[ f_t(x_t, v_t) - f_t(u_t, y_t) \right].$$1 over time	Monitoring convergence under broader equilibria or robust estimation
Variational inequalities	$$D\text{-}DGap\,(u_{1:T}, v_{1:T}) := \sum_{t=1}^T \left[ f_t(x_t, v_t) - f_t(u_t, y_t) \right].$$2 along iterations	Merit function with error bounds and linear convergence consequences

The current research record therefore places Dynamic Duality Gap at the intersection of online learning, min–max optimization, equilibrium computation, and merit-function theory. Its most precise modern incarnation is the OCCO comparator-sequence metric (Meng et al., 9 Sep 2025), but its broader significance lies in demonstrating that duality-gap ideas can be made adaptive, trajectory-aware, and algorithmically central across multiple problem classes.