--- title: 'AltGDA: Alternating Gradient Descent-Ascent' url: https://www.emergentmind.com/topics/alternating-gradient-descent-ascent-altgda type: topic --- # AltGDA: Alternating Gradient Descent-Ascent Searching arXiv for recent and foundational papers on Alternating Gradient Descent-Ascent. Alternating Gradient Descent–Ascent (AltGDA) is a class of single-loop first-order methods for minimax optimization and game dynamics in which the two blocks are updated sequentially rather than simultaneously. In its canonical form for a smooth two-player problem $\min_x \max_y f(x,y)$, the method performs a descent step in $x$ using $(x_t,y_t)$ and then an ascent step in $y$ using the fresh iterate $x_{t+1}$ and the stale iterate $y_t$. This alternation distinguishes AltGDA from simultaneous GDA and is the source of several theoretical phenomena reported across the literature: bounded regret and recurrent cycling in unconstrained bilinear zero-sum games [1907.04392], near-optimal local linear convergence in strongly convex–strongly concave problems [2102.09468], improved global iteration complexity over simultaneous updates in smooth strongly-convex-strongly-concave settings [2402.10475], and single-loop convergence guarantees in nonconvex minimax regimes with proximal, stochastic, adaptive, Riemannian, or zeroth-order variants [2007.13605]. ## 1. Canonical formulation and update structure The standard AltGDA iteration for an unconstrained smooth minimax problem \[ \min_{x\in\mathbb{R}^n}\max_{y\in\mathbb{R}^m} f(x,y) \] takes the form \[ x_{t+1} = x_t - \eta \nabla_x f(x_t,y_t), \qquad y_{t+1} = y_t + \eta \nabla_y f(x_{t+1},y_t), \] or, with separate step-sizes, \[ x_{t+1} = x_t - \eta_x \nabla_x f(x_t,y_t), \qquad y_{t+1} = y_t + \eta_y \nabla_y f(x_{t+1},y_t). \] This sequential structure appears verbatim in the strongly convex–strongly concave analysis of Alt-GDA [2102.09468] and in the global SCSC analysis comparing alternating and simultaneous updates [2402.10475]. In bilinear zero-sum games with payoff $f(x_1,x_2)=x_1^\top A x_2$, the alternating update becomes \[ x_1^{t+1} = x_1^t + \eta_1 A x_2^t,\qquad x_2^{t+1} = x_2^t - \eta_2 A^\top x_1^{t+1}, \] which was introduced as “sequential” or “alternating” play in the finite-regret analysis of Bailey, Gidel, and Piliouras [1907.04392]. In constrained matrix games, the same alternation is combined with Euclidean projection: \[ x^{t+1} = \Pi_{\mathcal X}(x^t-\eta A^\top y^t),\qquad y^{t+1} = \Pi_{\mathcal Y}(y^t+\eta A x^{t+1}), \] with convergence measured by the ergodic duality gap [2510.03855]. A major branch of the literature studies proximal AltGDA for composite objectives. For nonconvex–concave problems of the form \[ \min_x \max_y \Psi(x,y)=f(x)+\Phi(x,y)-h(y), \] the deterministic single-loop updates are \[ x_{k+1}=\prox_{\eta_x f}(x_k-\eta_x g_k),\qquad y_{k+1}=\prox_{\eta_y h}\bigl(y_k+\eta_y\nabla_y\Phi(x_{k+1},y_k)\bigr), \] with $g_k\in\partial_x\Phi(x_k,y_k)$ [2007.13605]. The same proximal sequential template recurs in later regularized and accelerated formulations, including Prox-AltGDAm [2112.11663] and AGDA+ [2406.14371]. The defining feature of AltGDA is therefore not a single objective class but the ordering of information flow: the ascent step is evaluated at the updated primal block. This suggests that AltGDA is best understood as an algorithmic pattern spanning Euclidean, constrained, proximal, stochastic, and manifold settings rather than a single fixed method. ## 2. Bilinear zero-sum games: finite regret, bounded orbits, and recurrence The 2019 analysis of unconstrained bilinear zero-sum games gave one of the sharpest early characterizations of alternation [1907.04392]. In the game \[ \max_{x_1\in\mathbb{R}^{k_1}} \min_{x_2\in\mathbb{R}^{k_2}} x_1^\top A x_2, \] AltGDA yields an energy–payoff identity for each block: \[ \frac{\|x_1^{t+1}\|^2-\|x_1^t\|^2}{\eta_1} = \langle x_1^{t+1}+x_1^t,\; A x_2^t\rangle, \] and \[ \frac{\|x_2^{t+1}\|^2-\|x_2^t\|^2}{\eta_2} = -\langle x_1^{t+1},\; A(x_2^{t+1}+x_2^t)\rangle. \] These identities imply a bounded-regret theorem: if Agent 1 plays AltGDA with fixed $\eta_1$, then against any sequence $x_2^t$, her regret after $T$ updates is bounded by \[ R_T \le \frac{\langle x_1^0-2x_1,\;x_1^0\rangle+\|x_1\|^2}{\eta_1}, \] regardless of how the opponent updates [1907.04392]. The result is notable because it holds with arbitrary fixed step-size rather than decaying step-sizes. When both agents use AltGDA, the same paper establishes an exactly conserved “perturbed energy” \[ E(t)\coloneqq \frac{\|x_1^t\|^2}{\eta_1}+\frac{\|x_2^t\|^2}{\eta_2}+\langle x_1^t,\;A x_2^t\rangle, \] which remains equal to $E(0)$ for all $t$ [1907.04392]. If $\sqrt{\eta_1\eta_2}\le 2/\|A\|$, then the iterates remain in a bounded region, and a complementary lower bound shows the joint norm is bounded away from $0$ whenever $E(0)>0$. The map is also volume-preserving because AltGDA is a composition of two block-triangular maps whose Jacobians each have determinant $1$; Poincaré’s recurrence theorem then implies that almost every initial condition returns arbitrarily close to its start infinitely often [1907.04392]. The constrained bilinear literature later shifted the emphasis from bounded cycles to ergodic convergence. For compact convex strategy sets, including matrix games on simplices, AltGDA with constant step-size achieves an $O(1/T)$ ergodic convergence rate when there is an interior Nash equilibrium, with \[ \operatorname{Gap}(\bar x,\bar y)\le \frac{9+4\eta\|A\|}{\eta T}, \] provided \[ \eta \le \frac{1}{(\|A\|\cdot \min_i x_i^*,\; \min_j y_j^*)} \] [2510.03855]. Without an interior equilibrium, the same work proves an $O(1/T)$ local convergence rate in a neighborhood $S_0$ under the constant-step condition $\eta\le 1/(2\|A\|)$ and a refined energy argument [2510.03855]. These two strands are complementary rather than contradictory. In unconstrained bilinear dynamics, AltGDA generically produces bounded recurrent orbits rather than convergence to equilibrium [1907.04392]. In compact constrained bilinear games, averaged iterates can nevertheless enjoy $O(1/T)$ ergodic gap decay [2510.03855]. A common misconception is that “alternation guarantees convergence” in bilinear games; the record is narrower. The unconstrained theory emphasizes boundedness and recurrence, whereas the constrained theory emphasizes averaged duality-gap convergence. ## 3. Why alternation differs from simultaneous GDA Several papers isolate the mechanism by which alternation changes the dynamics. In the bilinear setting, simultaneous GDA \[ x_1^{t+1}=x_1^t+\eta_1 A x_2^t,\qquad x_2^{t+1}=x_2^t-\eta_2 A^\top x_1^t \] is described as the naive “Euler” integrator of the continuous Hamiltonian system $\dot x_1=A x_2$, $\dot x_2=-A^\top x_1$. Its drawbacks are explicit: it does not conserve energy, iterates diverge from equilibrium, and the map is volume-expanding, leading to formally chaotic behavior and linear or unbounded regret. AltGDA is instead interpreted as a symplectic integrator of Verlet/leapfrog type: it exactly conserves a perturbed energy, preserves volume, and thereby avoids systematic drift away from equilibrium [1907.04392]. In strongly convex–strongly concave problems, the distinction appears through the Jacobian of the iteration map near the Nash equilibrium. For SCSC objectives with twice continuously differentiable $f$ satisfying \[ \nabla_{xx}^2 f(x,y)\succeq \mu_x I_n,\qquad \nabla_{yy}^2 f(x,y)\preceq -\mu_y I_m, \] and Hessian block norms bounded by $L_x,L_y,L_{xy}$, Alt-GDA with $\eta\le 1/(2L)$ satisfies \[ \rho(\nabla F_{\mathrm{alt}}(z^*))\le 1-\frac{1}{2\kappa}, \] where $L=\max\{L_x,L_y,L_{xy}\}$, $\mu=\min\{\mu_x,\mu_y\}$, and $\kappa=L/\mu$ [2102.09468]. Standard linearization then yields local linear convergence with iteration complexity $O(\kappa\log(1/\epsilon))$. The simultaneous counterpart admits a much slower worst-case spectral radius \[ \rho(\nabla F_{\mathrm{sim}}(z^*))\approx 1-\frac{1}{4\kappa^2}, \] corresponding to $O(\kappa^2\log(1/\epsilon))$ iterations [2102.09468]. A more global comparison is provided for twice-differentiable SCSC functions with Lipschitz gradients, where Alt-GDA uses step-sizes satisfying \[ \alpha \le \tfrac12\min\{1/L_x,\sqrt{\mu_y}/(L_{xy}\sqrt{L_x})\},\qquad \beta \le \tfrac12\min\{1/L_y,\sqrt{\mu_x}/(L_{xy}\sqrt{L_y})\}. \] Under these conditions, the iteration complexity is \[ K=O[(\kappa_x+\kappa_y+\kappa_{xy}(\sqrt{\kappa_x}+\sqrt{\kappa_y}))\log(1/\epsilon)], \] where $\kappa_x=L_x/\mu_x$, $\kappa_y=L_y/\mu_y$, and $\kappa_{xy}=L_{xy}/\sqrt{\mu_x\mu_y}$ [2402.10475]. The corresponding simultaneous bound is \[ \Theta[(\kappa_x+\kappa_y+\kappa_{xy}^2)\log(1/\epsilon)], \] and the paper states that the upper bound for Alt-GDA is strictly smaller than the lower bound for Sim-GDA [2402.10475]. The same work proposes Alternating-Extrapolation GDA (Alex-GDA), with extrapolation parameters $\gamma,\delta\ge 0$, subsuming Sim-GDA and Alt-GDA as special cases $(\gamma,\delta)=(0,1)$ and $(1,1)$ [2402.10475]. Alex-GDA attains \[ K=O[(\kappa_x+\kappa_y+\kappa_{xy})\log(1/\epsilon)] \] and is reported to match the extragradient lower bound while requiring two gradient evaluations per iteration instead of four [2402.10475]. This suggests that alternation can be viewed as an intermediate point in a broader design space of sequential and extrapolated schemes. ## 4. Nonconvex minimax theory: concavity, PL structure, and proximal variants Outside convex–concave settings, AltGDA is primarily analyzed through structure imposed on the maximization block. One influential line considers nonconvex–concave or nonconvex–PL objectives, where the value function \[ \Phi(x)\coloneqq \max_y f(x,y) \] or its composite analog is smooth or weakly convex enough to support first-order stationarity analysis. For nonconvex–PL minimax problems with Lipschitz gradient constant $l$ and PL constant $\mu$ in $y$, deterministic AGDA uses \[ x_{t+1}\leftarrow x_t-\eta_x\nabla_x f(x_t,y_t),\qquad y_{t+1}\leftarrow y_t+\eta_y\nabla_y f(x_{t+1},y_t). \] Choosing \[ \eta_x=\frac{1}{68\,l\,\kappa^2},\qquad \eta_y=\frac{1}{l},\qquad \kappa=l/\mu, \] yields \[ \|\nabla \Phi(\hat x)\|^2 = O(l\kappa^2\Delta/T), \] and hence $T=O(\kappa^2\epsilon^{-2})$ to reach $\|\nabla \Phi\|\le \epsilon$ [2112.05604]. The stochastic single-sample version attains sample complexity $O(\kappa^4\epsilon^{-4})$ under bounded variance [2112.05604]. The same paper emphasizes that AGDA matches the best $O(\kappa^2\epsilon^{-2})$ complexity under the weaker PL assumption instead of strong concavity. A related but stronger condition is the two-sided PL inequality. For unconstrained nonconvex-nonconcave $f$, if both blocks satisfy PL-type inequalities, AGDA with \[ \eta_y=\frac{1}{2\mu_2},\qquad \eta_x=\frac{\mu_1}{3673\,L^2} \] converges globally at a linear rate in the potential \[ \Phi(x,y)=g(x)-g^*+\gamma[g(x)-f(x,y)], \] with contraction factor \[ 1-\mu_1\mu_2/(3673L^2), \] where $g(x)=\max_y f(x,y)$ [2002.09621]. This extends linear convergence beyond strong convexity/concavity to a class of nonconvex-nonconcave minimax problems. For nonconvex–concave composite problems, the proximal single-loop AltGDA of Boţ and Böhm studies \[ \min_x\max_y \Psi(x,y)=f(x)+\Phi(x,y)-h(y), \] under $\rho$-weak convexity in $x$, concavity and $L_{\nabla\Phi}$-smoothness in $y$, bounded $\operatorname{dom}(h)$, and lower boundedness of $g(x)=\max_y \Psi(x,y)$ [2007.13605]. With \[ \eta_x=\min\Bigl\{\frac{\epsilon^4}{L_{\nabla\Phi}\rho^2 D_h^2 (L+L_f)^2},\; \frac{\epsilon^2}{\rho L^2}\Bigr\},\qquad \eta_y=\frac{1}{L_{\nabla\Phi}}, \] the deterministic algorithm finds an $\epsilon$-stationary point of the Moreau envelope $g_\lambda$ after \[ K=\mathcal O\Bigl(\max\Bigl\{\frac{\rho L^2}{\epsilon^2},\; \frac{\rho^2 L_{\nabla\Phi}D_h^2(L+L_f)^2}{\epsilon^4}\Bigr\}\Bigr) \] iterations [2007.13605]. In the stochastic case, with batch-size $M=O(\epsilon^{-2})$, one obtains $K=\mathcal O(\epsilon^{-4})$ for $\min_{k