Gradient Equilibrium (GEQ): A Multifaceted Overview

• Gradient Equilibrium (GEQ) is a multifaceted concept that uses equilibrium conditions to extract gradient information across online learning, physical dynamics, and galactic metallicity studies.
• In online learning, GEQ ensures that the long-run average of subgradients converges to zero, offering insights into bias correction, residual orthogonality, and the balance between constant and decaying step sizes.
• In physical and astrophysical contexts, GEQ underpins gradient extraction from equilibrium responses in oscillator networks and explains the Milky Way’s radial metallicity gradient via local equilibrium abundances.

Searching arXiv for the cited GEQ papers and closely related work. Search query: "Gradient Equilibrium in Online Learning: Theory and Applications" Gradient Equilibrium (GEQ) is a term used in multiple, non-equivalent senses in recent research. In online learning, it denotes the requirement that the average of gradients of losses along an online trajectory converges to zero. In equilibrium-propagation work on physical and energy-based dynamical systems, it denotes an exact or asymptotically exact correspondence between equilibrium displacement and a parameter gradient. In Galactic chemical evolution, it denotes an equilibrium explanation of the Milky Way’s radial metallicity gradient, where the observed spatial gradient is interpreted as the imprint of a radius-dependent local equilibrium abundance. The coexistence of these usages suggests that GEQ is presently a domain-local term rather than a single standardized concept (Angelopoulos et al., 14 Jan 2025, Ahmadi, 11 Apr 2026, Johnson et al., 2024).

1. Terminological scope

Across the cited literature, the common structural motif is equilibrium plus gradient information, but the objects involved differ sharply. In online learning, the relevant equilibrium object is the long-run average realized subgradient. In oscillator and equilibrium-propagation systems, the relevant object is the displacement between free and weakly nudged equilibria. In Galactic chemical evolution, the relevant object is the local equilibrium metallicity of the interstellar medium as a function of Galactocentric radius. Related mathematical literatures study “gradient-like” or GENERIC dynamics, where Lyapunov structure and relaxation to equilibria are central, but those works do not use the same GEQ definition (Angelopoulos et al., 14 Jan 2025, Ahmadi, 11 Apr 2026, Johnson et al., 2024, Kraaij et al., 2017, Benaim, 2014).

Usage	Equilibrium object	Gradient-related quantity
Online learning	Long-run online iterate sequence	$\frac{1}{T}\sum_{t=1}^T g_t(\theta_t)$
Kuramoto / equilibrium propagation	Free and nudged equilibria	Phase displacement as gradient w.r.t. $\omega_i$
Galactic chemical evolution	Local ISM equilibrium metallicity	Radial metallicity gradient

This suggests that encyclopedia treatment of GEQ must distinguish among at least three technically separate frameworks. Confusion most often arises when the phrase “gradient equilibrium” is taken to imply a single abstract theory spanning all of them. The primary cross-cutting similarity is not formal equivalence but the use of equilibrium structure to expose a gradient, gradient signal, or spatial gradient.

2. GEQ in online learning

In online learning, GEQ was introduced as a new perspective in which a sequence of iterates $\{\theta_t\}_{t\ge 1}$ achieves gradient equilibrium when

$$\frac{1}{T}\sum_{t=1}^T g_t(\theta_t)\to 0 \qquad \text{as } T\to\infty.$$

Here $g_t(\theta_t)$ is a generalized subgradient of $\ell_t$ at $\theta_t$. The condition is about the norm of the average gradient, not the average of gradient norms, and it is sequential and deterministic: it holds along a single data sequence, not in expectation (Angelopoulos et al., 14 Jan 2025).

A central claim of this framework is that GEQ is distinct from sublinear regret. The same paper proves and illustrates that GEQ does not imply no regret, and no regret does not imply GEQ. The absolute-loss example shows vanishing regret with persistent average subgradient, while the squared-loss example shows that even with optimal average squared loss, bias can remain large. This is one of the main reasons the framework is presented as a different objective from competition with the best fixed predictor.

The basic algorithmic mechanism is unusually simple. For online gradient descent with constant step size $\eta$,

$\theta_{t+1} = \theta_t - \eta g_t(\theta_t),$

the paper gives the exact telescoping identity

$$\frac{1}{T}\sum_{t=1}^T g_t(\theta_t) = \frac{\theta_1-\theta_{T+1}}{\eta T}.$$

Hence GEQ holds whenever the iterates are bounded, or more generally satisfy $\omega_i$0. The paper states the sharper characterization

$\omega_i$1

for constant-step-size gradient descent. This is the theoretical basis for the claim that constant step sizes, rather than decaying step sizes, can be sufficient for GEQ.

The theory of restorative losses explains when such iterate control is available. A loss is $\omega_i$2-restorative if

$\omega_i$3

for every generalized subgradient $\omega_i$4. In the univariate $\omega_i$5-Lipschitz and $\omega_i$6-restorative case, the paper gives

$\omega_i$7

In the multivariate zero-curvature setting it gives

$\omega_i$8

which is generally $\omega_i$9 and still sufficient for GEQ.

The statistical interpretations are one reason the framework attracted attention. For squared loss, GEQ becomes asymptotic unbiasedness of predictions. For quantile loss, it becomes coverage calibration,

$\{\theta_t\}_{t\ge 1}$0

For GLM-type losses, it yields orthogonality of residuals and covariates in aggregate. For pairwise preference prediction and Elo-style models, it yields unbiased long-run predicted win rates for each player. The same paper applies post hoc gradient updates to debias black-box predictions under arbitrary distribution shift, to calibrate predicted quantiles under distribution shift, and to obtain unbiased Elo scores for pairwise preference prediction (Angelopoulos et al., 14 Jan 2025).

3. GEQ, Blackwell approachability, regret minimization, and calibration

A later paper formalizes GEQ as a tuple $\{\theta_t\}_{t\ge 1}$1 with decision set $\{\theta_t\}_{t\ge 1}$2 and vector-field class $\{\theta_t\}_{t\ge 1}$3. At each round, the learner picks $\{\theta_t\}_{t\ge 1}$4, nature picks $\{\theta_t\}_{t\ge 1}$5, and the learner observes $\{\theta_t\}_{t\ge 1}$6. In the constrained setting, the paper defines GEQ by the existence of normal-cone corrections:

$\{\theta_t\}_{t\ge 1}$7

For unconstrained $\{\theta_t\}_{t\ge 1}$8, the normal cone is $\{\theta_t\}_{t\ge 1}$9, and the criterion reduces to the vanishing norm of the average vector field (Lee et al., 25 Jun 2026).

The central result is that GEQ is equivalent to Blackwell approachability in the algorithmic sense. A Blackwell approachability problem can always be solved using queries to a black-box GEQ oracle, with no asymptotic loss in the oracle’s error rate, and vice versa. Together with known equivalences between approachability, regret minimization, and calibration, this yields the chain

$$\frac{1}{T}\sum_{t=1}^T g_t(\theta_t)\to 0 \qquad \text{as } T\to\infty.$$0

at the level of black-box reductions. The paper is explicit, however, that this does not mean the objectives are identical: GEQ error and regret are incomparable objectives even though the frameworks are algorithmically equivalent.

The geometric bridge is especially transparent in the unconstrained case. The paper encodes GEQ as a BA instance with $$\frac{1}{T}\sum_{t=1}^T g_t(\theta_t)\to 0 \qquad \text{as } T\to\infty.$$1, $$\frac{1}{T}\sum_{t=1}^T g_t(\theta_t)\to 0 \qquad \text{as } T\to\infty.$$2, $$\frac{1}{T}\sum_{t=1}^T g_t(\theta_t)\to 0 \qquad \text{as } T\to\infty.$$3, and target set $$\frac{1}{T}\sum_{t=1}^T g_t(\theta_t)\to 0 \qquad \text{as } T\to\infty.$$4. Then GEQ becomes approachability of the singleton $$\frac{1}{T}\sum_{t=1}^T g_t(\theta_t)\to 0 \qquad \text{as } T\to\infty.$$5. Conversely, for conic target sets $$\frac{1}{T}\sum_{t=1}^T g_t(\theta_t)\to 0 \qquad \text{as } T\to\infty.$$6, a BA problem can be reduced to GEQ by taking $$\frac{1}{T}\sum_{t=1}^T g_t(\theta_t)\to 0 \qquad \text{as } T\to\infty.$$7 and defining vector fields from the BA payoffs. Non-conic targets are handled by conic lifting.

This equivalence has algorithmic consequences. The paper states that its reductions can be used to transfer refined guarantees, such as optimism and strong adaptivity, from regret minimization to GEQ. It also identifies Blackwell’s condition as the necessary and sufficient condition for GEQ once the problem is viewed through the BA lens, while restorativity remains a sufficient but not necessary condition for algorithmic tractability (Lee et al., 25 Jun 2026).

4. GEQ in equilibrium propagation and physical gradient computation

In a distinct line of work, GEQ denotes a gradient identity in dynamical systems at stable equilibrium. For a coupled Kuramoto oscillator network

$$\frac{1}{T}\sum_{t=1}^T g_t(\theta_t)\to 0 \qquad \text{as } T\to\infty.$$8

the paper proves that, under stable equilibrium and symmetric coupling, the physical phase displacement under weak output nudging is the gradient of the loss with respect to natural frequencies. The central identity is

$$\frac{1}{T}\sum_{t=1}^T g_t(\theta_t)\to 0 \qquad \text{as } T\to\infty.$$9

The equilibrium Jacobian is a weighted graph Laplacian, and when the reduced Jacobian is symmetric and nonsingular the implicit function theorem yields the same result as the energy-based equilibrium-propagation derivation. The paper reports cosine similarity exactly $g_t(\theta_t)$0 between the two-phase gradient and finite-difference gradient in the reported tables, and likewise for a computationally independent PyTorch autograd implementation. On sparse layered architectures, frequency learning outperforms coupling-weight learning among converged seeds at matched parameter count, $g_t(\theta_t)$1 versus $g_t(\theta_t)$2, while topology-aware spectral seeding removes the approximately $g_t(\theta_t)$3 convergence failure under random initialization, improving from $g_t(\theta_t)$4 to $g_t(\theta_t)$5 seeds on the primary task (Ahmadi, 11 Apr 2026).

A related paper extends Equilibrium Propagation from learning energy minima to learning energy gradients. Its mechanism, Gradient Equilibrium Propagation (GradEP), replaces hard input clamping with a spring potential

$g_t(\theta_t)$6

so that the visible units also evolve. At equilibrium, the displacement from the anchor encodes the learned velocity, with readout

$g_t(\theta_t)$7

This is then applied to flow matching for generative modelling in the system called FlowEqProp. The demonstration uses a two-hidden-layer MLP with $g_t(\theta_t)$8 parameters on the UCI Optical Recognition of Handwritten Digits dataset, trains with local equilibrium measurements and no backpropagation, and shows a flow matching loss decreasing smoothly from about $g_t(\theta_t)$9 to about $\ell_t$0 over $\ell_t$1 epochs. The paper also reports monotonic improvement across tested spring stiffness values, from $\ell_t$2 at $\ell_t$3 to $\ell_t$4 at $\ell_t$5, and notes that the time-independent energy landscape permits extended generation beyond the training horizon, with $\ell_t$6 producing sharper digit samples while very long integration such as $\ell_t$7 leads to over-sharpening (Gower, 9 Apr 2026).

These papers support a common interpretation of GEQ as native gradient extraction from equilibrium response. A common misconception is that such results are only approximate or only useful when optimization succeeds. The Kuramoto paper argues that random-seed failures are a loss-landscape or basin-of-attraction problem, not a gradient-computation problem, while the GradEP paper presents local equilibrium measurements as a substitute for explicit backpropagation in a velocity-field objective.

5. GEQ as an equilibrium explanation of the Milky Way radial metallicity gradient

In Galactic chemical evolution, the term is used differently. “The Milky Way Radial Metallicity Gradient as an Equilibrium Phenomenon: Why Old Stars are Metal-Rich” argues that the Milky Way’s radial metallicity gradient is best understood as a gradient equilibrium phenomenon: the interstellar medium at each Galactocentric radius rapidly evolves toward a local equilibrium metallicity, and the observed radial gradient is the spatial imprint of how that equilibrium varies with radius. Empirically, using the APOGEE DR17 AstroNN value-added catalog with selection cuts $\ell_t$8, $\ell_t$9, $\theta_t$0 K, $\theta_t$1 Gyr, $\theta_t$2 kpc, and $\theta_t$3 kpc, the paper studies a final sample of $\theta_t$4 red giant/red clump stars. Its core result is that the metallicity–radius relation is nearly unchanged for stars up to $\theta_t$5 Gyr old, with present-day slopes of $\theta_t$6 dex kpc$\theta_t$7 and $\theta_t$8 dex kpc$\theta_t$9. The median stellar age versus radius is fit with

$\eta$0

but the crucial empirical point is that age and metallicity do not show a strong joint evolution in normalization (Johnson et al., 2024).

The theoretical interpretation contrasts inside-out growth or non-equilibrium evolution with an equilibrium scenario in which each radius quickly reaches a local equilibrium abundance $\eta$1 and then evolves little after a few Gyr. A central relation in the near-equilibrium limit is

$\eta$2

and differentiation with radius gives

$\eta$3

The proposed mechanism is that stronger outflows at larger radius lower the local equilibrium metallicity and also shorten the processing timescale

$\eta$4

so the interstellar medium reaches equilibrium quickly, in about a few $\eta$5 to $\eta$6 Gyr depending on radius and model.

The fiducial multi-zone model is Exp2, with $\eta$7, $\eta$8 kpc, and $\eta$9. In this model, $\theta_{t+1} = \theta_t - \eta g_t(\theta_t),$0 quickly becomes nearly flat in time, the ISM metallicity reaches the equilibrium gradient $\theta_{t+1} = \theta_t - \eta g_t(\theta_t),$1 Gyr ago, and old stars therefore formed out of gas already close to the same radial metallicity pattern seen today. An outer-disk starburst or accretion-burst perturbation then relaxes back to the original equilibrium metallicity on a timescale of about $\theta_{t+1} = \theta_t - \eta g_t(\theta_t),$2 Gyr. The paper therefore concludes that the radial metallicity gradient is not primarily a fossil record of disk assembly timing but an equilibrium abundance pattern shaped by star formation, accretion, stellar mass return, and radius-dependent outflows (Johnson et al., 2024).

6. Related gradient-like equilibrium frameworks and conceptual boundaries

Two related papers clarify the broader mathematical landscape within which some GEQ usages can be situated. “Deriving GENERIC from a generalized fluctuation symmetry” shows that detailed balance at the level of path-space large deviations implies gradient flow, while a generalized fluctuation symmetry around a nonzero current $\theta_{t+1} = \theta_t - \eta g_t(\theta_t),$3 implies pre-GENERIC and, with an auxiliary energy variable, standard GENERIC:

$\theta_{t+1} = \theta_t - \eta g_t(\theta_t),$4

with degeneracy conditions

$\theta_{t+1} = \theta_t - \eta g_t(\theta_t),$5

In this framework, the free energy remains a Lyapunov functional for the dissipative part, and the reversible current transports the system without destroying the Lyapunov structure (Kraaij et al., 2017).

“On Gradient like Properties of Population games, Learning models and Self Reinforced Processes” studies simplex ODEs of the form

$\theta_{t+1} = \theta_t - \eta g_t(\theta_t),$6

and gives conditions under which the dynamics are gradient-like. The key ingredients are a strict Lyapunov function $\theta_{t+1} = \theta_t - \eta g_t(\theta_t),$7, an angle condition

$\theta_{t+1} = \theta_t - \eta g_t(\theta_t),$8

and identification of equilibria with critical points of $\theta_{t+1} = \theta_t - \eta g_t(\theta_t),$9. Under these assumptions, omega-limit sets and internally chain transitive sets consist of equilibria; in the real analytic case, every trajectory converges to an equilibrium; and in the reversible case the dynamics are $$\frac{1}{T}\sum_{t=1}^T g_t(\theta_t) = \frac{\theta_1-\theta_{T+1}}{\eta T}.$$0-close to a gradient vector field (Benaim, 2014).

These frameworks are not GEQ in the online-learning or equilibrium-propagation sense, but they show that equilibrium selection, Lyapunov descent, reversibility, and gradient structure recur across disparate domains. A plausible implication is that the phrase “gradient equilibrium” has become attractive wherever equilibrium behavior is rich enough to encode either asymptotic first-order optimality, a physically measurable gradient signal, or a stable spatial gradient. The main conceptual boundary is therefore terminological rather than merely notational: GEQ in online learning, GEQ in oscillator equilibrium propagation, and GEQ in Galactic chemical evolution should not be treated as interchangeable definitions.