Riemannian Subgradient Flow Analysis

• Riemannian Subgradient Flow is a continuous dynamical system for nonsmooth nonconvex optimization using a barrier-induced metric over constrained manifolds.
• The framework unifies Hessian-barrier and mirror descent methods as discretizations of the same interior dynamics with a clear geometric interpretation.
• Random perturbation strategies and strict complementarity conditions are employed to avoid spurious equilibria and ensure convergence to true stationary points.

Searching arXiv for the cited paper and closely related work so the article can be grounded in current literature. Riemannian subgradient flow denotes, in the formulation studied in (Ding et al., 21 Jul 2025), a continuous-time dynamical system for nonsmooth nonconvex optimization over constraints of the form $x\in \overline C\cap \mathcal M$. The construction starts from an open convex set $C$ equipped with a barrier-induced Riemannian metric and restricts the motion to a smooth manifold $\mathcal M$, yielding a differential inclusion that remains strictly within the interior $C\cap \mathcal M$. Within this framework, the Hessian barrier method and the mirror descent scheme appear as discrete approximations of the same continuous flow, while spurious stationary points are interpreted as stable equilibria that do not correspond to stationary points of the original constrained problem (Ding et al., 21 Jul 2025).

1. Geometric problem formulation

The optimization problem is

$$\text{minimize } f(x) \qquad \text{subject to } x\in \overline C\cap \mathcal M .$$

Here $C\subset \mathbb R^n$ is open, convex, and admits a $C^2$ barrier $\phi:C\to\mathbb R$ such that $\phi$ is Legendre on $C$, $C$0 is positive definite on $C$1, and $C$2 extends continuously to $C$3. The manifold is

$C$4

a $C$5 embedded manifold of dimension $C$6 given by equality constraints with LICQ everywhere. The objective $C$7 is locally Lipschitz and path-differentiable, so its Clarke subdifferential $C$8 is well-defined (Ding et al., 21 Jul 2025).

The barrier $C$9 induces the Riemannian metric

$\mathcal M$0

This metric is defined on $\mathcal M$1, and hence on $\mathcal M$2. In this sense, the geometry is neither Euclidean nor arbitrary: it is the Hessian-Riemannian geometry generated by the barrier. A plausible implication is that the interior-point character of the dynamics is not an ancillary numerical feature, but is encoded directly in the metric structure.

2. Differential inclusion and dual characterization

For $\mathcal M$3, the tangent and normal spaces are

$\mathcal M$4

The $\mathcal M$5-orthogonal projection onto the tangent space is denoted

$\mathcal M$6

The Clarke subdifferential is defined by

$\mathcal M$7

The Riemannian subgradient flow, called in the paper the “interior mirror-descent flow,” is the differential inclusion

$\mathcal M$8

for almost every $\mathcal M$9 (Ding et al., 21 Jul 2025).

Via the change of variable $C\cap \mathcal M$0, this is equivalently written as

$C\cap \mathcal M$1

This second inclusion highlights the Bregman-subgradient character of the dynamics. The equivalence is established by selecting $C\cap \mathcal M$2, observing that the minimal $C\cap \mathcal M$3-norm correction in $C\cap \mathcal M$4 to $C\cap \mathcal M$5 is exactly $C\cap \mathcal M$6, and then invoking KKT in the dual variable $C\cap \mathcal M$7. This formulation makes explicit that the flow is simultaneously a projected Riemannian subgradient system and a dual-space mirror-type inclusion.

3. Relation to Hessian-barrier and mirror-descent discretizations

The framework identifies two classical iterative constructions as discretizations of the same continuous flow (Ding et al., 21 Jul 2025).

The first is the Hessian-barrier method: $C\cap \mathcal M$8 When $C\cap \mathcal M$9 is smooth, this is the classical Hessian-barrier method; more generally, $$\text{minimize } f(x) \qquad \text{subject to } x\in \overline C\cap \mathcal M .$$0 is replaced by a subgradient.

The second is the mirror-descent scheme in the dual or Bregman viewpoint. The relevant Bregman divergence is

$$\text{minimize } f(x) \qquad \text{subject to } x\in \overline C\cap \mathcal M .$$1

and the update is

$$\text{minimize } f(x) \qquad \text{subject to } x\in \overline C\cap \mathcal M .$$2

with $$\text{minimize } f(x) \qquad \text{subject to } x\in \overline C\cap \mathcal M .$$3. Under smoothness of $$\text{minimize } f(x) \qquad \text{subject to } x\in \overline C\cap \mathcal M .$$4, this yields exactly the proximal mapping in dual space.

In both schemes the step-sizes satisfy

$$\text{minimize } f(x) \qquad \text{subject to } x\in \overline C\cap \mathcal M .$$5

The continuous formulation therefore serves as a unifying object: the Hessian-barrier method appears as an Euler-type discretization in the primal Riemannian geometry, whereas mirror descent appears in the dual Bregman geometry. This suggests that the two methods are not merely analogous, but are different coordinate expressions of one interior dynamical system.

4. Stable-set geometry and stationary points

A central distinction in the framework is between stable points of the flow and stationary points of the original constrained problem (Ding et al., 21 Jul 2025). The stable set of the flow is

$$\text{minimize } f(x) \qquad \text{subject to } x\in \overline C\cap \mathcal M .$$6

while the stationary set of the original problem is

$$\text{minimize } f(x) \qquad \text{subject to } x\in \overline C\cap \mathcal M .$$7

The paper proves that $$\text{minimize } f(x) \qquad \text{subject to } x\in \overline C\cap \mathcal M .$$8, and that strict inclusion is possible. In particular, examples based on the entropy barrier on $$\text{minimize } f(x) \qquad \text{subject to } x\in \overline C\cap \mathcal M .$$9 exhibit boundary points in $C\subset \mathbb R^n$0 that fail complementarity. These are the spurious stationary points observed in Hessian barrier and mirror descent schemes: they satisfy the projected-subgradient condition induced by the interior geometry, but not the KKT-type condition of the original constrained problem.

The long-term behavior of trajectories is described in terms of the $C\subset \mathbb R^n$1-limit set. Any limit point of a trajectory $C\subset \mathbb R^n$2 lies in $C\subset \mathbb R^n$3. At the same time, if $C\subset \mathbb R^n$4, then there exists a neighborhood $C\subset \mathbb R^n$5 of $C\subset \mathbb R^n$6 such that whenever $C\subset \mathbb R^n$7 enters $C\subset \mathbb R^n$8, it must exit $C\subset \mathbb R^n$9 in finite time. Consequently, a convergent trajectory cannot land on a spurious point $C^2$0, because it would then have to leave every sufficiently small neighborhood of its putative limit. The continuous-time viewpoint therefore reinterprets spurious fixed points of the discrete methods as stable-set points of the flow that fail KKT, rather than as purely discretization-specific artifacts.

5. Avoidance of spurious equilibria and random perturbation

Two sufficient conditions are given for excluding spurious $C^2$1-limit points, provided the relevant complementarity structure is present (Ding et al., 21 Jul 2025).

The first is a strict complementarity condition. At $C^2$2, one has

$C^2$3

which forces $C^2$4. The corresponding theorem states that under strict violation of complementarity the flow is repelled from any spurious $C^2$5 in all directions, hence no limit can occur there; if all limit points satisfy complementarity, then they lie in $C^2$6.

The second is an isolated-point condition: if $C^2$7 consists only of isolated points, then continuity together with connectedness of the $C^2$8-limit set forces the limit into $C^2$9.

When $\phi:C\to\mathbb R$0 contains continua, the paper proposes a random perturbation strategy. Both objective and constraints are perturbed according to

$\phi:C\to\mathbb R$1

with $\phi:C\to\mathbb R$2 chosen randomly in a small ball. By the Morse–Sard theorem, the new stable set $\phi:C\to\mathbb R$3 is almost surely discrete, so the isolated-point argument applies to the perturbed flow. The perturbed dynamics then converges to a true stationary point of the perturbed problem, and, back in the original problem, this yields an $\phi:C\to\mathbb R$4-approximate KKT point. A plausible implication is that perturbation is not merely a numerical regularizer here, but a device for converting non-isolated spurious structure into a regime where the dynamical exclusion arguments become effective.

6. Discrete Riemannian subgradient methods

The continuous theory induces two interior-point algorithms (Ding et al., 21 Jul 2025). In both, all iterates are kept in $\phi:C\to\mathbb R$5 via a retraction

$\phi:C\to\mathbb R$6

for example via geodesic or simple affine-space projection.

The Riemannian Hessian-barrier subgradient method takes inputs $\phi:C\to\mathbb R$7, step-sizes $\phi:C\to\mathbb R$8 with $\phi:C\to\mathbb R$9, tolerances $\phi$0, and martingale-noise $\phi$1. At iteration $\phi$2, it picks $\phi$3, observes noise $\phi$4, computes the corresponding projected direction in the barrier-induced metric, and retracts according to

$\phi$5

Under standard boundedness, closed-graph of $\phi$6, noise$\phi$7 conditions, and the weak Sard-type assumption on critical values of $\phi$8, stochastic-approximation theory yields two conclusions: cluster points lie in $\phi$9 and $C$0 converges; under complementarity or isolated conditions, any cluster point in fact lies in $C$1. If the small random perturbation is applied at the start, the perturbed iterates almost surely converge to a true stationary point of the perturbed problem, hence to an $C$2-approximate stationary point for the original problem.

The mirror-descent variant is discretized in the dual or Bregman form: $C$3 followed by retraction if needed. Although verifying the fine interpolation condition is more delicate near $C$4, the paper states that in many kernels, including separable entropy on $C$5, the discrete path is an asymptotic solution of the flow and therefore inherits the same avoidance and convergence guarantees under the same sufficient conditions.

In summary, the framework identifies a barrier-induced Hessian-Riemannian metric, defines a projected subgradient flow on $C$6, interprets Hessian-barrier and mirror-descent schemes as discretizations of that flow, and explains spurious fixed points as stable-set equilibria that fail KKT. The resulting algorithms remain interior-point in nature and inherit the asymptotic behavior of the continuous dynamics.