Mode-Weighted Transport Certificates for State-Dependent Reflected Switching Diffusions

arXiv:2608.01992 · cs.CE, cs.SY, eess.SY · Submitted 2026-08-03 · Read on arXiv

Listen

Transcript

Introduction to the show: ident: Robotics Radio. Generated commentary on the latest robotics and control papers.

Rosa: We're looking at the paper: "Mode-Weighted Transport Certificates for State-Dependent Reflected Switching Diffusions".

Dev: The gist: State-dependent switching diffusions can be contractive in distribution even when some modes are individually expansive,

Rosa: First, who's behind it and why it matters.

Title and authors: Rosa: So we're looking at the paper, "Mode-Weighted Transport Certificates for State-Dependent Reflected Switching Diffusions," and who wrote it—Yutong Zhu and Ye Zhang. The title tells you a lot about what they're doing with these switching diffusions.

Dev: Yeah, it sounds like this work is about finding a way to prove contraction even when some individual modes are expanding on their own. That’s the big hook there, Rosa.

Rosa: Exactly, because normally if one mode was expansive, you’d just expect the whole system to blow up. But these authors are showing you how to get contraction in distribution even with that complexity.

Taro: For me, I'm interested in how they handle that complexity when the world actually misbehaves during those switches. Does their proof hold up when things get really chaotic?

Dev: That’s a fair point, Taro. The paper says they use a computable transport-based condition to establish this contraction on R n and on compact convex domains with normal reflection.

Rosa: What that means is they're giving us a concrete way to check for stability, not just relying on some abstract theory. It’s about developing a computable transport-based condition for such contraction.

Taro: So it's like they built a safety check that works even when the dynamics are switching around wildly?

Dev: Pretty much. The core idea is to combine mode-dependent spatial weights with a discrete mode penalty while preserving spatial separation for cross-mode pairs, which they call the transport cost.

The paper's summary: Rosa: Okay, so the summary of this paper, "Mode-Weighted Transport Certificates for State-Dependent Reflected Switching Diffusions," is that they defined a hybrid state space X = times M, where is the continuous part and M is the finite set of modes.

Dev: They describe the continuous component as satisfying a normally reflected equation, which has terms for spatial drift, diffusion, and that reflection term from the boundary.

Rosa: And they define how these modes switch using conditional transition rates lambda ij(x), which depend on where you are in space and what mode you're currently in.

Taro: So it's not just a fixed set of switches; the rules for switching change based on position, which makes things much more complex to analyze than standard switching models.

Dev: Right. And then they talk about the generator core D N, and they show how to find elements phi = (phi one phi M) that satisfy certain boundary conditions on the domain <ref:2608.01992#pg1>.

Rosa: The key finding here is that this construction results in a conservative, nonexplosive strong Markov process when you interlace it with the bounded state-dependent jump clocks.

Taro: A conservative and nonexplosive process sounds good for stability tests because it means the system stays within its bounds over time.

Dev: That's the implication: they showed that using this hybrid construction, they can derive separate generator inequalities for same-mode and cross-mode configurations.

The paper's improvements: Rosa: Now let's talk about the specific improvements suggested by the authors in "Mode-Weighted Transport Certificates for State-Dependent Reflected Switching Diffusions." They focus on defining a transport cost c v,B(z,).

Dev: This cost combines mode-dependent spatial weights with a discrete mode penalty while preserving spatial separation for cross-mode pairs. It's defined as c v,B((x, i), (y, j)) = q v i x - q + beta ij + v j q - y.

Taro: The paper shows that the closed form of this cost simplifies nicely to a ij x - y + beta ij, where a ij is related to the minimum spatial weights. That’s a simplification we can use.

Rosa: Precisely. And then they have a discrepancy functional W v,B(mu,) that controls both the continuous first moment and the mismatch probability of the discrete modes.

Dev: They prove that if you have a Markov coupling kernel Q t that satisfies Z c v,B(z', ') Q t((z,), dz' d ') e-eta t c v,B(z,), then the discrepancy functional W v,B(mu t P t, mu P t) decays exponentially.

Taro: So they are essentially creating a test where if this cost function is small enough relative to some rate eta, you get exponential contraction of the measures. That’s a strong result for proving stability in these complex systems.

Conclusion: Rosa: To wrap up, the main result from "Mode-Weighted Transport Certificates for State-Dependent Reflected Switching Diffusions" is Theorem one <ref:2608.01992#pg1,Mode-Weighted Transport Certificates for State-Dependent Reflected Switching Diffusions>. It states that if there exist some positive v and a graph cost B such that certain conditions on the same-mode margin S i(v, B) and cross-mode cost C ij(x, y; v, B) hold, then the discrepancy functional decays exponentially.

Dev: They are constructing these certificates by using algorithms like Algorithm one which uses bisection to find the largest feasible eta by solving an linear programming problem <ref:2608.01992#pg1>.

Taro: What I see is that this moves beyond just a grid-only test; Proposition three gives a finite-mesh certificate, but then Proposition four shows how changes in parameters affect the residuals, which lets you extend it over the full domain.

Rosa: So they are using this finite-mesh condition to show something holds on all of because of how the interaction among these envelopes is exposed by Proposition four.

Dev: And they conclude that this mode-weighted cost turns state-dependent switching into an explicit pairwise contraction test, which is really useful for analyzing how complex systems behave.

Taro: It’s about turning a problem with state-dependent switching into a concrete, pairwise contraction test using the mode-weighted cost.

Yutong Zhu, Ye Zhang

Northwestern Polytechnical University

cs.CE, cs.SY, eess.SY

Submitted: 2026-08-03

Updated: 2026-10-17

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

The gist: The gist: State-dependent switching diffusions can be contractive in distribution even when some modes are individually expansive, and this paper develops a computable transport-based condition for

Key concepts

Hybrid State Process: This model combines a continuous stochastic process (like Brownian motion) with a finite-state mode that switches based on the continuous state. The switching rates depend on both the current location and which mode is active. This structure allows for complex dynamics where the environment changes as the system moves.
Mode-Weighted Transport Cost: This cost function measures how much it costs to move from one hybrid state $(x, i)$ to another $(y, j)$. It combines a spatial distance term with mode-dependent penalties. Crucially, it is designed to preserve spatial separation between different modes while penalizing mismatches between the continuous position and the discrete mode.
Generator Certificate: This is a mathematical tool used to prove contraction by analyzing the infinitesimal generator of the process. The paper constructs specific inequalities (a certificate) involving same-mode and cross-mode costs. Satisfying these inequalities guarantees that the probability distribution contracts exponentially over time.
Normal Reflection: This condition applies when the domain $\Omega$ has a boundary, like a convex shape. Instead of letting particles leave the domain, they are reflected back into it according to their normal direction at the boundary. This reflection term is mathematically incorporated into the governing equations to ensure solutions remain within the defined space.

Terminology

Generator — 1× in this paper · explained in 2 episodes

Summary

The gist: State-dependent switching diffusions can be contractive in distribution even when some modes are individually expansive, and this paper develops a computable transport-based condition for such contraction on R n and on compact convex domains with normal reflection.

Hybrid State Process

The analysis begins by defining the hybrid state space X = omega × M, where the continuous component satisfies the normally reflected equation dXt = fσt(Xt) dt + √2ν dWt − n(Xt) dKt, and the mode has conditional transition rates Pr(σt+∆ = j Ft, Xt = x, σt = i) = λij (x)∆ + o(∆), j!= i <ref:2608.01992#pg3>. The reflection term is absent for R n where Kt ≡ 0 <ref:2608.01992#pg3>. This construction results in a conservative, nonexplosive strong Markov process when interlaced with the bounded state-dependent jump clocks <ref:2608.01992#pg3>.

Generator and Weak Solutions

The generator core DN has elements φ = (φ1,..., φM) satisfying ∂nφi = 0 on ∂omega <ref:2608.01992#pg4>. On R n, the generator is (Aφ)i(x) = fi(x)⊤∇φi(x) + ν∆φi(x) + X j!=i λij (x)[φj (x) − φi(x)] <ref:2608.01992#pg4>. A narrowly continuous curve µt = (µ1,t,..., µM,t) is a weak measure solution if it satisfies the integral equation involving the generator core DN <ref:2608.01992#pg4>. Proposition 1 establishes that µt = µ0Pt satisfies this identity under Assumption 1 <ref:2608.01992#pg4>.

Mode-Weighted Hybrid Transport

The transport cost cv,B(z, z˜) is defined as cv,B((x, i),(y, j)) = inf q∈omega vi∥x − q∥ + βij + vj∥q − y <ref:2608.01992#pg4>. Lemma 1 shows that the closed form of this cost is cv,B((x, i),(y, j)) = aij∥x − y∥ + βij <ref:2608.01992#pg4>. The discrepancy functional Wv,B(µ, µ˜) controls both the continuous first moment and the mismatch probability of the discrete modes <ref:2608.01992#pg4>. Lemma 2 proves that if a Markov coupling kernel Qt satisfies Z cv,B(z', z˜')Qt((z, z˜), dz' d˜z') ≤ e−ηtcv,B(z, z˜), then Wv,B(µPt, µP˜ t) ≤ e−ηtWv,B(µ, µ˜) <ref:2608.01992#pg4>.

Generator Certificate for Exponential Contraction

The main result is Theorem 1: If there exist v > 0 and a graph cost B such that Si(v, B) ≤ −ηvi and Cij (x, y; v, B) ≤ −η(aij∥x − y∥ + βij), then for all µ0, µ˜0 ∈ P1(X), Wv,B(µ0Pt, µ˜0Pt) ≤ e−ηtWv,B(µ0, µ˜0) <ref:2608.01992#pg5>. The certificate involves the same-mode margin Si(v, B) and cross-mode cost Cij (x, y; v, B), which are constructed from envelopes defined in Assumptions 2 through 4 <ref:2608.01992#pg5>.

Certificate Computation and Conservatism

The synthesis procedure involves Algorithm 1, which uses bisection to find the largest feasible η by solving an LP of Proposition 2 <ref:2608.01992#pg8>. Proposition 3 provides a finite-mesh certificate, stating that if Fij (ˆx, yˆ) ≤ −Hij δ for every (ˆx, yˆ) ∈ S, then (23) holds on all of omega 2 <ref:2608.01992#pg9>. The interaction among the envelopes is exposed by Proposition 4 <ref:2608.01992#pg9>, which shows how changes in parameters affect the residuals.

Numerical Studies

The one-dimensional analytical certificate validates the distributional computation with specific drifts and switching rates, showing that certain families of weights cannot certify a positive rate <ref:2608.01992#pg10>. The planar three-mode certificate synthesis demonstrates the full procedure, yielding a final set of weights and costs that certify contraction over the entire domain <ref:2608.01992#pg11>.

Conclusion

The mode-weighted cost turns state-dependent switching into an explicit pairwise contraction test <ref:2608.01992#pg11>. Fixed weight order and fixed decay rate leave a semi-infinite linear feasibility problem, which is solved by enumerating orders, bisection, constraint generation, and the explicit mesh buffer to yield a full-domain certificate <ref:2608.01992#pg11>.

--- Page 1 ---

Mode-Weighted Transport Certificates for State-Dependent Reflected Switching Diffusions Yutong Zhu and Ye Zhang Abstract— State-dependent switching diffusions can be contractive in distribution even when some modes are individually expansive. We develop a computable transport-based condition for such contraction on R n and on compact convex domains with normal reflection <ref:2608.01992#pg2>. The transport cost combines mode-dependent spatial weights with a discrete mode penalty while preserving spatial separation for cross-mode pairs <ref:2608.01992#pg2>. Using synchronous coupling of the Brownian motions and maximal coupling of the state-dependent jump clocks, we derive separate generator inequalities for same-mode and cross-mode configurations <ref:2608.01992#pg2>. Convex normal reflection contributes a nonpositive finite-variation term, so the same conditions apply to the associated no-flux Fokker-Planck-Kolmogorov system <ref:2608.01992#pg2>. Their feasibility guarantees global pairwise exponential contraction of the Markov semigroup and weak measure solutions, with unit prefactor and an explicit rate <ref:2608.01992#pg2>. For a fixed ordering of the mode weights and a prescribed decay rate, the conditions are affine in the spatial weights and graph costs and form a semiinfinite linear feasibility problem <ref:2608.01992#pg2>. A finite-mesh condition with a Lipschitz margin certifies the inequalities over the full domain <ref:2608.01992#pg2>.

I. INTRODUCTION

Switching diffusions couple a continuous stochastic state to a finite-state mode <ref:2608.01992#pg3>. When the transition rates depend on position, the discrete environment and the continuous dynamics form a feedback loop <ref:2608.01992#pg3>. Well-posedness and Feller properties for such processes are developed in [1], [2], while Lyapunov, matrix, and recurrence criteria show that switching can stabilize a system even if some modes are not stable in isolation [3], [4], [5]. The state dependence that makes this mechanism useful also complicates a direct stability test: the rate matrix changes along the continuous trajectory, and two copies at different positions no longer share the same jump clocks <ref:2608.01992#pg3>.

II. REFLECTED SWITCHING DIFFUSION AND THE FPK SYSTEM

A. Hybrid state process

Let M = M and let omega be either R n or a compact convex domain with C2 boundary <ref:2608.01992#pg4>. For x ∈ ∂omega, let n(x) be the outward unit normal <ref:2608.01992#pg4>. The hybrid state space is X = omega × M <ref:2608.01992#pg4>. The continuous component satisfies the normally reflected equation dXt = fσt(Xt) dt + √2ν dWt − n(Xt) dKt, (1) and the mode has conditional transition rates Pr(σt+∆ = j Ft, Xt = x, σt = i) = λij (x)∆ + o(∆), j!= i <ref:2608.01992#pg4>.

--- Page 2 ---

and Kt is continuous, nondecreasing, and supported on the boundary: K0 = 0, Z t 0 1 for X s∈omega◦ dKs = 0 <ref:2608.01992#pg3>. For Ω = R n, the reflection term is absent and Kt ≡ 0 <ref:2608.01992#pg3>. The switching does not reset Xt. We set λii(x) = −P j!=i λij (x) <ref:2608.01992#pg3>.

B. Generator, no-flux equation, and weak solutions

Let DN be the generator core from Assumption 1 <ref:2608.01992#pg4>. Its elements φ = (φ1,..., φM) are twice continuously differentiable and satisfy ∂nφi = 0 on ∂omega <ref:2608.01992#pg4>. On R n, the boundary condition is void and a compactly supported smooth core may be used <ref:2608.01992#pg4>. On this class, the generator is (Aφ)i(x) = fi(x)⊤∇φi(x) + ν∆φi(x) + X j!=i λij (x)[φj (x) − φi(x)] <ref:2608.01992#pg4>.

--- Page 3 ---

The corresponding density equation is ∂tρi = −∇· (fiρi) + ν∆ρi + X M j=1 λjiρj, i ∈ M, (5) with no-flux boundary condition fiρi − ν∇ρi⊤n = 0 on ∂omega <ref:2608.01992#pg4>.

A narrowly continuous curve µt = (µ1,t,...

Improvements for AI systems

Bold headers for state-dependent switching diffusions allow AI systems to exhibit global pairwise exponential contraction of the Markov semigroup and weak measure solutions (Theorem 1, page 4), enabling robust convergence guarantees regardless of individual mode instability.
The transport cost structure, combining mode-dependent spatial weights with a discrete mode penalty while preserving spatial separation for cross-mode pairs, allows the AI to utilize a Wasserstein-type discrepancy that controls both the continuous first moment and the mismatch probability of the discrete modes (Lemma 1, page 3).
The certificate allows for explicit synthesis procedures, enabling AI systems to achieve full-domain certificate rather than a grid-only test (Section V), meaning stability proofs can be extended beyond computational meshes by solving a semi-infinite linear feasibility problem.
The system can be designed to satisfy the same-mode one-sided contraction condition, where the margin is defined as Si(v, B) = − civi + X k̸=i (vk − vi) + λik + (vk − vi)−λik (Equation 20), ensuring that even expansive modes are stabilized by cross-mode interactions.
The AI can learn transition rates depending on both state coordinates, as demonstrated in the planar three-mode example, allowing for highly complex, non-linear dynamics described by rate fields like λ12(x) = 3.20 + 0.50x1 + 0.35x2 (Equation 63).