Contractive Propagation Condition

• Contractive Propagation Condition is a set of sufficient criteria that guarantee the quantitative shrinking of metrics across diverse systems, ensuring global stability.
• It is applied in various domains—from scalar conservation laws and Hamilton–Jacobi equations to interconnected control systems and neural architectures—by enforcing local, geometric, or algebraic conditions.
• The condition utilizes mechanisms such as flux limiters, dual Lyapunov inequalities, and spectral penalties to propagate contractivity and achieve convergence and robustness in complex dynamical environments.

The contractive propagation condition (CPC) refers to a set of sufficient criteria, appearing across diverse mathematical and applied domains, ensuring that contractivity—quantitative shrinking of the relevant metric or functional—propagates through a system or across interfaces. Rigorous CPCs have been established for conservation laws and Hamilton–Jacobi equations at spatial junctions, for the stability of interconnected control systems, for inertial flows in machine learning (notably, diffusion and neural ODE architectures), for Filippov systems with switching, for stochastic particle systems (via coupling), and even for fixed-point theorems in metric spaces. Although the specific technical forms differ, all CPCs enforce (or propagate) contractive properties through local, geometric, or algebraic conditions, often yielding strong global convergence, stability, or robustness results.

1. Contractive Propagation in Scalar Conservation Laws and Hamilton–Jacobi Equations

Consider scalar conservation laws on the real line with a space-discontinuous flux,

$$H(x,p) = H^\ell(p)\,\mathbf{1}_{x<0} + H^r(p)\,\mathbf{1}_{x>0},$$

where $H^\ell$, $H^r$ are strictly concave, $C^1$, and Lipschitz, and each vanish at $0$ and $R$. For such laws, a semigroup $S_{CL}$ on $L^\infty_{loc}(\mathbb{R})$ is deemed to propagate contraction if:

• $L^1$-contractivity holds globally: for any initial data $\rho_0^1,\rho_0^2$,

$$\|S_{CL}(t,\rho_0^1)-S_{CL}(t,\rho_0^2)\|_{L^1} \leq \|\rho_0^1-\rho_0^2\|_{L^1}.$$

• Finite speed of propagation and scale invariance are enforced.
• Away from $x=0$ the dynamics match classical Kruzhkov entropy solutions.

The contractive propagation condition at the junction is formalized by the existence of a "maximal, $L^1$-dissipative, complete germ" $\mathcal{G}_A$, determined by a unique flux limiter $A \in [0,A_{\max}]$, ensuring all traces $(\rho(t,0^-),\rho(t,0^+))\in\mathcal{G}_A$ a.e.\ $t$. This germ enforces the Rankine–Hugoniot relation $H^\ell(q_-) = H^r(q_+)$ and a dissipation inequality ensuring the global $L^1$ contraction. Analogous results hold for Hamilton–Jacobi equations via flux-limited viscosity solutions determined by the same $A$; the entire structure of the solution semigroup is thus dictated by the propagation of contractivity across the junction (Cardaliaguet, 2024).

2. Contractive Propagation in Interconnected (Almost-ISS) Control Systems

In the context of feedback connections of input-to-state stable (ISS) systems, the CPC appears as a hybrid of local contraction (small-gain) and propagation through non-contractive regions:

• On each "contractive interval," corresponding to a loop gain $\gamma(s) < s$, derivative-based small-gain inequalities ensure local ISS with Lyapunov functions.
• In the gaps between intervals, where contractivity fails, one uses a Rantzer-type density-propagation (dual Lyapunov) inequality to ensure that almost all trajectories escape non-contractive regions and enter contractive bands.

The interplay of these conditions yields "almost ISS": for all but a measure-zero set of initial conditions,

$\limsup_{t\to\infty}|x(t)| \leq \gamma(|u|_\infty)$

for some $\mathcal{K}_\infty$-gain $\gamma$, i.e.\ contractive propagation yields global stability from local and dual conditions (Feketa et al., 2016).

3. Algorithmic Enforcements: Contractive Diffusion and Neural ODEs

Modern generative and control architectures deploy CPCs to ensure robustness:

• Contractive Diffusion Policies (CDP): In stochastic or ODE-based diffusion models for policy generation, CDP enforces at every time $t$ that

$$\lambda_\text{max}\left(J_F(x, t) + J_F(x, t)^\top\right) < 0$$

where $J_F$ is the Jacobian of the reverse SDE flow. This is reduced to a constraint on the symmetric part of the score Jacobian:

$$\lambda_\text{max}(J_s^\mathrm{sym}(x, t)) < -\frac{f(t)}{h(t)}$$

and realized in training as a spectral (or Frobenius-norm) penalty added to the loss, guaranteeing all flow trajectories contract and thus improving robustness and variance reduction (Abyaneh et al., 2 Jan 2026).

• Contractive Hamiltonian Neural ODEs (CH-NODEs): For NODEs, explicit Hamiltonian structure and scalar damping $\gamma$ are tuned to ensure

$\operatorname{sym}\,A(t, \xi) \preceq -\epsilon I$

for the Jacobian $A$. This uniform contractivity ensures non-exploding gradients and provable shrinkage of adversarial or random perturbations (Zakwan et al., 2022).

4. Propagation Across Piecewise-Smooth and Switching Systems

For multi-modal piecewise smooth (PWS) and Filippov systems, CPC requires that:

• Each smooth mode is infinitesimally contractive with respect to a (possibly non-Euclidean) metric.
• Every switching or sliding manifold admits a "jump-contractive" condition:

$$\mu_Q\left((f_j(x) - f_i(x)) \nabla H_{ij}(x)\right) \leq 0$$

• For intersecting switching manifolds, additional linear combinations and compatibility at intersection points must also be contractive.

The result is global exponential convergence of Filippov solutions, regardless of the switching sequence, provided the composite contraction/propagation criteria are met (Liu et al., 18 Dec 2025).

5. Contractive Propagation Conditions for Markovian and Metric Fixed-Point Systems

• Markov Couplings for Kinetic Systems: For the Nanbu $N$-particle system (Maxwell types), a coupling is constructed so that the expectation of squared distance contracts in expectation, modulo a sharp geometric inequality involving parallelogram areas. This yields, under $2+\alpha$ moment bounds, a power-law contraction in Wasserstein distance with no exponential regime: the system admits only quasi-contractive propagation due to moment limitations (Rousset, 2013).
• Contractive-Iterate Fixed-Point Theorems: In metric spaces, a mapping $T$ may not be globally contractive, but the existence, for each $x$, of a depth (possibly varying) after which $T^{n_1}$ is contractive on a neighborhood, suffices to propagate contraction and recover periodic points or unique limits under iterated application (Karaibryamov et al., 2011).

6. Extensions: Higher-Order and Structured Contraction Propagation

Recent work generalizes CPC to higher-order geometric objects:

• $k$-Contraction and Structured Interconnections: For systems admitting multistationarity, the propagation of $k$-contraction (e.g., $k=2$) entails that certain area (not just length) measures in tangent space contract. New small-gain theorems propagate $2$-contractivity from each subsystem and their coupling, expressed through additive compounds and Metzler-matrix inequalities, yielding convergence of all bounded orbits to equilibria even in the presence of multiple stable points (Ofir et al., 2024).

---

Across all these domains, the contractive propagation condition offers a powerful, unifying principle: the enforcement of local or interface-specific shrinking behavior, together with auxiliary structural or duality criteria, suffices to guarantee strong forms of asymptotic stability, robustness, and synchronization in nonlinear, hybrid, stochastic, and high-dimensional dynamics. The technical form of the CPC adapts to the governing equations, geometry, and metrics of the system, but always secures global behavior via the local-to-global propagation of contraction.