Intrinsic Stability in Complex Systems

• Intrinsic stability is defined by internal structural properties, utilizing tools like Lipschitz matrices, continuous stability matrices, and intrinsic geometries rather than external trajectory analysis.
• It is delay-independent and ensures exponential convergence in systems, preserving stability even in the presence of bounded time-varying delays or non-unique minimizer sets.
• The concept spans diverse applications—from dynamical networks and autoregressive reasoning to thermoacoustics and geometric stability—providing versatile criteria for system robustness.

Intrinsic stability is a domain-dependent term for stability properties defined at the level of a system’s internal structure, solution set, or intrinsic geometry rather than solely by a particular trajectory, selector, or external resonance. Across the cited literature, it denotes a quantitative contractivity condition for dynamical networks with arbitrary bounded time-delays, a set-valued continuity property for non-unique convex empirical risk minimization, a process-level limit on autoregressive reasoning, several internal instability mechanisms in flames and thermoacoustics, and geometric stability notions formulated through intrinsic coordinates or intrinsic flat convergence (Reber et al., 2019, Bounja et al., 25 Jan 2026, Liao, 6 Feb 2026, Varillon et al., 2023, Allen et al., 2020).

1. Terminological scope and recurring structure

The term does not have a single universal definition. In some works it is explicitly stronger than ordinary global stability; in others it is the correct stability notion only after replacing a point-valued output by a set-valued correspondence; in others it refers to instability mechanisms that persist independently of cavity modes or external forcing. A common pattern is that the stability object is elevated from a single observed trajectory to a more structural entity: a Lipschitz influence matrix, a minimizer correspondence, a stochastic transition kernel, an intrinsic coordinate system, or a Riemannian/current-space geometry (Reber et al., 2019, Bounja et al., 25 Jan 2026, Varillon et al., 2023, Aguirregabiria et al., 2014, Allen et al., 2020).

Context	Stability object	Criterion or mechanism
Dynamical networks	Lipschitz matrix of $(F,X)$	$\rho(A)<1$
Convex non-strict ERM	$S(D)=C_D=\arg\min L_D$	PK upper semicontinuity
Autoregressive reasoning	decision advantage $\rho_t$	$\rho(L)\le \rho_0 e^{-\gamma L}$
Delay differential equations	continuous stability matrix $\mathcal M$	$\alpha(\mathcal M)<0$
Intrinsic flat geometry	integral current spaces	volume preserving intrinsic flat convergence

This plurality is essential. In particular, “intrinsic stability” is not a synonym for asymptotic stability, spectral stability, or robustness in general. The non-unique ERM paper argues that algorithmic stability of a selected predictor can differ from intrinsic stability of the full minimizer correspondence (Bounja et al., 25 Jan 2026). The thermoacoustic literature uses “intrinsic” to distinguish a flame’s internal feedback loop from chamber acoustics (Varillon et al., 2023). Geometric papers use “intrinsic” to indicate coordinate systems or convergence notions defined on the manifold itself rather than through an ambient linearization (Aguirregabiria et al., 2014, Allen et al., 2020).

2. Contractive dynamical systems and delay-independent stability

In dynamical networks, intrinsic stability is introduced as a stronger form of global stability. For a continuous map $F:X\to X$ on a product of complete metric spaces, a matrix $A=[a_{ij}]$ is a Lipschitz matrix if

$$d_i(F_i(\mathbf{x}),F_i(\mathbf{y}))\le \sum_{j=1}^n a_{ij}\, d_j(x_j,y_j).$$

The network is intrinsically stable when $\rho(A)<1$0. This criterion is structural: it controls worst-case componentwise gain rather than only the asymptotic behavior of the undelayed system. The same work shows that intrinsic stability is preserved under delay lifting, so constant delays do not change whether the system is intrinsically stable, and extends the result to switched systems and arbitrary bounded time-varying delays through a joint spectral radius criterion (Reber et al., 2019).

The switched-system formulation replaces a single Lipschitz matrix by a Lipschitz set $\rho(A)<1$1 and uses the joint spectral radius $\rho(A)<1$2. If $\rho(A)<1$3, all orbits become exponentially close; if the switched system has a shared fixed point, every orbit converges to it. For time-varying delay families, the paper proves that if the original undelayed network is intrinsically stable with $\rho(A)<1$4, then every delayed instance has a globally attracting lifted fixed point, independently of whether the delays are periodic, stochastic, or otherwise, provided they are bounded (Reber et al., 2019).

A closely related but continuous-time formulation appears for nonlinear delay differential equations. There the central object is the continuous stability matrix

$\rho(A)<1$5

and the system is intrinsically stable if $\rho(A)<1$6, where $\rho(A)<1$7 is the spectral abscissa. This single matrix criterion yields global exponential attraction of a fixed point for any admissible bounded time-varying delays, including piecewise uniformly continuous delays with arithmetic discontinuities. The proof uses comparison with a positive linear DDE, approximation by grid-aligned delays, finite-dimensional switched matrices, and isospectral reduction to preserve spectral-radius information while reducing large block matrices to a smaller matrix close to $\rho(A)<1$8 (Leishman et al., 22 Aug 2025).

These two lines of work make a strong distinction between ordinary stability and intrinsic stability. In both, the latter is delay-independent and computationally anchored in a matrix criterion—$\rho(A)<1$9 or $S(D)=C_D=\arg\min L_D$0—rather than in a delay-specific Lyapunov construction (Reber et al., 2019, Leishman et al., 22 Aug 2025).

3. Learning, optimization, and sequential inference

For convex but non-strict empirical risk minimization, intrinsic stability is defined at the level of the full minimizer set

$S(D)=C_D=\arg\min L_D$1

not at the level of a chosen predictor. The paper identifies Painlevé–Kuratowski upper semicontinuity as the intrinsic stability notion: $S(D)=C_D=\arg\min L_D$2 Equivalently, if $S(D)=C_D=\arg\min L_D$3 and $S(D)=C_D=\arg\min L_D$4 with $S(D)=C_D=\arg\min L_D$5 strongly in $S(D)=C_D=\arg\min L_D$6, then $S(D)=C_D=\arg\min L_D$7. Mosco convergence of the losses together with local boundedness of minimizers yields PK-u.s.c., continuity of the minimal value $S(D)=C_D=\arg\min L_D$8, and consistency of vanishing-gap near-minimizers. With quadratic growth,

$S(D)=C_D=\arg\min L_D$9

the paper derives the deviation bound

$\rho_t$0

Strong convexity is presented as a canonical stabilization mechanism for selectors, but the paper explicitly distinguishes stability of a regularized selector from intrinsic stability of the original non-unique ERM correspondence (Bounja et al., 25 Jan 2026).

In autoregressive reasoning, intrinsic stability refers to preservation of directional alignment toward the correct conclusion along a single-path execution. The key quantity is the decision advantage

$\rho_t$1

Under a contraction assumption on the transition kernel,

$\rho_t$2

Theorem A gives

$\rho_t$3

and therefore a critical horizon

$\rho_t$4

The paper argues that stable long-horizon reasoning consequently requires discrete segmentation and graph-like execution structures such as directed acyclic graphs, because a purely linear uninterrupted autoregressive chain has a finite stability horizon. Synthetic tasks, a branch-free chain sensitivity analysis, and TextWorld experiments are reported as exhibiting performance cliffs consistent with this prediction (Liao, 6 Feb 2026).

A distinct but related line in deep reinforcement learning treats stability as an explicit training target. “Global stabilisation via Intrinsic Fine Tuning” introduces a Stabilising Markov Decision Process that preserves the original dynamics, observation space, and action space but replaces the task reward with an intrinsic reward defined from a reference trajectory: $\rho_t$5

$\rho_t$6

On MuJoCo Playground / DeepMind Control Suite locomotion tasks, stability is measured by the Maximal Lyapunov Exponent, and the reported pattern is that GIFT reduces MLE by roughly an order of magnitude while often preserving comparable task reward. The paper is explicit that reward does not uniformly improve: for example, PPO + GIFT on Humanoid Walk falls from $\rho_t$7 to $\rho_t$8 (Young et al., 25 Apr 2026).

4. Internal instability mechanisms in combustion and resonant media

In thermoacoustics, intrinsic instability denotes a mode that is not coupled to the natural acoustic resonances of the combustor. For a laminar premixed slit flame, the intrinsic thermoacoustic feedback loop is closed by vorticity generated in the slit boundary layer by acoustic waves penetrating the slit. The least stable eigenvalue of the Linearized Reactive Flow operator determines whether the base flow is intrinsically stable or unstable: $\rho_t$9 indicates instability, and $\rho(L)\le \rho_0 e^{-\gamma L}$0 gives the pseudo-frequency. Adjoint modes identify the receptive region; in this case the adjoint vorticity is large in the shear layer enclosing the boundary layer of the slit and along the flame sheet. Acoustic penetration correlates strongly with instability, with the acoustic flux entering the slit reported as $\rho(L)\le \rho_0 e^{-\gamma L}$1 for $\rho(L)\le \rho_0 e^{-\gamma L}$2 and $\rho(L)\le \rho_0 e^{-\gamma L}$3 for $\rho(L)\le \rho_0 e^{-\gamma L}$4 (Varillon et al., 2023).

A complementary flame literature studies intrinsic instabilities of the flame front itself. “Eigenvalue-based Linear Stability Analysis of Intrinsic Instabilities in Laminar Flames” formulates a generalized eigenvalue problem

$\rho(L)\le \rho_0 e^{-\gamma L}$5

for perturbations of a 1D base flame. The method reproduces the analytical Darrieus–Landau dispersion relation and, for a finite-thickness reactive Navier–Stokes flame, matches DNS dispersion relations and eigenmode structures while reducing computational effort by up to eight orders of magnitude. The reported timings per wavenumber are $\rho(L)\le \rho_0 e^{-\gamma L}$6 s on 8 cores for DNS 2D, $\rho(L)\le \rho_0 e^{-\gamma L}$7 s on 228 cores for DNS 3D, $\rho(L)\le \rho_0 e^{-\gamma L}$8 s on 1 core for GEVP-LSA 2D, and $\rho(L)\le \rho_0 e^{-\gamma L}$9 s on 1 core for GEVP-LSA 3D (Kaiser et al., 30 Mar 2026).

Near internal cavity resonance in intrinsic Josephson-junction stacks, stability of dynamic coherent states is limited by two mechanisms. The paper distinguishes a short-scale instability, associated with regions where the time-averaged Josephson coupling becomes negative, from a long-scale instability caused by resonance excitation of fast modes at finite wave vectors. The homogeneous state in modulated junctions is typically unstable with respect to short-scale alternating phase deformations unless the Josephson current is completely suppressed in one half of the stack. The alternating-kink state is stable with respect to such deformations because its effective sign-changing coupling removes the negative regions of the local restoring term. The unstable wave-vector range created by the long-scale mechanism shrinks as resonance is approached and as in-plane dissipation increases; for finite-height stacks, the stability frequency range near resonance increases with decreasing the height (Koshelev, 2010).

5. Geometry, relativity, and intrinsic-flat formulations

In general relativity, intrinsic stability can refer to the persistence of intrinsic conserved quantities under perturbation when evaluated in intrinsic coordinates. For the linearized Kerr metric, the vanishing intrinsic energy previously established for Schwarzschild is shown to remain stable under a slow rotational perturbation. The construction uses intrinsic Gauss coordinates satisfying the asymptotic Gauss/synchronous condition $\mathcal M$0, $\mathcal M$1, together with asymptotically rectilinear behavior. In these coordinates, the Weinberg linear and angular 4-momenta vanish on a fixed slice $\mathcal M$2: $\mathcal M$3 The result is described as linear stability of the creatable character of Schwarzschild under slow rotation (Aguirregabiria et al., 2014).

A separate geometric tradition uses intrinsic flat convergence to formulate stability of manifolds and rigidity theorems. For a compact, connected, oriented manifold with boundary endowed with a reference metric $\mathcal M$4 and continuous metrics $\mathcal M$5, one theorem states that if

$\mathcal M$6

the diameters are uniformly bounded, the volumes converge, and the boundary metrics converge in $\mathcal M$7, then $\mathcal M$8 in volume preserving intrinsic flat sense. The proof uses a $\mathcal M$9-doubling construction, almost-everywhere distance convergence, large good subsets, and an explicit flat-distance estimate (Allen et al., 2020).

This framework underlies stability versions of the positive mass theorem. For asymptotically flat graphical hypersurfaces in Euclidean space, if the ADM mass tends to zero, designated regions $\alpha(\mathcal M)<0$0 converge in intrinsic flat distance to Euclidean balls, and with basepoints $\alpha(\mathcal M)<0$1 one obtains pointed intrinsic flat convergence to $\alpha(\mathcal M)<0$2 (Huang et al., 2014). For asymptotically hyperbolic graphical manifolds, if $\alpha(\mathcal M)<0$3 in the class $\alpha(\mathcal M)<0$4, then the compact regions $\alpha(\mathcal M)<0$5 converge in intrinsic flat distance to $\alpha(\mathcal M)<0$6, and corresponding pointed balls converge as well (Pacheco et al., 2023). A later paper formalizes intrinsic flat convergence of points and its compatibility with Gromov–Hausdorff convergence, providing the point-convergence machinery needed to strengthen and correct parts of the graphical positive-mass arguments (Huang et al., 2020).

These geometric results show that “intrinsic” can mean either intrinsic coordinates for conserved quantities or intrinsic current-space convergence for manifolds. In both cases the stability statement is attached to the manifold itself, not to an ambient Euclidean approximation.

6. Materials, lattices, biochemical networks, and intrinsic geometry of engineered systems

In phosphorene nanotubes, intrinsic stability is governed by intrinsic strain built into the rolled-up phosphorene lattice. The strain energy per atom is

$\alpha(\mathcal M)<0$7

and larger $\alpha(\mathcal M)<0$8 means larger intrinsic strain, especially in the hoop direction. The paper reports that armchair PNTs are more stable than zigzag PNTs, larger diameter increases thermal stability and stiffness, and the compressive buckling mode of armchair tubes changes from column buckling to shell buckling as the diameter/length ratio increases. A striking example is that $\alpha(\mathcal M)<0$9 armPNT survives up to about $F:X\to X$0 K, whereas $F:X\to X$1 zigPNT survives only about $F:X\to X$2 K (Liao et al., 2015).

For intrinsic localized modes in a discrete nonlinear Schrödinger lattice with competing power nonlinearities, stability is classified in the anticontinuum limit by finite codes built from the smaller and larger nonzero site states $F:X\to X$3 and $F:X\to X$4. All-large same-sign codes are spectrally and orbitally stable. All-small alternating codes are spectrally stable for any $F:X\to X$5, but orbitally stable only for $F:X\to X$6. Mixed stacked codes satisfy explicit criteria such as

$F:X\to X$7

for spectral and orbital stability, while other spectrally stable codes carry eigenvalues of negative Krein signature and are therefore vulnerable to Hamiltonian-Hopf bifurcations under perturbation (Alfimov et al., 16 Nov 2025).

In metabolic-cycle analysis, intrinsic stability denotes topology-driven stability across broad kinetic realizations rather than fine tuning of rate constants. Structural Kinetic Modeling rewrites the Jacobian as

$F:X\to X$8

For a class of single-input, single-output non-autocatalytic cycles, the characteristic polynomial has all nonzero coefficients of the same sign, so all real eigenvalues are negative and any instability, if it occurs, must be oscillatory. Computational sampling found no unstable cases in about $F:X\to X$9 simulations per cycle length from 3 to 8 and for each cofactor position. By contrast, a small autocatalytic cycle is only conditionally stable and is analyzed by the Routh–Hurwitz condition $A=[a_{ij}]$0, $A=[a_{ij}]$1, $A=[a_{ij}]$2 for its cubic characteristic polynomial (Reznik et al., 2010).

A different use of the term appears in intrinsic geometric analyses of engineered parameter spaces. In power systems, the metric tensor is defined as a Hessian of effective power,

$A=[a_{ij}]$3

and the determinant and scalar curvature are interpreted as local and global indicators of reliability and voltage stability. Positive minors and positive determinant indicate stability, while divergence of curvature or vanishing determinant indicates a critical point (Gupta et al., 2010). In controller design, the Hessian metric of the controller response

$A=[a_{ij}]$4

encodes Gaussian fluctuations. Constant mismatch factor controllers define a flat intrinsic manifold with zero scalar curvature, whereas variable mismatch factor controllers are curved and develop a cusp-like singularity near $A=[a_{ij}]$5 (Bellucci et al., 2011).

Intrinsic stability is therefore best understood as a family of structural stability notions rather than a single invariant definition. The relevant stability object may be a fixed point, a minimizer set, a reasoning process, a flame eigenmode, a manifold, a lattice code, or a parameter-space metric. This suggests that comparisons across fields require identifying first what is being stabilized—trajectory, correspondence, dispersion relation, conserved quantity, or geometric current space—and only then asking how “intrinsic” the stability certificate is.