Many nonlinear PDEs, such as reaction–diffusion, Allen–Cahn, Henon and Kohn–Sham models, have several steady states under the same parameters and boundary conditions. Recent solvers are good at finding them. A homotopy method that uses physics-informed neural networks, meaning networks trained to satisfy the equation, recovers 9 solutions of the Henon equation and 25 Gray–Scott patterns. Almost none of these solvers says which solutions a real system would actually settle into. When stability is reported, it comes from eigenvalues computed in whatever representation the solver happened to use: a mesh, a spectral basis, network outputs, physical units. Control theory has shown that such verdicts can depend on the coordinates. One classical stability test for a Kalman filter reads 0.73 (stable) in one coordinate frame and 1.52 (unstable) after a change of coordinates, for the same system (An Improved Stability Condition for Kalman Filtering with Bounded Markovian Packet Losses). The verdicts are least reliable near bifurcations, the points where branches of solutions meet and multiplicity is born. There the deciding eigenvalue is close to zero and its sign is at the mercy of discretization.
We propose to attach a certificate to every computed steady state. The certificate gives the number of unstable directions and a guaranteed size of the region of attraction, and it is built so that it cannot change when the coordinates change. The tool is Lyapunov–Schmidt reduction: near a singular point, it splits the problem into a few "critical" directions, where the linearization nearly vanishes, and a well-behaved remainder. We handle the remainder with a metric-optimizing stability test that is invariant under changes of coordinates. We handle the critical directions with the exact low-dimensional reduced equation instead of its linearization.
Core Claim
A stability verdict for one of many coexisting steady states can be made intrinsic to the PDE rather than to the solver's chart. It can also stay informative right up to a bifurcation, where linearization-based certificates degenerate. Two design choices make this possible.
The first choice is how the problem is split. Take a computed state $u^*$ with linearization $L = D_uF(u^*)$. Let $\Pi$ be the spectral (Riesz) projection onto the eigenvalues of $L$ within distance $\delta$ of zero, meaning the projection built from $L$'s own eigenvectors and adjoint eigenvectors rather than an $L^2$-orthogonal projection. Under any invertible change of coordinates $T$, $\Pi$ transforms as $T\Pi T^{-1}$. As a result, both the split and the counts derived from it are properties of the system.
The second choice concerns the part of the spectrum away from zero. On the range of $I-\Pi$, the eigenvalues are separated from the imaginary axis by a certified gap. We count the unstable ones there, and we search for a Lyapunov metric $P$ by solving a linear matrix inequality (LMI):
$$ P \succ 0,\qquad (I-\Pi)^{\top}\!\left(L^{\top}P + PL + 2\gamma P\right)(I-\Pi) \preceq 0 . $$
Here $\gamma>0$ is the certified decay rate on the complement. Feasibility of this LMI is invariant under similarity, in the same sense as the Kalman result: if $P$ works in coordinates $x$, then $T^{-\top}PT^{-1}$ works in coordinates $Tx$.
On the $m$-dimensional range of $\Pi$, we do not use the eigenvalue sign. We use the reduced equation
$$ \dot a = g(a) := \Pi F\!\left(u^* + \Phi a + w^*(a)\right). $$
Here $\Phi$ is a basis of the critical subspace, $a\in\mathbb{R}^m$ are its coordinates, and $w^*(a)$ is the complement correction obtained from the implicit function theorem.
The quantitative prediction that makes the reduction worth its cost is this. At a supercritical pitchfork at distance $\mu$ from the bifurcation, the stable bifurcated state has critical eigenvalue of order $\mu$. A linearization certificate needs $|\lambda| r > C r^2$, so its certified basin radius scales like $r \sim \mu$. The reduced cubic normal form $\dot a = \mu a - a^3$ has an actual basin extending to the neighbouring saddle at distance $\sqrt{\mu}$. A certificate built from $g$ should therefore recover the exponent $1/2$ instead of $1$.
What Exists and What Is Missing
Solution-multiplicity work supplies the objects to certify but no stability verdicts. Three solvers belong here:
- HomPINNs tracks branches by homotopy continuation with physics-informed networks and reports no stability. - An Adaptive Orthogonal Basis Method finds 10 and 20 solutions of 2D Allen–Cahn/Gray–Scott problems with spectral collocation and trust-region solves, again with no stability. - Newton Informed Neural Operator learns Newton steps and explicitly breaks down where the Jacobian is singular.
We use these solvers as sources of candidate solutions. The fact that they represent solutions in very different ways is exactly what we exploit to test invariance.
The closest prior work pairs Lyapunov–Schmidt reduction with a stability judgment. Holst, Hu and Lu's DFT symmetry-breaking study reduces the Kohn–Sham equations at the critical exchange strength to a pitchfork. It then judges stability from Hessian eigenvalues on a fixed finite-element mesh. This is one bifurcation, in one chart, with no certificate and no region of attraction. We reuse their model as a test case where the ground truth is known.
Two further reduction papers supply method but not the result we want. The bubble-tower construction uses the reduction to prove existence of multiple solutions and leaves their dynamical stability open. The Moore–Greitzer reduction works in a fixed Fourier basis. Both confirm that the reduced equation carries the relevant dynamics. Neither asks whether conclusions survive a change of chart.
Coordinate-invariant stability work supplies the metric side. ComputeLDF shows that aligning coordinates with the system's real Jordan modes turns Euclidean over-approximation into visible contraction, cutting simulations on a 4D robot arm from 1159 to 81. It assumes a single equilibrium, however. The Kalman LMI paper gives the similarity-invariance argument we adopt, but only for linear systems. Barreau et al.'s drilling-string analysis shows that LMI Lyapunov conditions extend to PDE–ODE couplings after a well-chosen change of coordinates, with a unique equilibrium.
None of these works combines several coexisting states, a singular or near-singular linearization, and a verdict that is invariant under coordinate changes.
Research Program
The pipeline takes a candidate steady state from any solver. It first re-projects the state onto a reference high-order discretization and polishes it with Newton's method. This step is deliberately solver-agnostic, so that invariance is tested on the stability step and not on the solution itself.
It then computes a small cluster of eigenvalues nearest the imaginary axis, together with the corresponding left and right eigenvectors, which give $\Pi$. A residual-based enclosure then certifies that no other eigenvalue lies within $\delta$ of the imaginary axis. The threshold $\delta$ is chosen at the largest spectral gap, not fixed by hand.
On the complement, the pipeline counts unstable eigenvalues outside the cluster and solves the LMI above in a reduced Krylov subspace, which keeps the matrix dimension manageable. On the critical subspace, it computes $w^*(a)$ by a contraction iteration and fits the reduced map $g$ up to cubic order by finite differences in $a$. It then certifies either the local stability of $a=0$, or the sign structure of $g$, via a polynomial Lyapunov function found by sum-of-squares.
The two metrics are glued into a block Lyapunov function. The coupling terms are bounded using the contraction constant of $w^*$, which yields a certified basin radius in the intrinsic metric. If the gap or the contraction check fails, the pipeline abstains; it does not guess.
For the variational problems (Henon, Allen–Cahn, DFT), the unstable count is the Morse index, and the reduced step reduces to the Hessian of the reduced energy. For Gray–Scott, the linearization is non-self-adjoint and non-normal. This is where orthogonal projections and Euclidean contraction tests should fail most visibly, and where the spectral projection matters.
The analysis side has two goals. The first is to prove the invariance statement: the certificate's conclusions are unchanged under diffeomorphic changes of state coordinates and under congruent changes of the discretization basis. The second is to prove the $\sqrt{\mu}$ basin scaling for generic pitchforks and the corresponding exponent for folds.
Baselines
Naive reference. This is the current practice in the multiplicity literature: solutions reported with no stability label, as in HomPINNs and the adaptive basis method. Its "accuracy" is the base rate of the majority label. We report it so that gains from any labelling are visible.
Solver-native eigenvalues. For each solver, we compute the Jacobian or Hessian spectrum in that solver's own representation: collocation matrix without mass weighting, adaptive-basis Galerkin matrix, or the linearization of the network output at its training points. This is what a practitioner would do first, and it is the setting where coordinate artifacts should appear. It uses the same candidate solutions and the same eigensolver budget as our method.
Fine-mesh generalized eigenproblem. This is the strongest competitor for unstable counts. It solves $J v = \lambda M v$ with the correct mass matrix on a resolution four times finer than our reference discretization, which is the approach used in the DFT study. Away from bifurcations it should essentially always be right. Beating it means one of two things. Either we get strictly fewer wrong labels in the band of states within a small distance of a bifurcation, at equal or lower compute. Or we match it while also providing a certificate and basin radius it does not produce.
Direct time integration. We perturb each state by a random perturbation of size $\epsilon$ in the intrinsic norm and integrate to a long horizon. This is the physical ground truth for stable versus unstable, and it also estimates the empirical basin radius by bisection over $\epsilon$. We give it the same wall-clock budget as our full pipeline. The results we care about are our cost per state and our certified radius as a fraction of the empirical radius.
Contraction-certificate baselines. For basin radii, we run the ComputeLDF-style certificate in two versions. One uses the symmetric part of the Jacobian in Euclidean coordinates. The other uses real-Jordan-aligned coordinates, which is the strongest existing coordinate-transformation certificate in the file. Beating the aligned version means a larger certified radius with a better scaling exponent in $\mu$.
Ablations. The first ablation replaces the spectral projection $\Pi$ with the $L^2$-orthogonal projection onto the approximate kernel. This isolates the claim that the projection choice drives invariance for non-self-adjoint problems. The second ablation replaces the reduced equation with its linearization at $a=0$. This isolates the claim that the reduction is what buys the $\sqrt{\mu}$ scaling.
No reported stability-classification result was found on Lacuna for the Henon, Gray–Scott, Allen–Cahn or DFT benchmarks. The only reported reference numbers for these benchmarks are solution counts: 9 Henon solutions and 25 Gray–Scott patterns from 60 starts in HomPINNs, and 10 and 20 solutions in the adaptive basis method. The coordinate-sensitivity numbers are 1159 versus 81 simulations for ComputeLDF, and 0.73 versus 1.52 for the Kalman test.
Why not just…?
Why not just compute the generalized eigenvalues on a fine mesh? The spectrum of $M^{-1}J$ is already similarity invariant, and far from bifurcations this is a real threat: it will likely label those states correctly. We expect it to fail in two places. The first is near singular points, where the deciding eigenvalue is smaller than the discretization error and its sign flips with resolution. The second is that it provides no region of attraction, and a region of attraction is what distinguishes a practically stable pattern from a marginal one. Experiment 1 stratifies accuracy by distance to the nearest bifurcation precisely so that this baseline's regime of adequacy is measured, not argued away. If it matches us everywhere, the contribution shrinks to the basin certificate, and we will say so.
Why not just simulate the dynamics from perturbed states? Time integration is intrinsic and hard to fool. However, it reveals only whether a state is stable or unstable, not how many unstable directions a saddle has, and these high-index saddles are exactly what multiplicity solvers produce. It also suffers critical slowing down near bifurcations, because decay times grow like $1/\mu$. We use it as ground truth and compare cost per state, expecting the gap to widen as $\mu\to0$.
Why not just align coordinates with the Jordan modes, as ComputeLDF does? Jordan alignment fixes the rotation-induced over-approximation that ComputeLDF documents. It remains a linearization-based certificate, though, so near a pitchfork its certified radius is capped at order $\mu$, and a transformation computed from a nearly defective spectrum is itself ill-conditioned. The basin-scaling experiment tests directly whether the reduction's $\sqrt{\mu}$ advantage survives in practice.
Proposed Experiments
Experiment 1: invariance and correctness of unstable counts. We generate candidate solutions for the Henon equation on the unit disk and the Gray–Scott model on the heart-shaped domain with HomPINNs. For the 2D Allen–Cahn and Gray–Scott problems we use the adaptive orthogonal basis method, plus a finite-element Newton solver as a third source. Each state is then presented to every method in four representations: the solver-native form, a rescaled-units form, a Chebyshev re-expansion, and a smooth domain reparametrization.
The primary metric is the fraction of states whose unstable count agrees across all four representations. The secondary metrics are agreement with the ground truth from time integration and fine-mesh eigenvalues, and the abstention rate. All metrics are reported in bins of distance to the nearest detected bifurcation along homotopy paths.
The decisive check has two parts. Solver-native eigenvalues and the orthogonal-projection ablation should disagree across representations for a measurable fraction of the non-self-adjoint Gray–Scott states. Our certificate should show zero disagreements among non-abstained states. If the orthogonal ablation is equally invariant, the spectral-projection claim is wrong.
Experiment 2: basin scaling near a pitchfork with known ground truth. We use the H₂ Kohn–Sham model from the DFT study and sweep the exchange strength through $\alpha_c(R)$. On the broken-symmetry branch, we compare certified basin radii against the empirical radius from time integration, for the Euclidean, Jordan-aligned, linearized-reduction and full-reduction certificates. The decisive check is the fitted log–log slope of certified radius versus $|\alpha-\alpha_c|$: near 1 for the linearization-based certificates, near 1/2 for ours. A slope of 1 for our method falsifies the central quantitative claim.
Experiment 3: effect on downstream solution catalogues. For the 25 Gray–Scott patterns and the 9 Henon solutions, we report how many are certified stable, how many are certified saddles with a given index, and how many are abstentions. We also report how these counts change between solver-native labelling and our certificate. This gives the multiplicity community the first chart-independent stability census of benchmarks it already uses. It also tests whether the abstention mechanism stays rare enough, which we target at under a fifth of states, for the method to be useful rather than merely safe.
References
- An improved stability condition for Kalman filtering with bounded Markovian packet losses
- HomPINNs: Homotopy physics-informed neural networks for learning multiple solutions of nonlinear elliptic differential equations
- An Adaptive Orthogonal Basis Method for Computing Multiple Solutions of Differential Equations with Polynomial Nonlinearities
- Newton Informed Neural Operator for Computing Multiple Solutions of Nonlinear Partials Differential Equations
- Symmetry Breaking and the Generation of Spin Ordered Magnetic States in Density Functional Theory Due to Dirac Exchange for a Hydrogen Molecule
- Sign-changing bubble tower solutions for a Paneitz-type problem
- Hopf Bifurcations of Moore-Greitzer PDE Model with Additive Noise
- Bounded Verification with On-the-Fly Discrepancy Computation
- Exponential Lyapunov Stability Analysis of a Drilling Mechanism