Many trajectories in practice are only approximately correct. A numerical integrator leaves a small error at every step, a learned surrogate model mismatches the true physics, and a simulator runs on slightly wrong parameters. In each case the trajectory almost satisfies the true differential equation: the gap $\|\dot y - f(y)\|$, called the defect, is at most $\varepsilon$. Hyers-Ulam stability is the mathematical property that guarantees such an almost-solution lies within $K\varepsilon$ of some exact solution, where $K$ is a fixed "Ulam constant." What sets it apart from ordinary error bounds is the phrase "some exact solution." The certifying trajectory may start from a different initial state. For dynamics with unstable directions, such as saddles, hyperbolic orbits, and long horizons, this freedom is what keeps $K$ bounded. Forward error bounds tied to the original initial state grow exponentially in the same setting.
The weakness is that $K$ depends on the coordinates. Rescaling units or rotating the state changes the norm in which both the defect and the distance are measured. For non-normal systems, where eigenvectors are nearly parallel, the Euclidean constant can be inflated by orders of magnitude by an artifact of representation rather than by the dynamics. We propose to separate the two effects. The first is an intrinsic Ulam constant, the infimum over metrics, which is invariant under changes of coordinates. The second is a certified constant in the user's own coordinates, computed by choosing the best metric through a semidefinite program (SDP), meaning a convex optimization over positive-definite matrices. We then use both to certify long approximate trajectories on standard verification benchmarks. In those regimes, forward reachability bounds become uninformative.
Core Claim
For linear time-invariant and periodic vector systems, and locally along trajectories of nonlinear systems, the Ulam constant can be computed rather than merely shown to exist. The computation is a two-stage, coordinate-invariant procedure. The first stage finds a splitting of the state into decaying and growing directions, known as an exponential dichotomy. The second stage optimizes a separate metric on each part by SDP.
The resulting constants should match the known analytical minima in the scalar case. They should be exactly invariant under similarity transforms. On hyperbolic, stiff, and non-normal benchmarks, they should give certified distances to a true trajectory that are much tighter than defect-based Gronwall bounds, logarithmic-norm bounds, or Jordan-form discrepancy bounds.
Where This Sits
The closest existing bridge is Ulam Stability of Second-Order Periodic Linear Differential Equations with Periodic Damping. It reduces the scalar equation $y''+f y'+g y=0$ to a Riccati equation. It gives a necessary and sufficient condition for Ulam stability and proves $K_{\min}=1/|pq|$ for $y''-(p+q)y'+pqy=0$. That work supplies our scalar ground truth. It also leaves open the three things we target: vector (coupled) systems, the effect of changing coordinates on the constant, and numerical validation, which the authors explicitly defer.
Its nonhomogeneous open problem is easy for linear equations. Subtracting a particular solution reduces it to the homogeneous case. We state this as a remark and do not count it as a contribution.
Two papers motivate the coordinate side.
- An Improved Stability Condition for Kalman Filtering with Bounded Markovian Packet Losses shows that a scalar stability index flips from 0.73 (stable) to 1.52 (unstable) under a unit change. It then builds a similarity-invariant LMI (linear matrix inequality, a convex matrix constraint) test. That is the same disease we diagnose for Ulam constants, but for a yes/no stability question rather than a quantitative constant.
- Coordinate-invariant incremental Lyapunov functions defines incremental stability through the existence of some metric and proves converse theorems. It gives no construction of the metric, and it addresses asymptotic convergence rather than defect-to-distance bounds. We adopt its "exists a metric" stance but make the metric computable and quantitative.
Two papers supply methods we reuse and generalize.
- Nonquadratic Stability Condition of Continuous Fuzzy Systems certifies stability through logarithmic norms in transformed coordinates, $\mu_p(S^{-1}AS)<0$. Its authors leave the search for $S$ open. Our SDP is that search for $p=2$, extended to systems with both decaying and growing directions.
- Bounded Verification with On-the-Fly Discrepancy Computation bounds trajectory divergence by $\mu_2$ of the Jacobian and switches to real Jordan coordinates for oscillatory systems. On the 4D robot arm this cut the simulations needed from 1159 to 81. Its bounds concern exact trajectories from perturbed initial states and follow the original initial condition forward. Ours concern defects and allow the initial condition to move.
Finally, Stable Neural ODE with Lyapunov-Stable Equilibrium Points for Defending Against Adversarial Attacks obtains robustness by forcing contraction toward finitely many attractors. It notes that the model remains vulnerable between attraction basins. Our certificate needs no attractors, which is why learned surrogates are one source of approximate trajectories in our experiments.
Research Program
Definitions. Fix a system $\dot x=f(t,x)$ on a time interval $I$, a defect-injection matrix $B$, an output matrix $C$, and a positive-definite metric $M$. Define $K_M$ as the smallest constant such that every $y$ with $\|\dot y-f(t,y)\|_M\le\varepsilon$ on $I$ has an exact solution $x$ with
$$\sup_{t\in I}\|C(y(t)-x(t))\|\le K_M\,\varepsilon .$$
The $B$ and $C$ matrices let the definition recover the scalar setting of Li et al., in which the defect enters only the $y''$ equation and the distance is measured in $y$. We define two constants.
- The intrinsic constant is $K_{\mathrm{int}}=\inf_M K_M$. It is invariant under every linear change of coordinates $z=Tx$, because $T$ simply reparametrizes the set of metrics.
- The certified Euclidean constant is $K_{\mathrm{euc}}=\inf_M K_M\,\kappa(M)^{1/2}$, where $\kappa(M)$ is the condition number of $M$. This is the bound a user can actually apply in their own units.
Linear computation. For $\dot x=Ax$ with no eigenvalues on the imaginary axis, first compute an invariant splitting $T A T^{-1}=\mathrm{diag}(A_s,A_u)$, where $A_s$ holds the decaying directions and $A_u$ the growing ones. Then solve
$$A_s^\top P_s+P_sA_s\preceq-2\alpha P_s,\qquad A_u^\top P_u+P_uA_u\succeq 2\beta P_u,\qquad P_s,P_u\succ0,$$
bisecting over the decay rates $\alpha,\beta>0$. Integrating the stable part of the defect forward and the unstable part backward gives $K_M\le 1/\alpha+1/\beta$ in the block metric. Minimizing $\kappa(M)^{1/2}(1/\alpha+1/\beta)$ jointly over $T$, $P_s$ and $P_u$ gives the Euclidean certificate. The splitting $T$ is optimized as well rather than fixed to Jordan form.
The lower bound $K_{\mathrm{int}}\ge 1/\min_i|\mathrm{Re}\,\lambda_i|$, where $\lambda_i$ are the eigenvalues of $A$, follows from defects aligned with eigenvectors. We will measure how wide the gap between this lower bound and the SDP value is in practice.
For periodic $A(t)$, the matrices $P_s$ and $P_u$ become periodic and are parametrized by trigonometric polynomials. This yields the vector-valued extension that Li et al. leave open. It replaces their scalar Riccati equation by a periodic splitting.
Nonlinear, local. Along an approximate trajectory $y$, we linearize to obtain $\dot\xi=Df(y(t))\xi$. We then certify a time-varying dichotomy by applying the LMIs above piecewise in time, which gives a bound $K_{\mathrm{lin}}$ on the inverse of the linearized operator. A contraction-mapping argument in function space combines $K_{\mathrm{lin}}$, the measured defect, and a bound on the second derivative of $f$. It proves that a true trajectory exists within a computed radius. When the dynamics are contracting everywhere, only the stable block is present, and the method reduces to an optimized-metric logarithmic-norm bound.
Baselines
Gronwall bound. The naive reference point is the defect-based Gronwall bound with a global Lipschitz constant, keeping the initial condition fixed. Bounded Verification with On-the-Fly Discrepancy Computation already reports that this bound grows exponentially. It sets the floor.
Euclidean logarithmic-norm bound. The next baseline replaces the Lipschitz constant by $\mu_2(J)$ along the trajectory. This is the standard a posteriori error estimate and the discrepancy computation of Fan and Mitra, applied to defects.
Jordan-form discrepancy bound. The strongest existing method is the same logarithmic-norm bound computed in real Jordan coordinates, as in Fan and Mitra. The best figure the file reports for it is the reduction from 1159 to 81 simulations on the 4D robot arm. The file reports no result for defect-to-distance tightness, so there is no reported result found on Lacuna for shadowing-radius tightness on any of these benchmarks.
Fair comparison. All methods receive the same approximate trajectories and the same defect estimates, measured by dense re-evaluation of $\dot y-f(y)$. They also receive the same Jacobian access and the same wall-clock budget. We report both the certified radius and the runtime.
Ablations. Two ablations isolate the proposal's two ingredients.
- Free initial condition. Our metric optimization is run with the initial condition held fixed, so only the stable block is used. This isolates the Ulam ingredient.
- Optimized metric. The dichotomy construction is run with $M=I$, so the free initial condition is kept but the metric is not optimized. This isolates the coordinate ingredient.
What counts as beating the baselines. We beat the Jordan-form baseline if the certified radius is at least an order of magnitude tighter on hyperbolic or non-normal benchmarks over horizons where it exceeds the state's scale. We also require no loss on contracting benchmarks.
Why not just…?
Why not just use the logarithmic-norm error bound in good coordinates? On contracting systems with well-conditioned eigenvectors, our method reduces to this bound, and we expect no gain there. This is a real limit, and we report it as such. The bound fails in two places. On unstable or saddle directions it follows the original initial condition forward and grows exponentially. On non-normal systems, $\mu_2$ in any fixed coordinates overestimates transient growth. The first ablation measures the first effect directly, and the non-normal sweep below measures the second.
Why not just transform to Jordan or Floquet modal coordinates, as Fan and Mitra do? A modal transform is one particular metric. When eigenvalues are nearly repeated, its condition number blows up, and this inflates $K_{\mathrm{euc}}$. For diagonalizable, well-conditioned systems it may match our SDP, and that would be an honest null result. The non-normality sweep is designed to show where the two methods diverge, if they do.
Why not just find the nearby true trajectory numerically, by multiple shooting? Multiple shooting gives an estimate, not a certificate. It also has to be rerun for every trajectory, whereas our constant applies to all trajectories with defect at most $\varepsilon$ at once. We use high-precision multiple shooting as ground truth for how tight our bounds are, rather than as a competitor.
Proposed Experiments
Scalar ground truth. We test whether the SDP with $B$ and $C$ set to the companion-form defect and output recovers $K_{\min}=1/|pq|$ (Theorem 9 of Li et al.). We sweep the roots $p,q$ over a grid and test periodic $f,g$ against their Riccati-based minimal constants and their instability condition $\int_0^T p\,(\,\cdot\,)=0$. This is also the numerical validation those authors deferred. Failure to reproduce $1/|pq|$ within solver tolerance would falsify our formulation.
Coordinate sensitivity. Following the protocol of Wu et al., we take random similarity transforms of increasing condition number and of non-normal $2\times2$ to $20\times20$ systems. We report how the Euclidean $K_M$ at $M=I$ varies across transforms, compared with $K_{\mathrm{int}}$, which should be exactly constant. We also report the gap between the SDP value and the eigenvalue lower bound.
Certifying approximate trajectories. We use benchmarks from the file:
- the 4D robot arm and the 28-dimensional helicopter from Bounded Verification with On-the-Fly Discrepancy Computation, whose helicopter verification took about 67 seconds;
- coupled Van der Pol (CVDP23) and the stiff Robertson system (ROBE21) from ARCH-COMP24 Category Report: Continuous and Hybrid Systems with Nonlinear Dynamics, where JuliaReach verified CVDP23 in 1.2 seconds while other tools took over 800 seconds or failed;
- the linearized cart-pendulum of Robust and Exponential Stabilization of a Cart-Pendulum System via Geometric PID Control, run near its unstable upright equilibrium without control.
The approximate trajectories come from three sources: coarse explicit integrators, perturbed parameters, and a neural ODE surrogate trained on each system. For each, we measure four things over horizons up to the point where the Jordan baseline's radius exceeds the state's scale. These are the certified radius divided by the true nearest-trajectory distance (from multiple shooting), the fraction of trajectories certified, the runtime, and the defect $\varepsilon$ at which certification breaks down.
The decisive check. The decisive check is the pair of ablations on the cart-pendulum and CVDP23. If allowing the initial condition to move does not shrink the certified radius by a large factor on these saddle-containing dynamics, the Ulam ingredient adds nothing over existing discrepancy methods, and the proposal reduces to metric optimization alone. Likewise, if metric optimization does not tighten the bound beyond the Jordan baseline on the non-normal sweep, the coordinate ingredient adds nothing.
References
- Ulam Stability of Second-Order Periodic Linear Differential Equations with Periodic Damping
- An improved stability condition for Kalman filtering with bounded Markovian packet losses
- Coordinate-invariant incremental Lyapunov functions
- Nonquadratic Stability Condition of Continuous Fuzzy Systems
- Bounded Verification with On-the-Fly Discrepancy Computation
- Stable Neural ODE with Lyapunov-Stable Equilibrium Points for Defending Against Adversarial Attacks
- ARCH-COMP24 Category Report: Continuous and Hybrid Systems with Nonlinear Dynamics
- Robust and Exponential Stabilization of a Cart-Pendulum System via Geometric PID Control