Stability is a property of a system, not of the variables used to write it down. In practice, though, checking whether a linear partial differential equation (PDE) system decays, stays bounded, or blows up starts with an arbitrary choice. The analyst rewrites the equations as a first-order evolution problem, picks state variables, and inspects eigenvalues or builds an energy function. When the system carries constraints (a divergence-free condition, a gauge freedom, an algebraic relation between unknowns), there is no canonical way to make that choice. Different choices can add spurious modes or hide slow polynomial growth.
This note proposes that differential primary decomposition already supplies the canonical choice. Differential primary decomposition is an algebraic procedure that splits a linear constant-coefficient PDE system into elementary pieces, each with an explicit family of exponential-polynomial solutions. It plays the same role for PDE systems that the Jordan normal form plays for matrices. We will extract from it a stability exponent and a polynomial growth order. Both are invariant under every invertible differential change of unknowns, and we will certify them with sum-of-squares proofs, one per piece. The result is a constructive, presentation-independent stability test for the class of systems where the algebra is exact. We will also measure honestly where it beats the obvious Fourier-symbol eigenvalue check and where it does not.
Core Claim
Write a linear constant-coefficient system in time $t$ and space $x\in\mathbb{R}^n$ (or the torus) as a module $M$ over the ring $R=\mathbb{C}[\partial_t,\partial_1,\dots,\partial_n]$. Here $M$ is the quotient of $R^k$ by the rows of the operator matrix, and its solutions are the system's solutions. Linear PDE with constant coefficients implements the Ehrenpreis–Palamodov fundamental principle in Macaulay2. It decomposes $M$ into primary components, each with an associated prime $\mathfrak{p}_i$ (an algebraic variety $V_i$ of frequencies $(\lambda,\zeta)$) and finitely many Noetherian multipliers $B_{i,j}$ (polynomials in $t,x$ and the frequencies). Every solution is then a superposition
$$u(t,x)=\sum_{i,j}\int_{V_i} B_{i,j}(t,x,\lambda,\zeta)\,e^{\lambda t+\zeta\cdot x}\,d\mu_{i,j}(\lambda,\zeta),$$
where the $\mu_{i,j}$ are measures on $V_i$. Restricting to spatially oscillatory solutions ($\zeta=i\xi$ with $\xi$ real) gives two quantities. The first is the growth exponent
$$\omega(M)=\max_i\ \sup\{\operatorname{Re}\lambda:\ (\lambda,i\xi)\in V_i,\ \xi\in\mathbb{R}^n\}.$$
The second is the growth order $\kappa(M)$: the largest $t$-degree among the multipliers on the components that attain $\omega$.
The claim has three parts. The first is that $\omega$ and $\kappa$ depend only on the isomorphism class of $M$. They are therefore unchanged by any differential change of unknowns given by an invertible polynomial operator matrix, a class far larger than the smooth state diffeomorphisms considered in Coordinate-invariant incremental Lyapunov functions. The second is that, under a uniformity condition on the multipliers in $\xi$ that we will make explicit, the pair $(\omega,\kappa)$ gives the sharp decay or growth rate $t^{\kappa}e^{\omega t}$. The third is that the bound $\operatorname{Re}\lambda\le-\alpha$ on each real frequency slice is a polynomial inequality. A per-component sum-of-squares certificate can prove it, which turns an existence-only notion of a stability metric into a computed one.
For a linear ordinary differential equation $\dot x=Ax$, the module lives over $\mathbb{C}[\partial_t]$. The primes are the eigenvalues and the multipliers are the Jordan chains, so the whole construction reduces to the real Jordan form. Bounded Verification with On-the-Fly Discrepancy Computation chooses that Jordan form by hand to cut reachability conservatism. Our proposal extends that coordinate choice to constrained, overdetermined and higher-order PDE systems.
Relation to Prior Work
Four lines of work frame the proposal. Three supply its methods and one constrains its scope.
The method side comes from the Macaulay2 lineage. Linear PDE with constant coefficients computes the decomposition and the multipliers, but asks only for solution representations, never for stability. Noetherian operators and primary decomposition and the package Noetherian operators in Macaulay2 provide the symbolic and numerical ways to compute the multipliers. Local dual spaces and primary decomposition shows that embedded components (lower-dimensional pieces sitting inside larger ones) are captured by excess dual spaces. This matters for stability because embedded components are exactly where extra polynomial-in-time factors appear. That paper and the Noetherian-operators paper both leave numerical detection of embedded components open, so embedded components must be handled symbolically in our work.
Syzygies, constant rank, and beyond uses primary decomposition and syzygies to explain an analytic property, compensated compactness, that linear-algebra tests alone miss. We take this as evidence that the module carries analytic information a chosen realization does not. Our proposal applies the same idea to time decay.
The stability side motivates the work. Coordinate-invariant incremental Lyapunov functions defines stability through the existence of some metric, precisely because Euclidean tests change under coordinate changes, but gives no way to construct that metric. Our method gives a construction, restricted to linear constant-coefficient systems, where incremental stability and plain stability coincide. Several works pick their coordinates by insight rather than by algorithm: the Riemann coordinates in Exponential Lyapunov Stability Analysis of a Drilling Mechanism, the Jordan alignment of Fan and Mitra, and the two-stage change of variables in Robust and Exponential Stabilization of a Cart-Pendulum System via Geometric PID Control. Our method replaces that choice for the class it covers. It does not replace it for nonlinear or boundary-coupled systems.
The closest connecting work constrains the design. Algorithmic Verification of Linearizability for Ordinary Differential Equations uses Thomas decomposition to find linearizing coordinates, but never checks whether the resulting linear system is stable, and its decomposition-based test runs out of memory for equations of order 10 and above. Thomas decomposition, as implemented in The MAPLE package TDDS, splits a differential system into disjoint simpler subsystems, one per qualitative case. Thomas decompositions of parametric nonlinear control systems and Thomas Decomposition and Nonlinear Control Systems use it to split parameter space into regimes for flatness and observability, again without stability.
We differ from all three in the same way. Linearity is an assumption we make, not a goal we pursue, and stability is the output. We borrow only their parameter-splitting idea, for small parametric families. The out-of-memory result tells us to stay with linear, low-order systems and not to aim at nonlinear normal forms.
Research Program
The first piece is theory. We will prove invariance of $(\omega,\kappa)$ under module isomorphism. This part should be routine, since associated primes and multiplicity structure are isomorphism invariants. We will then state and prove the sharpness result under an explicit hypothesis: the Noetherian multipliers on the real slice grow at most polynomially in $\xi$. This is a Petrovsky-type well-posedness condition, the classical requirement that solutions not grow faster than polynomially in frequency, and it is the substantive analytic step. We expect the hypothesis to hold for hyperbolic and parabolic examples and to fail in instructive ways for others. On periodic domains the frequencies $\xi$ form an integer lattice, and the supremum becomes a maximum over lattice points plus a limit at infinity. This matters for damped waves, where low frequencies decay more slowly than high ones.
The second piece is certification. For each component we will search for a sum-of-squares certificate that $\operatorname{Re}\lambda+\alpha\le 0$ on the real slice of $V_i$, and bisect on $\alpha$. Assembling the per-component certificates gives an explicit decay rate. The multipliers then define a quadratic energy in what we call "primary coordinates", a canonical change of unknowns that block-diagonalizes the dynamics. That energy is the constructed metric. We will compare this decay rate with the exact $\omega$ obtained by solving the real-slice optimization in closed form, to measure how conservative the certificate is.
The third piece covers parametric families. When coefficients depend on parameters such as damping, the decomposition changes at special values. Critical damping, where two decay modes merge and a $t\,e^{\lambda t}$ factor appears, is the standard example. We will use Thomas-style case splitting over parameters to partition parameter space into cells with constant primary structure, and compute $(\omega,\kappa)$ symbolically on each cell. The output is an exact stability map with its singular boundaries marked.
The engineering builds on the NoetherianOperators package. We will add a real-slice optimization step and an interface to a sum-of-squares solver. The test generator produces equivalent presentations of the same system by applying random invertible polynomial operator matrices to both the unknowns and the equations.
Baselines
The naive reference point is what a practitioner does by default. They rewrite the system in first-order evolution form using a fixed textbook rule: introduce time derivatives as new states, and keep constraints as additional equations or drop them. They then Fourier-transform in space and declare stability if every eigenvalue of the symbol matrix has negative real part on a dense $\xi$ grid. This baseline is cheap and is the bar that any added machinery must clear.
The strongest existing competitor is the same symbol analysis performed carefully. An expert picks the reduction, projects out constraint modes by hand, and estimates Jordan growth from the eigenvector conditioning and the resolvent norm of the symbol. This is the strongest competitor because, for square explicit evolution systems, it is mathematically equivalent to our method. Its only weakness is that it depends on the chosen presentation. We will run it on exactly the same systems and the same random equivalent presentations, with the same frequency grid and the same wall-clock budget, and apply the same reduction rule to every presentation rather than re-tuning per instance.
For the ODE sanity case the competitor is the real Jordan transformation of Fan and Mitra. On their 4D robotic-arm benchmark it reduced the simulations needed from 1159 to 81 (Bounded Verification with On-the-Fly Discrepancy Computation). We will check that our coordinates reproduce their transformation on the linearized arm and give the same simulation count. Matching is the goal there, not beating.
For boundary-coupled distributed systems, the reference is the Riemann-coordinate LMI hierarchy (a sequence of increasingly fine linear-matrix-inequality conditions) of Barreau et al. They report a guaranteed decay rate of 1.000 at approximation order $N=2$ for the dynamic controller, against 0.2159 for the feedforward controller, with damping $d=0$ (Exponential Lyapunov Stability Analysis of a Drilling Mechanism). Our method does not handle boundary coupling, so this number is a scope marker, not a head-to-head target. On the periodic damped string ($d>0$), which both methods can treat, no reported result was found on Lacuna, and we will run both methods ourselves.
Two ablations isolate the key ingredient. The first keeps only the associated primes and drops the Noetherian multipliers. This yields $\omega$ but not $\kappa$, and shows whether the multipliers are needed to detect polynomial growth. The second drops embedded components, which tests whether the excess dual spaces of Chen, Härkönen and Leykin matter for stability rather than only for algebraic completeness.
Beating the baselines means three things. Our method must give one classification across all equivalent presentations while the strongest baseline varies. It must correctly classify cases with polynomial growth and constraint modes that the strongest baseline gets wrong. Its certified decay rate must come within a stated tolerance of the exact $\omega$.
Why not just…?
Why not just Fourier-transform and check eigenvalues of the symbol? For square systems already in explicit evolution form this gives the same answer, and we say so plainly. The threat is real, and our experiments are designed around it. The method's value is confined to systems without such a form: Maxwell-type systems with divergence constraints, Stokes and Oseen flow with incompressibility, descriptor systems of differentiation index two or three, and higher-order-in-time mixed systems. In these cases the reduction step is a choice. It can introduce zero modes from constraint violation or misreport the size of Jordan blocks. If the careful expert baseline agrees with us on every presentation in that suite, our contribution shrinks to automation and invariance bookkeeping, and we will report it that way.
Why not just build a quadratic Lyapunov functional with LMIs, as Barreau et al. do? LMIs are the right tool where they apply, and they currently reach boundary-coupled systems that we cannot. But the functional is written in a chosen state, the hierarchy is truncated, and the published drilling result covers only $d=0$. On the periodic damped string we will measure how far each method's certified rate sits from the exact rate as $d$ varies, including near critical damping, where we expect a fixed-order LMI to lose accuracy. A truncation hierarchy has no built-in mechanism for multiplicity changes, which is the effect our parametric cells are designed to expose.
Why not just compute the Jordan form numerically, as in ComputeLDF? Numerical Jordan forms are unstable exactly at the defective points that create $t^k e^{\lambda t}$ growth. A PDE also has a Jordan structure at every frequency rather than a single one. The parametric experiment tests this directly: we compare symbolic cell boundaries against numerical Jordan detection on a dense parameter sweep.
Why not just train a neural Lyapunov certificate? For linear constant-coefficient systems a learned certificate is pointless when exact quantities are available. It also inherits the sample-density limits noted in Learning Hybrid Control Barrier Functions from Data. We do not include it as a baseline.
Proposed Experiments and Decisive Checks
The benchmark suite has two parts. The first contains periodic-domain systems with closed-form solutions, so the ground-truth $(\omega,\kappa)$ is known: the damped wave and telegraph equations over a damping sweep through critical values; Maxwell's equations with conductivity, including the divergence constraints; linearized Stokes and Oseen flow; damped linear elasticity; and index-two and index-three descriptor systems. The second contains the linearized robotic arm and the periodic damped string, which connect to Fan–Mitra and Barreau et al. Each system will be posed in twenty random equivalent presentations.
We will measure five things: classification accuracy for stable, marginal and unstable; exact recovery of $\kappa$; the spread of each method's verdict across presentations; the gap between certified and exact $\alpha$; and runtime as a function of the number of unknowns and the operator order.
Three checks decide the outcome. The first is invariance: our verdicts must agree across presentations, while the naive baseline's spread is measured. The second is the kill test: we look for at least a few constrained or descriptor systems where the careful expert baseline is wrong or presentation-dependent and our method is right. Finding none ends the claim that the algebra is needed. The third is scaling: we find where symbolic decomposition stops being feasible, expecting a wall like the order-10 limit reported for TDDS, and we report that boundary as the method's honest scope.
References
- Linear PDE with constant coefficients
- Coordinate-invariant incremental Lyapunov functions
- Bounded Verification with On-the-Fly Discrepancy Computation
- Noetherian operators and primary decomposition
- Noetherian operators in Macaulay2
- Local dual spaces and primary decomposition
- Syzygies, constant rank, and beyond
- Exponential Lyapunov Stability Analysis of a Drilling Mechanism
- Robust and Exponential Stabilization of a Cart-Pendulum System via Geometric PID Control
- Algorithmic Verification of Linearizability for Ordinary Differential Equations
- The MAPLE package TDDS for computing Thomas decompositions of systems of nonlinear PDEs
- Thomas decompositions of parametric nonlinear control systems
- Thomas Decomposition and Nonlinear Control Systems
- Learning Hybrid Control Barrier Functions from Data