Many engineered systems are described by differential-algebraic equations (DAEs). These are models in which some variables evolve in time while others must satisfy instantaneous equations at every moment. Power grids are the standard case: generator states obey differential equations, while bus voltages must satisfy the power-flow equations exactly. Proving that such a system is stable, for example that a grid recovers after a fault is cleared, usually means handling those algebraic equations first. Engineers either solve them for the algebraic variables, which requires inverting a large, state-dependent Jacobian, or replace them with a fictitious fast dynamic that turns the DAE into an ordinary differential equation (ODE). Both routes produce certificates that depend on bookkeeping choices, such as how voltages are parametrized, rather than on the physics.
This note proposes a stability certificate stated directly on the surface of states the grid can actually occupy. It asks that nearby trajectories converge only for perturbations that respect the algebraic constraints. The constraint enters through a multiplier (Finsler's lemma), so no Jacobian is ever inverted and no fictitious dynamics are added. The resulting condition is a convex feasibility problem. In exact form it depends only on the geometry of the constraint surface and the flow on it. That geometric property makes one prediction testable: the certified critical clearing time should not change when the same grid is written with polar instead of rectangular voltage coordinates. The current lifted-ODE method has no such guarantee.
Core Claim
Consider a semi-explicit index-1 DAE, $\dot{x} = f(x,y)$ and $0 = g(x,y)$. Here $x$ are differential states (rotor angles, frequencies), $y$ are algebraic variables (bus voltages), and $\partial g/\partial y$ is invertible, which is what "index-1" means. A small perturbation $(\delta x, \delta y)$ that stays on the constraint surface must satisfy $G_x\delta x + G_y\delta y = 0$, where $G_x, G_y$ are the partial Jacobians of $g$. Because the system is index-1, $\delta x$ alone determines the tangent vector. So a metric $V = \delta x^\top M(x)\,\delta x$, with $M(x)$ positive definite, is a genuine Riemannian metric on the constraint surface.
Contraction at rate $\lambda$ means that the distance $V$ between any two neighbouring trajectories shrinks at least like $e^{-2\lambda t}$. It requires $\dot V \le -2\lambda V$ for every tangent perturbation. Finsler's lemma converts "negative for all vectors in the kernel of $[G_x\ G_y]$" into a matrix inequality with a free multiplier $L(x,y)$:
$$ \begin{bmatrix} M F_x + F_x^\top M + \dot M + 2\lambda M & M F_y \\ F_y^\top M & 0 \end{bmatrix} + L\,[G_x\ \ G_y] + [G_x\ \ G_y]^\top L^\top \preceq 0 . $$
Here $F_x, F_y$ are the Jacobians of $f$, and $\dot M$ is the derivative of $M$ along the flow. The inequality is linear in $(M, L)$, uses $G_y$ only through products and never through its inverse, and is imposed only on the constraint surface. For polynomial models, such as power flow in rectangular voltage coordinates, it becomes a sum-of-squares (SOS) program. SOS is a convex search over polynomial certificates, solved as a semidefinite program. Sublevel sets of the associated Lyapunov function then give a region of attraction and a sufficient critical clearing time (sCCT), the fault duration the grid is guaranteed to survive.
Relation to Prior Work
The closest work is Contraction Analysis of Nonlinear DAE Systems. It avoids inverting $\partial g/\partial y$ by building a "virtual extended" ODE in a lifted space and solving LMIs (linear matrix inequalities) for quadratic metrics that yield ellipsoidal regions of attraction and an sCCT. That paper motivates this proposal and also constrains it. It proves that the DAE always contracts faster than its virtual extension, and that some extension comes arbitrarily close to the DAE's rate. The second result warns that our gain in contraction rate may be small if the extension is tuned well. We therefore do not rest the proposal on a tighter rate alone. The claims we test are three: no extension has to be designed and tuned; the metric can depend on the state rather than being a single quadratic form; and the certificate does not depend on how the algebraic variables are parametrized. The lifted construction offers none of these.
The method borrows from Control Contraction Metrics: Differential L2 Gain and Observer Duality, which synthesizes state-dependent Riemannian metrics via SOS for explicit ODEs. We reuse its SOS machinery but restrict the contraction condition to the constraint tangent space, which that work does not model.
Two papers motivate the invariance test. An Improved Stability Condition for Kalman Filtering with Bounded Markovian Packet Losses shows a stability metric moving from 0.73 to 1.52 under a mere coordinate rescaling. Bounded Verification with On-the-Fly Discrepancy Computation shows that aligning coordinates with the true modes of the dynamics cut the required simulations from 1159 to 81 on a robot arm. Both show that coordinate choice can dominate certificate quality. Neither handles algebraic constraints.
The evaluation setting and the set-based baseline come from Benchmarks for the Formal Verification of Power Systems, which supplies CORA converters from MATPOWER and PSS/E RAW files into index-1 DAEs.
We deliberately exclude high-index DAEs. Reachability Analysis for High-Index Linear Differential Algebraic Equations handles them only for linear systems via projector decoupling. Geometric Methods for Adjoint Systems notes that reduction beyond index-1 is numerically fragile. Our index-1 argument, that $\delta x$ determines the tangent vector, does not carry over.
Research Program
Theory. We will prove that, before SOS relaxation, feasibility of the tangent-space condition is equivalent to contraction of the reduced flow on the constraint surface. We will also prove that this feasibility is unchanged under any smooth reparametrization $y = \phi(\tilde y)$ of the algebraic variables and any linear change of the $x$ coordinates.
We must be honest about one point. The SOS relaxation can break this invariance, because a polynomial of fixed degree in one parametrization need not be polynomial in another. We will therefore measure invariance empirically as well as prove it in the exact setting.
Metric derivative. When $M$ depends on $y$, $\dot M$ involves $\dot y$. Computing $\dot y$ would reintroduce $G_y^{-1}$. Instead we treat $\dot y$ as an auxiliary variable constrained by $G_x\dot x + G_y\dot y = 0$ and handle it with a second multiplier. The fallback is to restrict $M$ to depend on $x$ only.
Software. We will build the solver on top of the CORA-converted models so that every method reads identical DAEs.
Baselines
Time-domain simulation (the naive reference). We bisect over fault duration using simulation of the full DAE. This gives the true critical clearing time for a given fault. It is not a certificate, but it is the ground truth that any sufficient clearing time must stay below. The primary metric is the ratio sCCT/CCT. Methods are also compared on region-of-attraction volume fraction, estimated by Monte Carlo sampling on the constraint surface.
The virtual-extension LMI of Contraction Analysis of Nonlinear DAE Systems. This is the strongest competitor because it is the only existing inversion-free contraction certificate for nonlinear DAEs. We will reimplement it and give it the same solver and the same wall-clock budget. Its extension parameters will be tuned by grid search rather than left at defaults, so that its own "arbitrarily close" result works in its favour. The file does not give a numerical sCCT for its 2-bus, 14-variable benchmark: no reported result found on Lacuna for that value. We will therefore compare against our reimplementation. Beating it means one of two things. Either a higher sCCT/CCT ratio on the 2-bus and SMIB systems at equal compute, or a statistically indistinguishable ratio together with invariance under reparametrization that the baseline lacks.
CORA set-based reachability. On the same benchmarks, the file reports 6.9 seconds for the SMIB region-of-attraction case and 405 seconds for IEEE 30-bus transient stability, both from Benchmarks for the Formal Verification of Power Systems. We compare certified conclusions and runtime on identical fault specifications. A contraction certificate should at least match these conclusions. Because it is an infinite-horizon result, it should do so without per-scenario reruns.
Ablations. Three ablations isolate the key ingredients.
- A constant metric $M$ versus a state-dependent $M(x)$. This separates the benefit of the richer metric class from the benefit of the tangent-space formulation.
- The Finsler multiplier versus an explicit tangent basis computed numerically at sample points. The explicit basis requires null-space or inverse computations. This isolates the value of avoiding inversion, measured in runtime and scaling.
- Explicit elimination of $y$ followed by ODE contraction analysis. This is the substitute a skeptical reviewer would reach for first, discussed below.
Why not just…?
Why not just eliminate the algebraic variables with the implicit function theorem and run standard ODE contraction or SOS tools? For a two-bus system this is a real threat, and it is ablation 3. The limitation is structural. Even when $g$ is polynomial, the solved function $y(x)$ generally is not. SOS methods then need either Taylor truncation, which is only valid near the operating point and so weakest during large faults, or pointwise inversion. The virtual-extension work was designed specifically to avoid inverting $\partial g/\partial y$ on large networks. We will report where elimination remains competitive, expected on the smallest systems, and where it breaks down, expected at IEEE 14 and 30 buses.
Why not just use per-unit scaling, which power engineers already apply, with the existing virtual-extension LMI? Per-unit scaling is a diagonal linear rescaling, and any LMI with a free full matrix $P$ already absorbs linear rescalings. The dependence we target is nonlinear: the choice between polar and rectangular voltages, and the design of the fictitious extension dynamics. Per-unit scaling does not touch either. The decisive experiment below measures this directly. If the baseline turns out to be insensitive to parametrization in practice, we will say so and narrow the claim to the metric class and the absence of tuning.
Why not just run exhaustive simulation sweeps? Sweeps give the true CCT only for the sampled faults and operating points, and they give no guarantee between samples, as Benchmarks for the Formal Verification of Power Systems points out. We use simulation as the ground-truth ceiling, not as a competitor for certification.
Proposed Experiments
Decisive check 1: parametrization invariance. On SMIB and the 2-bus benchmark, we write power flow in rectangular voltages and in polar voltages, the latter polynomialized by trigonometric substitution. We then compute the sCCT with our method and with the virtual-extension baseline. The prediction is that our sCCT changes by no more than solver tolerance plus SOS-degree effects, while the baseline's shifts measurably. A failure of this prediction would falsify the invariance claim at the relaxation level.
Decisive check 2: conservatism versus ground truth. On SMIB, the 2-bus system, and the CORA IEEE 14-bus case, we compare sCCT/CCT and region-of-attraction volume across our method, the baseline, and the ablations, all at a matched SOS degree and time budget.
Scaling. On IEEE 14 and 30 buses we record SOS solve time as the network grows, alongside CORA's reported runtimes. If the full SOS program does not fit, we report the largest network that does and use the constant-metric variant beyond that point. We will not extrapolate.
References
- Contraction Analysis of Nonlinear DAE Systems
- Control Contraction Metrics: Differential L2 Gain and Observer Duality
- An improved stability condition for Kalman filtering with bounded Markovian packet losses
- Bounded Verification with On-the-Fly Discrepancy Computation
- Benchmarks for the Formal Verification of Power Systems
- Reachability Analysis for High-Index Linear Differential Algebraic Equations
- Geometric Methods for Adjoint Systems