Signed Equilibrium Ledgers for Mapping Multistability Across Dimensionless Design Space

Engineers who design snapping structures, such as origami tubes, kirigami shells and buckled bilayer sheets, need a "stability map." This map records, for every setting of the structure's dimensionless design ratios (thickness over length, prestretch, crease angle), how many stable shapes the structure has. Today such maps are built by brute force: simulate many parameter points, search each for equilibria, and hope no branch was missed. This proposal borrows a single rule from enumerative geometry. There, mathematicians count curves on a complicated surface by deforming the surface into simple pieces and checking that weighted counts add up exactly. Here, when equilibria are counted with a sign given by the curvature of the energy at each one, the total cannot change as the design ratios vary. Equilibria can only appear or vanish in canceling pairs. That one integer acts as a ledger. It certifies when a search at one parameter point has found every equilibrium, and it constrains how the stability map can change between sampled points. Both properties should make stability maps far cheaper to learn and much harder to get silently wrong.

The proposal treats the two research directions unequally. Degeneration tracking supplies the method: a conserved weighted count, plus the practice of evaluating that count in an extreme, simplified limit where it is easy to compute. Dimensionless scaling supplies the application regime. Because mechanical response collapses onto a few ratios, the stability map lives in a low-dimensional space where walls and conservation rules are tractable. Those same ratios also have natural extreme limits, such as a very slender shell or a nearly rigid crease, which play the role of the "degenerated" configuration.

Core Claim

Let $\pi$ be the vector of dimensionless design parameters, $E(\pi)$ the set of equilibria of the discretized energy at $\pi$, and $H(x;\pi)$ the energy Hessian (the matrix of second derivatives, whose eigenvalue signs tell whether an equilibrium is a minimum or a saddle) at equilibrium $x$. Two conditions are needed: equilibria must be nondegenerate, and none may escape to the boundary of configuration space. For a coercive energy, these follow from Morse theory, the classical result linking the critical points of a function to the shape of the space it is defined on. Under them,

$$\sum_{x \in E(\pi)} \operatorname{sign}\det H(x;\pi) \;=\; D,$$

where $D$ is an integer independent of $\pi$ (for a coercive energy on $\mathbb{R}^n$, $D=1$). The standard local behavior at the transitions follows from the same fact. Crossing a fold wall, the parameter surface where a stable shape and a saddle merge, changes the number of minima and the number of index-one saddles together, by exactly $\pm 1$. A symmetric pitchfork, where one shape splits into two mirror images, preserves the signed sum.

The mathematics is classical. The claim is that it can be used as an algorithmic constraint, and this is new. Every candidate stability map, whether computed or learned, must be an integer labeling of dimensionless space that satisfies these local balance rules, much as tropical counts must satisfy balancing at every face. The constant $D$ and a complete set of seed equilibria can be read off at an extreme limit of a dimensionless ratio, where the energy becomes piecewise simple and its equilibria can be enumerated combinatorially.

This template comes from three kinds of degeneration-tracking work. Chern classes and degenerations of hypersurfaces and their lines shows that the 27 lines of a cubic split exactly as 12 on a quadric plus 15 on a plane under degeneration. Tropical fans and the moduli spaces of tropical curves proves that weighted preimage counts of a fan morphism do not depend on the chosen point. Counting tropical elliptic plane curves with fixed j-invariant shows that when a wall is crossed, curve types vanish and appear while the weighted total is unchanged. We reuse the logic of these results: pick weights, prove invariance, compute the invariant in a degenerate limit. We do not reuse their algebraic machinery, because a mechanical energy landscape is real and generally non-polynomial.

The closest structural-mechanics bridge is Rational tensegrities through the lens of toric geometry. It identifies the self-stresses of a fixed rational framework with a Chow group of a glued toric variety. That result is static and restricted to rational coordinates. Ours is about how equilibria change as continuous dimensionless parameters vary, and it needs no rationality or toric structure.

Research Program

The work has three parts.

The first part builds the ledger for reduced mechanical models. For each model we compute $D$ and the degenerate-limit seeds analytically. The first model is the Kresling truss of Dynamics of Kresling origami deployment, parameterized by side count $n$, normalized radius $R_0$ and orientation angle $\delta_0$. Its four stability regimes and its zero-stiffness point at $\delta_0 = 67^\circ$ give known walls to recover; the zero-stiffness point is exactly where the nondegeneracy assumption fails, so the ledger must flag it. The second model is the chain of bistable ligaments in Propagation of pop ups in kirigami shells, parameterized by $\delta/l$ and unit count. There the number of metastable states grows combinatorially, and the stiff-ligament limit makes each unit an independent two-well system whose count is a simple product.

The second part is a ledger audit for any equilibrium finder, including full finite-element solvers. At a queried $\pi$, the finder returns equilibria. We compute each Hessian sign, using the stiffness matrix the solver already assembles. If the signed sum is not $D$, the search is declared incomplete. Search then continues, by deflation (a technique that modifies the equations to push the solver away from roots already found) or by continuation from neighboring certified points, until the sum balances. This check makes no polynomial assumption, which separates it from algebraic root counting.

The third part is active learning of the stability map under the ledger constraint. The map is represented as a piecewise-constant labeling of dimensionless space, giving the number of equilibria of each Morse index in each region, together with explicit wall surfaces. The learner proposes labelings consistent with the local balance rules. It queries the solver where the walls are most uncertain, and it rejects proposals that violate conservation, such as a single minimum appearing without a saddle. Dimensionless collapse matters here. We first check, in the style of Hierarchical Dimensionless Learning (Hi-π), that the map depends only on the reduced groups. That method assumes a single smooth response surface and cannot represent branch splitting. We use it only to choose coordinates, never to model the response.

Baselines

The naive reference point is a dense grid in dimensionless space. At each grid point we run a fixed number of random-restart Newton solves and classify each converged equilibrium by its Hessian. This baseline is easy to run, and it fails silently, because nothing tells it when a basin was never sampled. It runs on the same models with the same total solver-call budget as our method. A dense, high-restart version, using far more solver calls, also serves as ground truth for regime-map accuracy.

The strongest existing method for tracing branches across a parameter is surrogate-assisted numerical continuation in the style of Surrogate normal-forms for the numerical bifurcation and stability analysis of navier-stokes flows via machine learning. That work embeds simulation snapshots with proper orthogonal decomposition and fits a Gaussian-process latent dynamics model. It then runs MATCONT continuation to locate a symmetry-breaking pitchfork at $Re_{sb} \approx 43.12$ in a sudden-expansion channel, with field reconstruction error reduced to about 3%. It is the strongest competitor because it already detects bifurcations from a dimensionless parameter using limited simulation. We run it on our structural models and on the sudden-expansion channel with an identical simulation budget and the same hyperparameter-tuning budget. We count it as beaten if it misses branches that the ledger catches, or if it needs more simulations to reach the same wall-location error.

For multi-parameter design maps, the relevant competitor is active learning with Bayesian optimization. Rapid design of fully soft deployable structures via kirigami cuts and active learning reaches target shapes with SSIM around 0.9 using about 100 finite-element evaluations. That figure measures shape matching rather than regime mapping. For stability-map accuracy, there is no reported result on Lacuna for any of these models, so the comparison must be produced here. We implement an unconstrained Gaussian-process classifier over regime labels with the same acquisition rule and query budget as ours.

The key ablation removes the conservation constraint and the seeding from the degenerate limit, keeping everything else fixed. This isolates whether the ledger, rather than the active learner, is responsible for any gain. The proposal beats the field if, at equal solver budget, it achieves a lower missed-branch rate and a lower misclassified fraction of dimensionless area. A missed branch is an equilibrium present in ground truth but absent from the map.

Why not just…?

Why not just use polyhedral homotopy continuation, which is already the tropical method for counting roots? This is the most serious alternative. For the Kresling truss, after a trigonometric substitution, it could enumerate all complex equilibria at each grid point, so our seeding step may add nothing there. The proposal's value does not depend on beating it there. Homotopy continuation requires polynomial equations, which finite-element shell models, the regime where Study on the Cell Magnification Equivalent Method in Out-of-Plane Compression Simulations of Aluminum Honeycomb reports that failure modes switch at large magnification, do not provide. It also spends work on complex roots and starts from scratch at every point. We report a head-to-head comparison on the Kresling model; if homotopy wins there, the claim narrows to non-polynomial models, where only the signed audit applies.

Why not just run many more random restarts? Restarts reduce the miss rate but never certify completeness, and the cost grows with the number of basins. The kirigami ligament chain, where states multiply with unit count, is designed to show this. We plot missed-branch rate against restart count for both methods.

Why not just trust standard continuation such as MATCONT? Continuation follows branches connected to its starting point. It cannot find a disconnected branch, called an isola, that appears inside a region. The signed sum detects isolas only indirectly, because an isola contributes zero net sign, so it does not settle this weakness. The degenerate-limit seeds and the constraint that minimum counts change only at fold walls do address it, because a stable state can only arise where a wall has been crossed. We test this on a Kresling setting chosen to produce an asymmetric bistable branch.

Proposed Experiments

The decisive checks are as follows. On the Kresling truss, recover the four stability regimes and the 67° singularity, and compare missed-branch rate and query count against all baselines at equal budget. On kirigami pop-up chains with up to many units, check that ledger-audited search finds all metastable states that dense enumeration finds, at a fraction of the cost. On the sudden-expansion channel, verify that the steady-state degree is conserved through the pitchfork and that the wall lands near the reported $Re_{sb} \approx 43.12$. This tests transfer to non-gradient steady-state maps, where Brouwer degree, the analogous signed count of zeros of a general vector field, replaces Morse theory. On the aluminum honeycomb compression models from the magnification study, test whether a ledger imbalance flags the reported shift from progressive folding to global buckling before force-curve correlation degrades. If audits at fixed budget do not reduce missed branches relative to the constraint-free ablation, the central claim fails.