When engineers optimize the shape of a joint, seal, or interlocking part to remove a stress peak, they almost always simulate contact with a smoothed stand-in for the real physics. Typical stand-ins are a penalty that pushes overlapping surfaces apart, or a barrier that blows up as surfaces approach. Each has a smoothing parameter that should go to zero, and each becomes numerically ill-behaved as it does. This proposal borrows a trick from game theory, where the same mathematical problem is called the trembling-hand limit. There, Practical exact algorithm for trembling-hand equilibrium refinements in games showed that one does not need to compute at a vanishingly small perturbation. One can solve at a moderate one and then verify, with a series expansion in the perturbation, that the solution's combinatorial structure stays the same all the way to the limit. We carry this idea over to contact mechanics. The method solves the smoothed contact problem where it is well-conditioned, reads off which surface points are touching, and certifies in ordinary double precision that this contact set is the exact limit. It then returns the exact limit stress field and its shape sensitivity.
The reason this matters specifically for stress mitigation is a point we think has been overlooked. Minimizing stress concentration at a contact interface drives the design toward contact patches whose pressure fades smoothly to zero at their edges. At those edges, points are simultaneously touching and carrying no force, a "weakly active" or degenerate contact. That is exactly where penalty gradients are biased, where active-set solvers flip back and forth, and where floating-point precision is lost. So the optimizer pushes itself into the regime its own simulator handles worst. The game-theoretic analogue is an action played with zero probability and zero regret, the case that refinements like trembling-hand perfection exist to resolve.
Core Claim
For small-strain elastic bodies with unilateral contact, a quadratic-penalty solution with a fixed candidate contact set is a rational function of the penalty parameter. Let $u(\epsilon)$ be the displacement, $\epsilon$ the penalty compliance, $K$ the stiffness matrix, $f$ the load, and $G_A$ the gap Jacobian of the candidate active set $A$. Then $u(\epsilon)$ solves
$$\left(\epsilon K + G_A^\top G_A\right) u(\epsilon) = \epsilon f + G_A^\top g_A^0,$$
where $g_A^0$ is the initial gap of the points in $A$. This matrix can be singular at $\epsilon = 0$, for example with redundant face-to-face constraints or with parts held only by contact, such as the tail of a dovetail. Its solution still has a Laurent expansion in $\epsilon$. The limit contact set is certified when two sign conditions hold for all sufficiently small $\epsilon$. Every point in $A$ must have a nonnegative contact force, which we check through the first nonzero coefficient of its force series. Every point outside $A$ must have a nonnegative gap, which we check through the first nonzero coefficient of its gap series. This is the contact analogue of the basis-stability oracle in the trembling linear program.
A point whose force and gap both vanish at order zero is degenerate. Its higher-order coefficient tells which side of the kink the penalty path approaches from, and that gives the correct one-sided shape derivative of a peak-stress objective. We claim this procedure has three properties. It recovers limit stresses and gradients to near machine precision without extended-precision arithmetic. It flags degenerate contacts that stress optimization creates. Used inside a gradient optimizer, it reaches lower true peak stress than penalty, barrier, and plain active-set pipelines at comparable cost.
Relation to Prior Work
On the game side, three papers supply the method and the cautionary tales.
- Method we transfer. Computing an Extensive-Form Perfect Equilibrium in Two-Player Games handles infinitesimal perturbations symbolically inside a complementarity problem. Its finding that refinement adds only about 8% pivoting work over an unrefined solve suggests that tracking the limit is cheap once precision is controlled.
- The closer template. The trembling linear program of Practical exact algorithm for trembling-hand equilibrium refinements in games replaces symbolic arithmetic with a series-expansion stability check at moderate $\epsilon$. It found stable bases near $10^{-5}$ to $10^{-6}$ even though theoretical bounds suggested $10^{-42}$ or smaller. That paper is confined to linear programs over game trees. We move the oracle to a symmetric positive-semidefinite complementarity system arising from finite elements, add frictional stick–slip states, and use it for sensitivities, not just for the forward solve.
- Constraint on design. Quasi-Perfect Stackelberg Equilibrium and the survey A survey on algorithms for Nash equilibria in finite normal-form games show that the alternative, arbitrary-precision arithmetic with GMP, buys exactness at a steep and growing time cost. We use it only to build ground truth.
- Why a finite-precision certificate is plausible. Universal Complexity Bounds Based on Value Iteration and Application to Entropy Games proves that finite precision suffices above an algebraically determined threshold. This is the same spirit as a certificate that works in double precision.
On the mechanics side, three papers supply the application and motivate the gap.
- Gradient-Based Dovetail Joint Shape Optimization for Stiffness differentiates through a softplus penalty with adjoint sensitivities. Its sharpness parameter faces the dilemma we target: soft settings allow interpenetration and wrong stresses, while stiff settings overflow or make the adjoint system near-singular. Its objective was stiffness, not peak stress.
- Topology Optimization with Frictional Self-Contact uses a diverging barrier (incremental potential contact) inside density-based topology optimization. Its authors concede that grid-aligned boundaries create artificial corner stress spikes, and they rely on a heuristic strain-limiting fix. We treat their barrier as a competitor, not as the method.
- Single-Level Differentiable Contact Simulation relaxes complementarity to a product of gap and force bounded by $\epsilon$, and drives $\epsilon$ down empirically. It reports that its interior-point solver cannot be warm-started. Our certificate is a direct answer to "how small must $\epsilon$ be?", and an identified active set is a natural warm start. That paper treats rigid bodies and so says nothing about internal stress.
Empirical stress-redistribution studies explain why edge-of-contact peaks are the quantity worth controlling. In Octopus-Inspired Biomimetic Annular Sealing Grooves, seal grooves were tuned by parametric finite element sweeps to a 21.71% improvement in average von Mises stress. In Experimental Study on Anchoring Performance of Short-Lapped-Rebar Splices, redesigned rebar splices removed end-of-anchorage peaks. Neither uses gradient optimization, but both show that design improvement lives at contact edges.
Research Program
We will build a 2D, and later 3D, small-strain finite element contact solver with node-to-segment gaps and the four components below.
- Probe. The solver is run at a sequence of moderate penalties, starting from $\epsilon$ with a condition number around $10^6$ and halving when certification fails. This is the same outer loop as the trembling linear program.
- Stability oracle. The oracle computes the leading Laurent coefficients of displacement, force, and gap for the candidate set. It does this with a structured expansion that projects out the null space of $G_A^\top G_A$, so the singular matrix is never factored.
- Exact limit solve. On the certified set, the solver produces the limit stress field.
- Adjoint sensitivity. A smooth aggregate of von Mises stress (a p-norm) is differentiated on the certified set. At degenerate points flagged by the oracle, the solver returns the one-sided derivative selected by the next nonzero series coefficient, or optionally both one-sided derivatives, so the optimizer can take a nonsmooth step.
Coulomb friction makes the system nonsymmetric and adds stick/slip labels. Here the oracle checks sign conditions on tangential slack as well.
We will also analyze theoretically when a finite-precision certificate exists. Our conjecture, following the trembling-linear-program argument, is that the certifying $\epsilon$ is bounded below by a quantity depending only on the geometry's algebraic data. We will measure in practice how large it is.
Baselines
The naive reference is the dovetail paper's pipeline, a fixed softplus penalty with adjoint gradients (Gradient-Based Dovetail Joint Shape Optimization for Stiffness). It will be run with its sharpness tuned by grid search on a held-out geometry, using the same mesh, optimizer, and iteration budget. The only reported numbers for this pipeline are stiffness gains: Complex Dovetail rising from 518.1 to 1036.3 N/mm, a 70% physical stiffness increase for the Single Dovetail, and 10–20 minute runtimes. There is no reported result on Lacuna for peak von Mises stress on the dovetail benchmarks, so we will establish that number ourselves.
The barrier competitor is incremental-potential-contact barrier smoothing as used in Topology Optimization with Frictional Self-Contact. It is reported to be 5–10 times faster than air-mesh methods, with no reported compliance values on Lacuna. We will run it with its activation distance tuned under the same budget.
The relaxed-complementarity competitor follows Single-Level Differentiable Contact Simulation, an interior-point solve with relaxation driven to a fixed small value. Its reported 0.1% failure rate is for rigid box dropping, not deformable stress, so it is not directly comparable.
The strongest existing method is an exact semismooth-Newton or primal–dual active-set solver for the discrete contact problem. It is paired with implicit-function sensitivities computed on whatever active set it returns, given the same double precision, a warm start from the previous design, and the same per-iteration time. Ground truth for all accuracy metrics comes from GMP-rational solves of the same discrete problem.
The key ablation keeps our probe-and-shrink loop but replaces the series oracle with a heuristic check: declare stability if the active set is unchanged between $\epsilon$ and $\epsilon/2$. This isolates whether the Laurent expansion, as opposed to mere continuation, is what earns the gains.
Beating the field means three things on degenerate-contact cases, at equal or lower wall time:
- lower error in limit peak stress, and lower error in its gradient, against the GMP reference;
- a lower failure rate in the optimization loop, counting divergence, chattering active sets, and interpenetration;
- lower final true peak stress after optimization.
Why not just…?
Why not just shrink the penalty gradually with continuation and stop when the stress stops changing? This is what the ablation tests. A quadratic penalty has error of order $\epsilon$ in displacement, and conditioning grows like $1/\epsilon$, so double precision loses roughly $\log_{10}(1/\epsilon)$ digits. Stress converging to a few digits can still hide a wrong contact label at a weakly active node. At such a node the gradient is wrong in sign, not merely in magnitude. The trembling-linear-program results show the stable structure appears at moderate $\epsilon$. Continuation alone cannot tell the user that it has. If the heuristic ablation matches the oracle on degenerate benchmarks, we will report that the series machinery is unnecessary.
Why not just use an exact active-set or semismooth Newton solver, which needs no penalty at all? This is the most serious threat. For nondegenerate frictionless contact we expect it to tie with our method, and we will say so. Its weakness is the degenerate regime. With zero gap and zero force at the same point, active-set iterations can cycle, a phenomenon analogous to pivoting cycles in Lemke-type paths. The implicit-function gradient on the returned set is then one arbitrary branch. The experiments therefore stratify results by the number of oracle-flagged degenerate nodes. A win concentrated in the degenerate stratum supports our claim, while a uniform tie refutes it.
Why not just run everything in extended or arbitrary precision? It fixes the forward solve but not the gradient-branch question. It also carries the cost penalty documented for GMP in Quasi-Perfect Stackelberg Equilibrium. Quad precision is cheaper, so we include a quad-precision penalty pipeline in the timing comparison. If it matches our accuracy within a small constant factor of runtime, the precision argument weakens, and the remaining contribution is the degeneracy-aware gradient alone.
Proposed Experiments
Verification. We start with problems whose limit is known. These include a Hertz-type cylinder on a plane, where pressure vanishes smoothly at the patch edge and gives naturally degenerate edge nodes, and a flat punch with an edge singularity. On each we measure:
- the certifying $\epsilon$;
- relative error in peak von Mises stress and in contact-patch width against GMP ground truth;
- gradient error against GMP finite differences.
Joint benchmarks. We then reuse the Single, Complex, and Double Dovetail parameterizations from Gradient-Based Dovetail Joint Shape Optimization for Stiffness. The objective changes from stiffness to peak contact-edge stress under a stiffness constraint. We report final true peak stress, stiffness (checked against their 1036.3 N/mm Complex Dovetail figure as a sanity anchor), optimizer failures, and wall time.
Seal benchmark. A 2D axisymmetric annular groove seal mirrors the eight bionic configurations of Octopus-Inspired Biomimetic Annular Sealing Grooves. We first check that our forward solver reproduces their qualitative ranking, in which groove count dominates and depth matters little. We then ask whether gradient optimization from their "bionic6" design further reduces average von Mises stress beyond their reported 21.71% improvement.
Frictional extension. We repeat the self-contacting beam scenario of Topology Optimization with Frictional Self-Contact on a smooth level-set boundary. We compare corner stress spikes and Newton convergence against their barrier formulation.
Decisive check. The test that decides the proposal is this: on designs where the optimizer has created at least several degenerate contact nodes, does the certified-limit gradient reach lower GMP-verified peak stress than both the exact active-set baseline and the continuation ablation, within the same time? If not, the game-theoretic transfer is a curiosity rather than a tool.
References
- Practical exact algorithm for trembling-hand equilibrium refinements in games
- Computing an Extensive-Form Perfect Equilibrium in Two-Player Games
- Quasi-Perfect Stackelberg Equilibrium
- A survey on algorithms for Nash equilibria in finite normal-form games
- Universal Complexity Bounds Based on Value Iteration and Application to Entropy Games
- Gradient-Based Dovetail Joint Shape Optimization for Stiffness
- Topology Optimization with Frictional Self-Contact
- Single-Level Differentiable Contact Simulation
- Octopus-Inspired Biomimetic Annular Sealing Grooves: Design and Performance Optimization Under Extreme Conditions
- Experimental Study on Anchoring Performance of Short-Lapped-Rebar Splices with Pre-Set Holes and Spiral Hoops