Truncating Elastic Interaction Series Without Breaking Them: A Connectivity Criterion from Homotopical Localization

When small rigid particles sit inside a soft elastic gel, pushing or twisting one particle deforms the gel and moves its neighbours. Physicists predict these gel-mediated interactions with series in the inverse distance between particles. The series is built from Faxén relations, which are formulas giving the force and twist a particle feels from the surrounding deformation, and from the method of reflections, in which each particle's disturbance bounces off its neighbours and back again. The most detailed version of these series now runs to sixth order in inverse distance. Any practical use of it, however, must cut the series off somewhere, and nobody has asked whether the cut-off version still obeys the rules the exact physics must obey. Those rules are that the coupling matrix is symmetric (reciprocity: the effect of A on B equals the effect of B on A), that it stores positive energy, and that three-particle terms agree with what the two-particle pieces compose into.

Algebraic topology has studied this problem in the abstract for decades. It asks when "localizing" or truncating an object keeps its multiplication intact. The answer found by Casacuberta and Gutiérrez is that structure survives if the truncation commutes with degree shifts, or if everything involved is connective, meaning it has nothing in negative degrees. We propose to carry that criterion over to elastic inclusion series as an engineering rule. Cutting the series off preserves the way reflections compose exactly when every term has non-negative degree in a single small parameter. When a wall, close contact, or a second length scale introduces terms that behave like negative degrees, the series must first be reorganized into a "connective" form and only then truncated. The output is a truncation procedure for the existing sixth-order matrices, tested against full numerical solutions, with a clear prediction about where naive truncation breaks.

Core Claim

Displacement field around a rigid sphere in a compressible elastic environment, corresponding higher-order Faxén relations, as well as higher-order displaceability and rotateability matrices derives the two Faxén relations beyond the stresslet for compressible gels. It extends the matrices that map applied forces and torques to particle displacements and rotations (the "displaceability and rotateability matrices") up to order $r^{-6}$. At orders five and six it finds translation–rotation and rotation–rotation couplings that depend on Poisson's ratio $\nu$. That paper supplies the objects we study. Write the grand coupling matrix for $n$ particles of radius $a$ as

$$M(\varepsilon) = \sum_{k \ge 0} M_k \, \varepsilon^k, \qquad \varepsilon = a/r,$$

where $M_k$ collects all reflection terms of order $k$ in the inverse separation $r$. Reflection terms are produced by composing single-particle operators, so a term of order $k_1$ composed with one of order $k_2$ lands in order $k_1+k_2$. Truncation $T_N$, which discards all orders above $N$, then respects composition, $T_N(AB) = T_N(T_N A \, T_N B)$, provided no term carries a negative degree. This is the elementary content of the connectivity hypothesis in Homotopical localizations of module spectra. That paper shows that a localized ring spectrum keeps a unique multiplication when either the localization commutes with suspension or both the spectrum and its localization are connective. The authors trace failures to dimension shifts that push products out of the range the truncation sees.

Our claim is that three physically realistic situations make the elastic series non-connective in exactly this sense. In each, naive truncation at $r^{-6}$ produces matrices that violate reciprocity, lose positive definiteness, or give three-body predictions inconsistent with composed two-body ones. In each, a specific reorganization fixes the problem, playing the role a connective cover plays in topology. The three situations are these. First, a confining wall at distance $h$ introduces a second parameter $\delta = a/h$, so terms $\varepsilon^k\delta^m$ are graded inconsistently if one truncates in $k$ alone. Second, near-contact separations $r \to 2a$ make $\varepsilon$ not small, and gap-type corrections enter with negative powers of $(r-2a)/a$. Third, in the viscoelastic extension the memory kernel adds time derivatives that shift the effective order of terms.

We are explicit that the topology serves as a lens, not as a transferred theorem. Composition consistency follows from the grading argument. Positive definiteness does not follow automatically, and we treat it as an empirical question with a two-body remainder bound as the only target proof.

Research Program

The first step is bookkeeping. We will re-express the order-by-order results of Forces and torques on rigid inclusions in an elastic environment: Resulting matrix-mediated interactions, displacements, and rotations and the sixth-order paper as a graded composition algebra with two kinds of inputs, forces and torques. Every reflection term gets an explicit degree and a record of which single-particle operators produced it. This is the concrete content of the operadic picture in Localization of algebras over coloured operads. That work proves that localized algebras over a coloured operad (a bookkeeping device for operations with several typed inputs) keep their structure when the localization is compatible with the product. It also notes that truncation-type localizations need connectivity hypotheses, which is the case we are in.

We do not aim to prove operad-level theorems. The aim is a symbolic implementation that can generate the reflection series to arbitrary order and record degrees, so that truncation consistency can be checked exactly rather than by inspection.

The second step designs the reorganizations. For walls, we regrade by total degree $k+m$ in a single combined parameter and truncate on that. For near contact, we split off the singular gap dependence into a separately computed block and expand only the regular remainder in $\varepsilon$. This is analogous to building a connective cover by removing the negative part. For viscoelastic media, following Memory-based mediated interactions between rigid particulate inclusions in viscoelastic environments, which couples elastic and viscous responses through the memory parameter $\mathcal{V} = \gamma\eta/\mu$, we check whether truncation keeps the time-dependent coupling passive, meaning it never generates energy. This third item is an extension, attempted only after the static cases succeed.

The third step is validation against full numerical solutions of the compressible Navier–Cauchy equations, the linear equations of small-deformation elasticity, around rigid spheres. We use boundary-element computations, which we must build ourselves because the research file contains no public multi-inclusion elasticity dataset.

Baselines

The do-nothing reference is the naive sixth-order truncation as published in the sixth-order Faxén paper, used directly for all separations, Poisson ratios, and geometries. This is what a practitioner would do today. The question is whether it ever fails in the ways we predict. If naive truncation turns out never to violate reciprocity, positive definiteness, or composition consistency in the regimes we test, the proposal's premise is wrong, and we will report that.

The strongest analytical competitor is the fourth-order truncation of Forces on Rigid Inclusions in Elastic Media and Resulting Matrix-Mediated Interactions. It is the only version validated experimentally. Nickel microspheres in a silicone gel were rotated by a magnetic field, and the shear modulus fitted from 2-, 3- and 4-particle configurations came out at 83.0 ± 14.3 Pa, 76.3 ± 11.7 Pa and 85.7 ± 12.6 Pa. That paper reports this cross-configuration consistency as its evidence that the series captures the physics. Beating it means two things. On the same boundary-element ground truth, a connective sixth-order truncation must give lower displacement and rotation error at moderate separations ($r/a$ between about 2.5 and 5), where the fourth-order series should begin to fail. And if the raw tracking data can be obtained from the authors, which we have not confirmed, the fitted moduli across particle numbers must have smaller spread.

The strongest structure-repairing competitor is generic post-hoc repair. One symmetrizes the naive truncated matrix and clips negative eigenvalues to project it onto the positive semidefinite matrices, as is common practice for truncated mobility matrices in suspension simulations. This is the ablation that isolates our key ingredient. If post-hoc repair matches connective truncation in accuracy, the regrading step adds nothing.

Full boundary-element solutions are the ground truth, not a competitor. We will report their cost so the speed side of the comparison is visible. A learned graph-network surrogate in the style of Learning Mesh-Based Simulation with Graph Networks, trained on the same boundary-element runs, is a secondary baseline. For it, no reported result was found on Lacuna for multi-inclusion elastic interactions.

All analytical variants share identical configurations, $\nu$ values and reference solutions, and none has tunable parameters. The learned baseline gets the same training data and a fixed hyperparameter budget.

Why not just…?

Why not just run finite-element or boundary-element solvers for every configuration? They are the ground truth here, and for a handful of static configurations they suffice. The inclusion papers note, however, that full-field methods must re-mesh the matrix for every configuration change. Machine learning acceleration of simulations of Stokesian suspensions reports days of cluster time for high-fidelity suspension runs, even in 2D. Dynamic magnetic-elastomer simulations with many particles need a closed-form coupling matrix, and the question is whether that matrix can be trusted when cut off.

Why not just stop at fourth order, which already matched experiment? The sixth-order paper shows that the rotation–rotation and translation–rotation couplings, and much of the $\nu$ dependence, first appear at orders five and six. Fourth order therefore cannot describe torque-driven systems in compressible gels at all. Its experimental success was established for magnetic forces at moderate separation, and our separation sweep measures directly where it loses accuracy.

Why not just symmetrize and clip eigenvalues after truncating? This is a real threat, because it is cheap and guarantees two of the three properties. It does not restore composition consistency, so three-body predictions can still disagree with composed two-body terms. Clipping also changes the matrix in an uncontrolled way that may cost accuracy. The ablation above settles the question. If clipped naive truncation matches connective truncation on error and on three-body consistency, we will recommend clipping and treat the topology as having contributed only a diagnosis.

Why not just train a neural surrogate? PDEBench: An Extensive Benchmark for Scientific Machine Learning finds that standard learned PDE models frequently violate boundary conditions and conservation laws despite low pointwise error. A learned coupling matrix has no built-in reciprocity either. We include one as a baseline, but a surrogate cannot answer when the analytical series is structurally safe to truncate.

Proposed Experiments

We will compute boundary-element reference solutions for 2-, 3- and 4-sphere configurations. Separations will range from $r/a = 2.1$ to 10 and Poisson's ratio from 0 to 0.5, with point forces and torques applied to individual particles. Wall configurations will vary $h/a$ from 1.5 to 10.

For every configuration and every truncation scheme we will measure four things. The first is relative error in displacements and rotations. The second is reciprocity violation, measured as the norm of the antisymmetric part of the grand matrix relative to the whole. The third is the smallest eigenvalue of the grand matrix, with the fraction of configurations that are indefinite. The fourth is three-body consistency, measured as the mismatch between the truncated three-particle matrix and the truncated composition of its two-particle reflections.

The decisive check is a predicted map of failure. Naive truncation should keep composition consistency in unbounded media at all separations where the series converges. It should break consistency once a wall is present. It should lose positive definiteness near contact. Connective truncation should remove the wall failure entirely and shrink the near-contact failure region, at equal or lower error than eigenvalue clipping.

If naive truncation shows no failures, or if clipping does as well, the connectivity lens is not useful for this problem, and that negative result is itself worth reporting. A second check uses the magnetic-gel benchmark. If the original tracking data can be obtained, we refit the shear modulus with fourth-order, naive sixth-order and connective sixth-order matrices and compare the spread across 2-, 3- and 4-particle configurations against the published values.