When a soft outer layer grows faster than the tissue beneath it, the layer buckles into folds. This is "differential growth," and it is the leading physical account of how the cortex acquires its gyri and sulci. Buckling is also notoriously sensitive. Tiny perturbations decide which of several competing fold patterns appears. Yet real brains fold in species-consistent ways, so something must be steering the physics. One candidate is the set of discrete anchors under the cortex, such as axonal bundles and radial fibers, which pull inward on the growing surface. We propose to ask the inverse question directly. Given a target fold pattern, what is the smallest set of tethers that makes that pattern appear reliably across random perturbations? We pose this as a search over graphs, where the nodes are surface and core locations and the edges are tethers. The score of a candidate graph is estimated by running a stochastic buckling simulator several times.
This search sits where three known failure modes meet. First, greedy edge-by-edge search gets trapped, because useful configurations often need several edges added together. Second, the simulator's "noise" is not small jitter around a mean. It is a discrete jump between fold modes. Third, every evaluation is an expensive mechanics solve. Our central bet is that one technique handles all three. We score candidates by the probability of producing the target mode. We compare candidates on shared perturbation seeds, which cancels most of the branch-switching noise. We admit a tether only when a paired statistical test says it helps. We look ahead over tether pairs, because single tethers often do nothing on their own.
Core Claim
Define a tether graph $G$, a perturbation seed $\omega$ (random imperfections in the initial mesh and growth field), a simulator $S(G,\omega)$ that returns the buckled surface, and a classifier $m(\cdot)$ that labels the resulting fold mode. The quantity to optimize is the reproducibility of the target mode $m^*$, penalized by the number of tethers:
$$p(G) = \Pr_{\omega}\big[m(S(G,\omega)) = m^*\big], \qquad \max_G \; p(G) - \lambda |G|,$$
where $\lambda$ is the cost per tether. Every candidate is evaluated on the same seeds $\omega_1,\dots,\omega_K$. A proposed change from $G$ to $G'$ is then judged by its paired outcomes, that is, by counting the seeds on which the change flips the mode toward or away from $m^*$. Seeds that land in the same branch under both graphs contribute no variance. This is the "common random numbers" idea applied to bifurcation noise.
We claim that this paired, mode-probability objective has two consequences, and that it matters more than the choice of outer search algorithm. It lets a conservative acceptance rule reject spurious tethers. It also makes short lookahead affordable. In the regime where target patterns require coordinated tethers, as conjunction-type 3-hinge gyri plausibly do, we expect the full method to recover planted tether graphs that greedy search with seed averaging misses at the same simulator budget. If greedy search with paired seeds and averaging matches it everywhere, the claim is wrong, and the contribution reduces to the objective alone. The experiments are designed to tell these outcomes apart.
Prior Work and Where This Sits
The mechanics side motivates the problem. The circular bilayer model of Mechanism Exploration of 3-Hinge Gyral Formation and Pattern Recognition shows that fold shape balances outward expansion against inward axonal pull. Its tether configurations, however, are set by hand across forward runs. We automate exactly that step and run it in reverse. Role of data-driven regional growth model in shaping brain folding patterns fits regional growth laws to dHCP and BCP MRI data using symbolic regression and simulates them forward in finite elements. It cuts sulcal depth error from over 50% to 13.41% in one region. That pipeline is forward-only and contains no tethers. We take its growth fields as fixed inputs and search only over anchors. Dynamic patterns of cortical expansion during folding of the preterm human brain shows that migrating expansion hotspots act as a blueprint. This competes with our hypothesis, because growth heterogeneity alone might explain reproducibility, and we test that directly.
The search side supplies the method and also the warnings. Less Greedy Equivalence Search rejects edge insertions that contradict a conditional-independence test even when the noisy score improves. Its authors report roughly halved structural Hamming distance relative to GES at 150 or more variables. Our paired test is the mechanical counterpart: a tether that does not change mode outcomes on shared seeds is refused. Tree search in DAG space with model-based reinforcement learning for causal discovery attacks the horizon effect, where a needed move first lowers the score, using Monte Carlo tree search. It beats hill-climbing with far fewer score calls. We borrow shallow lookahead but replace its deterministic BIC score with a noisy Bernoulli one. Two papers constrain our design. Permutation-based Causal Inference Algorithms with Interventions argues that naive greedy search over discrete graphs can fail even with unlimited data. Improving Causal Discovery By Optimal Bayesian Network Learning documents precision saturation, where early greedy mistakes are reinforced rather than corrected. Both imply that adding seeds to a greedy search will not rescue it once a wrong tether has been accepted. That is a testable prediction, not an assumption.
The noise side supplies the budget logic. Depth over Fidelity in Fixed-Budget Noisy Evolution Strategies finds that under fixed budgets, spending evaluations on more updates often beats spending them on cleaner rankings. We expect paired seeds to change this trade-off, because pairing makes each comparison cheap to denoise, and we measure this. High Probability Complexity Bounds for Line Search Based on Stochastic Oracles builds its acceptance tests on zero-mean, bounded noise. Branch switching violates that assumption, which is why we test discrete mode outcomes rather than relaxing a descent condition. BoGrape: Bayesian optimization over graphs with shortest-path encoded and GIBBON: General-purpose Information-Based Bayesian OptimisatioN are the strongest model-based competitors, and we run them as baselines.
Research Program
We will build a fast two-dimensional bilayer simulator with geometry matching the circular model of the 3-hinge paper. It will have a stiff, growing outer ring on a softer core, with tethers modeled as linear springs from chosen surface nodes to chosen core nodes. It is solved quasi-statically with seeded imperfections. A three-dimensional spherical-shell version is a second stage, run only if the 2D results justify it. The mode classifier is defined geometrically. It counts and locates sulci by curvature sign changes and labels hinge junctions, so that "same pattern" has an explicit, checkable meaning.
The search operates over a candidate set of a few dozen allowed tether sites. Each step proposes single-tether and tether-pair additions and deletions. Pairs are restricted to geometrically nearby sites, which keeps the branching factor manageable. Candidates are scored on a shared pool of $K$ seeds. Pool sizes are adaptive: candidates that are clearly losing are dropped after a few seeds, in the style of successive halving. An addition is accepted only if a paired sign test on mode flips rejects "no effect" at a fixed level. We also study identifiability. Several tether graphs may produce the same mode distribution, which is a mechanical analogue of Markov equivalence classes. We therefore report recovery both against the planted graph and against the set of graphs whose estimated $p(G)$ lies within error of it.
Baselines
The naive reference point uses no tethers: the bilayer grows with uniform growth, and separately with the hotspot growth fields, and nothing else. This tells us how reproducible the target mode is from growth alone. If $p$ is already near one for the hotspot field, tethers are unnecessary for that target, and the question becomes moot there.
The first search baseline is greedy hill-climbing over single tethers. We run it in two forms: one simulator run per candidate, and the strongest form, with the same shared seeds and the same total simulator budget as our method. This is the real competitor, and both forms receive the same seeds, sites, and compute. We count our method as beating greedy only if it achieves higher held-out $p(G)$, or lower structural Hamming distance to the planted graph, with a confidence interval across independent target instances that excludes zero.
The second baseline is a direct transplant of the tree search of Tree search in DAG space with model-based reinforcement learning for causal discovery. It uses independently sampled seeds and no paired test, and gets the same budget. It isolates whether lookahead alone is enough.
The third baseline is model-based: graph Bayesian optimization as in BoGrape: Bayesian optimization over graphs with shortest-path encoded, with a GIBBON-style information-gain acquisition from GIBBON: General-purpose Information-Based Bayesian OptimisatioN in place of expected improvement. Our candidate site sets stay under the roughly 30-node range in which BoGrape's mixed-integer solve is reported to be tractable.
The fourth baseline is a continuous relaxation. Every allowed tether gets a stiffness weight with an L1 penalty, the weights are tuned by finite differences or adjoint gradients averaged over seeds, and the result is thresholded.
Three ablations isolate our key ingredient. Each removes one component: paired seeds (replaced by independent seeds), the conservative paired test (replaced by accept-on-improvement), and pair lookahead (singles only).
There is no reported result found on Lacuna for tether-graph recovery or mode-reproducibility on any buckling benchmark, so no published number exists to beat on the main task. The closest reported figure on real anatomy is the sulcal depth MAPE of 13.41% in Region 9 (supramarginal) from Role of data-driven regional growth model in shaping brain folding patterns. We use it only as a loose reference in the 3D stage, because our model geometry differs.
Why not just…?
Why not just run greedy search and average over more seeds? This is the most serious threat. If precision saturation is real here, as Improving Causal Discovery By Optimal Bayesian Network Learning found in causal discovery, more seeds will sharpen greedy search's commitment to a wrong first tether without undoing it. The budget-matched greedy baseline with shared seeds tests exactly this. If it matches our method on pair-requiring targets, we will drop the search contribution and present the paired mode-probability objective as the result.
Why not just relax tethers to continuous stiffnesses and use gradients? Near a bifurcation the mode label is a step function of the stiffnesses. Gradients of any smooth surrogate say little about which branch will be taken. The continuous-relaxation literature already concedes non-convex traps, as in Gradient-Based Neural DAG Learning. Our relaxation baseline measures whether thresholded continuous solutions land in the right equivalence class. It may well win on targets that involve no mode competition, and we will report where it does.
Why not just use Bayesian optimization on the simulator? Expected improvement chases spurious peaks under noise, and information-based acquisition only partly fixes this. A Gaussian-process surrogate over graphs must also learn a discontinuous, multimodal response from few runs. This is a real competitor at small site counts, and the GIBBON-plus-BoGrape baseline settles the question empirically.
Why not just rely on growth hotspots and skip tethers? Dynamic patterns of cortical expansion during folding of the preterm human brain suggests that hotspots may suffice. The no-tether reference with hotspot growth measures how much reproducibility growth heterogeneity buys before any tether is added.
Proposed Experiments
The first experiment is planted recovery in 2D. We sample hidden tether graphs with 2 to 8 tethers over 20 to 40 sites, in two groups. In one group every tether individually shifts the mode. In the other, tethers act only in pairs. Each hidden graph defines a target mode, and every method gets a fixed budget of simulator runs. The metrics are held-out $p(G)$ on fresh seeds, structural Hamming distance to the planted graph and to its estimated equivalence class, and the number of runs needed to reach $p \ge 0.9$. The decisive check is the gap between our method and seed-shared greedy search in the pair-only group. We predict a large gap there and a negligible gap in the single-effect group. A uniform gap, or no gap, falsifies the horizon-effect explanation.
The second experiment measures noise structure and budget. We sweep $K$ and the perturbation amplitude and record how often branch switching occurs. We then compare paired and independent seeds at equal budget, testing whether pairing reverses the depth-over-fidelity preference reported in Depth over Fidelity in Fixed-Budget Noisy Evolution Strategies.
The third experiment uses anatomical targets. We extract 2D cortical sections from dHCP surfaces and use the regional growth fields of Role of data-driven regional growth model in shaping brain folding patterns as growth inputs. We then ask whether a sparse tether graph raises reproducibility of the observed sulcal pattern above the growth-only reference, and whether the recovered tether sites fall near known hinge locations. Agreement there would be suggestive, not proof. Only if the 2D results hold will we move to a 3D shell and report sulcal depth MAPE against the 13.41% reference, stating clearly that the setups differ.
References
- Mechanism Exploration of 3-Hinge Gyral Formation and Pattern Recognition
- Role of data-driven regional growth model in shaping brain folding patterns
- Dynamic patterns of cortical expansion during folding of the preterm human brain
- Less Greedy Equivalence Search
- Tree Search in DAG Space with Model-based Reinforcement Learning for Causal Discovery
- Permutation-based Causal Inference Algorithms with Interventions
- Improving Causal Discovery By Optimal Bayesian Network Learning
- Depth over Fidelity in Fixed-Budget Noisy Evolution Strategies
- High Probability Complexity Bounds for Line Search Based on Stochastic Oracles
- BoGrape: Bayesian optimization over graphs with shortest-path encoded
- GIBBON: General-purpose Information-Based Bayesian Optimisation
- Gradient-Based Neural DAG Learning