A rupture on a frictional interface, such as a fault, a glacier bed or a lab slider, is usually pictured as one front sweeping along the surface. Simulations show something messier. Fronts arrest and restart, slow slip creeps in behind a stalled tip, and separate slip patches merge. A single physical event can therefore occupy several disconnected stretches of the interface at once. Computational linguistics has a precise language for exactly this. A constituent that is split into $k$ disconnected pieces has fan-out $k$. Grammars that cap fan-out at a fixed number are called mildly context-sensitive. They are expressive enough for crossing, discontinuous structure but can still be parsed in polynomial time.
This proposal asks a concrete question with a yes-or-no answer. When a rate-and-state interface is made longer, does the fan-out of its slip events stay bounded, or does it grow with system size? Rate-and-state friction is the standard law in which resistance depends on slip speed and on a slowly evolving "state" that remembers contact history. We will measure fan-out across the phase diagram of a minimal rupture model. We then test whether the answer is useful in practice: a parser whose fan-out cap is set by the measurement should split sparse slip records into causally distinct events better than contiguity-based or unconstrained alternatives. If fan-out is bounded, interval-grammar parsing becomes a tractable tool for event catalogs. If it grows, grammar-based approaches are ruled out for this physics.
Core Claim
We hypothesize that, under homogeneous loading, causally defined slip events have fan-out at most 2, independent of chain length. Steady regimes (slow, sub-Rayleigh, supershear, expanding crack, slip pulse) should have fan-out 1. The unsteady start-stop regime and coalescence-driven fronts should reach fan-out 2. Fan-out above 2 should appear only when prestress is heterogeneous. If this holds, two consequences follow. Rupture structure falls in the same bounded-discontinuity class that the parser of Span-Based LCFRS-2 Parsing was designed for. Fan-out also serves as an order parameter that front speed alone does not supply.
Grounding and Prior Work
The physics supplies the generator. Minimal model for slow, sub-Rayleigh, supershear, and unsteady rupture propagation uses a one-dimensional chain of spring-coupled blocks with velocity-strengthening friction, run over 40,000 simulations. It maps a full set of rupture modes onto two dimensionless knobs: prestress $\bar\tau$ and interfacial viscosity $\bar\alpha$. It identifies the start-stop regime but classifies runs only by front kinematics. We reuse its setup and add a structural measurement it does not make. Two companion works motivate the hypothesis that discontinuity is intrinsic rather than incidental. The moment duration scaling relation for slow rupture arises from transient rupture speeds shows that slow fronts are long-lived transients, never steady. Secondary acceleration of slip fronts driven by slow slip event coalescence shows that separate slip regions merge and re-accelerate. The rate-and-state law itself is the one used in Rate-and-state friction explains glacier surge propagation, with state evolving as $\dot\theta = 1 - V\theta/d_c$. Here $V$ is slip speed and $d_c$ is a characteristic slip distance. We use this aging law in a variant of the chain, which gives repeated stick-slip cycles rather than single ruptures.
The linguistics supplies the measurement and the parser. The criteria in Factoring Recursion and Dependencies define mild context-sensitivity: polynomial parsing, constant growth, and bounded crossing. "Bounded" is the property we test empirically on a physical system. Characterizing Structural Descriptions produced by Various Grammatical Formalisms shows that linear, non-erasing interval rewriting (LCFRS) is polynomial for fixed fan-out. Parsing cost scales with fan-out, which is why the cap matters. The method comes from two sources. Span-Based LCFRS-2 Parsing scores a fan-out-2 item as a left span, a gap and a right span. Unsupervised Discontinuous Constituency Parsing with Mildly Context-Sensitive Grammars supplies the $O(\ell^5)$ restricted LCFRS-2 chart and tensor-factored rule scores.
Two results constrain the design. The span-based oracle analysis shows that fan-out 2 leaves a recall ceiling on real syntax, so we must measure the ceiling on physical data rather than assume one. The unsupervised induction work fails to find discontinuity when it is rare (about 1.8% of English constituents). For that reason we train the parser with supervision from simulator labels and do not induce grammars.
The nearest physics-side neighbor is heterogeneous cellular automata for foreshocks. It uses local heuristic rules and makes no structural measurement. Thread Automata suspends and resumes parsing threads, which looks superficially like start-stop fronts. We do not borrow that analogy. Our link is fan-out as a measurable quantity, not a metaphor between mechanisms. No work on Lacuna connects the two fields.
Research Program
Step 1: generate data. We simulate the velocity-strengthening chain on a grid over $(\bar\tau, \bar\alpha)$, matching the cited phase-diagram setup. Chain lengths run from $N=500$ to $N=16{,}000$ blocks. We also run a rate-and-state aging-law variant under slow tectonic-style loading, which produces long multi-event catalogs. A third set adds spatially random prestress with controlled correlation length.
Step 2: label causal events. Ground-truth event identity is computed by causal tagging in the simulator. Each nucleation carries a label. A block's slip is attributed to the event whose stress perturbation pushed it past threshold, tracked through the elastic coupling. A block reactivated by a different nucleation receives that new label.
Step 3: measure fan-out. At each instant, an event's slipping set is a union of disjoint intervals along the chain. The event's fan-out $k$ is the maximum number of such intervals over its lifetime. Gap degree and well-nestedness (whether two events' gaps interleave without crossing) are computed exactly as in treebank analyses.
Step 4: test boundedness. For each phase-diagram cell we fit how the 99th-percentile fan-out scales with $N$. Bounded fan-out predicts a plateau. Unbounded growth would show up as a power law or a logarithmic trend.
Step 5: build the parser for the application. The task mimics real monitoring. The input is a sparse, noisy, time-downsampled slip record, for example 64 virtual sensors with additive noise. The output segments slip into events, where each event may be discontinuous. We adapt the span-gap-span scoring of the LCFRS-2 parser to space-time "constituents." We also build an LCFRS-3 variant using the restricted rule shapes, to test whether the extra fan-out ever pays for its cost.
Baselines
The naive reference point is connected-component segmentation. Every spatially contiguous slipping region that persists in time counts as one event, so the method is fan-out 1 by construction. It is what a catalog builder would do by default. It is run on identical sensor records with its thresholds tuned on the validation split. Beating it means higher event-level discontinuous F1, with event-level F1 defined like the discontinuous F1 of parsing evaluation. Discontinuous F1 measures how well the method recovers events that span several disjoint intervals. On those events this baseline should score zero by definition, just as PCFGs do in Unsupervised Discontinuous Constituency Parsing.
The strongest competitor is an unconstrained neural segmenter. It is a transformer over the sensor sequence that predicts pairwise "same-event" links, which are then clustered. It can represent any fan-out and needs no grammar, so it is the real threat. It gets the same training data, the same parameter budget within 10%, and the same hyperparameter search budget as our parser. To count as beating it, the grammar-constrained parser must achieve higher discontinuous F1 or equal F1 with clearly better robustness when sensors are halved. Whether the parser also produces structurally valid outputs is reported separately.
The ablation that isolates the key ingredient is the same span parser with the fan-out cap varied from 1 to 2 to 3, with everything else fixed. If our measurement is right, accuracy should jump from cap 1 to cap 2 and stay flat at cap 3 under homogeneous loading. Under heterogeneous loading it should keep rising at cap 3. A second ablation removes the explicit gap-span score, which tests whether modeling the arrested region carries information.
For reported numbers, there is no reported result on Lacuna for any rupture-event segmentation task. As a sanity anchor for our parser reimplementation, Unsupervised Discontinuous Constituency Parsing reports NEGRA F1 46.1 and discontinuous F1 8.0 for TN-LCFRS with 4,500 nonterminals, against PCFG F1 40.8. the Chomsky-Schützenberger MCFG parser reports supervised NEGRA F1 of about 74.3%. We reproduce the supervised span-parser regime on NEGRA before switching domains.
Why not just…?
Why not just classify the regime from front speed, as the phase-diagram paper already does? Front speed labels the mode but does not say which slipping blocks belong to which event. In start-stop and coalescence runs, a speed-based summary merges or splits events wrongly. We test this directly by asking whether fan-out explains variance in event moment and duration, within cells of equal mean front speed. If it adds nothing there, we report fan-out as a re-description, not a new order parameter.
Why not just use an unconstrained neural segmenter? It may match the parser on accuracy, which is a genuine risk. If it does, the application claim weakens to robustness and interpretability. The boundedness measurement in Steps 3 and 4 still stands on its own, because it is a property of the physics, not of any model. Low-sensor and length-extrapolation tests are where a fan-out prior should matter most, and they settle the question.
Why not just use classical static/kinetic friction and skip rate-and-state? With a single drop from static to kinetic friction, the interface has no velocity strengthening and no slow state evolution. Instabilities at frictional interfaces shows that without high-speed strengthening, fronts run away to the elastic wave speed. That removes the arrest and creep that generate discontinuity in the first place. We include a static/kinetic control run and predict it yields fan-out 1 almost everywhere.
Decisive Checks
The hypothesis fails if 99th-percentile fan-out grows with $N$ under homogeneous loading in any cell of the phase diagram. It also fails if the cap-2 parser shows no gain over cap 1 on start-stop runs. Either failure ends the grammar-based application. It succeeds if fan-out plateaus at 2 for homogeneous loading and increases with prestress heterogeneity. It also requires the cap-2 parser to beat connected components on discontinuous F1 and to match or exceed the unconstrained segmenter when sensors are halved.
References
- Span-Based LCFRS-2 Parsing
- Minimal model for slow, sub-Rayleigh, supershear, and unsteady rupture propagation along homogeneously loaded frictional interfaces
- The Moment Duration Scaling Relation for Slow Rupture Arises From Transient Rupture Speeds
- Secondary acceleration of slip fronts driven by slow slip event coalescence in subduction zones
- Rate-and-state friction explains glacier surge propagation
- Factoring Recursion and Dependencies: an Aspect of Tree Adjoining Grammars (Tag) and a Comparison of Some Formal Properties of Tags, GPSGs, Plgs, and LPGS
- Characterizing Structural Descriptions produced by Various Grammatical Formalisms
- Unsupervised Discontinuous Constituency Parsing with Mildly Context-Sensitive Grammars
- Study on Simulation of Foreshock Activity Properties before Strong Earthquake Using Heterogeneous Cellular Automata Models
- Parsing Mildly Context-Sensitive Languages with Thread Automata
- Implementation of a Chomsky-Schützenberger n-best parser for weighted multiple context-free grammars
- Instabilities at frictional interfaces: Creep patches, nucleation, and rupture fronts