When Parallel Training Slows Down: Graph Surrogates That Time Ecological Tipping Points

Ecosystems such as rainforests, lakes and grasslands can switch abruptly into a degraded state once a slowly rising pressure crosses a threshold. The standard warning sign is "critical slowing down": as the threshold nears, the system recovers more slowly from small knocks, so its fluctuations become larger and more strongly correlated in time. These warnings are computed one sensor at a time and assume the pressure changes slowly. Real tipping, though, plays out on networks whose links are themselves changing, as when deforestation removes the moisture-recycling links of the Amazon. It is often driven fast enough that the classical warnings never appear. And what managers need is not "a transition is coming" but "how long until it cannot be reversed."

We propose to learn a graph-structured model of how each node of an ecological network relaxes after a disturbance, with the recovery rate written into the model as an explicit, time-varying quantity. Fitted to data, this model gives two things. It reads off each node's recovery rate and the speed of the external pressure. Plugged into a known scaling law for time-dependent bifurcations, these two numbers yield a forecast of the point of no return. The methodological twist is that training such a model quickly over long records requires solving the recurrence in parallel across time, and the iterative parallel solvers used for spiking networks converge in a number of sweeps that grows like the system's recovery time. They therefore break down exactly as the system approaches tipping. We turn that failure into a design constraint and a diagnostic.

Core Claim

Time-dependent saddle–node bifurcation shows analytically that a system pushed through a fold does not collapse when its stable state vanishes, at the "breaking time" $\hat\tau$. It lingers in a bottleneck until a later point of no return $\tau^*$. The gap between the two obeys

$$\tau^* - \hat\tau \approx 2.338\,(\lambda V)^{-1/3},$$

where $V$ is the rate at which the driver changes and $\lambda$ is the local curvature of the fold. The law is deterministic and was derived for a scalar model. Nobody has estimated $\lambda$, $V$ and $\hat\tau$ from data on a spatial network. Our claim is that a learned non-autonomous graph model can do this, provided it represents recovery explicitly, and that its estimates time the point of no return better than classical statistics or reservoir-based detectors. This should hold especially under fast forcing, where Huang et al. show that variance and autocorrelation give no warning.

The second, narrower claim concerns training. Consider fixed-point parallel-in-time solvers, of the kind introduced for spiking networks by Fixed-Point Parallel Training, which converges in about $K\approx3$ sweeps, and equilibrium-based training as in Dy-SIGN. Both rely on strong contraction. Take a causal recurrence $h_t = F_t(h_{t-1})$ whose step Jacobian has spectral radius $\rho = 1-\gamma\Delta t$, where $\gamma$ is the slowest recovery rate and $\Delta t$ the time step. Then the error after $k$ Jacobi-style sweeps decays roughly like $\rho^k$. Reaching tolerance $\epsilon$ therefore takes

$$K \approx \frac{\log(1/\epsilon)}{\gamma\,\Delta t},$$

which is proportional to the recovery time and diverges at the fold. Leaky spiking neurons sit far from this regime, which is why FPT's small $K$ works. An ecological surrogate fitted near a tipping point does not. A parallel trainer for this setting must therefore adapt its sweep budget, and the sweep count it needs per time window is itself an estimate of slowing down.

Research Program

The model treats each spatial unit (a moisture box on the Amazon transect, a lake, an oscillator in a coupled system) as a node with a low-dimensional state. Its edge weights at each time come from observed or simulated connectivity, such as moisture transport or hydrological links, and can appear, weaken or vanish. Node dynamics take the form

$$\dot h_i = \big(W_t - W_t^{\top}\big) h_i - \gamma_i(t)\, h_i + \sum_j A_{ij}(t)\,\phi(h_i,h_j) + B\,u_i(t),$$

where $h_i$ is node $i$'s state, $W_t$ a learned matrix, $\gamma_i(t)\ge$ a small negative floor is a learned per-node recovery rate, $A_{ij}(t)$ the time-varying adjacency, $\phi$ a learned coupling, and $u_i(t)$ exogenous drivers such as climate, land use or nutrient load.

The anti-symmetric term comes directly from Anti-Symmetric DGN, which uses $(W-W^\top)$ to give message passing purely imaginary Jacobian eigenvalues and stable gradients across depth, and which was tested only on static graphs. That paper also acts as a constraint here. A purely anti-symmetric model pins the real part of the spectrum at zero and so cannot represent a recovery rate that drifts toward zero. Adding the explicit dissipative term $\gamma_i(t)$ is the minimal change that makes slowing down visible in the parameters.

Driver inputs enter separately from the intrinsic dynamics, so that human disturbance is treated as forcing rather than mistaken for loss of resilience. The LAGOS lake study motivates this choice: across 24,452 lakes, disturbance masked climate-driven shifts, which appeared most clearly in undisturbed watersheds.

From a fitted model we read three quantities. The first is $\gamma_i(t)$ and its trajectory. The second is the time $\hat\tau_i$ at which an extrapolated $\gamma_i$ crosses zero. The third is the curvature $\lambda$ and forcing velocity $V$, obtained by fitting the local normal form to the learned vector field along the driver. Substituting these into the 1/3 law gives a forecast of $\tau_i^*$ with uncertainty from ensembles of fits.

Training unrolls the model over long records with an FPT-style parallel sweep, modified in two ways. The sweep count adapts per segment of time until the residual meets tolerance. Where the learned contraction is too weak, the sequence is split into chunks that are solved in parallel internally and carried sequentially between them, so that training stays well-posed past $\hat\tau$, where equilibrium-based implicit gradients as in Dy-SIGN are undefined. The per-window sweep count $K(t)$ is logged as a by-product. Spiking units are deliberately not used; what we borrow from that literature is the parallel-in-time solver, not the binary events.

Baselines

The naive reference is classical early warning on each node: rolling-window variance and lag-1 autocorrelation with a Kendall-trend alarm, applied to the same series after detrending by the same exogenous drivers our model sees. A second version uses no detrending. No reported result was found on Lacuna for these indicators on our synthetic networks. We will compute them ourselves with window length tuned on a validation set using the same search budget as our model. Beating this baseline means a lower error in the forecast point of no return and earlier lead time at a matched false-alarm rate.

The strongest existing detector is Tipping Point Detection Using Reservoir Computing. RC-TPD passes series through a fixed random recurrent network and regresses tipping position from changes in its readout weights. On the coupled Lorenz system it reports a detection-position error of 38.89 under random reconnection and 28.32 under a sparse-to-dense structural transition, and 23.39 on Lorenz63. We will rerun it on its own coupled-Lorenz protocol, so that our numbers are comparable to these, and on our fold-type networks, giving it the same input series and the same hyperparameter budget as our model. To be fair to a method built on a static reservoir, we will also feed it the adjacency as extra input channels. Beating it means lower position error on its own coupled-Lorenz settings, and a clearly lower error in timing the point of no return on fold networks with varying forcing rates.

The strongest dynamic-graph learner is Coden. Coden pairs parameter-free PageRank propagation with a continuous state-space model and triggers updates lazily, and it reports up to 44.80× training speedup over TGL and ASTGCN on node-classification benchmarks. We give it a regression head trained to predict $\tau^*$ directly, with matched parameters and compute. It tests whether any good temporal graph model learns the timing without explicit recovery rates.

Three ablations isolate our ingredients. The first sets $\gamma\equiv0$, giving pure A-DGN dynamics. The second removes the 1/3 law and regresses $\tau^*$ directly from the learned features. The third trains the identical model with a fixed $K=3$ solver, which shows whether naive FPT-style training degrades near tipping, and with sequential backpropagation through time, which serves as the accuracy reference for the parallel solver.

Why not just…?

Why not just compute variance and autocorrelation per node? They are transparent and need no training. They also fail precisely where timing matters. Huang et al. show they give no precursor under rate-induced tipping. They cannot separate $\hat\tau$ from $\tau^*$, which differ by an amount that depends on $V$. And they treat nodes in isolation, while the South American monsoon study shows deforestation in the east cutting western precipitation by up to 40%. Our forcing-rate sweep tests this directly. If classical indicators match us at slow rates, as we expect, and lose only at fast rates, the advantage is real but limited to that regime, and we will report it as such.

Why not just extend reservoir computing with graph inputs? This is a serious threat, since RC-TPD needs no gradient training and already handles coupled systems. Its readout-weight drift, however, signals that something is changing; it does not yield $\lambda$, $V$ or a recovery rate, so it cannot apply the scaling law or extrapolate to forcing rates it was not trained on. The graph-augmented RC-TPD baseline, evaluated on held-out forcing rates, settles this. If it times $\tau^*$ as well as we do on unseen rates, the mechanistic readout adds nothing, and we would say so.

Why not just read eigenvalues off any learned model with automatic differentiation, and why bother with parallel training at all? For the first part, this is a fair objection to treating $K(t)$ as a warning. We therefore make $K(t)$ a secondary result and test it against the model's own Jacobian spectrum. It is worth reporting only if it tracks the spectrum at negligible extra cost. For the second part, sequential training is adequate for 34 annual lake observations. It becomes the bottleneck for daily reanalysis over decades on a 100-box transect, and for the thousands of forcing-rate ensembles needed to calibrate uncertainty. The decisive reason for the solver work is still well-posedness, which the fixed-$K$ ablation tests, rather than speed.

Why not just fit the known mechanistic model? When the equations are known, as for the monsoon box model, parameter estimation should win, and we include it as an oracle upper bound. Our case rests on systems where no such model exists, such as lakes and soil communities.

Proposed Experiments

The primary testbed is the moisture-transport box model of the monsoon paper. Its published reference points are bistability at amplification at or above 2.75, a fold near 40 deforested boxes, and irreversibility at 3.00. We simulate it with stochastic forcing, deforestation advancing at rates $V$ spanning two orders of magnitude, and edges removed as boxes are deforested. Ground-truth $\hat\tau$ and $\tau^*$ are known for every run. We train on a band of rates and test on held-out slower and faster rates. The metrics are absolute error in $\tau^*$, lead time at 5% false alarms, and the fitted exponent of $\tau^*-\hat\tau$ against $V$, which should approach $-1/3$ if the model has captured the bottleneck.

The second testbed is RC-TPD's coupled Lorenz system with random reconnection and sparse-to-dense transitions, scored by RC-TPD's position-error metric for direct comparison. The third is a field study on LAGOS-US chlorophyll-a, using watershed links as edges and land use and nutrients as drivers. There is no ground-truth tipping time, so this is a check of plausibility rather than accuracy. The test is whether separating disturbance as a driver makes declining $\gamma$ appear in disturbed watersheds at rates closer to those the LAGOS study found in undisturbed ones.

Four decisive checks decide the proposal. First, on held-out fast forcing, our $\tau^*$ error must beat both graph-augmented RC-TPD and the Coden head; otherwise the scaling-law readout is not earning its place. Second, the $\gamma\equiv0$ ablation must fail to localize tipping, which confirms that explicit dissipation matters. Third, the fixed-$K$ solver must visibly lose accuracy relative to backpropagation through time as runs approach $\hat\tau$, while the adaptive solver does not. Fourth, $K(t)$ should correlate with the inverse of the model's leading Jacobian eigenvalue, with the expected $1/\gamma$ form. If it does not, the solver diagnostic is dropped and the method stands or falls on the first two checks.