Paying for Attribution Only Where Dynamics Bend: Error-Bounded Path Attribution Through Nonlinear Rollouts

When a controller misses a target or breaks a safety limit, engineers want to know which deviations caused it. The deviations of interest are the initial-state error, the disturbance at step 12, and the actuator noise on one channel. Integrated Gradients is a method that answers this kind of question by averaging gradients along a straight line from a reference input to the actual input. It is the standard way to get attributions that add up exactly to the change in output. Its cost is the problem. Each point on the line needs a full forward and backward pass, and practitioners use 50 to 300 points. For a control system, each "pass" is a backward pass through an entire unrolled trajectory. That is too expensive to run online, and when it is cut to a fixed handful of points, as in Q-value Path Decomposition, there is no indication of how wrong the answer is.

This note rests on one observation. The cost of path attribution is a direct measure of how nonlinear the system is along the path. For a linear system, $x_{t+1}=Ax_t+Bu_t$, superposition holds and a single gradient gives the exact attribution. Every additional quadrature point is spent on curvature. We propose an attribution method for nonlinear rollouts that uses this observation. It places integration points only where the trajectory's sensitivity changes, it reports a per-feature error bound with every attribution, and it reduces to the one-gradient answer when the dynamics are locally linear. Nonlinear control supplies both the regime where the cost matters and a natural reference point: the planned trajectory, run with nominal (zero) disturbances.

Core Claim

Let $J(w)$ be a scalar outcome of a closed-loop rollout. Examples are the terminal cost or the minimum distance to a constraint. Here $w$ stacks the initial-state deviation and all per-step disturbances over horizon $H$. Take the reference $w'$ to be the nominal plan, meaning zero disturbance. With $\Delta = w - w'$, the attribution to component $i$ is

$$A_i = \Delta_i \int_0^1 h_i(\alpha)\, d\alpha, \qquad h_i(\alpha) = \frac{\partial J}{\partial w_i}\big(w' + \alpha\Delta\big),$$

and these attributions sum exactly to $J(w)-J(w')$. One reverse-mode pass through the rollout returns $h_i$ for every $i$ at once. The cost therefore depends on the number of points along $\alpha$, not on the dimension. Suppose $h_i$ changes at rate at most $L_i$ on a subinterval of length $\ell$. Then the trapezoid rule on that subinterval has error at most $|\Delta_i| L_i \ell^2/4$. This is the standard bound for a Lipschitz integrand. $L_i$ is a Hessian-of-rollout quantity, and it is zero for linear dynamics.

The method splits subintervals until the summed bound falls below a user tolerance. It has two tiers. The cheap tier estimates $L_i$ with one Hessian-vector product per subinterval, which gives an error estimate. The certified tier, available when the dynamics have known analytic form, bounds $L_i$ by interval arithmetic over the subinterval, which gives a guaranteed bound.

Two endpoints are close to free. The nominal-trajectory gradient at $\alpha=0$ is often already computed by a gradient-based planner, and the realized trajectory at $\alpha=1$ is executed anyway. The simplest trapezoid rule therefore costs one extra backward pass. Refinement is triggered only near saturation, contact, or switching. The claim we will test is that, on standard nonlinear control tasks, this method matches a 1024-point reference attribution to within tolerance using a small fraction of uniform Integrated Gradients' passes. We also claim that its stated bound holds empirically, and that its cost tracks a controllable nonlinearity parameter.

Relation to Prior Work

Integrated Gradients itself (Axiomatic Attribution for Deep Networks) supplies the axioms and the completeness identity we keep. It uses uniform straight-line quadrature in static input space and gives no error control. Two papers motivate and constrain the adaptive part. Integrated Decision Gradients concentrates samples where the output changes fastest along the path, and it improves AIC on ResNet101 from 0.571 to 0.701. It is the closest method in spirit. It allocates samples by the size of the first derivative rather than by curvature, gives no error bound, and has no formulation for unrolled dynamics. Our step rule differs: a steep but straight region needs no refinement, while a flat region with a sharp bend does.

The iterative refinements (Iterative Search Attribution, Greedy PIG, Guided IG, GGIG) change the path or the objective, and they multiply cost to fix correlation or noise in images. We deliberately keep the straight path and spend effort on accuracy rather than on redefining the attribution.

The one paper that runs Integrated Gradients along real trajectories is Q-value Path Decomposition. It integrates a learned centralized critic along the observed state-action trajectory with a fixed $m=5$ to split team value among agents. We differ in three ways. We differentiate through the dynamics rollout rather than through a critic. Our reference is the nominal plan rather than a terminal state assumed to have zero value, which can be unsafe in constrained control. And step count is set by a bound rather than fixed.

The control papers play the role of motivation and testbed. Köhler et al. obtain fixed-complexity robust tubes only because linear dynamics let one Jacobian describe how errors spread, and they state that the nonlinear extension is open. ABC-LMPC handles stochastic nonlinear dynamics with derivative-free sampling but cannot say which deviation drove a trajectory out of the safe set. Yin et al. certify neural controllers only on linear plants. We do not propose a new tube-based MPC. We supply the cheaper, error-bounded sensitivity object that such an extension would need, and we evaluate it as an attribution method.

Research Program

We will first implement differentiable closed-loop rollouts in JAX. These will cover the ABC-LMPC tasks (point mass, 7-link reacher, inverted pendulum swing-up with process noise) and the Imitation Learning with Stability and Safety Guarantees tasks (inverted pendulum, Generic Transport Model aircraft, vehicle lateral control). They will also include the mass-spring-damper of Köhler et al. with an added cubic spring term whose coefficient $\kappa$ is a nonlinearity knob, where $\kappa=0$ recovers the linear case. Controllers will be the original methods' controllers where reproducible, and simple MPC or learned policies otherwise. Actuator saturation is kept, because saturation is where local gradients vanish.

Second, we will build the adaptive quadrature. It starts from the two endpoints, estimates or bounds $L_i$ per subinterval, and bisects the subinterval with the largest bound contribution until the tolerance is met. The certified tier will use interval bounds on the Jacobians of the analytic dynamics, composed through the horizon. We expect this bound to be loose for long horizons. We will report its tightness honestly and treat the estimate tier as the practical one.

Third, we will analyze the method in two ways. We will measure how required passes scale with $\kappa$ and with horizon $H$. We will also identify the trajectory events where refinement concentrates, such as saturation onset, swing-up through the unstable region, and contact in the reacher.

Baselines

The naive reference point is the single-gradient (Jacobian) attribution. This is $\Delta_i$ times the rollout gradient taken at the nominal trajectory, or at the realized trajectory. It is what linearization-based control already computes, and it costs one backward pass. It is also the strongest competitor where the dynamics are mildly nonlinear. It will be run on identical rollouts. Beating it means lower per-feature error against the reference wherever its error exceeds tolerance, and matching it in cost where it does not.

The accuracy ceiling is uniform Integrated Gradients with 1024 Gauss-Legendre points, which serves as ground truth. Uniform runs with $m=5$ (the setting used in Q-value Path Decomposition) and $m=50$ are the practical baselines.

The strongest existing adaptive method is the Integrated Decision Gradients sampling rule, ported to rollouts. It allocates steps by the size of $\partial J/\partial\alpha$ under an identical pass budget. Its only reported number is on images: AIC 0.701 versus 0.571 for standard Integrated Gradients on ResNet101 (Integrated Decision Gradients). There is no reported result found on Lacuna for attribution accuracy on any control benchmark, so all control comparisons will be newly measured. Beating it means lower maximum per-feature error at equal backward passes.

A black-box baseline is sampling-based Shapley value estimation with the same number of rollouts. It is included because derivative-free reasoning is how ABC-LMPC already treats these systems.

The key ablation keeps our bisection procedure but replaces the curvature criterion with uniform bisection. This isolates whether curvature-driven allocation, rather than adaptivity in general, delivers the savings.

All methods receive identical rollouts, identical references, and a matched budget of backward passes. Methods with tunable parameters get the same small grid search on held-out trajectories.

Why not just…?

Why not just use the Jacobian, as linearization-based control does? This is a serious threat. For smooth tasks such as vehicle lateral control near the nominal path, it may be essentially exact, and our method should then cost the same, because it starts from those endpoints. The case against it is saturation. When an actuator clips or a barrier penalty flattens, the local gradient drops toward zero for the very input that caused the violation. This is the Sensitivity failure identified in Axiomatic Attribution for Deep Networks. The $\kappa$ sweep and the pendulum swing-up will show where the Jacobian's error leaves tolerance. If it rarely does on these benchmarks, we will report that the extra machinery is unnecessary there.

Why not just run Integrated Gradients with a fixed small number of steps, as Q-value Path Decomposition does? A fixed $m$ cannot know where the curvature is. We expect errors with a heavy tail: most trajectories are fine at $m=5$, while a few, near switching, are badly wrong with no warning. We will report the full error distribution and the rate at which the bound flags these cases, not only mean error.

Why not just estimate attributions by sampling perturbations? Sampling needs no gradients, but the number of inputs is $H$ times the disturbance dimension, which reaches hundreds even for the point mass. Shapley sampling variance grows with that count, while one backward pass returns all components. The matched-budget comparison measures this directly.

Why not lift the system with a Koopman model and attribute linearly? Koopman methods learn a representation in which the dynamics act linearly, as in MamKO. In the lifted space, one-gradient attribution would be exact for the surrogate, but the surrogate's truncation error then becomes attribution error with no bound attached. Because this requires training a separate model per task, we will test it only on the pendulum and the cubic spring, comparing its attributions against the reference.

Proposed Experiments

The primary metric is the maximum and mean per-feature absolute error against the 1024-point reference, plotted against backward passes. Secondary metrics are wall-clock time per attribution, the completeness residual, and the empirical coverage of both the estimated and certified bounds.

As a task-level faithfulness check, we will reset the top-$k$ attributed disturbance components to nominal and re-run the rollout. We then record how much of the constraint violation disappears as $k$ grows. This is the control analogue of the deletion curve used on ImageNet, and it tests whether cheaper attributions change what an engineer would conclude.

As a portability check, we will apply the curvature-driven step rule to standard Integrated Gradients on ResNet101 over the 5,000-image ImageNet protocol of Integrated Decision Gradients. We will report AIC at matched passes against its 0.701 and Integrated Gradients' 0.571.

The decisive checks are three. First, required passes should be one at $\kappa=0$ and grow with $\kappa$. If cost does not track nonlinearity, the central claim fails. Second, on swing-up and saturation-heavy rollouts, the method should match the reference's deletion curve using fewer than a fifth of the passes of uniform $m=50$, while the Jacobian visibly diverges. Third, the certified bound must never be violated, and the estimate bound's violation rate must be reported. If curvature allocation does not beat uniform bisection at matched budget, we will drop that ingredient and report adaptivity alone as the contribution.