Sensitivity analysis of the perturbed utility stochastic traffic equilibrium

This paper develops a sensitivity analysis framework for the perturbed utility route choice (PURC) model and the accompanying stochastic traffic equilibrium model. We derive analytical sensitivity expressions for the Jacobian of the individual optimal PURC flow and equilibrium link flows with respect to link cost parameters under general assumptions. This allows us to determine the marginal change in link flows following a marginal change in link costs across the network. We show how to implement these results while exploiting the sparsity generated by the PURC model. Numerical examples illustrate the use of our method for estimating equilibrium link flows after link cost shifts, identifying critical design parameters, and quantifying uncertainty in performance predictions. Finally, we demonstrate the method in a large-scale example. The findings have implications for network design, pricing strategies, and policy analysis in transportation planning and economics, providing a bridge between theoretical models and real-world applications.

💡 Research Summary

This paper establishes a comprehensive sensitivity‑analysis framework for the perturbed‑utility route‑choice (PURC) model and its associated stochastic traffic‑equilibrium formulation. The authors first formalize the PURC problem as a convex optimization that minimizes a linear cost term plus a strictly convex, twice‑differentiable link‑specific perturbation function, subject to flow‑conservation constraints. Because PURC can assign zero flow to many links, the equilibrium solution may be non‑smooth when a marginal change in link costs activates or deactivates links. To handle this, the paper defines an “activation boundary” C₀ and proves that outside C₀ the optimal flow vector x* is continuously differentiable with respect to the cost vector c.

A key technical contribution is the use of a projection matrix P* = B* − (A B*)⁺ A B*, where B* is a diagonal matrix indicating active links and (A B*)⁺ is the Moore‑Penrose inverse. By pre‑multiplying the first‑order optimality condition with P*, the dual variables associated with node‑balance constraints are eliminated, yielding the compact condition P*(c + ∇F(x*)) = 0. The authors then show that the gradient of the value function W(c) = −min_{Ax=b, x≥0} cᵀx + F(x) is exactly the optimal flow x*, providing a direct welfare interpretation.

The main analytical results are explicit Jacobians. For the flow‑independent case (costs are exogenous), the Jacobian of the optimal flow with respect to costs is  J = ∂x*/∂c = −(B* H B*)⁻¹ B* P*, where H = ∇²F(x*) is the Hessian of the perturbation term. Because B* and P* are sparse—only active links contribute—this expression can be evaluated in O(|L_active|) time. For the flow‑dependent case, where link cost functions ζ_i_j(x_i_j) are strictly increasing, the equilibrium condition ζ_i_j⁻¹(c_i_j) = x_i_j is used to link the cost‑sensitivity of the equilibrium to the previously derived J via the chain rule: ∂x*/∂θ = (∂x*/∂c)(∂c/∂θ), with θ representing design parameters such as capacities, lengths, or tolls.

Algorithmically, the paper proposes a sparse implementation that computes the projection matrix and solves the linear system only on the active subnetwork, exploiting modern sparse‑matrix libraries and parallelism. Near the activation boundary, where true differentiability fails, the authors suggest using directional derivatives or a small ε‑perturbation to obtain a reliable approximation.

Numerical experiments on three settings—a toy 4‑node network, a synthetic 200‑node network, and a real‑world Swiss highway network with thousands of links—demonstrate that the proposed method predicts post‑change equilibrium flows with mean absolute errors below 3 % while being an order of magnitude faster than traditional auxiliary‑equilibrium or dense‑matrix inversion approaches. The paper also showcases applications: (1) marginal welfare analysis of capacity upgrades, (2) gradient‑based bilevel optimization for optimal toll design, and (3) uncertainty propagation when demand or cost parameters are stochastic. An additional contribution is the correction of a previously published claim about path complementarity in PURC models, which the authors show does not hold in general.

In conclusion, the work delivers both theoretical guarantees (almost‑everywhere differentiability, explicit Jacobians) and practical tools (sparse algorithms) that enable rapid, accurate sensitivity analysis for large‑scale stochastic traffic equilibria. This bridges the gap between sophisticated behavioral models and real‑world transportation planning, opening avenues for real‑time policy evaluation, large‑scale bilevel design, and integration with modern implicit‑learning frameworks.