Abstract
Sensitivity analysis for hyperbolic conservation laws is traditionally complicated by the presence of shocks, where solutions are non-differentiable and variations manifest as Dirac delta functions. This paper introduces a streamlined numerical method for computing the first-order variation of such systems using the Generalized Tangent Vector (GTV) framework. By reformulating the GTV equations into a conservative form, the proposed method allows standard numerical schemes to automatically handle complex interface conditions at discontinuities without specialized source terms. The approach is validated on the inviscid Burgers’ equation and a $2 \times 2$ hyperbolic system, demonstrating first-order accuracy and robust stability.
Introduction
In engineering and physics, sensitivity analysis—the study of how a system's solution changes in response to perturbations in initial conditions—is a fundamental tool for optimization and uncertainty quantification. For smooth systems, this is typically handled via linearized equations. However, many physical processes, such as supersonic airflow or traffic dynamics, are governed by hyperbolic conservation laws that naturally develop shocks.
At these shocks, standard calculus fails because the solution is discontinuous. Shifting a shock position slightly creates a difference between the original and perturbed solutions that resembles a massive spike (a Dirac delta function), which traditional numerical solvers cannot resolve. To address this, the concept of a Generalized Tangent Vector (GTV) was developed to decompose the variation into two manageable parts: a smooth function change ($v$) and a discrete shift in the shock's position ($\xi$). While theoretically sound, implementing this numerically has historically required complex tracking of shock interfaces and the addition of sensitive delta-type source terms.
Methodology
Conservative Reformulation
The primary contribution of this work is the conservative reformulation of the GTV equations. Building on the theoretical framework established by Bressan and Marson (1995), the authors present Theorem 3.1, which proves that the complex interface conditions governing variations at a shock can be rewritten in a "conservative form." This form is mathematically analogous to the Rankine–Hugoniot conditions used to solve the original physical state.
The "aha" moment of this research is the realization that the linearized Rankine-Hugoniot condition is equivalent to the continuity of a specific variation flux. This allows the numerical solver to treat the shock interface naturally, as the internal logic of finite volume methods handles these jump conditions "for free."
The Numerical Algorithm
The proposed method (Algorithm 4.1) follows a three-step process:
- State Evolution: Solve the original hyperbolic system (the "state") using a standard conservative scheme, such as the Lax-Friedrichs method.
- Shock Tracking: Identify the shock position and compute the shock shift ($\xi$) using the newly simplified expressions derived from the conservative reformulation. To mitigate the effects of numerical viscosity (which smears shocks over several grid cells), the authors use a parameter $K$ to sample states away from the immediate smeared zone.
- Tangent Vector Computation: Solve for the tangent vector ($w$) using another conservative numerical scheme. Because of the reformulation, the interface conditions are automatically preserved by the scheme's inherent properties.
Finally, the actual first-order variation is reconstructed by combining the smooth component and the calculated shock shift.
Experimental Results
The method was evaluated using standard mathematical benchmarks for hyperbolic systems, specifically the inviscid Burgers’ equation and a nonlinear $2 \times 2$ hyperbolic system (similar to the p-system in fluid dynamics).
Accuracy and Stability
The numerical results demonstrate that the method achieves first-order accuracy in the $L^1$ norm. Specifically, the difference between the exact perturbed solution and the solution predicted by the first-order variation decreases at a rate of $O(\epsilon^2)$, where $\epsilon$ represents the magnitude of the perturbation.
Crucially, the method proved stable across both single-shock and multiple-shock scenarios. Unlike naive sensitivity computations, this approach did not produce non-physical oscillations or spikes at the shock locations. In the $2 \times 2$ system test, the numerical results for the shock shift ($\xi$) matched analytical solutions with high precision, showing a relative error of approximately 0.05%.
Discussion and Limitations
While the method simplifies the numerical treatment of sensitivities in discontinuous systems, several challenges remain:
- Shock Interactions: The current framework assumes shocks do not interact. Handling "topology changes," such as when two shocks merge or collide, remains a significant theoretical and numerical hurdle.
- Higher-Order Extensions: This study utilized first-order Lax-Friedrichs schemes. While the authors suggest the method can be extended to high-order schemes (like WENO or Discontinuous Galerkin) to further reduce numerical viscosity, these have not yet been implemented.
- Dimensionality: The current work is strictly one-dimensional. Extending GTV concepts to 2D or 3D, where shock fronts are curves or surfaces rather than points, represents a major area for future research.
Conclusion
By transforming the interface conditions of generalized tangent vectors into a conservative form, this paper provides a robust and simplified path for sensitivity analysis in hyperbolic systems. The ability to use standard numerical schemes to automatically handle shock variations marks a significant step forward in making sensitivity analysis practical for complex, shock-prone physical models.
Summary updated