Calculating Value Function Gradients Across Capital Trajectories
Value function gradients across capital trajectories compute the marginal sensitivity of terminal discounted returns to resource shifts at each state coordinate.
adjoint state vector
Meaning
Mathematical representations of the sensitivity of a cost function to changes in the state variables along a dynamic path define the derivative of a future objective with respect to current state variations. The adjoint state vector provides the necessary coefficients to compute these sensitivities without re-evaluating the system equations for each parameter change. By representing marginal values of state variables at each point in time, the adjoint state vector enables efficient optimization over complex multi-period trajectories.
Optimization Multiplier
Pontryagin’s maximum principle utilizes auxiliary variables to link state dynamics with optimization goals. The adjoint state vector represents these auxiliary variables, acting as a dynamic Lagrange multiplier that tracks the shadow price of each state over time. As the system moves, these multipliers evolve according to a differential equation driven by the state equations and the objective function.
Such evolution allows the optimizer to assess the downstream impact of localized decisions.
Gradient Feedback
Numerical optimization of control paths requires the calculation of derivatives across thousands of steps. Instead of using finite differences for every variable, the adjoint state vector resolves the entire gradient in a single backward pass. The process begins at the terminal boundary and runs in reverse, accumulating sensitivities sequentially.
Such efficiency allows the optimizer to scale to larger parameter dimensions.
Trajectory Divergence
Discrepancies in the backward sweep can cause the sensitivity values to diverge if the underlying system exhibits chaotic behavior. The adjoint state vector remains valid only within a localized neighborhood of the reference trajectory. In highly nonlinear regimes, the linear approximations used in the backward pass degrade quickly, requiring frequent updates to the reference state path.