Multi-Step Perturbation Solution of Nonlinear Rational Expectations Models

This paper develops and illustrates the multi-step generalization of the standard single-step perturbation (SSP) method or MSP. In SSP, we can think of evaluating at x the computed approximate solution based on x0, as moving from x0 to x in "one big step" along the straight-line vector x-x0. By contrast, in MSP we move from x0 to x along any chosen path, continuous, curved-line or connected-straight-line, in h steps of equal length 1/h. If at each step we apply SSP, Taylor-series theory says that the approximation error per step is 0(e) = h^(-k-1), so that the total approximation error in moving from x0 to x in h steps is 0(e) = h^(-k). Thus, MSP has two major advantages over SSP. First, both SSP and MSP accuracy declines as the approximation point, x, moves from the initial point, x0, although only in MSP can the decline be countered by increasing h. Increasing k is much more costly than increasing h, because increasing k requires new derivations of derivatives, more computer programming, more computer storage, and more computer run time. By contrast, increasing h generally requires only more computer run time and often only slightly more. Second, in SSP the initial point is usually a nonstochastic steady state but can sometimes also be set up in function space as the known exact solution of a close but simpler model. This "closeness" of a related, simpler, and known solution can be exploited much more explicitly by MSP, when moving from x0 to x. In MSP, the state space could include parameters, so that the initial point, x0, would represent the simpler model with the known solution, and the final point, x, would continue to represent the model of interest. Then, as we would move from the initial x0 to the final x in h steps, the state variables and parameters would move together from their initial to final values and the model being solved would vary continuously from the simple model to the model of interest. Both advantages of MSP facilitate repeatedly, accurately, and quickly solving a NLRE model in an econometric analysis, over a range of data values, which could differ enough from nonstochastic steady states of the model of interest to render computed SSP solutions, for a given k, inadequately accurate. In the present paper, we extend the derivation of SSP to MSP for k = 4. As we did before, we use a mixture of gradient and differential-form differentiations to derive the MSP computational equations in conventional linear-algebraic form and illustrate them with a version of the stochastic optimal one-sector growth model.