next up previous contents
Next: Vali: Validation of Up: The major tools Previous: Combinations

PoCo: Polynomial Continuation

PoCo supports two different types of homotopies. The user can give either an artificial-parameter homotopy, by explicitly giving a start system, or a natural-parameter homotopy, in the form of a polynomial system with one unknown more than the number of equations. In both cases a list of start solutions must be provided.

The tracing of the solution paths is organized in two stages. During the first stage we have t < 1. No singularities are likely to occur in this stage. Moreover, we are not interested in the exact solutions for t<1. So we can afford ourselves to track the paths more loosely. The second stage is at the end of each path, when . Here we may have to deal with paths converging to singular solutions and with paths running to infinity, as . In this stage, we have to track the paths more carefully.

The solution paths are not likely to turn back as the continuation parameter t increases. Therefore, instead of pseudo arc-length, only increment-and-fix predictor corrector methods have been implemented. This means that after each increase of t, t remains fixed while correcting the solution . There are four different predictor methods available. They are listed in Table 2, with a short description.

  
Table 2: predictors for and .

The complex predictor for t is based on an idea of Drexler, see [6], who proposed to take t in complex space when singularities during continuation occurred.
The differences between secant and tangent predictor are illustrated in Figure 7.

  
Figure 7: The secant and tangent predictor with as step length.

In Figure 7, we see that the tangent predictor is able to follow the solution path more accurately. This justifies the extra work, such as computing derivatives and solving a linear system.

We distinguish four different sets of parameters:

  1. Monitoring the path tracker. We can follow a number of paths with the same values of the continuation path, opposed to the more sequentially oriented path following. This number is the first parameter to be set. There are different choices of the predictor for and for t. The number of steps along a path is bounded. The distance from the target to the beginning of the end game can be adjusted. Finally, the automatic mechanism of re-computing clustered paths is controlled by the maximum number of re-runs.

  2. Step-size control. The parameters for the predictor are bounds on the minimum and maximum step size. Because the step size is changed during continuation, reduction and expansion factors determine the variation of the step size respectively in case of failing or succeeding convergence of the corrector. A threshold value on the number of consecutive successful steps controls the frequency of enlarging the step size.

  3. Path closeness. The Newton corrector attains convergence if either the desired precision for the residual is reached or if the correction step becomes small enough. A distinction is made between relative and absolute precision. The maximum number of allowed corrector steps can be used either to enforce quadratic convergence or to loosen this demand.

  4. Solution tolerances. To decide whether a solution has to be considered as singular, at infinity or as clustered we provide bounds respectively on the inverse of the condition number, the magnitude of the norm of the solutions and the closest distance of two separate solutions.

Initially, default variables are provided and shown to the user for adjustment. To facilitate tuning, there is a central parameter that models the condition of the homotopy. Increasing this parameter results in a modification of the other parameters, in the sense that paths will be tracked more accurately. Practical experiences have shown that when we force the path tracker to stay very close to the path, we risk of getting stuck at a singularity.

Distinguishing path failures either due to singular solutions or to solutions at infinity has been a hard task until recently. Taking Bernshtein's second theorem into account, by means of an inexpensive side-calculation the directions of diverging paths can be approximated up to a sufficient level of accuracy. The path directions of diverging paths yield the outer normals to the faces of the Newton polytopes, which cause the deficiency. This feature of PoCo provides a certificate of deficiency from diverging solution paths.

There are several output levels available to put the data processed during continuation on a log file. The user can choose to see only the solutions at the end of the path or to have all intermediate solutions on file. Furthermore, information about the performance of predictor and corrector can be delivered at all intermediate steps. The output file of PoCo could in this way become an input file to routines to visualize the solution paths.



next up previous contents
Next: Vali: Validation of Up: The major tools Previous: Combinations



Jan Verschelde
Thu Nov 21 10:50:01 MET 1996