R: Detailed iteration report of nleqslv

nleqslv-iterationreport {nleqslv}

R Documentation

Detailed iteration report of nleqslv

Description

The format of the iteration report provided by nleqslv when the trace component of the control argument has been set to 1.

Details

All iteration reports consist of a series of columns with a header summarising the contents. Common column headers are

Iter: Iteration counter
Jac: Jacobian type. The Jacobian type is indicated by N for a Newton Jacobian or B for a Broyden updated matrix; optionally followed by the letter s indicating a totally singular matrix or the letter i indicating an ill-conditioned matrix. Unless the Jacobian is singular, the Jacobian type is followed by an estimate of the inverse condition number of the Jacobian in parentheses as computed by Lapack. This column will be blank when backtracking is active.
Fnorm: square of the euclidean norm of function values / 2
Largest |f|: infinity norm of f(x) at the current point

Report for linesearch methods

A sample iteration report for the linesearch global methods (cline, qline and gline) is (some intercolumn space has been removed to make the table fit)

Iter        Jac Lambda        Ftarg         Fnorm   Largest |f|
   0                                 2.886812e+00  2.250000e+00
   1 N(9.6e-03) 1.0000 2.886235e+00  5.787362e+05  1.070841e+03
   1            0.1000 2.886754e+00  9.857947e+00  3.214799e+00
   1            0.0100 2.886806e+00  2.866321e+00  2.237878e+00
   2 B(2.2e-02) 1.0000 2.865748e+00  4.541965e+03  9.341610e+01
   2            0.1000 2.866264e+00  3.253536e+00  2.242344e+00
   2            0.0298 2.866304e+00  2.805872e+00  2.200544e+00
   3 B(5.5e-02) 1.0000 2.805311e+00  2.919089e+01  7.073082e+00
   3            0.1000 2.805816e+00  2.437606e+00  2.027297e+00
   4 B(1.0e-01) 1.0000 2.437118e+00  9.839802e-01  1.142529e+00

The column headed by Lambda shows the value of the line search parameter. The column headed by Ftarg follows from a sufficient decrease requirement and is the value below which Fnorm must drop if the current step is to be accepted.

The value of Lambda may not be lower than a threshold determined by the setting of control parameter xtol to avoid reporting false convergence. When no acceptable Lambda is possible and the Broyden method is being used, a new Jacobian will be computed.

Report for the pure method

The iteration report for the global method none is almost the same as the above report, except that the column labelled Ftarg is omitted. The column Lambda gives the ratio of the maximum stepsize and the length of the full Newton step. It is either exactly 1.0, indicating that the Newton step is smaller than the maximum stepsize and therefore used unmodified, or smaller than 1.0, indicating that the Newton step is larger than the maximum stepsize and therefore truncated.

Report for the double dogleg global method

A sample iteration report for the global method dbldog is (some intercolumn space has been removed to make the table fit)

Iter        Jac   Lambda    Eta   Dlt0   Dltn        Fnorm  Largest |f|
   0                                           2.886812e+00 2.250000e+00
   1 N(9.6e-03) C        0.9544 0.4671 0.9343* 1.699715e-01 5.421673e-01
   1            W 0.0833 0.9544 0.9343 0.4671  1.699715e-01 5.421673e-01
   2 B(1.1e-02) W 0.1154 0.4851 0.4671 0.4671  1.277667e-01 5.043571e-01
   3 B(7.3e-02) W 0.7879 0.7289 0.4671 0.0759  5.067893e-01 7.973542e-01
   3            C        0.7289 0.0759 0.1519  5.440250e-02 2.726084e-01
   4 B(8.3e-02) W 0.5307 0.3271 0.1519 0.3037  3.576547e-02 2.657553e-01
   5 B(1.8e-01) N        0.6674 0.2191 0.4383  6.566182e-03 8.555110e-02
   6 B(1.8e-01) N        0.9801 0.0376 0.0752  4.921645e-04 3.094104e-02
   7 B(1.9e-01) N        0.7981 0.0157 0.0313  4.960629e-06 2.826064e-03
   8 B(1.6e-01) N        0.3942 0.0029 0.0058  1.545503e-08 1.757498e-04
   9 B(1.5e-01) N        0.6536 0.0001 0.0003  2.968676e-11 5.983765e-06
  10 B(1.5e-01) N        0.4730 0.0000 0.0000  4.741792e-14 2.198380e-07
  11 B(1.5e-01) N        0.9787 0.0000 0.0000  6.451792e-19 8.118586e-10

After the column for the Jacobian the letters indicate the following

C: a fraction (\le1.0) of the Cauchy or steepest descent step is taken where the fraction is the ratio of the trust region radius and the Cauchy steplength.
W: a convex combination of the Cauchy and eta*(Newton step) is taken. The number in the column headed by Lambda is the weight of the partial Newton step.
P: a fraction (\ge1.0) of the partial Newton step, equal to eta*(Newton step), is taken where the fraction is the ratio of the trust region radius and the partial Newton steplength.
N: a normal full Newton step is taken.

The number in the column headed by Eta is calculated from an upper limit on the ratio of the length of the steepest descent direction and the length of the Newton step. See Dennis and Schnabel (1996) pp.139ff for the details. The column headed by Dlt0 gives the trust region size at the start of the current iteration. The column headed by Dltn gives the trust region size when the current step has been accepted by the dogleg algorithm.

The trust region size is decreased when the actual reduction of the function value norm does not agree well with the predicted reduction from the linear approximation of the function or does not exhibit sufficient decrease. And increased when the actual and predicted reduction are sufficiently close. The trust region size is not allowed to decrease beyond a threshold determined by the setting of control parameter xtol; when that happens the backtracking is regarded as a failure and is terminated. In that case a new Jacobian will be calculated if the Broyden method is being used.

The current trust region step is continued with a doubled trust region size when the actual and predicted reduction agree extremely well. This is indicated by * immediately following the value in the column headed by Dltn. This could save gradient calculations. However, a trial step is never taken if the current step is a Newton step. If the trial step does not succeed then the previous trust region size is restored by halving the trial size. The exception is when a trial step takes a Newton step. In that case the trust region size is immediately set to the size of the Newton step which implies that a halving of the new size leads to a smaller size for the trust region than before.

Normally the initial trust region radius is the same as the final trust region radius of the previous iteration but the trust region size is restricted by the size of the current Newton step. So when full Newton steps are being taken, the trust region radius at the start of an iteration may be less than the final value of the previous iteration. The double dogleg method and the trust region updating procedure are fully explained in sections 6.4.2 and 6.4.3 of Dennis and Schnabel (1996).

Report for the single (Powell) dogleg global method

A sample iteration report for the global method pwldog is (some intercolumn space has been removed to make the table fit)

Iter        Jac   Lambda   Dlt0   Dltn         Fnorm  Largest |f|
   0                                    2.886812e+00 2.250000e+00
   1 N(9.6e-03) C        0.4671 0.9343* 1.699715e-01 5.421673e-01
   1            W 0.0794 0.9343 0.4671  1.699715e-01 5.421673e-01
   2 B(1.1e-02) W 0.0559 0.4671 0.4671  1.205661e-01 4.890487e-01
   3 B(7.3e-02) W 0.5662 0.4671 0.0960  4.119560e-01 7.254441e-01
   3            W 0.0237 0.0960 0.1921  4.426507e-02 2.139252e-01
   4 B(8.8e-02) W 0.2306 0.1921 0.3842* 2.303135e-02 2.143943e-01
   4            W 0.4769 0.3842 0.1921  2.303135e-02 2.143943e-01
   5 B(1.9e-01) N        0.1375 0.2750  8.014508e-04 3.681498e-02
   6 B(1.7e-01) N        0.0162 0.0325  1.355741e-05 5.084627e-03
   7 B(1.3e-01) N        0.0070 0.0035  1.282776e-05 4.920962e-03
   8 B(1.8e-01) N        0.0028 0.0056  3.678140e-08 2.643592e-04
   9 B(1.9e-01) N        0.0001 0.0003  1.689182e-12 1.747622e-06
  10 B(1.9e-01) N        0.0000 0.0000  9.568768e-16 4.288618e-08
  11 B(1.9e-01) N        0.0000 0.0000  1.051357e-18 1.422036e-09

This is much simpler than the double dogleg report, since the single dogleg takes either a steepest descent step, a convex combination of the steepest descent and Newton steps or a full Newton step. The number in the column Lambda is the weight of the Newton step. The single dogleg method is a special case of the double dogleg method with eta equal to 1. It uses the same method of updating the trust region size as the double dogleg method.

Report for the hook step global method

A sample iteration report for the global method hook is (some intercolumn space has been removed to make the table fit)

Iter         Jac       mu  dnorm    Dlt0   Dltn         Fnorm   Largest |f|
   0                                             2.886812e+00  2.250000e+00
   1  N(9.6e-03) H 0.1968 0.4909  0.4671 0.9343* 1.806293e-01  5.749418e-01
   1             H 0.0366 0.9381  0.9343 0.4671  1.806293e-01  5.749418e-01
   2  B(2.5e-02) H 0.1101 0.4745  0.4671 0.2336  1.797759e-01  5.635028e-01
   3  B(1.4e-01) H 0.0264 0.2341  0.2336 0.4671  3.768809e-02  2.063234e-01
   4  B(1.6e-01) N        0.0819  0.0819 0.1637  3.002274e-03  7.736213e-02
   5  B(1.8e-01) N        0.0513  0.0513 0.1025  5.355533e-05  1.018879e-02
   6  B(1.5e-01) N        0.0090  0.0090 0.0179  1.357039e-06  1.224357e-03
   7  B(1.5e-01) N        0.0004  0.0004 0.0008  1.846111e-09  6.070166e-05
   8  B(1.4e-01) N        0.0000  0.0000 0.0001  3.292896e-12  2.555851e-06
   9  B(1.5e-01) N        0.0000  0.0000 0.0000  7.281583e-18  3.800552e-09

The column headed by mu shows the Levenberg-Marquardt parameter when the Newton step is larger than the trust region radius. The column headed by dnorm gives the Euclidean norm of the step (adjustment of the current x) taken by the algorithm. The absolute value of the difference with Dlt0 is less than 0.1 times the trust region radius.