User Guide

Worst-case analysis
315
Performing worst-case analysis with tight tolerances on
the parameters produces sensitivity and worst-case
results (in the output file). You can use these to decide
how the parameters should be varied to achieve the
desired response. You can then make adjustments to the
nominal values in the circuit file, and perform the
worst-case analysis again for a new set of gradients.
Monte Carlo analysis
Monte Carlo (.MC) analysis may be helpful when
worst-case analysis cannot be used. Monte Carlo analysis
can often be used to verify or improve on worst-case
analysis results. Monte Carlo analysis randomly selects
possible parameter values, which can be thought of as
randomly selecting points in the parameter space. The
worst-case analysis assumes that the worst results occur
somewhere on the surface of this space, where parameters
(to which the output is sensitive) are at one of their
extreme values.
If this is not true, the Monte Carlo analysis may find a
point at which the results are worse. To try this, replace
.WC in the circuit file with .MC <#runs>, where <#runs> is
the number of simulations you want to perform. More
runs provide higher confidence results. The Monte Carlo
summary in the output file lists the runs in decreasing
order of collating function value.
Next, add the following option to the .MC statement, and
simulate again.
OUTPUT LIST RUNS <worst_run#>
This performs only two simulations: the nominal and the
worst Monte Carlo run. The parameter values used
during the worst run are written to the output file, and the
results of both simulations are saved.
Using Monte Carlo analysis with YMAX is a good way to
obtain a conservative guess at the maximum possible
deviation from nominal, since worst-case analysis usually
cannot provide this information.
Parametric sweeps (.STEP),
l
i
k
e t
h
e one
performed in the circuit file shown in
Figure 85, can be used to augment this
procedure.
To save
d
is
k
space,
d
o not speci
f
y any
OUTPUT options.
Pspug.book Page 315 Wednesday, November 11, 1998 1:14 PM