This example shows how to modify the uncertainty description to avoid discontinuities and get reliable estimates of the margin of uncertainty.

Computing the smallest perturbation that causes instability at a given frequency is the cornerstone of most robustness analysis algorithms in Robust Control Toolbox™. To estimate, for example, the robust stability margin over the entire frequency range, the function `robuststab`

performs this basic computation for a finite set of frequencies and looks at the worst case over this set.

Under most conditions, the robust stability margin is continuous with respect to frequency, so this approach gives good estimates provided you use a sufficiently dense frequency grid. However in some problems with only parameter uncertainty (UREAL), the migration of poles from stable to unstable can occur at isolated frequencies (generally unknown to you). As a result, any frequency grid that excludes these particular frequencies fail to detect worst-case perturbations and give over-optimistic stability margins.

In this example, we consider a spring-mass-damper system with 100% parameter uncertainty in the damping coefficient and 0% uncertainty in the spring coefficient. Note that all uncertainty here is of `ureal`

type:

m = 1; k = 1; c = ureal('c',1,'plusminus',1); sys = tf(1,[m c k]);

As the uncertain element `c`

varies, the only place where the poles can migrate from stable to unstable is at s = j*1 (1 rad/sec). No amount of variation in `c`

can cause them to migrate across the jw-axis at any other frequency.

When computing the robust stability margins with the `robuststab`

function, almost any frequency grid will exclude f=1, leading to the incorrect conclusion that the margin of uncertainty on c is infinite.

```
omega = logspace(-1,1,40); % one possible grid
sysg = ufrd(sys,omega);
[stabmarg,du,rep,info] = robuststab(sysg);
stabmarg
```

stabmarg = LowerBound: 5.0348e+03 UpperBound: Inf DestabilizingFrequency: 0.1000

When we look at the `mussv`

bounds computed by the `robuststab`

function, the structured singular value mu is zero at all test frequencies in `omega`

:

opt = bodeoptions; opt.MagUnits = 'Abs'; opt.PhaseVisible = 'off'; opt.YLim = [0 1]; bodeplot(info.MussvBnds(:,1),'b*',info.MussvBnds(:,2),'g*',opt) title('Original \mu bounds, 40 frequency points') legend('Upper bound','Lower bound')

**Figure 1:** Original mu bounds, 40 frequency points.

Note that making the grid denser would not help. Only by adding f=1 to the grid will we find the true margin.

f = 1; stabmarg = robuststab(ufrd(sys,f))

stabmarg = LowerBound: 1.0000 UpperBound: 1 DestabilizingFrequency: 1

The example above shows that the robust stability margin can be a discontinuous function of frequency. In other words, it can have jumps. We can eliminate such jumps by adding a small amount of uncertain dynamics to every uncertain real parameter. This amounts to adding some dynamics to pure gains. Importantly, as the size of the added dynamics goes to zero, the estimated margin for the modified problem converges to the true margin for the original problem.

In the spring-mass-damper example, we model `c`

as a `ureal`

with the range [0.05,1.95] rather than [0,2], and add a `ultidyn`

perturbation with gain bounded by 0.05. This combination covers the original uncertainty in `c`

and introduces only 5% conservatism.

cc = ureal('cReal',1,'plusminus',0.95) + ultidyn('cUlti',[1 1],'Bound',0.05); sysreg = usubs(sys,'c',cc);

Now let's recompute the robust stability margin:

[stabmarg,du,report,info] = robuststab(ufrd(sysreg,omega)); stabmarg report

stabmarg = LowerBound: 2.0016 UpperBound: 2.3630 DestabilizingFrequency: 1.0608 report = Assuming nominal UFRD system is stable... Uncertain system is robustly stable to modeled uncertainty. -- It can tolerate up to 200% of the modeled uncertainty. -- A destabilizing combination of 236% of the modeled uncertainty was found. -- This combination causes an instability at 1.06 rad/seconds. -- Sensitivity with respect to the uncertain elements are: 'cReal' is 35%. Increasing 'cReal' by 25% leads to a 9% decrease in the margin. 'cUlti' is 63%. Increasing 'cUlti' by 25% leads to a 16% decrease in the margin.

Now the calculation determines that the margin is not infinite. The value 2.36 is still greater than 1 (the true margin) because the density of frequency points is not high enough.

If we plot the mu bounds (reciprocal of robust stability margin) as a function of frequency, a peak clearly appears around 1 rad/s:

bodeplot(info.MussvBnds(:,1),info.MussvBnds(:,2),opt) title('Regularized \mu, 40 frequency points, 5% added dynamics'); legend('Upper bound','Lower bound')

**Figure 2:** Regularized mu, 40 frequency points, 5% added dynamics

The sharpness of the peak of the `mussv`

plot as well as its shape suggests that a denser frequency grid may be necessary. We'll increase the frequency-grid density by a factor of 5, and recompute:

OmegaDense = logspace(-1,1,200); [stabmarg,du,report,info] = robuststab(ufrd(sysreg,OmegaDense)); stabmarg report

stabmarg = LowerBound: 1.0026 UpperBound: 1.0056 DestabilizingFrequency: 0.9885 report = Assuming nominal UFRD system is stable... Uncertain system is robustly stable to modeled uncertainty. -- It can tolerate up to 100% of the modeled uncertainty. -- A destabilizing combination of 101% of the modeled uncertainty was found. -- This combination causes an instability at 0.988 rad/seconds. -- Sensitivity with respect to the uncertain elements are: 'cReal' is 95%. Increasing 'cReal' by 25% leads to a 24% decrease in the margin. 'cUlti' is 6%. Increasing 'cUlti' by 25% leads to a 2% decrease in the margin.

The computed robust stability margin has decreased significantly. It helped to use a denser grid, and the margin estimate is now close to the true value 1. The mu plot shows a pronounced peak at 1 rad/sec:

bodeplot(info.MussvBnds(:,1),info.MussvBnds(:,2),opt) title('Regularized \mu, 200 frequency points, 5% added dynamics'); legend('Upper bound','Lower bound')

**Figure 3:** Regularized mu, 200 frequency points, 5% added dynamics

For the modified uncertainty model, the robust stability margin is 1, which is equal to that of the original problem. In general, the stability margin of the modified problem is less than or equal to that of the original problem. If it is significantly less, then the answer to the question "What is the stability margin?" is very sensitive to the uncertainty model. In this case, we put more faith in the value that allows for a few percents of unmodeled dynamics. Either way, the stability margin for the modified problem is most trustworthy.

The command `complexify`

automates the procedure of replacing a `ureal`

with the sum of a `ureal`

and `ultidyn`

. The analysis above can be repeated using `complexify`

obtaining the same results.

sysreg = complexify(sys,0.05,'ultidyn'); [stabmargc,duc,reportc,infoc] = robuststab(ufrd(sysreg,OmegaDense)); stabmargc reportc bodeplot(info.MussvBnds(:,1),'b+',info.MussvBnds(:,2),'b+',... infoc.MussvBnds(:,1),'r',infoc.MussvBnds(:,2),'r',opt) title('Regularized \mu, 200 frequency points, 5% added dynamics'); legend('Upper bound','Lower bound','(complexify) Upper bound',... '(complexify) Lower bound','Location','NorthEast');

stabmargc = LowerBound: 1.0026 UpperBound: 1.0056 DestabilizingFrequency: 0.9885 reportc = Assuming nominal UFRD system is stable... Uncertain system is robustly stable to modeled uncertainty. -- It can tolerate up to 100% of the modeled uncertainty. -- A destabilizing combination of 101% of the modeled uncertainty was found. -- This combination causes an instability at 0.988 rad/seconds. -- Sensitivity with respect to the uncertain elements are: 'c' is 95%. Increasing 'c' by 25% leads to a 24% decrease in the margin. 'c_cmpxfy' is 6%. Increasing 'c_cmpxfy' by 25% leads to a 2% decrease in the margin.

**Figure 4:** Regularized mu, 200 frequency points, 5% added dynamics

The continuity of the robust stability margin, and the subsequent computational and interpretational difficulties raised by the presence of discontinuities are considered in [1]. The consequences and interpretations of the regularization illustrated in this small example are described in [2]. An extensive analysis of regularization for 2-parameter example is given in [2].

[1] Barmish, B.R., Khargonekar, P.P, Shi, Z.C., and R. Tempo, "Robustness margin need not be a continuous function of the problem data," Systems & Control Letters, Vol. 15, No. 2, 1990, pp. 91-98.

[2] Packard, A., and P. Pandey, "Continuity properties of the real/complex structured singular value," Vol. 38, No. 3, 1993, pp. 415-428.

Was this topic helpful?