2
$\begingroup$

I’m analysing the discrete-time feedback system

$$ T(z) = \frac{1}{1 – H(z)}, \\ H(z) = \sum_{k=1}^{N} h[k] z^{-k}, h[k] \in \mathbb{R}, N ≥ 1 $$

Thus, $H(z)$ is a strictly causal FIR with real taps. So the closed-loop poles are the roots of $$ 1 – H(z) = 0. $$

What is already clear

  • Small-gain condition: If $\|H(e^{j \omega})\| < 1$ for every $ω$, the Neumann (geometric) series converges and the system is BIBO-stable.

The question

If all taps are real (so the Nyquist plot is symmetric about the real axis):

Does the weaker condition $\mathrm{Re}\{H(e^{j\omega})\} < 1$ for every $ω$ already guarantee closed-loop stability?

Or can someone construct a counterexample: a real-coefficient FIR $H(z)$ that satisfies $\mathrm{Re}\{H(e^{j\omega})\} < 1$ everywhere yet places a pole on or outside the unit circle?

And a follow-up is: if $H(z)$ is a matrix of real FIRs, then the stability criterion would be applied to the eigenvalues $\lambda_i(z)$ of $H(z)$. Can $\mathrm{Re}\{ \lambda_i(z) \} < 1$, while the system is still unstable?

$\endgroup$

2 Answers 2

3
$\begingroup$

Since you are defining the closed loop transfer function as

$$ T(z) = \frac{1}{1 – H(z)}, $$

implies that you are dealing with positive unit feedback. For many results about closed loop control negative unit feedback is used instead. However, to convert between the two, one can just negate $H(z)$.

For the stability analysis of negative unit feedback one could use the Nyquist stability criterion. All poles of $H(z)$ lie at the origin, so are all stable and therefore this criterion in this case comes down to $-H(e^{j\omega})$ not encircling the point $-1+0j$ in the complex plane. However, if it is ensured that $\Re(H(e^{j\omega}))<1$ and thus $\Re(-H(e^{j\omega}))>-1$ means that no encirclements of $-1+0j$ can occur and thus that closed loop $T(z)$ is stable.

$\endgroup$
4
  • $\begingroup$ Thanks! Would that argument still be true if H(z) is a stable IIR filter? $\endgroup$ Commented Jun 18 at 3:20
  • 1
    $\begingroup$ @Jiro Yes, as long as all poles of the IIR filter lie inside the unit disk (and thus the open loop is stable) and the same constraint holds on the real part. $\endgroup$ Commented Jun 18 at 6:59
  • $\begingroup$ Doesn't the Nyquist criterion operate on the characteristic equation $1-H(z)$ and not just $H(z)$? $\endgroup$ Commented Jun 18 at 23:21
  • $\begingroup$ @RandyYates Checking the encirclements of -1 for $-H(e^{j\omega})$ is equivalent to checking the encirclements of 0 for $1-H(e^{j\omega})$. $\endgroup$ Commented Jun 19 at 7:38
2
$\begingroup$

It's interesting to consider this problem from a complex analysis point of view. Of course, Nyquist's stability criterion as used in fibonatic's answer relies on results from complex analysis.

Let's first define the winding number. Let $f(z)$ be a complex function that is analytic in the interior of a closed contour $C$ except for a finite number of poles (i.e., $f(z)$ is meromorphic). We assume that $f(z)$ is analytic and non-zero on $C$. The winding number (with respect to the origin) is the number of times $w=f(z)$ winds around the origin of the $w$-plane as $z$ moves along the contour $C$ once. The winding number is given by the integer value

$$\frac{1}{2\pi}\Delta_C\arg f(z)\tag{1}$$

where $\Delta_C\arg f(z)$ denotes the difference between the final and the initial arguments of $f(z)$ as $z$ describes the contour $C$ one time.

The argument principle states that the winding number equals the number of zeros $Z$ minus the number of poles $P$ of $f(z)$ inside the contour $C$ (counting multiplicities):

$$\frac{1}{2\pi}\Delta_C\arg f(z)=Z-P\tag{2}$$

I adopt the more common notation $1\color{red}+H(z)$ for the denominator of the closed-loop transfer function. Since we're interested in stability, we choose the unit circle $|z|=1$ as our contour $C$.

If we assume that $H(z)$ is (strictly) causal and stable, it must have all its $N$ poles inside the unit circle, where $N$ is the filter order. From the argument principle $(2)$ we get

$$\frac{1}{2\pi}\Delta_C\arg \big\{1+H(z)\big\}=Z-P=Z-N\tag{4}$$

where $Z$ and $P$ are the numbers of zeros and poles inside the unit circle, respectively. Since the total number of zeros equals the filter order $N$, the absolute value of $(4)$ equals the number of zeros outside the unit circle.

For stability of the closed-loop transfer functions we require all zeros of $1+H(z)$ to be inside the unit circle. Hence, we require

$$\frac{1}{2\pi}\Delta_C\arg \big\{1+H(z)\big\}=0\tag{5}$$

as $z$ moves along the unit circle. The requirement that the argument of $1+H(z)$ returns to its original value after $z$ has moved around the unit circle once is equivalent to requiring that the trace of $1+H(z)$ in the complex plane does not encircle the origin, because each encirclement of the origin would mean an increase (or decrease) of the argument by $2\pi$.

The trace of $1+H(z)$ will not encircle the origin if $|H(z)|<1$ for $|z|=1$. This is a sufficient condition which we could also have obtained by applying Rouché's theorem.

A less restrictive, but still sufficient condition ensuring that $1+H(z)$ does not encircle the origin is $\textrm{Re}\big\{1+H(z)\big\}>0$ for $|z|=1$, i.e.,

$$\textrm{Re}\big\{H(e^{j\omega})\big\}>-1,\quad\omega\in[0,2\pi)\tag{6}$$

Note that $(6)$ is not a necessary but a sufficient condition. A necessary and sufficient condition would be

$$H(e^{j\omega})>-1,\qquad \textrm{for }\omega\in[0,2\pi)\textrm{ such that }H(e^{j\omega})\in\mathbb{R}\tag{7}$$

because requiring the trace of $1+H(z)$ to only cross the positive real axis but never the negative real axis guarantees that there are no encirclements of the origin.

Examples:

I'll give examples of FIR transfer functions $H(z)$ all of which lead to stable closed-loop systems. All transfer functions satisfy at least one of the three stability criteria discussed above:

  1. $\big|H(e^{j\omega})\big|<1$
  2. $\textrm{Re}\big\{H(e^{j\omega})\big\}>-1$
  3. $H(e^{j\omega})>-1$ for all $\omega$ such that $H(e^{j\omega})\in\mathbb{R}$

Criterion 1 is the strictest, and criterion 3 is the least strict. Criteria 1 and 2 are sufficient but not necessary whereas criterion 3 is necessary and sufficient.

The first filter is given by

$$H_1(z)=0.4z^{-1}-0.4z^{-2}-0.3z^{-3}+0.2z^{-4}$$

It satisfies $|H_1(e^{j\omega})|<1$, and, consequently, it also satisfies the other two criteria. All zeros of $1+H_1(z)$ are inside or on a circle with radius $0.84$.

The second filter is

$$H_2(z)=0.5z^{-1}-0.4z^{-2}-0.5z^{-3}+0.2z^{-4}$$

It violates criterion 1 but it satisfies criterion 2 (and 3). Its maximum magnitude on the unit circle is $1.21$, and all zeros of $1+H_2(z)$ are inside or on a circle with radius $0.92$.

The last filter is

$$H_3(z)=0.5z^{-1}-0.4z^{-2}-0.6z^{-3}+0.3z^{-4}$$

It violates criteria 1 and 2 but it satisfies criterion 3. Its maximum magnitude on the unit circle is $1.37$, and its minimum real part on the unit circle is $-1.08$. All zeros of $1+H_3(z)$ are inside or on a circle with radius $0.99$.

The figure below shows the complex traces of the three corresponding denominator polynomials $1+H_k(z)$, $k=1,2,3$. enter image description here

Clearly, none of them encircles the origin, as required for stability. It can be seen that the imposed stability conditions become less strict from left to right. The real part of $1+H_3(z)$ even becomes negative, but the trace crosses the real line at positive values, as required by condition 3.

$\endgroup$

Start asking to get answers

Find the answer to your question by asking.

Ask question

Explore related questions

See similar questions with these tags.