Describing Function Analysis for Nonlinear Systems

While linear control theory provides powerful tools like Bode and Nyquist plots, real-world engineering systems are inherently nonlinear. Actuators saturate, gears have backlash, and relays exhibit hysteresis. Describing Function Analysis bridges this gap, offering an approximate but invaluable frequency-domain method to predict the behavior of systems containing a single, isolated nonlinearity. It allows you to extend the intuition of linear frequency response to predict phenomena unique to nonlinear systems, most notably the existence and characteristics of limit cycles—self-sustained oscillations.

The Challenge of Isolated Nonlinearities

A key insight of this method is its targeted application. It is designed for systems where a single nonlinear element (like a saturation block, a relay, or a dead-zone) is connected in a feedback structure with otherwise linear dynamics. The core challenge is that superposition does not hold for nonlinear elements; you cannot simply multiply their input by a transfer function. However, if we assume the input to the nonlinearity is a sinusoidal signal, we can analyze its output in the frequency domain. The goal is to replace the nonlinear block with a quasi-linear representation: an equivalent gain that depends not on frequency, but on the amplitude of the input sinusoid. This equivalent gain is the describing function.

Defining and Computing the Describing Function

The Describing Function, $N(A)$, is formally defined as the complex ratio of the fundamental harmonic component of the nonlinear element's output to its sinusoidal input. For a sinusoidal input $x(t)=A\sin (\omega t)$, the output $y(t)$ will be periodic but non-sinusoidal. We can represent $y(t)$ by its Fourier series. The describing function considers only the first harmonic (the component at frequency $\omega$), ignoring higher harmonics.

Mathematically, it is calculated as: $$N(A)=\frac{Y_1}{A}{e}^{j{\varphi }_1}$$ where $Y_1$ is the amplitude of the fundamental output component and ${\varphi }_1$ is its phase shift relative to the input. In many common nonlinearities with odd symmetry (symmetric about the origin), the phase shift ${\varphi }_1$ is zero, and $N(A)$ is a real number.

Example: Ideal Saturation Nonlinearity Consider an ideal saturation element with linear slope $k$ and saturation limits $\pm \delta$. For input amplitudes $A\le \delta$, the output is linear: $y(t)=kA\sin (\omega t)$. Thus, $N(A)=k$.

For $A>\delta$, the output waveform is a clipped sine wave. Computing the fundamental component via Fourier analysis yields: $$N(A)=\frac{2k}{\pi }\left[{\sin }^{-1}\left(\frac{\delta }{A}\right)+\frac{\delta }{A}\sqrt{1-{\left(\frac{\delta }{A}\right)}^2}\right]$$ for $A>\delta$. Crucially, $N(A)$ is now a function of amplitude $A$, starting at $k$ when $A=\delta$ and decreasing toward zero as $A$ increases.

Predicting Limit Cycles: The Nyquist Criterion Extension

The primary application of describing function analysis is the prediction of limit cycles. In a standard negative feedback loop with linear plant $G(s)$ and nonlinearity $N(A)$, a limit cycle is predicted to exist if the following harmonic balance equation is satisfied: $$G(j\omega )N(A)=-1$$ This can be rearranged as: $$G(j\omega )=-\frac{1}{N(A)}$$

This formulation leads to a powerful graphical technique. You plot two things on the same complex plane:

1. The standard Nyquist plot of the linear plant, $G(j\omega )$, which varies with frequency $\omega$.
2. The negative inverse describing function plot, $-1/N(A)$, which varies with amplitude $A$.

An intersection between these two curves indicates a solution $(\omega ,A)$ that satisfies the harmonic balance equation. The frequency at the intersection point (read from the $G(j\omega )$ curve) predicts the limit cycle frequency. The amplitude at the intersection point (read from the $-1/N(A)$ locus) predicts the oscillation amplitude at the input to the nonlinearity.

Assumptions and Validity of the Method

The accuracy of this approximate method rests on several critical assumptions:

1. The nonlinearity is time-invariant and separated. It must be a single, isolated block.
2. The linear part, $G(j\omega )$, acts as a low-pass filter. This is essential for the "filtering hypothesis." The linear dynamics must attenuate the higher harmonics generated by the nonlinearity sufficiently so that the signal fed back to the nonlinear input is predominantly sinusoidal. If $G(j\omega )$ does not have good low-pass characteristics, the results may be unreliable.
3. The system may be accurately modeled by a single sinusoid. The method predicts symmetric, periodic oscillations.

The stability of a predicted limit cycle can also be inferred from the plots. If, as amplitude $A$ increases slightly, the $-1/N(A)$ locus moves away from the encirclement region of the $G(j\omega )$ curve, the limit cycle is typically stable. If it moves further into the encirclement, the limit cycle is predicted to be unstable.

Common Pitfalls

Ignoring the Filtering Hypothesis: The most common error is applying describing function analysis to a system where the linear plant $G(j\omega )$ does not provide sufficient attenuation of higher harmonics. For example, if $G(j\omega )$ has a resonance near a harmonic frequency, the analysis will likely give inaccurate predictions. Always check the frequency response of $G(s)$ to confirm it acts as a low-pass filter relative to the limit cycle frequency.

Misinterpreting Multiple Intersections: A single $-1/N(A)$ curve can intersect the $G(j\omega )$ plot at more than one point. Each intersection represents a potential limit cycle. However, you must analyze the stability of each predicted point. Often, only one will be a stable limit cycle observable in practice, while others are unstable operating points the system will not sustain.

Applying the Method to Non-Isolated or Complex Nonlinearities: Describing function analysis becomes cumbersome and less accurate for multiple, interconnected nonlinearities or for nonlinearities that are not memoryless (like hysteresis, which has memory). The standard method is rigorously defined only for the classic single, isolated, memoryless nonlinearity structure. Attempting to force more complex setups into this framework often leads to incorrect conclusions.

Summary

• Describing Function Analysis is an approximate frequency-domain method that extends linear techniques to systems with a single, isolated nonlinearity by representing it with an amplitude-dependent equivalent gain, $N(A)$.
• The primary use is predicting limit cycles. This is done graphically by finding the intersection of the linear plant's Nyquist plot, $G(j\omega )$, and the negative inverse describing function locus, $-1/N(A)$. The intersection coordinates predict the oscillation frequency and amplitude.
• The method's validity hinges on the filtering hypothesis: the linear part of the system must sufficiently attenuate the higher harmonics generated by the nonlinearity so that the feedback signal remains approximately sinusoidal.
• It is a powerful engineering tool for initial analysis and insight but requires careful consideration of its assumptions. Misapplication to systems without proper low-pass filtering or with complex nonlinear interactions is a frequent source of error.