An Approach for the Qualitative Graphical Representation of the Describing Function in Nonlinear Systems Stability Analysis Davide Tebaldi, Roberto Zanasi Abstract The describing function method is a useful tool for the qualitative analysis of limit cycles in the stability analysis of nonlinear systems. This method is inherently approximate; therefore, it should be used for a fast qualitative analysis of the considered systems. However, plotting the exact describing function requires heavy mathematical calculations, reducing interest in this method especially from the point of view of control education. The objective of this paper is to enhance the describing function method by providing a new approach for the qualitative plotting of the describing function for piecewise nonlinearities involving discontinuities. Unlike the standard method, the proposed approach allows for a straightforward, hand-drawn plotting of the describing function using the rules introduced in this paper, simply by analyzing the shape of the nonlinearity. The proposed case studies show that the limit cycles estimation performed using the standard exact plotting of the describing function yields the same qualitative results as those obtained using the proposed qualitative method for plotting the describing function. I INTRODUCTION Persistent oscillatory behaviors, typically referred to as limit cycles, are an inherent property of a large variety of systems of different nature. Limit cycles are indeed present in engineering systems belonging to different energetic domains, including mechanical systems [ 1 ] such as robotic systems, electrical systems [ 2 ] such as electrical oscillators, and hydraulic systems [ 3 ] such as hydraulic actuators. One of the most effective tools to study the existence and the characteristics of limit cycles is the describing function method [ 4 ] . With reference to the system of Fig. 1 , limit cycles can exist under certain assumptions if the so-called self-sustaining equation F(X)\,G(j\omega)=-1 exhibits solutions, where G(j\omega) is the frequency response function of the linear system G(s) , and F(X) is the describing function of the considered nonlinearity. Despite being non-rigorous, the describing function method finds several applications, including photovoltaic systems [ 5 ] , disturbance rejection in nonlinear active disturbance rejection control [ 6 ] , and the identification of the structural nonlinearities in multi-degree-of-freedom nonlinear systems [ 7 ] . 10,1)(33,8) -3,-2)(3,2) Fig. 1: Considered nonlinear autonomous feedback system. Through the years, many different modifications or additions to the describing function method have been made. In [ 8 ] , a stability criterion is provided, giving an interesting justification of the quasi-static stability criterion upon which the describing function method is based. The describing function method is studied in [ 9 ] for limit cycles analysis in switched circuits, while a refinement of the approximated solution given by this method is proposed in [ 10 ] and an extension to systems exhibiting more than one nonlinearity in given in [ 11 ] . Thanks to computer aided control system design software, interactive tools have been developed for limit cycles analysis using the describing function method [ 12 ] . However, the use of graphical tools such as the fast hand-drawing of the describing function can be an important ability for control engineers. Indeed, despite being an effective method, the main limitation associated with the describing function, especially from the point of view of control education, is the derivation of the analytical expression of the describing function, which often requires a detailed analysis of the considered nonlinearity and is derived using heavy mathematical calculations [ 13 ] . Considering that the describing function method is non-rigorous, since it is based on an approximated analysis [ 14 ] , the complexity of calculation diminishes the easiness and interest of application of this method. The objective of this paper is therefore to enhance the standard describing function method by providing an approach for the approximate hand-drawing of the describing function itself, with no need to perform complex calculations, making the method more appealing from an educational point of view. The proposed method applies to piecewise nonlinearities involving discontinuities, and consists in a series of rules for drawing the qualitative describing function by hand very quickly, by simply looking at the nonlinear characteristic. Furthermore, the proposed method is also translated into an algorithm involving simpler calculations than the standard complex analytical expression of the describing function F(X) . The considered case studies show that the analysis and limit cycles estimation carried out using the qualitative describing function plotting give, qualitatively, the same results as those obtained using the standard describing function plotting. The remainder of this paper is structured as follows. The describing function of two standard nonlinearities are recalled in Sec. II , while an algorithm for plotting the describing function of piecewise nonlinearities with discontinuities is given in Sec. III . The new approach for the qualitative plotting of the describing function is proposed in Sec. IV , and applied to the limit cycle estimation of two case studies in Sec. V . The conclusions are given in Sec. VI . II Describing function Let x(t)\!=\!X\,\sin(\omega t) be the sinusoidal input of the nonlinear function y(x) shown in Fig. 1 . The corresponding output signal y(t) is a periodic signal that can be expanded in a Fourier series as follows: y(t)=\sum_{n=1}^{\infty}Y_{n}\,\sin\,(n\omega t+\varphi_{n}). (1) The constant term Y_{0} is missing in ( 1 ) due to Assumption 1 . Assumption 1 Reference is made to the nonlinear feedback structure shown in Fig. 1 , where the nonlinear function y(x) is purely algebraic, symmetric with respect to the origin, i.e. y(-x)\!=\!-y(x) , and independent of the frequency \omega of the input signal x(t) . The use of the describing function is based on the following approximation: all the higher order harmonics (i.e. for n\geq 2 ) in the Fourier series ( 1 ) are neglected, while the first harmonic y(t)\simeq Y_{1}(X)\ \sin\,\left(\omega t+\varphi_{1}(X)\right) is considered. The describing function F(X) of the nonlinear function y(x) links the sinusoidal input x(t)\!=\!X\,\sin(\omega t) to the first harmonic Y_{1}(X)\ \sin\,\left(\omega t+\varphi_{1}(X)\right) of the output signal y(t) [ 4 ] and, in the general case, is defined as follows: F(X)=\frac{Y_{1}(X)}{X}\,\,e^{j\varphi_{1}(X)}, (2) where \textstyle Y_{1}(X)=\sqrt{a_{1}^{2}(X)+b_{1}^{2}(X)} , \textstyle\varphi_{1}(X)=\arctan\,\frac{a_{1}(X)}{b_{1}(X)} , \begin{array}[]{l}\textstyle a_{1}(X)=\frac{1}{\pi}\!\int_{-\pi}^{\pi}\!y(X\sin\theta)\,\cos\theta\,d\theta,\\[5.69054pt] \textstyle b_{1}(X)=\frac{1}{\pi}\!\int_{-\pi}^{\pi}\!y(X\sin\theta)\,\sin\theta\,d\theta,\end{array} (3) and \theta=\omega t . Due to the symmetry hypothesis in Assumption 1 , the coefficient a_{1}(X) in ( 3 ) is zero, and the describing function F(X) simplifies as follows [ 4 ] : F(X)=\frac{b_{1}(X)}{X}=\frac{4}{\pi\,X}\!\int_{0}^{\frac{\pi}{2}}\!y(X\sin\theta)\,\sin\theta\,d\theta. (4) II- 1 Dead Zone The describing function F_{d}\!\left(\!\textstyle m,X_{1},X\!\right) of the dead zone y_{d}(x)=y_{d}\!\left(\!\textstyle m_{1},\!\textstyle X_{1},x\!\right) shown in Fig. 2 (a) is the following [ 13 ] : F_{d}(X)\!=\!F_{d}\!\left(\!\textstyle m,X_{1},X\!\right)\!=\!\left\{\begin{array}[]{@{}lcl@{}}0&\mbox{if}&X\!<\!X_{1},\\ m\,\Phi\!\left(\!\frac{X}{X_{1}}\right)&\mbox{if}&X\!\geq\!X_{1},\end{array}\right. (5) where: \textstyle\Phi\!\left(\!\frac{X}{X_{1}}\!\right)=1-\frac{2}{\pi}\left(\arcsin\frac{X_{1}}{X}\!+\!\frac{X_{1}}{X}\sqrt{1\!-\!\left(\frac{X_{1}}{X}\right)^{2}}\right). (6) The shape of function F_{d}(X) in ( 5 ) is shown in Fig. 2 (b). The derivative of function F_{d}(X) with respect to X is the following: \frac{F_{d}(X)}{dX}=\frac{4mX_{1}(X^{2}-X_{1}^{2})}{\pi X^{3}\sqrt{X^{2}-X_{1}^{2}}}, from which it follows that the slope of function F_{d}(X) is zero for X=X_{1} and is positive for X>X_{1} . This means that function F_{d}(X) always increases for X>X_{1} and reaches its maximum value m when X\rightarrow\infty . -6,-2)(14,9.5) -6,-2)(14,9.5) Fig. 2: (a) Dead zone function y_{d}(x) ; (b) Describing function F_{d}(X) . -6,-2)(14,9.5) -6,-2)(14,9.5) Fig. 3: (a) Relay with threshold y_{r}(x) ; (b) Describing function F_{r}(X) . II- 2 Relay with threshold The describing function F_{r}\!\left(\!\textstyle Y_{1},X_{1},X\!\right) of the relay with threshold y_{r}(x)\!=\!y_{r}(Y_{1},X_{1},x) shown in Fig. 3 (a) is [ 13 ] : F_{r}(X)\!=\!F_{r}\!\left(\!\textstyle Y_{1},X_{1},X\!\right)\!=\!\left\{\begin{array}[]{@{}l@{\;}c@{\;}l@{}}0&\mbox{if}&X\!<\!X_{1},\\ Y_{1}\Psi\!\left(\!X_{1},X\!\right)&\mbox{if}&X\!\geq\!X_{1},\end{array}\right. (7) where: \textstyle\Psi\!\left(\!X_{1},X\!\right)=\frac{4}{\pi X}\sqrt{1\!-\!\left(\frac{X_{1}}{X}\right)^{2}}. (8) The shape of function F_{r}(X) is shown in Fig. 3 (b). The derivative of function F_{r}(X) with respect to X is: \frac{F_{r}(X)}{dX}=\frac{4Y_{1}(2X_{1}^{2}-X^{2})}{\pi X^{3}\sqrt{X^{2}-X_{1}^{2}}}, from which it follows that the slope of function F_{r}(X) is infinite for X=X_{1} . The maximum F_{r}^{*} of function F_{r}(X) occurs at X=X_{r}^{*}=\sqrt{2}X_{1} and is given by F_{r}^{*}=\frac{2Y_{1}}{\pi X_{1}} . -2,-1.5)(21,9.5) -2,-4.55)(21,7.5) -2,-0.1)(21,0.85) Fig. 4: (a) Piecewise nonlinear function y(x) ; (b) Components of y(x) in ( 9 ); (c) Describing function F(X) in ( 10 ) computed using Algorithm 1 . III Describing function F(X) of piecewise nonlinear functions y(x) The describing functions of several classes of nonlinearities have been determined [ 13 ] . In this paper, we focus on the describing function F(X) of a generic piecewise nonlinear function y(x) . Property 1 The describing function F(X) of a piecewise nonlinear function y(x) with discontinuities can always be obtained as a linear combination of a certain number of describing functions F_{d}(X) and F_{r}(X) of the two nonlinear functions y_{d}(x) and y_{r}(x) shown in Fig. 2 and Fig. 3 . An an example, reference is made to the case study of Fig. 4 (a). In this case, the function y(x) can be obtained as the sum of six functions of type y_{d}(x) and y_{r}(x) , as in Fig. 4 (b): \begin{array}[]{r@{\,}c@{\,}l}y(x)&=&y_{d}^{1}\!\left(\!m_{1},\!X_{1},\!x\!\right)\!+\!y_{d}^{2}\!\left(\!m_{21},\!X_{2},\!x\!\right)\!+\!y_{r}^{2}\!\left(\!Y_{2},\!X_{2},x\!\right)\\[2.84526pt] &&\!\!\!\!+y_{d}^{3}\!\left(\!m_{32},\!X_{3},\!x\!\right)\!+\!y_{r}^{3}\!\left(\!Y_{3},\!X_{3},\!x\!\right)\!+\!y_{d}^{4}\!\left(\!m_{43},\!X_{4},\!x\!\right),\end{array} (9) where m_{21}\!=\!m_{2}\!-\!m_{1} , m_{32}\!=\!m_{3}\!-\!m_{2}\!=\!-m_{2} and m_{43}\!=\!m_{4}\!-\!m_{3}=m_{4} . From ( 9 ), it follows that the corresponding describing function F(X) , shown in Fig. 4 (c), is the sum of six describing functions of type F_{d}(X) and F_{r}(X) : \begin{array}[]{@{\!\!\!\!}r@{}c@{}l@{}}F(X)&=&\textstyle F_{d}\!\left(\!\textstyle m_{1},\!X_{1},\!X\!\right)\!+\!F_{d}\!\left(\!\textstyle m_{21},\!X_{2},\!X\!\right)\!+\!F_{r}\!\left(\!\textstyle Y_{1},\!X_{2},\!X\!\right)\\[2.84526pt] &&\!\!\!\!\!\!\!+F_{d}\!\left(\!\textstyle m_{32},\!X_{3},\!X\!\right)\!+\!F_{r}\!\left(\!\textstyle Y_{3},\!X_{3},\!X\!\right)\!+\!F_{d}\!\left(\!\textstyle m_{43},\!X_{4},\!X\!\right).\end{array}\! (10) This decomposition can be done for any type of generic piecewise nonlinear function y(x) with discontinuities. An algorithm for computing the describing function F(X) of a piecewise nonlinear function y(x) with discontinuities is reported in Algorithm 1 . Let p_{i}\!=\!(x_{i},y_{i}) , for i\in\{1,\,2,\,\ldots,\,r\} , denote the points where function y(x) changes its slope or has a discontinuity. The algorithm takes as input vectors {\bf x}=[x_{1},\,x_{2},\,\ldots,\,x_{r}] and {\bf y}=[y_{1},\,y_{2},\,\ldots,\,y_{r}] , defining the position of points p_{i} , and vector {\bf X}=(0:\delta_{X}:X_{m}) , containing the desired values of amplitude X of the input sinusoidal signal. The algorithm produces as output a vector {\bf F}({\bf X}) containing the samples of the resulting describing function F(X) in correspondence of the points in the input vector {\bf X} . The Matlab code of Algorithm 1 is freely available in the repository [ 15 ] . The points p_{i} corresponding to vectors {\bf x}=[2,5,5,9,9,13,19] and {\bf y}=[0,4,2,4,6,6,8] used for the example of Fig. 4 are highlighted in red in Fig. 4 (a), while the describing function F(X) given by Algorithm 1 is shown in Fig. 4 (c). Algorithm 1 Describing Function {\bf F}({\bf X}) 0: 0: 1: Initialize {\bf F}({\bf X})=(0,0,\ldots,0)\in\mathbb{R}^{n} , where n=\mbox{dim}({\bf X}) ; 2: Initialize i=2 ; 3: while i<\mbox{dim}({\bf x}) do 4: Find set {\cal S}=(k,k+1,\ldots,n) s.t. X_{k}>{\bf x}_{i} ; 5: if x_{i}=x_{i+1} then 6: Compute discontinuity Y_{k}=y_{i+1}-y_{i} as in Fig. 3 (a); 7: Compute \Psi(x_{i},{\bf X}({\cal S})) as in ( 8 ); 8: Compute {\bf F}({\bf X})={\bf F}({\bf X})+Y_{k}\Psi(x_{i},{\bf X}({\cal S})) as in ( 7 ); 9: Remove x_{i} from {\bf x} ; Remove y_{i} from {\bf y} ; 10: end if 11: Compute the two slopes m_{i} and m_{i+1} of the two segments [(x_{i-1},y_{i-1}),(x_{i},y_{i})] and [(x_{i},y_{i}),(x_{i+1},y_{i+1})] ; 12: Compute \Phi\!\left(\!\frac{{\bf X}({\cal S})}{x_{i}}\!\right) as in ( 6 ); 13: Compute {\bf F}({\bf X})={\bf F}({\bf X})+(m_{i+1}-m_{i})\Phi\!\left(\!\frac{{\bf X}({\cal S})}{x_{i}}\!\right) as in ( 5 ); 14: Increment index i=i+1 ; 15: end while 16: return {\bf F}({\bf X}) IV Describing function F(X) : qualitative graphical representation Despite the describing function method being non-rigorous, the exact calculation of the describing function F(X) usually requires complex analytical calculations. This numerical complexity greatly limits the use of this method, especially from the point of view of control education, as the compromise between complexity and precision of the results is not advantageous. On the contrary, if an approximated drawing of the describing function plot could be easily done by hand, it would enhance the interest in this method for an initial fast qualitative analysis of limit cycles. In this section, an intuitive and fast approach for the qualitative hand-drawn graphical representation of function F(X) is therefore proposed. Property 2 Qualitative graphical behavior \tilde{F}(X) of function F(X) . Let y(x) be a nonlinear function with discontinuities; let X_{j} , for j\in\{1,2,\ldots,r\} , be the values of x where function y(x) has a discontinuity or changes its slope from m_{j-1} to m_{j} ; let vectors \overline{{\bf X}} and {\bf m} be defined as follows \overline{{\bf X}}=[0,X_{1},\ldots,X_{r}] and {\bf m}=[m_{0},m_{1},\ldots,m_{r}] ; let \overline{{\bf X}}_{d}\in\overline{{\bf X}} be a subset of vector \overline{{\bf X}} containing all the values X_{j} where function y(x) has a discontinuity; and let \overline{{\bf Y}}_{d} be the vector of all the discontinuity amplitudes Y_{j} that occur at X=X_{j}\in\overline{{\bf X}}_{d} . The qualitative hand-drawn graphical behavior \tilde{F}(X) of function F(X) can be determined using the following rules: 1) Function \tilde{F}(X) is always continuous for X>0 . 2) If function y(x) is discontinuous for x=0 , then the initial value of function \tilde{F}(X) is infinite: \tilde{F}(X)|_{X=0}=\infty . 3) For X\leq X_{1} , the value of function \tilde{F}(X) is equal to the slope m_{0} of the first segment of function y(x) . 4) For X\rightarrow\infty , the value of function \tilde{F}(X) is equal to the slope m_{r} of the last linear segment of function y(x) . 5) Let \tilde{F}_{j}(X) denote the qualitative graphical behavior of function F(X) within the range X\!\in\![X_{j},\,X_{j+1}] where function y(x) has slope m_{j} . The sequence of qualitative functions \tilde{F}_{j}(X) , for j\in\{0,1,\ldots,r\} , can be plotted using Algorithm 2 . For the hand-drawing approach, note that the following rules apply: a) \tilde{\Phi}\!\left(\!\frac{X}{X_{j}}\!\right) is an exponential-like function, having a shape similar to that of function F_{d}(X) shown in Fig. 2 (b): it is equal to zero until X=X_{j} , where it has null slope, and then tends to one when X\rightarrow\infty ; b) \tilde{\Psi}\!\left(\!X,X_{j}\!\right) is an impulsive function, having a shape similar to that of function F_{r}(X) shown in Fig. 3 (b): it is equal to zero until X=X_{j} , where it has infinite slope, reaches its maximum \frac{2}{\pi X_{j}} for X=\sqrt{2}X_{j} , and then tends to zero when X\rightarrow\infty . The value of the first function \tilde{F}_{0}(X) , for X\!\in\![0,\,X_{1}] , is equal to the first slope m_{0} of function y(x) . The qualitative function \tilde{F}_{j}(X) , for j\in\{1,2,\ldots,r\} , starts from the value \tilde{F}_{j-1}(X_{j}) and then tends to the value m_{j} for X\rightarrow\infty . If a discontinuity occurs at X\!=\!X_{j} , then an additional term Y_{j}\tilde{\Psi}\!\left(\!X,X_{j}\!\right) must be added to function \tilde{F}_{j}(X) . The obtained function \tilde{F}_{j}(X) must be considered only within the range X\!\in\![X_{j},\;X_{j+1}] . The value \tilde{F}_{j}(X_{j+1}) is then used as starting point for plotting the qualitative function \tilde{F}_{j+1}(X) in the next range X\!\in\![X_{j+1},X_{j+2}] . \hfill\Box -2,-1)(26,8) -2,-0.1)(26,1.15) Fig. 5: (a) Piecewise nonlinear function y(x) . (b) Describing function F(X) of function y(x) , corresponding qualitative graphical representation \tilde{F}(X) , and exponential-like function \tilde{F}_{j}(X) for j\in\{0,\,1,\,2,\,3\} . Qualitative proofs of the aforementioned rules: 1) Function \tilde{F}(X) is always continuous because the describing functions F_{d}(X) and F_{r}(X) , composing function F(X) , are continuous, see Prop. 1 , Fig. 2 and Fig. 3 . 2) If y(x) is discontinuous for x=0 , then \tilde{F}(X)|_{X=0}=\infty because, when X\rightarrow 0 , the amplitude Y_{1}(X) in ( 2 ) of the first harmonic of signal y(t) is finite: Y_{1}(X)=\frac{4Y_{1}}{\pi} . 3) For X\leq X_{1} , function y(x) is linear, the output signal y(t)=m_{0}\,x(t) is proportional to x(t) , and \tilde{F}(X)=m_{0} because the ratio Y_{1}(X)/X in ( 2 ) is equal to the slope m_{0} . 4) When X\rightarrow\infty , the slope m_{r} of the last segment of function y(x) becomes increasingly important for the shape of the output signal y(t) : when X\rightarrow\infty , y(t)\simeq m_{r}x(t) and, therefore, \tilde{F}(X)|_{X\rightarrow\infty}=m_{r} . 5) Reference is made to the nonlinear function y(x) shown in Fig. 5 (a), defined by vectors {\bf x}=[2,7,20,20,25] and {\bf y}=[0,4.5,7.21,4.21,5.25] . The corresponding describing function F(X) (red solid line) and qualitative graphical representation \tilde{F}(X) (green solid line) are shown in Fig. 5 (b). Algorithm 2 Qualitative Function \tilde{{\bf F}}({\bf X}) 0: 0: 1: Find set {\cal S}_{0}=(0,1,\ldots,n) s.t. {\bf X}({\cal S}_{0}))\leq X_{1} ; 2: Compute \tilde{F}_{0}({\bf X}({\cal S}_{0}))=m_{0} ; 3: Compute \tilde{{\bf F}}({\bf X}({\cal S}_{0}))=\tilde{F}_{0}({\bf X}({\cal S}_{0})) ; 4: for j\leftarrow 1 to \mbox{dim}(\overline{{\bf X}}) do 5: Find set {\cal S}=(k,k+1,\ldots,k+n) s.t. X_{j}<{\bf X}({\cal S}))\leq X_{j+1} ; 6: Compute \tilde{F}_{j0}=\tilde{F}_{j-1}(X_{j}) ; 7: Compute \tilde{F}_{j}({\bf X}({\cal S}))=\tilde{F}_{j0}\!+\!(m_{j}\!-\!\tilde{F}_{j0})\tilde{\Phi}\!\left(\!\frac{{\bf X}({\cal S})}{X_{j}}\!\right) ; 8: if X_{j}\in\overline{{\bf X}}_{d} then 9: Compute \tilde{F}_{j}({\bf X}({\cal S}))=\tilde{F}_{j}({\bf X}({\cal S}))+Y_{j}\tilde{\Psi}\!\left(\!{\bf X}({\cal S}),X_{j}\!\right) ; 10: end if 11: Compute \tilde{{\bf F}}({\bf X}({\cal S}))=\tilde{F}_{j}({\bf X}({\cal S})) ; 12: end for 13: return \tilde{{\bf F}}({\bf X}) The steps for plotting the qualitative function \tilde{F}(X) are: a) For X\leq X_{1} , the qualitative function is \tilde{F}_{0}(X)=0 , because the slope of the first segment of function y(x) is zero. b) Within the range X\in[X_{1},\;X_{2}] , a correct qualitative description of function F(X) is described by the exponential-like function \tilde{F}_{1}(X) starting from point (X_{1},0) and reaching the value m_{1} for X\rightarrow\infty , see the blue dashed line in Fig. 5 (b). c) Within the range X\in[X_{2},\;X_{3}] , function F(X) is qualitatively described by function \tilde{F}_{2}(X) , starting from point (X_{2},\tilde{F}_{1}(X_{2})) and reaching value m_{2}=m_{3} for X\!\rightarrow\!\infty . d) For X\!>\!X_{3} , function F(X) is qualitatively described by function \tilde{F}_{3}(X) , starting from point (X_{3},\tilde{F}_{2}(X_{3})) and reaching value m_{2}\!=\!m_{3} for X\!\rightarrow\!\infty . Remark 1 When applying the describing function method, it is an interesting and useful feature to firstly determine whether stable/unstable persistent oscillations - known as limit cycles - are present in the system in a fast way, before delving into the calculations to determine their amplitude and frequency. While the standard describing function method does not allow for such initial fast analysis, the new proposed approach for plotting the so-called qualitative describing function does. In fact, while the qualitative function \tilde{F}(X) in Fig. 5 (b) was plotted using Algorithm 2 with \tilde{\Phi}\!\left(\!\frac{X}{X_{j}}\right)=1-\frac{X}{X_{j}}\! and \tilde{\Psi}\!\left(\!X,X_{j}\!\right)=\Psi\!\left(\!X,X_{j}\!\right) , a qualitatively similar plot can also be effectively drawn by hand by directly applying steps from 1) to 5) in Sec. IV . Furthermore, the qualitative describing function \tilde{F}(X) captures the behavior of the actual describing function F(X) with no need to perform the complex calculations in ( 5 ), ( 8 ) and ( 10 ). Therefore, a simple hand-drawn sketch of the qualitative describing function \tilde{F}(X) is sufficient to identify stable/unstable limit cycles in the system with sufficient detail, as discussed in the next section. V Describing function: limit cycles estimation It is known that an estimation of the limit cycles present in the nonlinear feedback system of Fig. 1 can be performed by using the following complex self-sustaining equation: F(X)\,G(j\omega)=-1 , which equation is is generally solved graphically by plotting both the functions G(j\omega) and -1/F(X) on the complex plane [ 4 ] . The intersection points P_{i}(\omega_{i},X_{i}) between G(j\omega) and -1/F(X) provide a good estimation of the frequencies \omega_{i} and the amplitudes X_{i} of the limit cycles. Due to Assumption 1 , all the nonlinear functions y(x) considered in this work are characterized by a positive real describing function F(X) , and therefore all the intersection points P_{i}(\omega_{i},X_{i}) are on the negative real semiaxis. Furthermore, it is well known that [ 4 ] :   1) an intersection point P_{i}(\omega_{i},X_{i}) corresponds to a stable limit cycle if, for increasing X , the point -1/F(X) exits from the Nyquist diagram of function G(j\omega) , and to an unstable limit cycle in the opposite case; 2) the frequency and amplitude \omega_{i} and X_{i} of the limit cycle correspond to the frequency where function G(j\omega_{i}) intersects the real negative semiaxis and to the value X_{i} such that F(X_{i})=\bar{K} , respectively, where \bar{K} is the gain margin of the transfer function G(s) . Fig. 6: Case study of Sec. V-A . (a) Qualitative and actual describing functions F(X) and \tilde{F}(X) of the nonlinear function y(x) in Fig. 5 (a). (b) Graphical solution of equation F(X)\,G(j\omega)=-1 on the Nyquist plane. The limit cycle generated by signal y(t) in the state space of system G(s) can be approximately determined as follows. Let {\cal S}=({\bf A},\,{\bf B},\,{\bf C},\,{\bf D}) be a state space realization of the transfer function G(s) . The Laplace transform {\bf X}(s) of state space vector {\bf x}(t) of {\cal S} can be expressed as follows: {\bf X}(s)={\bf H}(s)\,Y(s),\hskip 17.07164pt{\bf H}(s)\!=\!(s\,{\bf I}-{\bf A})^{\mbox{\tiny-1}}{\bf B}, (11) where Y(s) is the Laplace transform of the input y(t) . If only the first harmonic y(t)\!=\!y_{1}(t)\!=\!Y_{1}\sin(\omega t\!+\!\varphi_{1}) is considered, the steady-state vector \tilde{{\bf x}}(t) at steady-state is: \tilde{{\bf x}}(t)\!=\!\lim_{t\rightarrow\infty}{\bf x}(t)={\bf A}(\omega)\sin(\omega t+\boldsymbol{\varphi}(\omega)). (12) where {\bf A}(\omega)\!=\!Y_{1}\left|{\bf H}(j\omega)\right| and \boldsymbol{\varphi}(\omega)\!=\!\arg({\bf H}(j\omega))+\varphi_{1} . Property 3 The steady-state vector \tilde{{\bf x}}(t) in ( 12 ) belongs to the 2-dimensional subspace {\cal X}_{\omega}=\mbox{span}\left(\tilde{{\bf x}}(0),\,\tilde{{\bf x}}(\frac{\pi}{2\omega})\right) , where \tilde{{\bf x}}(0)=\tilde{{\bf x}}(t)|_{t=0} and \tilde{{\bf x}}(\frac{\pi}{2\omega})=\tilde{{\bf x}}(t)|_{t=\frac{\pi}{2\omega}} . Proof. From ( 12 ), vector \tilde{{\bf x}}(t) can be expressed as follows: \tilde{{\bf x}}(t)=\sin(\omega t)\underbrace{{\bf A}(\omega)\sin\boldsymbol{\varphi}(\omega)}_{\tilde{{\bf x}}(0)}+\cos(\omega t)\underbrace{{\bf A}(\omega)\cos\boldsymbol{\varphi}(\omega)}_{\tilde{{\bf x}}(\frac{\pi}{2\omega})} that is, for t>0 , vector \tilde{{\bf x}}(t) can be expressed as a linear combination of the two vectors \tilde{{\bf x}}(0) and \tilde{{\bf x}}(\frac{\pi}{2\omega}) . \hfill\Box Fig. 7: Case study of Sec. V-A . (a) Trajectories in case (b) of Fig. 6 (b); stable limit cycle C_{S} (red closed line) defined by point P_{b2}=(\bar{\omega},X_{2}^{b}) , and unstable limit cycle C_{U} (magenta dashed closed line) defined by point P_{b1}=(\bar{\omega},X_{1}^{b}) . (b) Trajectories in case (c) of Fig. 6 (b); unstable limit cycle C_{U} (magenta dashed closed line) defined by point P_{c}=(\bar{\omega},X_{1}^{c}) . V-A First Case Study Reference is made to the feedback system of Fig. 1 , where G(s)=\frac{k\,(2-s)}{s(s+1)} and y(x) is the nonlinear function defined in Fig. 5 (a). The Nyquist diagrams of function G(s) for k=1 , k=2.5 and k=6 are shown in Fig. 6 (b), see labels (a) , (b) and (c) , respectively. The intersection points of function G(s) with the negative real semiaxis are denoted by \textstyle P_{a}\!=\!-\frac{1}{\bar{K}^{a}} , P_{b}\!=\!-\frac{1}{\bar{K}^{b}} and P_{c}\!=\!-\frac{1}{\bar{K}^{c}} , and are characterized by the same frequency \bar{\omega}=\sqrt{2} and by the gain margins \bar{K}^{a}=1 , \bar{K}^{b}=0.4 and \bar{K}^{c}=0.166 . The actual and qualitative describing functions F(X) and \tilde{F}(X) are shown in Fig. 6 (a). The points P_{2} and P_{M} in Fig. 6 (b) are defined as follows: \textstyle P_{2}\!=\!-\frac{1}{m_{2}} and \textstyle P_{M}\!=\!-\frac{1}{\bar{K}_{M}} . The points P_{a} , P_{c} and P_{c} correspond to the following three working conditions: a) \bar{K}^{a}>\bar{K}_{M} : function -\frac{1}{F(X)} is outside the Nyquist diagram of function G(j\omega) , therefore the origin of the feedback system is globally asymptotically stable. b) \bar{K}_{M}\!>\!\bar{K}^{b}\!>\!m_{2} : function -\frac{1}{F(X)} intersects G(j\omega) in two points: P_{b1}\!=\!(\bar{\omega},X_{1}^{b}) and P_{b2}\!=\!(\bar{\omega},X_{2}^{b}) . The state space trajectories in the vicinity of the origin are shown in Fig. 7 (a), where x(t)=-x_{2}(t) is the output of system G(s) . The simulated amplitudes X_{1}^{b} and X_{2}^{b} of Fig. 7 (a) are quite similar to the amplitudes X_{1}^{b} and X_{2}^{b} in the analysis of Fig. 6 (a) when F(X)=\bar{K}^{b} . c) \bar{K}^{c}