| Idea! r C 6 is motor L—Friction ‘Amplifier 7 Bsgnd Friction band Error, € Figure 1-20. A servomechanism with Coulomb friction and its error responses. friction. That this should be so is apparent from the equation of motion which is derived from consideration of the block diagram as follows, 16, = T= —Bsgn 0, + Ke, «=6,—0, (1-40) = —Bsgn@,—K6, when 0, = 0 (1-41) 0 8, + san 6, + © 6 or defining K/J as w? and —B//as f, 6+o0%°+f=0, 8<0 G+0%—f=0, 450 (1-42)40 INTRODUCTION The friction level + frepresents a bias on the pure cosine wave which is the solution of 6 + w2 = 0. The output of the servomechanism stops moving at the end of any given half cycle whenever the error is small enough so that the motor torque minus the inertial torque is less than the frictional torque. As this may well occur at any level of the error within the friction “band,” the number of possible equilibrium points is infinite. A somewhat similar situation is obtained if the amplifier or error detector in a servomechanism has a threshold or deadband nonlinearity. The system can then come to rest anywhere in the deadband. 5. Nonexponential Time Response It is worth noting that, in the example of Figure 1-20, the response curves not only show the possibility of several equilibrium points but also exhibit Logarithmic decrement Constant decrement Friction band (a) (b) Figure 1-21. Oscillations with damping. (a) Linear damping. (b) Nonlinear Coulomb damping. an approach to rest in which the amplitude of each half cycle is decreased by a constant amount. This constant, instead of logarithmic decrement, is typical of responses with Coulomb friction. In a linear system each mode of aperiodic or oscillatory motion can only increase or decrease within or along exponential envelopes. This is not always true in nonlinear systems. Figure 1-21 presents the exponential (logarithmic) damping of a linear system contrasted to the nonlinear system with Coulomb friction damping. Another example of nonlinear response unlike the behavior of the normal (exponential) modes of a linear lumped constant system is the response of a mechanical system with a “cubic” spring. If the spring characteristic has the form K(a) = 2°, the period of oscillation increases as the amplitude decays under the influence of damping. (This kind of a spring is called a “shard” spring. If the spring “constant” decreased with deflection the spring would be referred to as a “soft” spring.)BEHAVIOR OF NONLINEAR PARAMETER SYSTEMS 4l 6. Limit Cycles Limit cycle is the name given to a nonlinear oscillation of fixed frequency and amplitude determined by the nonlinear properties of the system. Limit cycles are one of the most frequently encountered modes of behavior peculiar to nonlinear systems. Everyday examples of limit cycles include the action of the human heart, the squealing of chalk on a blackboard, the flashing of automobile turn signals, and the operation of an escapement clock. Limit cycles are distinguished from linear oscillations in that their amplitude of oscillation is independent of initial conditions. If a limit Amplifier Motor Hysteresis. 1 1 K T+ >>rR > Limit cycle —> Figure 1-22. Limit cycle oscillations of a servomechanism with hysteresis. cycle is itself stable, the system will tend to fall into this condition of oscil- lation if the oscillation amplitude approaches that of the limit cycle, no matter what the initial condition or forcing function may have been. Limit cycles are very easily recognized as closed curves in the phase plane repre- sentation where velocity is plotted against position. For example, Figure 1-22 shows the response to small and large initial conditions of a servo- mechanism with hysteresis. Note that the final oscillation, the limit cycle, has the same frequency and amplitude in each case. Figure 1-23 presents the same data in the phase plane where velocity is plotted against displacement and time is a parameter along the trajectory curve. It sometimes happens that the motion of a system will tend to fall into a limit cycle from any starting condition. This behavior is termed soft self- excitation and is typical of devices which are deliberately intended as42 INTRODUCTION oscillators. On the other hand, there are systems whose motion is only forced into a limit cycle by some (perhaps rare) combination of circum- stances, such as a particularly large step input. Here the term hard self- excitation is applied. In connection with control systems, limit cycles are undesirable, and soft self-excitation is very nearly intolerable under any circumstances. A servomechanism configuration proposed and actually built by Lewis* and investigated by Caldwell and Rideout illustrates several features of nonlinear behavior. In this case the nonlinearity is intentional. The idea is Output Rate, 6, Limit cycle Output, 8, Figure 1-23. Phase plane representation of the limit cycle. to have a positioning servomechanism whose damping is negative or at least very small when the error is large. This tends to insure a more rapid response to large errors than in the corresponding linear system. This “desirable” effect is accomplished by a nonlinear velocity generator feed- back. A block diagram of the servomechanism is presented in Figure 1-24. Figure 1-25(a) shows the step responses for several input magnitudes, and a comparison is made between these and the best possible response with linear damping. These very rapid and apparently stable responses illustrate how attractive intentional nonlinearities can be. If excited with a double pulse input, however, the output of this servo may diverge, indicating that the system is unstable for this input. The unstable response is illustrated in * J.B. Lewis, ‘‘The Use of Nonlinear Feedback to Improve the Transient Response of a Servomechanism,” Trans. AIEE, Pt. Il, vol. 71 (1952), pp. 449-453. 1 R. R. Caldwell, V. C. Rideout, ‘A Differential-Analyzer Study of Certain Non- linearly Damped Servomechanisms,” Trans. AIEE, Pt. Il, vol. 72 (1953), pp. 165-169.9% ‘Absolute Multiplier value unit Figure 1-24. Block diagram representation of the Lewis servomechanism. 4 \ Output : pL tnput v2 \ = Linear BL VAR response ae ¢ = 065 0 j_, __1 0 7 2r Time, ¢ 0 | 1 - 2 0 7 2 Time, t : (0) f= 05,a= 10 (a) § =22,a=07 I ft h rout frequency Input frequency, : f = 4.92 cps (c) Figure 1-25. Response of the Lewis servomechanism to a variety of inputs. (a) Error responses to a step function. (b) Output response to a square doublet. (c) Frequency response at two different frequencies. (Partly adapted from Caldwell and Rideout, op. cit.) 4344 INTRODUCTION Figure 1-25(b). The response to sinusoidal inputs can also be surprising. Figure 1-25(c) illustrates the steady-state sinusoidal response at two dif- ferent frequencies. At a frequency of 1.57 cps the system is fairly well behaved. At the higher frequency of 4.92 cps it no longer responds at the frequency of the input and it exhibits frequency demultiplication. The Lewis servo thus exhibits new frequencies, and stability depends on the input shape and amplitude. 1.4 DIFFICULTIES OF NONLINEAR ANALYSIS Such factors as frequencies in the output unlike those in the input, jump resonance, limit cycles, the effects on stability, response shape, and final equilibrium of the size of the initial conditions and the input cannot be predicted by the methods of linear constant-parameter analysis. To cor- rect this deficiency, other methods of analysis are required to cope with nonlinear parameter problems. Unfortunately, nonlinear systems are not only essentially different physically from linear constant-parameter ones but they are also essentially different mathematically. The difference in their mathematical description introduces formidable analytical diffi- culties. In principle at least, linear constant-coefficient differential equations can always be solved, in a uniform fashion, using a general technique. In contrast, solutions to nonlinear differential equations are relatively rare, and are found by techniques almost as various as the known solutions. Even when the equations are “‘solved” by an electronic computer, the control engineer has not resolved his difficulties. The physical under- standing which would permit the extrapolation of results to new cases is often totally absent. The principal supports of linear constant-parameter system analysis are inapplicable to nonlinear problems. Stability as a property of the system, the frequency response concept, transform calculus, and the principle of superposition will not be usable, at least unless they are drastically modified. In connection with nonlinear systems, the essential concept of stability needs re-examination. Stability, or the lack of it, is a property of the linear constant-parameter system, and the stability is predictable on the basis of the homogeneous equation of motion of the system. Initial con- ditions and forcing functions have no effect on the stability. On the other hand, a nonlinear system may typically be stable for one input or initial condition and unstable for another. Thus, while a stable situation can still be defined, attention must be focused upon system and input or output combinations rather than on the system alone.DIFFICULTIES OF NONLINEAR ANALYSIS 45 With linear components it should be quite apparent that the principle of superposition applies. Twice the cause produces twice the effect and, in general, the total output is the sum of the outputs produced by each of the several elements of the input. The principle of superposition also applies on a dynamic basis to linear systems. The amplitude ratio function in a linear system frequency re- sponse, for example, is always the same no matter what the amplitude of the input wave may be. The transient responses of a linear system to steps of different magnitudes all “look” the same. In fact, the only difference is in a “stretching” or “contraction” of the response scale. The total transient is a sum of exponential terms each of whose magnitudes is linearly related to the magnitude of the forcing function. Each term can be treated sepa- rately so that the analysis can be broken into simple parts. Furthermore, since any input can be approximated by a series of small steps, an extremely important consequence of the principle of superposition in linear systems is that if the response to one input is known it is at least theoretically possible to know the response to any input. Since the principle of superposition does not apply to nonlinear systems, the whole problem must be treated as an entity, and the knowledge of the response to one particular input is just that. In fact, as has already been pointed out, nonlinear machines are full of surprises. A nonlinear servo- mechanism adjusted for the best performance in following a step of a certain magnitude may be unstable for a larger step, and may produce frequency demultiplication if excited with a sine wave. The nonlinear analog of the correlation between responses in the time and frequency domains for linear systems is particularly weak. To further complicate the matter, the Laplace transformation, which is especially useful in problems where the principle of superposition applies, becomes inapplicable to most problems in nonlinear dynamics. In general the sum of two solutions to a nonlinear differential equation is not a solution. Unfortunately, no general method of analysis comparable to the opera- tional calculus has been developed to treat systems with nonlinearities. In nonlinear (and time-varying-parameter analysis) the usual engineering analysis methods and the applicability of the results both have extremely limited generality. Where and how to use the methods can be defined only for specific cases. Yet, in spite of the analytical obstacles encountered in nonlinear problems, the control analyst is often confronted with nonlinear characteristics and must do his utmost to understand and predict their interesting effects.2 GENERAL TECHNIQUES FOR SOLVING NONLINEAR CONTROL PROBLEMS There are very few nonlinear differential equations whose solutions are known. Furthermore, even the few equations which have been solved are seldom encountered in control engineering.* Several methods of analysis, however, can be applied in specific cases in control engineering to obtain an approximate solution or other useful information, such as the stability of small motions. A correct estimate of the dynamic stability can always be made for the small motions of systems with simple nonlinearities that have a finite first derivative. Also, at least in principle, any ordinary differential equation, or set of simultancous equations, can be integrated by a step by step process. A variety of methods employing both graphical and numerical techniques are available for this purpose. If an approximate solution can be found, by any method, the accuracy of the solution can be refined to an arbitrary degree by means of an iterative process. Many nonlinear systems are characterized by being piecewise linear. That is, the operation of the machine obeys a certain linear differential equation in one region of operation and obeys another linear differential equation or equations in other regions. Significant use can then be made of known solutions to linear differential equations by piecing them together to give a solution of the original nonlinear equation. The end conditions of * Anentire text, for engineers, on the solution of problems in nonlinear mechanics is: W. J. Cunningham, Introduction to Nonlinear Analysis, McGraw-Hill Book Co., New York, 1958. 46LIAPOUNOFF STABILITY 47 one solution segment are used as the initial conditions of the next segment. In these cases superposition, with ail its attendant advantages, can be applied in the various linear segments. An elementary example of this method has already been presented in connection with the friction damped servomechanism of Figure 1-20. Even when the nonlinear system is not actually piecewise linear, it can often be approximated as such to a satis- factory degree of accuracy. This may be especially true when the inputs are suitably restricted in their characteristics. There are, finally, two methods of engineering analysis for nonlinear control systems which also have an appreciable degree of generality. The describing function approach is most useful in complex systems where the effects of the nonlinearities are significant but “small.” This approach is most commonly used with sinusoidal inputs, but the concept can be ex- tended to other input functions, including those described statistically. The Phase plane method, on the other hand, is most useful in connection with large nonlinearities in simple systems. Although it can be viewed in quite general terms, its primary usefulness in engineering practice is in situations having only initial conditions of displacement or velocity, or step or ramp inputs. In a pragmatical sense, these two engineering methods complement each other. Together with computer solutions they constitute the principal tools of the control engineer in attacking nonlinear problems. It should be noted here that the use of the mathematical model set up on a computer is the most powerful method of all for specific problems. This, however, is the proper subject of books on computer applications* and is beyond the scope of the present work. 2.1 LIAPOUNOFF STABILITY The question of the stability of small motions of a nonlinear system con- taining only nonlinearities which possess continuous derivatives has been throroughly investigated by M. A. Liapounoff.+{ If the nonlinearity is * See footnote on page 13 and also: H. M. Paynter (Ed.), “A Palimpsest on the Electronic Analog Art,” Geo. A. Philbrick Researches, Boston, Mass. 1955. + M. A. Liapounoff, “Probléme général de la stabilité du movement,” Annals of Mathematical Studies, No. 17, Princeton University Press, Princeton, N.J., 1947. The second, or direct, method of Liapounoff for the determination of the stability of systems described by linear or nonlinear differential equations is best understood in connection with trajectories which are phase space representations of the solutions to the differential equations. Discussion of this matter is therefore deferred to the point where it can be taken up in the appropriate context of the trajectories in the phase space. + Because the Russian language uses an alphabet completely different from the Latin one, and because English is a notably unphonetic language, there is tremendous variety48 GENERAL TECHNIQUES FOR NONLINEAR PROBLEMS single-valued and has derivatives of every order in the vicinity of a point, a, the nonlinear function y = f(x) can be represented by a Taylor series: v= f(a) = ula) + (x — a() + Fe — ap (£4) oe *( £4) et (£4) - +5 ah Tt — Oey, + The first two terms of the series represent the operating point and the linear (or first) approximation to the actual nonlinearity. Liapounoff’s research resulted in the theorem: “If the real parts of the roots of the characteristic equation corresponding to the differential equations of the first approximation are different from zero, the equations of the first approximation always give a correct answer to the question of stability of a nonlinear system.” * According to this theorem, if all the real parts of the roots of the charac- teristic equation of the linear approximation of the differential equation are negative, the nonlinear system is stable about the point in question, and any small temporary disturbance in the input will result in a temporary disturbance in the output. If, however, any of the real parts of the roots of the characteristic equation of the linear approximation of the differential equation are positive, the nonlinear system is unstable about the operating point, and any small temporary disturbance at the input will result in an output which will diverge from this unstable point. If any of the roots of the linear approximation of the differential equation about the equilibrium point have zero real parts, the theorem may not be used to give a definitive answer to the question of stability. Zero roots may Tesult in a situation in which the stability might depend on the direction of the disturbance. In control engineering this situation would usually be as undesirable as an absolutely unstable situation, hence the fact that the theorem does not apply is of little consequence. It should be pointed out that, although the linear approximation of the differential equations may indicate a stable system for all amplitudes of disturbances, the theorem actually applies only to small disturbances within in the transliteration of proper names from the Russian, and no “correct” spelling. The author's name, for example, may be transliterated Liapounoff, Ljapunov, Lyapunov. and so forth. Since the main feature of library indexing is by authors’ names listed alphabetically, this confusion is a burden to the student and research worker. In this book the attempt is made to preserve the transliteration of the best recognized, or most available, translation or interpretation. * N. Minorsky, Introduction to Non-linear Mechanics, J. W. Edwards, Ann Arbor, Mich., 1947, p. 52.LIAPOUNOFF STABILITY 49 the range of validity of the Taylor series written about the selected operating joint. i To illustrate the application of the Liapounoff theorem, consider the differential equation: aay (2) 4 ke =@ (2-1) gaia dt This is van der Pol’s equation* and when « < 0 it describes, among other physical situations, a type of electronic oscillator. When the acceleration and velocity are zero, the value of x defines the point of equilibrium; that is, kx = Q; x = Q/k at the equilibrium point. In order to define the nonlinear function in the vicinity of an equilibrium point, X, let: w=X+2, where — = — = Then, by substitution: @& + ull — (XP + 2XE + Ze KK +]=0 22) The first (linear) approximation leads to the equation Px dx = 1-X*7]=+kze=0 2-3 wet oL ] Gate (2-3) and the characteristic equation is: ev +n —X%)s+k=0 (2-4) Since the Routh-Hurwitz stability condition applied to the characteristic equation is wl — X2)>0 (2-5) Equation 2-S indicates that when u > 0 the system is stable for all X less than unity. As X approaches unity, the coefficient u(1 — X*) becomes small and the system becomes poorly damped. The theorem is not appli- cable when X = | since this condition results in zero real parts for the roots of the characteristic equation. For X greater than unity the system is unstable. The concept of Liapounoff stability pertains solely to the dynamic sta- bility in the immediate vicinity of an equilibrium point. A related, but less precise, concept can be employed to predict any possible instability for systems which contain a simple nonlinearity. * B. van der Pol, “On Relaxation Oscillations,” Phil. Mag., series 7, vol. 2 (July-Dec. 1926), pp. 978-992.50 GENERAL TECHNIQUES FOR NONLINEAR PROBLEMS If the nonlinearity is single-valued and continuous and has a finite first derivative (but not necessarily finite higher order derivatives), a general statement can be made which applies to both small and large motions. The nonlinearity can be viewed as a gain-changing element, and any nonlinear- ity of this type can be approximated, arbitrarily closely, by a series of straight-line segments. Each of the straight-line segments corresponds to an incremental gain, K;. If the system is stable for all such K,, it will be stable even in the presence of the nonlinearity. Kalman* has suggested that a necessary but not sufficient condition for instability is a gain K; which would produce instability in an equivalent linear system. K ‘Actual nonlinear function Amplifier Output, K(e, Approximation by straight-line segments Error, € Figure 2-1. Approximation of a gain-changing nonlinearity with straight-line segments. Of course, conventional constant-coefficient linear analysis methods, for example, root loci and Bode diagrams, can be employed to examine the stability as a function of these incremental gains. Figure 2-1 shows a single-valued nonlinearity and the approximation by straight-line segments. If this nonlinearity were in the forward loop of the closed loop system illus- trated in Figure 2-2(a), and the system were stable for the maximum gain (slope) of the nonlinear function, it would be stable under all conditions. Figure 2-2(b) shows the root locus plot for the system of Figure 2-1. The system is indicated to be stable for the maximum gain K,. If G(s) were of such a form that the system were conditionally stable, the stability would have to be examined for all possible values of K;. This concept of a neces- sary (but not sufficient) condition for instability can be applied to a system of any order with any number of sirnple nonlinearities. *R. E. Kalman, ‘Physical and Mathematical Mechanisms of Instability in Non- linear Automatic Control Systems,”” Trans. ASME, vol. 79 (1957), pp. 553-563.LIAPOUNOFF STABILITY 51 The determination of stability according to Liapounoff can be viewed either as a powerful justification for the use of linear-analysis or as a pre- liminary step in the approximate nonlinear analysis. The validity of the linear approximation is of particular interest when it is used to predict the stability of the system. If the solution of continuous nonlinear equations indicated the possibility of a condition of instability not revealed by the € 1 9. > KO) | 08) = ae > Simple Linear transfer nonlinearity function (6) Figure 2-2. A servomechanism with the gain-changing nonlinearity in the forward loop and the root locus plot. linear approximation, the linear approximation would be practically worth- Jess. However, for simple nonlinearities in general, and in all cases where Liapounoff’s conditions are met, the linear approximation can always be depended upon to determine the stability in the vicinity of an equilibrium point. Furthermore, according to Kalman, no instability is possible for small or large motions without an incremental gain which would produce instability in the equivalent linear system.52. GENERAL TECHNIQUES FOR NONLINEAR PROBLEMS 2.2 DIRECT SOLUTIONS Where the mere knowledge of stability is insufficient or where the restric- tions placed on the given methods for examining stability are too severe, it may be desirable to attempt an exact or an approximate solution to the equations of motion. A few nonlinear equations have known solutions. The equation of motion of a simple pendulum, for example, can be solved in terms of elliptic integrals of the first kind. Although this problem has been treated extensively elsewhere,* the rudiments of an approximate solution are presented here. This serves the triple purpose of providing an illustrative example, showing a technique to be used again later, and presenting a classic nonlinear parameter problem, classic both historically and in the fact that it has been solved completely! The equation of motion of the simple pendulum is 2 SS + o,f sin 6 =0 (2-6) where § = deflection of the pendulum from the vertical ,° = g/l = acceleration of gravity/length of the pendulum t = time The first integration, to obtain the velocity, is set up by multiplying by dO/dt: ° 7 fom 1a (ap _ Alvar say Multiplication by df leaves the left-hand side as an exact differential. Integrating, and inserting the initial conditions 6 =0 and d0/dt = wy when ¢ = 0, the velocity is: dé 20n) 6 ae on - (222) sine (3) (2-8) A final integration then yields: -iUtpiog * See, for example; T. v. Karman, M. A. Biot, Mathematical Methods in Engineering, McGraw-Hill Book Co., New York, 1940. N. W. McLachlan, Ordinary Nonlinear Differential Equations in Engineering and Physical Sciences, University Press, Oxford, 1956. 7 do ne sin 8) ae (2-7) (2-9)DIRECT SOLUTIONS 53 This expression can be manipulated so as to produce a form which is an elliptic integral of the first kind as tabulated in tables of elliptic integrals. While three fundamentally different forms of motion are possible, depending upon the magnitude of 2w,,/m9, the possible oscillatory motion (2@,,/@o > 1) is the most interesting for the purpose of drawing analogies to control systems. In this case, the radical will be real only when 2(@,,/@ 9) sin (6/2) < 1. Consequently, the oscillatory motion must have a maximum amplitude, 6,,, such that: sin (8/2) = @o/2@, OF =~ Aq = 2 Sin? (wp/2,) The period of the motion can be found by noting that a quarter period corresponds to the time taken in going from 6 = 0 to @ = 6,,. Thus: 46 ad Tefal) The period is then easily computed if the integral is modified to have limits between zero and 7/2. This can be accomplished by replacing 2(w,/@o) sin (6/2) by sin 4, since the upper limit will then correspond to ¢ =7/2. Since, then, sin ¢ = 2(w,,/@) sin (8/2) differentiation yields cos ¢ df = (w,,/a») cos (0/2) db and the modified expression for the period becomes: (2-10) = cos ddd ol ihieas Tayo If w/2m, is much less than unity, the integrand can be expanded by the binomial theorem to give a series expression which converges rapidly, that is: Pos Po [a5 (af ome e+ 2i(m =) sot iat ‘Jes in! 2, + Hap +S (2) —] (2-42) p=4 (2-11) For the small values of w/w, assumed, w,/2m, = sin (6,,/2) © (6,,/2), So @o/w,, is approximately 6,,. In terms of 0,, the period is then: pa [i4%ey. “| (2-13) @,54 GENERAL TECHNIQUES FOR NONLINEAR PROBLEMS Besides the very few nonlinear equations with known solutions, one occasionally comes across an equation (almost invariably, or reducible to, a first-order equation) which can be integrated readily by standard methods. The most useful method in practice requires the separation of the variables. When one has | of the form dy _ HO 2-14 Shy) = 2) (2-14) these become: By) dy = $(t) dt (2-15) or fo dy -{ao a+C (2-16) The first integration of the pendulum problem utilized the fact that the variables could be separated. As another example consider the motion of a mass particle, m, subjected to a constant force, T, and operating against quadratic friction. The non- linear equation of motion is ee ea Z(t ne *) 2-17) dt m m m T where v = velocity b = damping coefficient Separating the variables: . dv T [ ——. = - | at 2-18 f 1-—(b/T)® mJo ean Integrating: 1+ Jd/T v2 In —_— = — J/bTt 2-19) 1— Jo/T vom Vv ae) or 1+ Jb/Tv (2 — ) ~~ = = JbTt 2-20) 1— /b/Tv ad m v (2-20) Solving for v: ae a Sm fF ean, (2-21) exp bT t) +1 Homogeneous equations, such as arf) omAPPROXIMATE SOLUTIONS IN SERIES 55 can be put into a form in which the variables are separable by means of the substitution y = vt. Consider, for example, the equation: dy_ yt? _y it at 244 - 2-23 . dt 2ty 2t a 2y ( ) Letting y = vt: a 7 a ly yp ot == t—= 2-24) ee eae em) oe 2 phi! (2-25) Be t 2v 2v dv dt = (2-26) 1-v’ ¢ Ci which is readily integrated by quadratures. The use of an integrating factor, which makes the equation an exact differential (for example, the multiplication by d6/dt in the pendulum problem), is another elementary method which is occasionally applicable to the problems of control system analysis. The techniques reviewed here by no means exhaust the possibilities ap- plicable to nonlinear equations in general.* For control engineering prob- lems, however, these techniques represent a fair cross section of the direct solution methods which have been found useful. In view of this last state- ment and the severely restricted types of problems to which these elementary direct methods apply, it should be evident that the chance of encountering a control engineering equation of motion which can be integrated directly is actually very small. 2.3 APPROXIMATE SOLUTIONS IN SERIES If exact solutions are not possible, approximate solutions to differential equations can be found in two different ways. A function, x = f(t), is to be found which satisfies the equation for a given range of values of the inde- pendent variable, ¢. Either 1. the expression for x can be found in terms of functions of ¢, and values of x are then obtained by substitution; or, alternatively, 2. tabulated values of « can be found which correspond to tabulated values of t. * A useful tabulation of particular nonlinear equations for which solutions are known, or which have been extensively studied, can be found in: E. Kamke, Dif- ferentialgleichungen, Losungsmethoden und Losungen, Chelsea Publishing Company, New York, 1948,56 GENERAL TECHNIQUES FOR NONLINEAR PROBLEMS An equivalent of the tabulated values is, of course, a graph of x against t, The practical application of these methods involves either finding an in- finite series which is an approximate solution or carrying out a step by step numerical or graphical integration. The material which follows is not meant to be a complete discussion of numerical or graphical methods of solution for ordinary differential equa- tions. It is intended, however, to illustrate what can be accomplished by several elementary methods, and to give the reader tools with which to attack what might otherwise be a severely difficult problem. More com- plete discussions and extended methods are presented by Levy and Baggot, and by Willers, among others.* The method of Picard is used to obtain a solution in terms of a power series. Unlike approaches which depend upon the Taylor series, this method can be used even when one or more of the derivatives of x are infinite at the initial point. Suppose the equation is: dz a TI) (2-27) and the initial conditions are x = xq, t = fo. If the origin is changed to the initial point, the equation is unchanged except that the variables are now understood to represent the departures from the initial conditions. If an initial condition on the dependent variable is given, this may very well be taken as the first approximation to a solution. Otherwise, assume for small values of ¢ that: du sae «x=at" and therefore oman a If these expressions are substituted in the original differential equation, the constants a and n can be determined. A first approximation «is calculated. 2, is substituted in the expression for f(x, t), and the equation is integrated to give a more accurate expression 2». If xg is substituted in the expression for f(x, t) and the equation is integrated again, a still more accurate expres- sion x, is the result. This process can be repeated as many times as may be desirable. It illustrates a general process of iteration which may be used to refine an approximate solution. As an example of the method of Picard, consider the equation of motion of a body opposed by quadratic friction accelerating from a standstill under * H. Levy, E. A. Baggot, Numerical Solutions of Differential Equations, Dover Publica- tions, New York, 1950. Fr. A. Willers, Practical Analysis, Dover Publications, New York, 1948.APPROXIMATE SOLUTIONS IN SERIES 57 the influence of a constant force. This problem was solved in exact form in the previous section. wl ibeyt (2-28) dt m m where v = vy = Owhent = 0. Let v, = at” Then ant = — 2 gyn E (2-29) m m for small ¢. Assuming that the term t?” may be neglected ant™ t= + ZI (2-30) m n — 1 = 0 by equating exponents therefore n=1 an = T/m by equating coefficients (2-31) therefore a=T/m and therefore finally v=a"= = t m Now ‘dv = —dt 2-32) van+[ 2-3) and from the original equation: See (2-33) tacit So the second approximation is obtained from the first one by substituting and integrating: a aao+] [Z-2s] dt olm om t 2 = [ [z zs #(Z) a| at (2-34) ioLm m\m -(}- (ETE m, m/ \m/ 358 GENERAL TECHNIQUES FOR NONLINEAR PROBLEMS By repeating the process of squaring the approximation, multiplying by —b/m, integrating the series with respect to time, and adding the result to (T/m)t (which would reappear after each integration), one may obtain: 2 3 2 5 “M4? ENTE ASCE ens m m! \m/ 3 m/ \m/ 3-5 m/ \m/7-9 2 8 2 5 " 7 BMT BCH m m/ \m/ 3 m/ \m/ 3-5 m/ \m/3-3-5-7 “Ps 9 6 YS s5- Oat m/ \m/3-3-5-+7-9 m/ \m/3-5-5-7-9- 11 tos ates o m/ \m/3+5-7-9-13 m/ \m/7-7-9-9-15 The exact solution to this problem, given by Equation 2-21, is: o= fF ton, b m 2 48 2 5 =T i (*)(2FE 4 2(8) (FE m m/ \m/ 3 m/ \m/ 15 ay? t 9 = (2 (24 + (”) (7)8.--- m/ \m/ 315 m/ \m/ 2835 Equation 2-36 may be seen to be identical to the exact solution through the fourth term. In any series approximation like the one of Equation 2-36, if the first neglected term is taken as approximately equal to the allowable error, then the range of values of the independent variable over which the series is Valid can be calculated. Thus, for example, if one chooses to neglect the term in 1° and subsequent terms in the series of Equation 2-36, and the maximum allowable error in v is 0.01, then ae 33(2) (Zf2 <0.01 ml \ml 5-7-9°9 2835(0.01) ie aes with appropriate units and numerical values for b/m and T/m. As the calculation of the coefficients in the series required to maintain a given degree of accuracy becomes excessively tedious, it is possible to start a new series which would be valid over a new range of values of the (2-37)APPROXIMATE SOLUTIONS IN SERIES 59 independent variable. In the example given above a new series could be developed for the values of t beyond f,,,x. Such a process, however, could rapidly become more of a task than the application of one of the methods of numerical or graphical integration. The method of Picard can also be applied to solve higher order and simul- taneous equations. With regard to the higher order equations, it should be noted that a differential equation of order may be reduced, byappropriate substitutions, to an equivalent system of n simultaneous equations, each of the first order. For example, the second-order equation 43 + 60=0 2-38 de Mic is equivalent to the set of two simultaneous equations: dx “= 2-39) ae (2-39) ® 6x —% (2-40) dt These equations, with a change in time scale, represent the simplified situation of an airplane following the localizer beam to an automatic landing. If zy = 1, (dz/dt)) = z = 1, and ty = Oare the initial conditions, then: t t zemtfednt+[ea 0 0 Hi z= % + [on —32)dt=1 + [6 — 32) dt By substituting the initial conditions as a first approximation and carrying out the indicated integrations, the first approximation is obtained: m=1tt (2-41) z=1-3t The second approximation is found by substituting the first approximation and carrying out the indicated integrations: 3f at =1+ Paaai+1— 2 : 32 (2-42) m%=it (—6myt — 3a) dt = 1 — 34+ — 2iP 060 GENERAL TECHNIQUES FOR NONLINEAR PROBLEMS The third approximation is determined in the same way: t 32 nat+| (1 — 34 3 _ 29) dt 0 2 Seu ate eee gee eects a Dea Deeg: 7 mel +f (—62yt — 324) dt : : (2-43) =14[' (6-68 + 98-3 491-8 + of) at 0 3°78 | 1504 Seep oeeroe Hinata . af ft ar | [ da=it1—-2 45-4 oe nee eee err and so forth. A power series which is an approximate solution to a differential equa- tion can often be found in terms of a Taylor series. For example: & = fe,» (2-44) Att=aandz=4), d oF = a = f(a, b) Differentiating Equation 2-44, : ee (=, t, 4) (2-45) de dt and evaluating at ¢ = a and « = b, a9" = g(a, b, x) and so forth. The process is continued and the results inserted into a Taylor series to obtain the solution: (= b)=(t—a)ey’ +L (¢—a)?xy’ + Lt — aay + tL ayy” 21 31 nl The use of the Taylor series implies the restriction that all derivatives, d"x/dt", must exist in the region about t = a. If the equation is of higher order than the first, the same procedure can be followed because the initial conditions on all the derivatives will be known, and the higher power terms in the solution can be calculated by successive differentiation.APPROXIMATE SOLUTIONS IN SERIES 6l Consider Equation 2-38, here for convenience, Px a mt +34 _ = + 6xt =0 (2-38) subject to the initial conditions ¢ = "6, daldt = 1.0, and x = 1.0. Then x" = —32' — bat a —3a" — 6a't — 6x a’ = —32" — 62"t — 122’ o = va — 6x"t — 182" aN! = —30" — 60" — 240” and x" = 3 my” =+9-6=3 ao’ = —9 —12 = —21 a’ = +63 + 54= +117 ay\| = —351 — 72 = —423 Therefore: 38 diet tT 42308 SUeTEage (Esmee eon te nc (@-)= This series coresponds term for term with the one developed by the method of Picard, except for the last term in the previous series which is the one which would change at the next step. The range over which the series is valid can again be approximated by equating the first neglected term to the maximum allowable error in the dependent variable. The Taylor theorem approach can be used with simultaneous equations by solving them in a parallel fashion. The process can be illustrated as follows: a SF = fe 2) dx (2-47) ee aya) az The initial conditions are x = 2», Y¥ = Yo, Z = 203 Yo and 29 are deter- mined by the original equations. Differentiating: y= fle H2 92) (2-48) 2" = Bolx, 2, Y', 2’) Now yo” and 2,” can be calculated by substituting known quantities on the right-hand sides of these equations. Proceeding to differentiate and sub- stitute, y)” and z)” and higher derivatives at the initial point are determined.62 GENERAL TECHNIQUES FOR NONLINEAR PROBLEMS Then the solution is given by the two equations: rgd oil ” Y= Yo + (® — a)uo! + 5; ( — Ba)” + 5 (@ — a)Y0” + ; : (2-49) 2 tot (U — tyleo! +55 ( — aa)en" + 37 (He — aa)e0l” +o 2.4 STEP BY STEP INTEGRATION When an approximate solution in terms of an infinite series is impractical, a solution can always be obtained by means of step by step integration. =_— & g > " —_~ HS BS Be — ~— 0 :— Figure 2-3. Graphical solution of a differential equation. Either graphical or numerical methods can be employed. The procedures are actually identical in principle. Since the graphical methods are useful as a means of visualizing the numerical procedures they will be explained first. Consider the first-order differential equation: d oe = in TIO (2-50) For any initial values of x and y, at the point Pp, there is a slope (dy/dx), which is obtained from the equation. A short line with this slope can be laid off from the initial point to the point P;. Vide Figure 2-3. The coordinates ofSTEP BY STEP INTEGRATION 63 P, determine a new pair of values x, and y;, and from them a new slope (dy/dz), can be calculated. This slope is now laid off from P, along a line segment ending at P,. There a new slope is determined and laid off, and so forth. Increments in the independent variable # need not be constant. In fact, the increments should be small and the line segments short where the slope is changing rapidly. They may be longer where the slope is changing slowly. Table 2-1 Numerical Values for the Graphical Integration dy x y i 0 1.0 0 0.1 1.0 —0.100 0.2 0.99 —0.198 0.3 0.97 —0.290 0.4 0.94 —0.375 0.5 0.90 —0.450 0.6 0.85 —0.510 0.8 0.75 —0.60 1.0 0.63 -0.63 1.2 0.50 —0.60 1.4 0.38 —0.530 1.6 0.275 -0.44 17 0.23 —0.392 1.8 0.19 —0,342 19 0.16 —0.304 2.0 0.13 =0.26 21 0.10 -0.21 2.2 0.08 —0.176 24 0.045 = —0.108 2.6 0.02 —0.052 This is the simplest form of the step by step integration of differential equations. It is relatively crude and errors tend to accumulate rapidly. Of course, if a first approximation to the solution can be found, the approxi- mation can be refined by iteration in the same way as with the solution in series. The. process is illustrated in detail by the example of the equation eye 2-51 aa xy (2-51) with the initial condition y = 1.0 when x = 0. A useful table, such as Table 2-1, giving numerical values, is maintained in conjunction with the64 GENERAL TECHNIQUES FOR NONLINEAR PROBLEMS graphical construction shown in Figure 2-4, The values of 2 and y at the end of each line segment are inscribed in the table and a record of the calcu- lated values of dy/dz is kept. The process starts with a line segment of zero slope from the point y = 1.0, = 0. An interval in x of 0.1 is chosen as convenient. At the point «= 0.1, y = 1.0 the slope is —(0.1)(1.0) = —0.100. A line with this slope is constructed forward from the point = 0.1, y = 1.0. Along this line at x = 0.2, y is read from the graph and is found to be 0.99. The slope here is dy/dx = —(0.2)(0.99) = —0.198, and so forth, 10 1 = rst approximation—step by step with | 06 Points from the SF taight-tine segments $2 = — xy, 2 analytical salution| | _ [A] S$yy90= 1: #0 = 0 04 eee 4 Second approximation In 02 mt [tee from first approximation RE 0 Renee gag x Figure 2-4. Graphical solution of the equation dy/dx = —xy. to the third, fourth, fifth, and sixth points on the curve. There it may be observed that the slope is no longer changing so rapidly. It is now per- missible to lengthen the increments in the independent variable, x, until the slope begins to change rapidly again. A solution obtained in this way might be accurate enough. More accu- racy could be obtained by plotting the slopes at smaller intervals. On the other hand, a second approximation (more accurate than the first) may be obtained by plotting the values of dy/dz (which are available in Table 2-1) and integrating the dy/dxcurve. The integration can beaccomplished graphi- cally, numerically, or mechanically (with a planimeter). The curves for dy/dx from the first approximation, and the second approximation y = S(«) obtained by integrating dy/dx (using the rectangular approximation), are plotted in Figure 2-4.STEP BY STEP INTEGRATION 65 It is apparent, of course, that Equation 2-51 can be integrated directly by separating variables. Points from the exact solution are plotted in Figure 2-4. for purposes of comparison. In this particular case, the second approximation is considered to be more than good enough. If it were not, however, a new set of slopes could be calculated from pairs of points on the second approximation solution curve. If these slopes were plotted and integrated, the result would be a third approximation and so forth. Simultaneous equations (and therefore higher order equations) can also be solved by this same method. Here, of course, it is necessary to carry the process forward simultaneously with more than one curve. Assume, for example, a second-order equation reduced by substitution of a new variable to two simultaneous first-order equations: oY = fla m2) 2-52) £ = o(x,4,2) dx ™ The values x», yy and zy may be substituted to obtain (dy/dx), and (dz/dr),. Line segments with these values are constructed on separate graphs of y vs. x and z vs. x. The end points of the line segments must now be taken at identical values of x. They determine a new set of values 2, y, and 2. These are substituted in the two equations and determine two new slopes (dy/dx), and (dz/dz),. The construction of the two graphs of the solution proceeds together in this manner, as shown in Figure 2-5. If the derivatives are plotted and integrated, a second approximation can be obtained exactly as in the case of the single first-order equation. It has already been pointed out that the simple method of step by step integration tends to accumulate errors. A considerable gain in accuracy is achieved by using an average slope between its two end points for each individual line segment. This amounts to taking half a step backward for each step forward, but the benefits in accuracy may well repay the extra labor. A graphical construction using the average slopes is illustrated in Figure 2-6. It starts in a way identical to the simple method with a slope (dy/dz)y calculated from the initial values of x and y. The slope (dy/dz),, calculated from the values x, and y, at the end of the first line segment, is also plotted proceeding from 2, Yo. The average slope between the two is then em- ployed to construct a segment of the actual solution. This segment will ter- minate at a point 2}, yy, different from 2,, y,. The pair of values +, 9}, is used to calculate a new slope (dy/dx),, which is projected to x, y, and this66 GENERAL TECHNIQUES FOR NONLINEAR PROBLEMS 3% es 0 0 x— > Figure 2-5. Graphical solution of two simultaneous first-order equations. pair of values determines (dy/dz),. The average slope between (dy/dz),, and (dy/dz), is now used to construct a second segment of the solution. The same procedure is applied repetitively until the interesting range of values of the independent variable has been covered. In some cases it might be discovered that there is no appreciable dif- ference between the first and second approximations to the value of the derivatives, such as (dy/dx), and (dy/dzx),,, at a given value of the independ- ent variable, such as x. If this is true, the graphical construction using average slopes can be carried out very rapidly since each slope is only calcu- Jated once, as in the simple method. Each slope, however, is plotted twice and improved accuracy is achieved. On the other hand, it may be advan- tageous, although tedious, to determine a third approximation to theSTEP BY STEP INTEGRATION 67 0 2—> Figure 2-6. Improving the graphical solution by using average slopes. dependent variable (y,,;, for example) at each step. Thus y,,, would be the end point (at x = z,) of a line segment drawn from xp, yy with a slope dy|de = 3{(dy[dx)y + (dyldz)s1). The procedures just described for the graphical solution of differential equations can alternatively be carried out completely numerically without any plotting. What has been called the simple method is known in its completely numerical form as the method of Euler. In Euler’s method, if the equation is dy/dxe = f(x,y), then the initial values x and yo, are used to calculate (dy/dx)y. This slope is multiplied by an incre- ment Az to give an increment in the dependent variable Ay. The new values x, = 2%) + Ax and y,; = yy + Ay are now substituted to find (dy/dzx),, which is in turn multiplied by an increment Az to find a new incre- ment Ay,. The calculation is repeated over and over in order to cover the range of values of x which is of interest. Procedures for integrating simultaneous equations, and for reducing higher order equations to simultaneous equations, are exactly analogous to the procedures used in conjunction with the graphical constructions. The method of Euler can also be modified to make use of average slopes as in the more elaborate graphical procedure. Suppose there is a differen- tial equation in the form: d"y _ (“2s dy dy ) dx” da gk dn (2-53). 90 8r'I— 88'1— IvI- ayeunnsy 86°0— I~ s61—- 8L:0- elz— 8L:0— r0 8L':0— Loz— 8L0- aqeunsy, 8s0— |~90'2 8e0- soz— 8e0— v0 8e0— 661 9€°0— areunsy, 61'0— [68 0 o8'I— 0 0 “G) “@) |) ‘G) o=>+? plo — atv +2 :SUOIIPUOD [eNTUy UL e071 sadojg adesiaay Buss, poyrapl daag jerjsaunyy ayy TT PIFEL 68STEP BY STEP INTEGRATION 69 The first estimate, (d"y/dx"),, for the slope at the starting point is obtained by substituting the initial conditions. Multiplying (d"y/dx") by Az gives a first approximation to (d"-ly/de"-1), Multiplying (d"“1y/de"“) by Ax gives a first approximation to (d"~*y/dx"-*), and so forth, to the first approximation for y,. y, and its derivatives are now substituted in the functional relationship to determine the second estimate for the slope (ay/ dx"), The average slope }[(d"y/dx")y + (d"y/dx"),], which is a final estimate, is then multiplied by Ax to give a second approximation to (d"y/dx"), This slope is averaged with (d"-1y/dx"), and multiplied by Az to give the second approximation to (d"~*y/dz"~*), and so on. Table 2-2 illustrates the process of step by step numerical integration by presenting the first few steps in the solution of the normalized equation of motion of the servomechanism whose block diagram is presented in Figure 1-24, The methods for the graphical and numerical solution of differential equations which have been presented are general and flexible. They are readily grasped, and are relatively easy to carry out. They give fairly accu- rate results with a minimum amount of labor. Other methods are available where a high degree of precision is required; and there are still other special methods where the equation has some particular form. For a discussion of these matters the reader is referred to any of the several excellent books on numerical methods.* There is one common engineering method of approximate graphical solution of first- and second-order ordinary differential equations which has not been discussed here. This is the method of isoclines. The discussion of this technique is deferred to Chapter 7, where it can be presented in a more appropriate setting in conjunction with the phase plane. The methods which have been presented should be employed when other methods fail or are inapplicable. Perhaps the solutions obtained for a * H. Levy, E. A. Baggot, Numerical Solutions of Differential Equations, Dover Publi- cations, New York, 1950. I. S. Sokolnikoff, E. S. Sokolnikoff, Higher Mathematics for Engineers and Physicists, McGraw-Hill Book Co., New York, 1941. A summary discussion of numerical methods aimed at control engineers is in E. S. Smith, Automatic Control Engineering, McGraw-Hill Book Co., New York, 1944. In addition to the general methods there have been several presentations of thenumber series methods which are analogous to numerical convolution. The number series method, however, unlike convolution, may be applied to nonlinear control systems. See, for example, A. Tustin, “A Method of Analyzing the Behavior of Linear Systems in Terms of Time Series,” Journal IEE, Pt. la, vol. 94 (1947), pp. 130-142; A. Tustin, “A Method of Analyzing the Effect of Certain Kinds of Non-linearity in Closed-Cycle Control Systems,” Journal IEE, Pt. Ha, vol. 94 (1947), pp. 152-160, and A. Madwed, Number Series Method of Solving Linear and Nonlinear Differential Equations, Report No. 6445-T-26, Instrumentation Laboratory, MIT, Cambridge, Mass. April 1950.70 GENERAL TECHNIQUES FOR NONLINEAR PROBLEMS very limited number of parameter and initial condition combinations may illuminate the whole problem. At worst, a limited amount of information, properly considered, is much to be preferred over an attack which entirely disdains the analytical approach. The examples which have been presented should indicate that a control engineer has no particular cause for despair even in the event thatconventional analytical approaches to nonlinear param- eter problems are unsuccessful. A direct attack aimed at a solution of the equation is likely to, quite literally, solve the problem. Step by step methods which have been presented here are, in principle, the same techniques by which digital computers solve the equations of motion of nonlinear control systems. 2.5 PIECEWISE LINEAR SOLUTIONS Many nonlinear systems are “piecewise linear,” or may approximate this condition. The term “piecewise linear” is meant to imply that, in its over-all operation, the system obeys two or more sets of linear constant- coefficient differential equations. Each of the several equation sets is asso- ciated with different regions of operation. The performance of the system can be examined by “piecing” together solutions to the individual equations. Final values of one solution piece, as the operation of the system crosses the boundary between operating regions, are used as the initial conditions for the next one. Some simplified representations of relay servomechanisms can be analyzed as if they were piecewise linear. A separate exposition of relay servomechanisms is presented in Chapter 9. Another example of a piecewise linear system is furnished by the inverted pendulum with a vertically oscillating pivot when its operation is described by the Meissner equation: * ao ade This equation is an approximation to the Mathieu equation, 6 + (« — A cos t)§ = 0, which actually describes the performance of the inverted pendulum with an oscillating pivot. The cosine term in the Mathieu equation has now been replaced by a square wave, and thus a linear time- varying equation has been replaced by a nonlinear equation. Transition + («4 po =0 (2-54) * E. Meissner, “Ueber Schuettelerscheinungen in Systemen mit periodisch veraender- licher Elastizitaet,” Schweitzerische Bauzeitung, vol. 72, no. 11 (Sept. 14, 1918), pp. 95-98. J. P. Den Hartog, Mechanical Vibrations, McGraw-Hill Book Co., New York, 1940.PIECEWISE LINEAR SOLUTIONS 7 from the plus value to the minus value of 8 and back takes place when t = 1/2, 327[2, Si7/2, 77/2, +++. The nature of the approximation is indicated in Figure 2-7. A solution is obtained by piecing together at times t = 7/2, 3/2, 5/2, 7/2, +--+, solutions to the two linear constant-coefficient equations: #6 — + («+ p)6=0 qe t eth and (2-55) &O —+(—-—p6=0 et @-# "A cost +8 o4 -B z 3x sr In on 0 2 2 2 2 2 Time, t —> Figure 2-7. The approximation employed in the Meissner equation. The solution segments are given by: a. sin Ja + Bt and (2-56) 6(t) = O)cos Ja — Bt Hoga tt — Bt A(t) = Oy cos fa + Bt+ To illustrate the fact that an inverted a with oscillating pivot can be “stable,” consider a configuration where « = —0.05, B = +0.35, and the initial conditions are 0, = 0.1 radian, (d0/dt)) = 0. 6 is the angle of the pendulum measured from vertical. When 9 = 0 the bob of the in- verted pendulum is standing directly above the vertically oscillating pivot. Inserting these values into Equation 2-56, 6(1) = 0.1 cos (0.548)t and (2-57) d6/dt = —0.1(0.548) sin (0.548)t72 GENERAL TECHNIQUES FOR NONLINEAR PROBLEMS characterize the first solution segment. When t=n7/2 6 = +0.0652 d6/dt = —0.0415 If these values are now used as the initial conditions for the second solution segment, 6(4) = 0.0652 cos (0.6321) — 22415 sin (0.632/2) 0.632) or O(¢) = 0.0652 cosh (0.6324) — 2-415 sinh (0.6321) 0.632 (2-58) and dO/dt = 0.0652(0.632) sinh (0.6321) — 0.0415 cosh (0.6324) See. Pendulum Deflection, 6 (radians) yPrte Figure 2-8. Response of the inverted pendulum described by the Meissner equation; a = —0.05, B = +0.35. characterize the second solution segment. These equations evaluated at t = 3n/2 provide the initial conditions for the third segment, and so forth. The solution is illustrated in Figure 2-8. It may be observed that the inverted pendulum does not diverge, as might be expected, from its “unstable” equilibrum position.* This is a very interesting phenomenon to observe. * Conversely, a simple pendulum with a vertically oscillating support can be “un- stable,” that is, execute diverging oscillations about its vertical (bob down) equilibrium position.EVALUATION OF METHODS 73 The concept of nonlinear or time-varying-parameter systems which are actually or appproximately piecewise linear is a very convenient one. It will be used over and over again, not so much to obtain solutions to non- linear differential equations but particularly in conjunction with the phase plane method as a means of visualizing, in approximate form, the main features of the behavior of nonlinear systems. Since, in principle anyway, the solutions to linear equations are known, the system trajectories of a piecewise linear system in the phase plane are composed of trajectory segments which are familiar representations of linear system behavior. 2.6 EVALUATION OF METHODS OF NONLINEAR CONTROL ANALYSIS The bulk of this chapter has been concerned with obtaining solutions to nonlinear differential equations subject to arbitrary forcing functions or initial conditions. In progressing to this point, the reader has been exposed to some of the rudiments of nonlinear theory which can often profitably be employed in the analysis of nonlinear control systems. Little mention has yet been made, however, of the key techniques used in the great majority of problems involving nonlinear controls. These are, of course, the subject of the remaining chapters, but some anticipation is pertinent here to indicate their connection, as problem-solving tools, with the so-called general methods previously discussed. For many engineering purposes a complete solution for the response of a system subjected to arbitrary forcing functions or initial conditions is not essential for adequate understanding of system behavior. For example, in many cases a linearized analysis of system performance will have been made. It may well be sufficient to modify this analysis by making suitable assumptions concerning the nature and extent of certain likely nonlinear- ities, and then to calculate their effect on system performance. A suscep- tibility to oscillation will be of particular interest to control engineers, and the possible amplitude and frequency of the oscillation can be of para- mount importance. Very likely the effect of raising or lowering the gain in one or more control loops will have already been estimated for the linear case. Consequently, if the effect of the nonlinearity can be reduced to these terms, the requisite analysis can be accomplished quite handily. The conditions for oscillation and the speed of response in terms of bandwidth can be estimated. It is the objective of the sinusoidal describing function method of non- linear analysis to reduce the representation of the actual nonlinearity to an equivalent linear gain (and phase angle). The representation of the non- linearity is simplified by assuming that the output of the nonlinear element,74 GENERAL TECHNIQUES FOR NONLINEAR PROBLEMS in response to a sine wave input, can be described in terms of the funda- mental (sine wave) component of the distorted waveform. Precisely because of this, the first approximation fundamental sinusoidal describing function technique does not account for harmonic or subharmonic oscil- lations. If some of the harmonic terms in the distorted waveform are retained, the analysis is more accurate and complete and the existence of harmonic and subharmonic osciliations may then be predicted. For these purposes a more general periodic input describing function is used. Since a nearly sinusoidal periodic wave having zero mean value is assumed to exist somewhere in the system, the possibility of more than one equilibrium point cannot be a direct result of a single analysis. Further- more, the result of a sinusoidal describing function analysis is a represen- tation of the system in the frequency domain. The generally satisfactory correlation between the time and frequency domains in linear analysis is much less precise in connection with nonlinear analysis. On the other hand, most servomechanism systems are analogous to a low-pass filter, and the higher frequency components of a distorted waveform are heavily attenuated during their trip around the control loop. For this reason, which is apt to be even truer in higher order systems, the sinusoidal de- scribing function technique is often impressively successful. The close analogy between describing function analysis and linear analysis especially recommends it. The synthesis of nonlinear systems, particularly the adjustment of gain and compensation of the phase characteristics, can be carried out with the sinusoidal describing function technique in much the same way as is done in connection with linear systems. In summary, the technique is most useful where the effect of the nonlinearity is small and the system is of relatively high order, and where a knowledge of the response to sine wave inputs has some significance. ‘Viewed in a general way, the describing function method is not restricted to oscillatory phenomena. Itis possible to define describing functions of non- linear elements for any kind of input. Of course, the describing functions of a particular nonlinearity will be different for each type of input considered. Probably the most valuable ones, other than the sinusoidal type, are the describing functions for statistical inputs. These allow the analyst to extend, within limits, the linear analysis and synthesis methods based upon statistical inputs to the similar treatment of nonlinear problems. Other types of describing functions are based upon step functions, which extend the indicial (step function) and weighting function response concepts, to some degree, into the nonlinear domain. The phase space representation is a most powerful tool in nonlinear analysis. Most engineers, however, have difficulty in interpreting or evenEVALUATION OF METHODS 75 imagining phase diagrams in more than two dimensions. Diagrams in the phase plane may represent the behavior of first- or second-order systems. The system, therefore, must be relatively simple in order to be practically amenable to this type of analysis. It often happens, however, that even when the system is of higher order a second-order mathematical model can be set up which is capable of representing the main features of system per- formance. A deadtime is useful as a representation of higher order lags in switching systems. In the use of the phase plane method there are no restrictions on the nature or extent of the nonlinearity. Initial displacements and velocities in any combination and, in some cases, steps and ramps can be used to excite the system. The phase plane diagrams can be interpreted directly in terms of the time behavior of the system. Finally, limit cycles and multiple equilibrium points are evident. On the other hand, the phase plane dia- grams do not readily indicate the steps that need to be taken to correct system performance, and synthesis can only be accomplished by cut and try procedures. It may thus be appreciated that the describing function technique and the phase plane method are uniquely complementary. One is useful where the other is not. It is necessary to master both. Table 2-3 sets forth the several methods of analyzing nonlinear control systems, and presents in concise form an evaluation of the area of practical application and difficulty of each method. None of the methods is as easy to apply as the corresponding tech- niques of linear analysis, and the easier ones are limited in their area of application. In control engineering practice, nonlinear problems are often solved by machine methods. Such a procedure strongly recommends itself for ob- taining quantitative results. Computing machines of either the analog or digital type can greatly extend the capability of design engineers. In fact, the machines can usually produce solutions to linear or nonlinear equations much faster than the operator or analyst can assimilate the results. It is for precisely this reason that the study of the theory of linear and nonlinear control systems is so important. Even if the analyst lets the machine relieve him of the burden of calculating solutions—theory, like experience, pro- vides guides for interpreting the results. The machines, dumb and robot- like, answer questions. The operator-analyst must pose the questions intelligently. Very often the direction of the line of questioning has to be set by a preliminary paper and pencil analysis. In any event, some know- ledge of the expected answer is required in order to be sure that that part of the question which is represented by the machine setup has been posed as intended.76 GENERAL TECHNIQUES FOR NONLINEAR PROBLEMS Table 2-3 Methods of Nonlinear Control System Analysis Method Application Difficulty Liapounoff stability criterion At an equilibrium point where nonlinearity is continuous and has continuous derivatives Relatively simple to apply Direct solutions Limited number of equations May be very difficult to cast practical problem into proper form Graphical solutions Numerical solutions Any transient response Any transient response Grows rapidly with sys- tem complexity and time interval of interest. Should be used as a last resort Periodic input describing function Prediction of periodic phenomena Represents extensions of familiar techniques Gaussian, random input describing Approximate perfor- mance with statistical Represents extensions of familiar techniques function inputs. Only “small” nonlinearities are permissible Phase plane Transient response of No undue difficulty nonlinear first- and second-order systems3 INTRODUCTION TO QUASI LINEARIZATION AND THE DESCRIBING FUNCTION TECHNIQUE A fundamental problem in dynamic analysis, already mentioned in Chapter 1, is the mathematical characterization of the cause and effect relationships for general system elements. The direct way to specify such information would be to determine the element responses resulting from a large variety of inputs, and then to catalog these results as input-response pairs. Except for simple elements acting on a small number of inputs, this direct procedure is extremely unwieldy. In order to achieve a simpler specification of the system element, the basic question “What are the effects due to various causes?” might be changed to “What is the operation of the system element in modifying a given cause into an effect?” To answer the dynamic analysis question expressed in this latter operational form, the analyst desires a “mathematical model” which responds to an input in a fashion which closely approximates the response of the actual physical element. For systems comprising only elements which behave in a manner de- scribable by linear constant-coefficient differential equations, the second question is answered simply by specifying the system’s weighting or trans- fer function. If the system were represented by the block diagram of Figure 3-1, where the weighting function, w(t), is the time response of the system when an impulse function is applied at zero time, then the relation- ship between the response and the input would be given by the so-called Superposition or convolution integral: yt) -|" w(r)a(t — 7) dr =|" w(t — 7)x(r) dr G-1) 7778 INTRODUCTION TO QUASI! LINEARIZATION Since the response, y(t), to any input, 2(7), can be found from Equation 3-1 if w(x) is known, no gigantic tabulation of input-response pairs is required to describe the system’s operation. The analyst need know only the weighting function, w(t). Because algebraic operations are easier to use than integrals, such as in Equation 3-1, it is usually preferable to work with transforms of the input, system Characterized by: Input ‘i 7 Output Characterized by: welchiine finetion Characterized by: |_Characterized by: 1. x(t) as a time W(s) or Wl joo) 1. y(t) as a time function transfer function function 2. X(s) as the Laplace 2. ¥(s) as the Laplace iransform of x(t) transform of y(t) 3. X(ja) as the Fourier 3. YG) as the Fourier transform of x(t) transform of y(t) Figure 3-1. Linear system representation. response, and weighting functions rather than with the time functions. If the Fourier transform is used, the transformation of Equation 3-1 becomes © Y(jo) = i y(tye7#* dt = Wjw)X (ja) (3-2) where Y(jw), X(ja), and W(ja) are the Fourier transforms of y(t), «(¢), and w(t) respectively. When the conventional unilateral Laplace transform is used, the convolution integral of Equation 3-1 is first modified to: yD = [wae —)dr, t>0 (3-3) 0 This form of the convolution integral can then be Laplace-transformed to give i ¥(s) =| y(de~" dt = W(s)X(s) (3-4) 0 where ¥(s), X(s), and W(s) are the Laplace transforms of y(#), x(r), and w(t). The Fourier transfer function W(jo) is essentially the same thing as the transfer function W(s), with s replaced by jw. The introduction of the weighting function, or its transform the transfer function, allows a linear system to be described as an operational entity which is independent of both the input and the response. In a nonlinear system, on the other hand, some of the system “parameters” depend upon the values of the dependent variables which define the system’s response. In this case, the system cannot be characterized as a separate entity, the convolution integral is not valid, and the behavior of the system is a functionTHE DESCRIBING FUNCTION TECHNIQUE 79 of the particular inputs and initial conditions. In a general sense, this situa- tion puts the analyst right back at the beginning—forcing him to define system operation by means of input-response pairs. This contrast in the behavior of linear and nonlinear systems has been emphasized in previous chapters by both word and example. Nevertheless, some of the examples shown previously, and a great many other nonlinear systems of interest, have specific input-response pairs which appear to be similar to input- response pairs for linear systems. This similarity leads to the notion that the performance of some nonlinear elements, for certain specific inputs, could be divided into two parts: (1) the response of a linear element driven by the particular input, and (2) an additional quantity called the remnant. From this general idea, there is evolved the concept of quasi linearization, which emphasizes the similarities, rather than the differences, between linear and nonlinear systems. A quasi-linear system is one in which the relationships between pertinent measures of system input and output signals are linear in spite of the existence of nonlinear elements. A quasi-linear system is an exact representation of the nonlinear system for specific inputs. A particular quasi-linear system is found from the actual nonlinear input-system combination by replacing the nonlinear elements with “equivalent” linear elements characterized mathematically by describing functions and remnants. Each equivalent linear element is derived from consideration of the response of the corresponding nonlinear element toa specific input. The new linear system has the same response to the input in question as the original nonlinear system. The describing function concept becomes most useful when it can be generalized for a whole category or class of inputs, when it can be shown that the describing function of a given nonlinear clement is unique, and when it can be shown or at least believed that the effect of the remnant on the input to the nonlinear element is negligible. Under these circumstances it is possible to develop a practical catalog of nonlinear element, input-response pairs in terms of describing functions, and the describing functions act as linear operators. The per- formance of the true or an approximate quasi-linear system may then be discovered by means of the powerful methods of linear analysis. The replacement of a nonlinear element by a quasi-linear equivalent tailored to a specific input is not restricted to any particular type of input. Indeed, in principle at least, the partition of an output response into two components, one linearly connected with the input and the other a remnant, can be accomplished for almost any combination of nonlinear elements and inputs. In control systems, however, interest is centered upon the system as a whole, and not on its constituent elements. Consequently, the fact that a quasi-linear equivalent can be found for a nonlinear element is80 INTRODUCTION TO QUASI! LINEARIZATION of little value unless this knowledge can be converted into a precise, or approximate, quasi-linear system model. The difficulties encountered in this process, of course, stem from the presence of feedback loops within the system. The actual input signal to a particular nonlinear element is required in order to construct the quasi-linear representation for the ele- ment. Because of the feedbacks, the input to the nonlinear element can only be found by solving for the system response. This was the problem in the first place, so little would appear to be gained by knowing the quasi- linear description of a given nonlinear element excited by a particular input signal. In a few cases, however, the signals existing within the system at the inputs to the nonlinearities are exactly the ones for which quasi-linear element representations are easily found. In these circumstances the quasi- linear representation for the nonlinear elements can be substituted for the actual elements to obtain an exact linear representation of the nonlinear input-system combination. In many other cases the signals at the inputs to the nonlinearities are quite similar to the ones for which describing func- tions and remnants are known or can be found. Then the quasi-linear representations of the nonlinear clements are substituted for the actual nonlinearities in order to obtain a close first approximation to the quasi- linear system. Corresponding to the three main categories of test input functions which are employed in linear analysis, there are three main types of describing functions: (1) transient (actually step) input describing functions, (2) periodic (usually sinusoidal) input describing functions, and (3) stationary random input describing functions. The transient input describing functions are of more conceptual than practical interest. They cannot be developed in as general and as satisfactory a form as the others. Sinusoidal input describing functions have been quite thoroughly explored for many, but by no means all, interesting nonlinearities. They have a surpassing import- ance in the analysis of nonlinear control systems because they are par- ticularly amenable to the determination of stability. It has been repeatedly shown that their use produces good results with a modicum of effort. Furthermore, the use of sinusoidal input describing functions permits the extension to nonlinear control systems of the well-known harmonic or frequency response method of designing equalizing or compensating net- works. This is a tremendous and almost unique advantage. Finally, the stationary random input describing functions can be developed in a fairly general way if the input has a Gaussian amplitude distribution and the effect of the nonlinear element on the performance of the system is not too large. The real significance of random input describing function analysis where these conditions are not met is very little understood. When theseTHE DESCRIBING FUNCTION TECHNIQUE 81 conditions are met, however, and the describing functions have been formulated analytically, or measured experimentally, the Gaussian, ran- dom input describing function may be employed to predict the response and accuracy of the system under dynamic conditions. 7 \xo(t)| x y) x aad La Figure 3-2. Block diagram of a simple feedback system containing a limiter. x1(t) +> €lt) In order to illustrate these remarks on quasi linearization and the use of describing functions it is helpful to consider a very simple example.* A closed loop system which comprises a unit, or normalized, limiter and an integrator in the forward path is shown in Figure 3-2. n— Kt €(t) e(t) Figure 3-3. The limiter with its actual input and output. (Chen, op. cit.) If the input, x,(1), is a step function of magnitude, applied at time t = 0. where 7 is greater than unity, then the limiter output immediately following t = 0 will be unity, and the output, y(t), will be equal to Kt until a time, ¢, equal to(n — 1)/K. Therefore between the times given by 0 < t < (n — 1)/K, the actual input to the nonlinear element is ¢(t) = x,(¢) — y(t) = n — Kt. This can be considered to be the proper specific input signal to use in the development of a describing function relating the input and output of the nonlinearity. The specific input and actual output are shown in Figure 3-3. Now the limiter can be replaced by an equivalent linear operator, the describing function, and an additional quantity, the remnant. This is shown in Figure 3-4, where the linearized transfer characteristic of the limiter (its * K. Chen, “Quasi-Linearization Techniques for Transient Study of Nonlinear Feed- back Control Systems,” Trans. AIEE, Pt. Il, vol. 75 (1956), pp. 354-363.82 INTRODUCTION TO QUASI LINEARIZATION a-} Remnant r(t) Figure 3-4. Equivalent limiter consisting of describing function, 1, and remnant, r(t). (Chen, op. cit.) describing function) is its gain in the unlimited region, unity; and the remnant signal, r(), accounts for the discrepancy when the limiter operates in the saturated region. The remnant signal, r(¢), can be put into a form which may be derived directly from the original input. The linear operator which produces r(¢) from the input, x,(¢), will be called G,(s). Itis developed as a transfer function as follows 1 r=(—-1)-Kt, O K Now a()=n, 1>0 and er] Gs) = == . 2[a()] Sea te) n = (C24) 2 -eo[-(24)]THE DESCRIBING FUNCTION TECHNIQUE 83 When the nonlinear element (limiter) in the system of Figure 3-2 is replaced by its describing function and remnant, the quasi-linéar limiter represen- tation of Figure 3-5 is the result. Then the system block diagram of Figure 3-6 may be rearranged according to the rules of block diagram algebra so as to appear in the form shown in Figure 3-7. 40 4D ed) x2(t) y(t) n ) Ot > Figure 3-5. Equivalent limiter with remnant derived from a linear operation on the input. (Chen, op. cit.) It is now seen that the nonlinear system of Figure 3-2 may be represented by the quasi-linear systems of Figures 3-6 and 3-7. Either one of these completely linear representations is an exact representation for the whole class of step function inputs to the system. It may be noticed that the 20) Figure 3-6. The quasi-linear system. (Chen, op. cit.) x(t) + K ) 1-6, 2° Ys) __K RO Tek U- Gr(6)] Figure 3-7. Equivalent block diagram of the quasi-linear system. (Chen, op. cit.) effects of the nonlinearity, in this case, are represented by inserting a linear, albeit transcendental, operator into the system outside the control loop. It may further be observed that the characteristics of this linear transfer func- tion depend on both the input, n, and the characteristics of the system, in this84 INTRODUCTION TO QUASI LINEARIZATION case K. In essence, the effect of nonlinearity has been represented by a closed loop linear system operating into a linear transfer characteristic which depends upon the nonlinearity, the input, and system parameters. This kind of representation will always be possible with systems having single, simple, piecewise linear elements within the loop. Sometimes this kind of representation may be applied when more than one nonlinear element is present, although the computational complexity in applying the method increases rapidly with the number of nonlinear elements. In connection with transient input describing functions and remnants, the representation is not unique for a given nonlinearity. Everything depends on the position of the nonlinearity within the loop. If the positions of the limiter and integrator wete interchanged in this example the appropriate quasi-linear system would be quite different. The basically difficult problem in the transient input describing function method, however, is involved in finding the transfer function, G,(s), which relates the output of the linearized closed loop system to the actual approxi- mate output, that is, to find the mathematical model of a linear element which gives an effect “equivalent” to the effect of the nonlinearity. This may become extremely onerous. While the simple analysis given above is exact, in more elaborate situations it is thoroughly impractical to account for more than gross effects, such as dominant modes; and when the non- linearity is a complex one, such as hysteresis or backlash, even though it can still be represented by piecewise linear characteristics, the task of finding a justifiable transient input quasi linearization is much more severe. It is for these reasons that the use of the transient input describing function is confined here merely to the illustration of the important concept of quasi linearization. The same concept of quasi linearization turns out to be much more powerful when it is employed with periodic input or stationary random input describing functions. This is because unique describing functions and remnants, developed for given nonlinearities considered as isolated elements, are employed directly to construct first approximation quasi- linear system models. This may seem strange by comparison with the transient input describing function technique where the performance of the whole system had to be taken into account, but in a very real sense it is only in such circumstances that the concept of quasi linearization has practical consequences for the design of feedback control systems. The isolated element sinusoidal input describing function is derived from consideration of the harmonic response of a nonlinearity to a sinusoidal input at various frequencies and amplitudes. In a constant-coefficient linear system, or element, excited by a sinusoid, a portion of the output will, of course, be a sinusoid of the same frequency, although the amplitudeTHE DESCRIBING FUNCTION TECHNIQUE 85 and phase angle of the sinusoidal response may differ from the amplitude and phase angle of the input. Now suppose that a sine wave is applied to the input of a nonlinear element having a single input and a single output. The output will very likely be a nonsinusoidal periodic wave with the same period as the input wave. If the output waveform is analyzed in terms of its Fourier components, the fundamental component will bear a relationship to the input sine wave which can be described in amplitude ratio and phase angle terms. A sinusoidal input describing function is defined as the complex ratio of the fundamental component of the output to the input. (This is the same thing as is done in connection with the definition of a Fourier transfer function for a linear system.) The remnant, in this case, consists of all the higher harmonics, and the output is then the sum of the describing function times the input, plus the remnant. Both the describing function and the remnant depend upon the input amplitude in such a fashion that the output of the quasi-linear element is identical to the output of the actual isolated nonlinear element. As a specific example of a quasi-linear replacement for a nonlinear element subjected to a sinusoidal input, consider again the limiter of the system of Figure 3-1 whose performance with a sine wave input was also discussed previously in connection with Figure 1-13. If the gain of the linear portion of the limiter characteristic is taken to be unity, the response to a sinusoidal input, «(#) = A sin wt, similarly to the result of Equation 1-28, will be: @ x{t)=b,sinot+ > 5b, sinnot (3-7) where whee by =1andb, = 2A fey (a a af } t= 74 [snr (4) + (Gt F] | 4A a) cos nB a | neta let Ji- Baa] naan naseta) 1 4 sin n. A>a | B=sin? (¢) | A The nonlinear element and its quasi-linear equivalent are shown in Figure 3-8. In this example the “phase angle” of the sinusoidal input describing function is zero, and the describing function is a pure gain which varies with input amplitude alone. This will always be true when the output is dependent only upon the instantaneous value of the input. (This type of nonlinearity is defined as a simple nonlinearity.) In general, of course, a A amen" | lim 5— if ein ondt at} (4-13) m0 n=—0 2-0 2T I-74 The value of the integral term in braces in Equation 4-13 can be evaluated by considering two cases, w,, # w, and w,, = w,. For the first of these, © #~ Wp, the term in braces is: 1 fT, 1 J eil@m-ondt 72 lim — eon ont dt = lim — ("| =0 (414) 7 . P0 2T 20 2T Lio, — o,)J-2 When w,, = o,, the same term becomes simply eae a 1 [? lim a foment dt = lim a‘ dt=1 (4-15) T+02T/-7 T202TJ-7 Consequently, the cross-correlation function of equation 4-13 becomes: Raft) =X on B ye" (4-16) na 0 It will be noted that Equation 4-16 shows only those frequencies which98 SINUSOIDAL DESCRIBING FUNCTIONS exist in both «() and y(¢), and that the amplitude of each term in the series depends upon both the amplitude and phase of the individual frequency components in x(t) and y(t). The autocorrelation for x(t) is obtained by inserting the expression of Equation 4-11 into the defining relation, Equation 4-6: 7 Ret) = tim [ s ate] S aqeront| dt D0 2T I-27 Ln=-0 m=— 0 SES cata em tim 4 [eton- eo ar} (417) Sto mito” ™ T-+02TJ-7 The term in braces can be treated in the same fashion as in Equations 4-14 and 4-15, so the autocorrelation function for the periodic time function x(t) becomes: Ret) =X |ttql? 7" (4-18) nite Examination of Equation 4-18 reveals that all of the frequency com- ponents in 2(¢) appear in the autocorrelation function. The coefficients, |a,|®, for each frequency component in the autocorrelation function are real, and have the same nature as a power. In this way the autocorrelation indicates the “power” at each frequency in the time function 2(1). Now, in order to specialize the discussion to the sinusoidal input describ- ing function for an open loop nonlinear element, consider the case when x(t) = A cos wt, and y(t) is the output. The output of the nonlinear ele- ment will then be given by a Fourier series, similar to Equation 4-9 or 4-10: a VOD) = XC, COs (Ot — Py) no © = > eye (4419) nico where Cy = S (cos Yn —J SiN Y,)s Con = & (cos y, + j sin y,) Assuming that the mean value of the output, y(¢), is zero, that is, cy = Cy = 0, the cross-correlation function between input and output will be: a R(t) = £ 2 ener = A fee + ee] = AG cos Wa (gir 4 gm tory som (eer om| 7 - cos (wr — yi) (4-20)CORRELATION CONCEPTS 99 Only the fundamental of the output, characterized byc, or C, and y,, appears in the expression for the cross correlation, R,,(r). It alone is linearly corre- lated with the input. All the higher harmonics in the output are uncor- related with the input, since the cross correlations between them and the input are zero. In essence, the multiplication and time-averaging processés involved in the cross correlation have “averaged” out the effects of those harmonics in the output which do not appear in the input signal. Linear coherence, or linear correlation, implies that a fixed phase angle exists between two signals at a given frequency. If the element under consideration were linear, with a transfer function W(s), the output of the element for a sinusoidal input x(t) = A cos wt would be: Ut) = A |W(ja)| cos [or — AWGo)] (4-21) The cross-correlation function in these circumstances is: Ae : Ralt) = > |W(ja)| cos [or — AW(o)] (4-22) Comparison of Equations 4-20 and 4-22 shows the similarity between the sinusoidal describing function, which would appear in operator form as (C,/A)e~*", and the transfer function of a linear system. The autocorrelation function, R,,(7), When x(t) = A sin wt, will become: A R,.(7) = ie cos wr (4-23) The cross-correlation function for this case can then, by analogy to Equation 4-21, be interpreted as the result of thedescribing function operat- ing upon the input autocorrelation function. In general, for inputs that are not necessarily single sinusoids, this same interpretation takes a form which is analogous to the convolution integral, given previously by Equation 3-1. Thus, if w(t) is the weighting function of a constant-coeffi- cient linear element, Ral) = ij © Rar — 1) du (4-24) All of the concepts relating the describing function to notions of linear correlation can also be depicted in frequency terms by using Fourier trans- forms of the autocorrelation and cross-correlation functions. For this pur- pose, so-called power- and cross-spectral densities can be defined as twice the Fourier transforms of the autocorrelation and cross-correlation func- tions, respectively. Thus, the power-spectral density is ©,,{j0) af" Rite!" dr (4-25)100 SINUSOIDAL DESCRIBING FUNCTIONS which, since R,,(r) is an even function, becomes: ©,,(0) = 4f "Ra cos wr dr (4-26) Similarly, the cross-spectral density is: 0,,(jo) = 2 | Rater" dr (4-27) As an example, consider again that the input and output of a nonlinear element are periodic, and are characterized mathematically by the Fourier series of Equations 4-11 and 4-12. Then, using the resulting autocorre- lation function for the input signal given by Equation 4-18, the power- spectral density, ®,,(w), of the input will be: o(o)=2f" $ | © n= @ [2 enor de ~ ~ =2> leat" eumaedn (4-28) The evaluation of the integral in Equation 4-28 requires the definition of a special function, the delta or impulse function, 6(v). This function (or “measure,” to be precise) has the properties: a nae eo 4-29 M0 — Po) =| neo, v= oy ie and / * av — ») dv =1 (4-30) (03 Uy, > Ug > Up OF Ug > Vy > Vy i | ie il "Alo)a(v — v) de 7 Af(vo)s vy > vq and either : im fe (431) Fi a = % Lf (00): Dy > Up > Va From Equation 4-31, the Fourier transform of 6(v — v9) will be: ® 8(v — rye #0" dv = e7 #0% (4-32) The inverse Fourier transform of e~* is: ALP ensereg+i00 day = L |” o*2000-%0) dy = 8(v =v) (4-33) In Jo In Jo The integral left unevaluated in the expression for the power-spectral density, Equation 4-28, can now be expressed as: | a ae 2nd(w — @,) (4-34)CORRELATION CONCEPTS 101 The power-spectral density for the periodic input signal then becomes: © ©,(0) = 47 > |a,|? o — o,) (4-35) n=a@ The nature of the power-spectral density for periodic functions is seen to be akin to that of a comb, with tooth areas proportional to the power at each frequency, and with teeth knocked out in the regions where no signal power is present. While it is not theoretically precise, it is common practice to make the height of each “tooth,” or “spectral line,” proportional to the value, |2,|, of the “power.” An illustrative example of the power-spectral density, using a square wave as the periodic time function, is presented in Figure 4-1. In a similar manner, the cross-spectral density between the nonlinear element’s periodic input and output, «(¢) and y(z), can be found by Fourier transforming the cross-correlation function of Equation 4-16. Accordingly nea ©,,(jo) = 2 | YT Byaqhernte 2" dr =2 S aant[” etoer de os 7 =40 Y Bit, *(o — o,) (4-36) we In Equation 4-36, the spectral line character of the cross-spectral density for periodic signals is indicated by the delta functions. Also, the cross spectra are seen to preserve phase characteristics in the same way as the cross-correlation function. As an illustrative example Figure 4-2 presents the cross-spectral density function obtained when «(¢) is a square wave and y(t) is a sawtooth wave. Power-spectral and cross-spectral densities for periodic functions may alternatively be defined directly in terms of Fourier transforms of the sig- nals involved. Thus, ®,,(0) = { Jim # [XGo)X(— jo)]|rate = 0) (4-37) ©,,(jo) = ( Jim aX jo) Yo) Ho — Oy) (4-38) where X(j) and Y(jew) are Fourier transforms defined in the following special way: i X(jo) =| a(the ** dt, a(t) = 0 except for -T (en) [sj - en + v| a (b) Figure 4-1. Time history and power-spectral density for a square wave. (a) Time history. (b) Power-spectral density. Specializing the above general periodic signal results to the case where the input signal to an element is a(t) = A cos wot and the output is U0) = 3 C, 605 (not — 4) ~ > c,eineot the power- and cross-spectral densities of Equations 4-35 and 4-36 become: (0) = rA*[3(@ + 0) + H(o — 0)] © (jo) = wAC Le’ 6(@ + co) +e 6(@ — a )] (4-39)CORRELATION CONCEPTS 103 (8) cd Ene 5] eet "| or [oe 0 = wot x(0) = “4 (cos@ - 4A ZS (a yyn cos ante oO" ent sin 20 4 sin30_ 2a aa =-2 Sep at (0) +90-- day ( 7 (degrees) 90 Figure 4-2. The cross-spectral density between a square and a sawtooth wave. (a) Square wave. (b) Sawtooth wave. {c) Cross-spectral density.104 SINUSOIDAL DESCRIBING FUNCTIONS It will be noted that the cross-spectral density has a real and an imaginary part, while the power-spectral density is totally real. The spectra exist only at the frequencies @ = +g. This fact, in connection with the cross-spectral density, again reflects the concept of linear coherence, that is, only that portion of the output which is linearly correlated with the input (at the same frequency) will emerge in the cross-spectral density. The most significant connection between the various correlation notions and the sinusoidal input describing function is now revealed by taking the ratio of the cross- and power-spectral densities for real frequencies: Dif jo) _ TAC” (w — 00) _ Cr g- 501 (4-40) ®,,(@) 7A? S(w — Wo) A This most unusual “ratio” which is presumed to “exist” only when the delta functions are not zero, is seen to be the sinusoidal input describing function. This relationship is much more general than is noted here. For instance, it is equally valid for each frequency of the more general situation for periodic inputs and outputs. In a linear system it essentially provides the transform equivalent of Equation 4-24, that is, if W(j) is the Fourier transform of the weighting function (the Fourier transfer function), then: ®,,(j0) = W(jo) Pp) (4-41) While the operation shown in Equation 4-40 is dubious mathematically, that of Equation 4-41 is quite legitimate. Besides providing the most general basis for the sinusoidal input describing function, cross-corre- lation or cross-spectral concepts can also be used for the experimental measurement of describing functions. This possibility is discussed further in Chapter 6. 4.2 SINUSOIDAL DESCRIBING FUNCTIONS OF SIMPLE NONLINEARITIES Sinusoidal input describing functions of simple nonlinearities are not frequency-sensitive. It is, therefore, only necessary to analyze the input- output relationship of the nonlinear element at a single frequency. In general, the outputs may be represented by a Fourier series written in terms of the magnitude of the nonlinearity and the amplitude of the input sine wave. Figure 4-3 shows the input-output relationship for several commonly interesting simple nonlinearities, together with an illustration of the output wave form for one input amplitude. When the output-input relationship is given graphically, the output waveform can be obtained very simply by projecting the amplitudes of anSIMPLE NONLINEARITIES 105 assumed input sine wave at successive instants of time up to the nonlinear characteristic and across to the derived output wave. Alternatively, since the projection of an input wave on a line with a slope which is the ratio of output to input scale factors is the same whether it is made from below or from the left, the construction can be made so as to maintain the sense of the transfer characteristic. The latter construction is illustrated in Figure 4-3, Once the shape of the output waveform is known, any one of several analytical, numerical, or graphical methods can be applied to evaluate the coefficients of the representative Fourier series.* For symmetrical simple nonlinearities, subjected to sine form inputs, the series for the output need contain only odd harmonic sine terms, since all the cosine and the even harmonic sine terms will be zero. Consequently, the Fourier series for the output of such a nonlinear element, for an input defined as A sin wt, is yt) = b, sin wt + by sin 3ot + bs sin Swt +--- (4-42) and the coefficients are given by 4 (PF Deg-4 = pe y(t) sin (2n — Let dt (4-43) Of course, the coefficients of the Fourier series of the output wave must be evaluated for several amplitudes of the input sine wave. If the Fourier coefficients should be negative, it indicates a “phase angle” of 180 degrees with respect to the fundamental component. “In Figures 4-4 through 4-9} there are presented the “amplitude ratios” b,/A of the fundamental, and b,/b, and b;/, of the third and fifth harmon- ics, as well as the “phase angles”, for the nonlinearly distorted waveforms of Figure 4-3. The amplitude ratio and phase angle of the fundamental represent the sinusoidal describing function of the nonlinearity in question. The amplitude ratio is scaled in decibels so as to be compatible for later use with Bode diagrams. Amplitudes of the third and fifth harmonics of the output, relative to the fundamental amplitude, b,, are also given in the figures. These remnant data are necessary in order to measure the accuracy of the quasi-linearization technique in a particular analysis. In a first approximation analysis of a closed loop quasi-linear system the effect of the harmonics when propagated around the control loop must be small * Fr. A. Willers, Practical Analysis, Dover Publications, New York, 1948. Reference Data for Radio Engineers, 4th ed., Federal Telephone and Radio Corpora- tion, New York, 1956. + Figures 4-5, 4-6, 4-7, and 4-18 are taken from Methods of Analysis and Synthesis of Piloted Aircraft Flight Control Systems, BuAer Report AE-61-41, Northrop Aircraft, Inc. U.S. Navy Bureau of Aeronautics, Washington, D.C., 1952. The original computations were performed by S. J. Press.