ht. 1. NaLIkro Mb. Vol. 1. pp. 311-324. lbrpmon Ram 1969.

Riatd in Gmat Bdtain

PHASE PLANE ANALYSIS OF NON-LINEAR SYSTEMS

USING WEIGHTED LINEARIZATION

M. L. ADELBWG and H. H. DENMAN

Departmentof Physics,Wayne State University, Detroit, Michigan

Ah&ad--An analysis based on weighted linear approximations is developed for certain non-linear ordinary
differential equations, and is applied to the description of singular points and trajectories in the phase plane.
This technique is compared with the analysis of Poincark based on Taylor series linearization where the latter
is applicable. Numerical results are obtained for a quadratic friction problem, and compared with the first-order
Krylov-Bogoliubov-Mitropolsky result. The new technique is also successfully applied to problems of Lienard,
Keil and Mabytrini, and to a special case of a non-isolated singular point, where Taylor series linearkations are
inapplicable or yield qualitatively incorrect results

1.INTRODUCTION

WE shall consider systems governed by the real differential equations

1 = X(x, Y), (l.la)

3 = YkY), (l.lb)
where the origin 0 = (0,O) is a singular point (s.p.). Unless otherwise specified, X and Y
are assumed C’ in some region about 0 and the s.p. 0 is isolated. Then (1.1) may be written

i = ClX + c,y + X,&,Y), (l.2a)

3= c3x + c,Y + yNL(x, Yh (1.2b)
where cI = X,(0,0), etc., and X NL,Y,, are the non-linear remainders in this decomposition.
Taylor series linearization ( Ti) of (1.1) simply discards X, and Y,, in (1.2).
Let the determinant and half-trace of the matrix in (1.2) be denoted, respectively, by

Cl c2
J=
I I
c3 c4
, I = (cl + c,)/Z.

Then, for J # 0, the solutions of the linear parts in (1.2), the T1 approximation,
(1.3)

fall into
four classes, with the usual designations for the s.p. :
1. J<O Saddle (1.4a)
2. 0 < J < l2 Node (1.4b)
3. 0 < I2 < J Focus (l.4c)
4.O<J, I=0 Center (l&l)
Tl correctly indicates the stability of the solutions to (1.1) in the first three of these classes,
although the approach to 0 when J = Z2 may be incorrectly described [l]. In the fourth
class, the stability is not always predicted correctly [l] ; the trajectory in the phase plane
311
312 M. L. ADQSEW and H. H. DENMAN

(x, y) is, in general, a focus, while TI predicts a center. Further, 7” is inapplicable when
J = 0, as it indicates a line (or plane) of s.p., while the complete system (1.1)has an isolated
s.p, at 0.
T1is au approx~ation at a point ; it ignores changes from linearity in X and Y at points
far from 0, and even near 0 may be quite inaccurate in some cases For this reason, we
shall introduce weighted linearizations, U,(J), which approximate X and Y better over a
region containing 0. We shah show that whenever Tl gives the correct description of the
s.p., U,(2) does also, but that the latter technique gives correct results for certain classes
of problems where TI fails or is not appli~b~. A number ofspecific examples are examined.

2. WmGHTED LUWARUATION

We wish to approximate X and Y in (1.1) linearly over a region containing the s.p. 0
(which may or may not be isolated). Let the weight function w(z,,I) be a positive (except,
perhaps, on a set of zero measure), even function of z(x or y), such that (a < 0)

$fw(xJ)dz, n=O,f,&... (2.Q

exist. (The limits a, b may be fmite or Semite, but will be taken as fmite in what follows,
since we are primarily interested in bounded motions around the s.p+).The weighted
linearization U,(J) approximates the function X by+
X* = c,(& A,, a) x + c&t,, Ag, n)y, (-2.2)
where

f lXta,&Y’) YWY',
E 4 WlAd’WA
:
- 4 W/A &4,
dXto, - AsY') Y'W, -As<y<O, (2.34

where N(1) = i (x’)%(x’,2) <ix’.Y(x,y) is similarly approximated, with linear coefl’icients

c&4,, A*, 2) and c&As, A,, 4. The At are chosen to include some range of the motion
about the ap. 0. In particular, they are taken (as closely as possrble) to be the turning
points of a cycle (half-cycle) when the trajectory is a center (focus).
The particular choice for the weight function,
w(z,L) = (1 - 22)1-f, -1G.7~1, (2.4)
corresponds to the expansion of X(x, 0) in ultraspherical polynomials, if X(x, 0) is an odd
function of x, etc. If X(x, 0) is even, the linearixation (2.3a, b) will also be even, composed of

t Note that, in general, the approximations (2.2) and (2.3) are composed of straight line segments of different
slope crossing at 0.
 Phaseplanetmafysis of nm-hem sys&wts wing weighted hearimtion 313

two straight line segments of opposite slope. We also note that, for the ultraspherical
polynomial weight function (2.4), the limit A -+ co corresponds to the Tt approximation, as
pointed out by Lanczos [2].
The approximate equations resulting from U,(A) are then

2 = c&%, A,, A)x + c,(& Aa, 2) Y (2Sa)

j = c,(& Aa, A)x + c&4,, A,, 2) Y. (2Sb)
Let J(A,, 2) and I(A, A) be the determinant and half-trace, respectively, of the matrix in
(2.5). It can easily be shown [3] that, as the Al -+ 0,

c&, 2) -. ci, i = 1,23,4. (2.6)

This implies that, in this limit,

f&t 2) --) J, f(Aj, n) --, f. (2.7

Thus, the U,(A) results will agree with those of T, in all cases except, possibly, when J = Z2,
or when I = 0. But these are exactly the cases where Tl may predict the behaviour of the
system incorrectly. In the examples which follow, we examine some of these cases in detail.

3.THE TsCENTER
If, in (ll),

X(x, Y)= --X(x,-Y) (3.la)

Y(x, Y)= w, -Y), (3.lb)
and I = 0, J > 0, then 0 is always a center [l, p. 95J. For example, the set

2 = g(Y) (32a)
3 = -h(x) + k(y), (3.2b)
where g is odd and k is even in y, g’(0) > 0, and h is restoring near the origin (h’(O)> 0),
satisfies the above conditions, so that 0 is a center. Since k(y) has no linear part, Tl gives
the matrix

LyO)
3. (3.3)

which correctly predicts the center.

The systems (3.1) and (3.2) have trajectories symmetric across the x axis (since the trans-
formations y + -y, t -t -t leave the system unaltered); thus we take A3 = A, in U,(A).
Then the U,(A) approximation to g(y) will be odd, and to k(y) will be even, so that the
U,(A) equations obey the above conditions and thus also yield a center.
However, suppose Y(x, y) is not even in y. Let

YCGYf = y,b, Y) + Y&AY), (3.4)

where

Y, = acyky) + w, -y)], r, = W(X,Y) - m, -Y)]. (3.5)

314 M. L. Ammmw and H. H. DENMAN

It can be shown [33 that the s-p. is now a stable (unstable) focus if Y,(x, y) < 0 (>O) for
0 < y < 6. If Y0has no linear part, Ti ignores it and again predicts, incorrectly, a center,
while U,(L) predicts the focus, if Y has the form (3.2b).

4. CONSERVATIVE SY!B’EMS

We define a conservative system as one governed by equations of the form

et=y (4. la)

3 = -f(x), (4.lb)
corresponding to the second-order equation
f +f(x) = 0. (4.2)
A first integral of the system exists,
$2” + V(x) = E, (4.3)

where V(x) = j f(z) dz, and E is a constant, determined, for example, by the initial con-
ditions. The classical turning points are roots, if any exist, of
E - V(x) = 0, (4.4)
and the motion is periodic between an adjacent pair of turning points, bounded but of
infinite period (for a separatrix), or unbounded.
Ti approximates (4.1) by
I=y (4Sa)
$=; -f ‘(0) x, (4Sb)
where~(0) is assumed 20, so that J # 0. (The -f(O) = 0 is included in the discussion
of Keils’ second problem in Sect. 7.) From (4.5), 7’r predicts a saddle (J < 0) or a center
(J > 0). U,(A) gives the matrix

1
0 1
[ -e&%, (4.6)
A,, A) 0
which also predicts a saddle or center. We shall discuss three important cases:
Case 1. “Hardening”f(x). Iff(x) = 0 only at x = OJ’(0) > 0, andf’(x) & 0 for x # 0,
the singular point is a center, and is correctly described by Ti. However, 7’i yields the simple
harmonic oscillator equation, so that the predicted period for every trajectory about 0
has the same value (27r/y(O)*), independent of the amplitude of the motion, which is
incorrect iff(x) has a non-linear part. In the U,(A) result (4.6), c,(A,, A,, A) will be >O
for this case, which indicates the center. But the-period of oscillation will, in general,
depend on A, and A2 (through cj), as it should. (Actually, A, and A, are not independent
parameters if they are taken as the exact turning points, since they then are related by:
V(A,) = V(A,) = E.)
Exanap~e: f(x) = x + 2, n an odd integer > 1.
T1 gives a period of 27rfor all the trajectories. Since the trajectories are symmetric about
the y axis, we take A, = A2 = A in U,(J), and cj = c,(A, A). Explicitly, for w(x,,l) =
(1 - 9) i-i, corresponding to ultraspherical ~lynomial linearization, U,(J) gives
 Phase plane analysis of non-linear systems using weighted linearization 315

(Jr+). . .(;>
CJ(A,A) = 1 + A”_’
(i+!!3(n+qp,.,,. (4*7)

Thus, the period of oscillation (27r/c,(A, A)*) is a fiction of the amplitude A. In fact, it
has been shown [4] that, as A + 0, the period resulting from (4.7) agrees with the exact
period to and including terms of order A’, for 1 = 0, for all oddf(x).
Case 2. “Softening”f(x). Iff’(0) > O,f(x) = 0 only for x = 0 and x = a # 0 (we take
a > 0 without loss of generality), andf(x) monotonic decreasing for x > a, then the point
(a, 0) is also an isolated s.p. In this case, the trajectories arc closed paths about the origin
for small amplitudes. But for the trajectories passing through (u,O), the period becomes
i&mite; these trajectories form a separatrix. Trajectories outside this separatrix are un-
bounded. T1 again gives a center, as does U,(n) for small amplitudes. But for some A, > Q,
@AZ, 4, for x > 0, becomes negative, indicating a change to a saddle point. In this way,
I.J,(J.) indicates, for sufficiently large A,, the existence of another s.p. and unbounded
trajectories.
Emple.f(x) = x(u - x) (1 + x2), a > 0. ?I’ireplaces f(x) by IIX,which yields a center,
with period 2wu-* for all trajectories. Ultraspherical polynomial linearization yields

where cl is used for x 8 0, A+ = A*, etc, and T(x) is the gamma function. The period of
oscillation is a function of amplitude through this expression. For I = -3, and A+ = a,
4 = 0, which indicates an infmite period for x: > 0; for A+ > a, c: < 0, and the period
becomes imaginary. Thus, ultraspherical polynomial linearization indicates a separatrix
for A, = a, if II = - f; (It can be shown that this is a general property for softening systems
andtZ.= -3.)
Case 3. “~ftening-hardening”~(x). Letf’(0) < 0, andf(x) -+ & cc as x -P + CQ.Then
at least two other s.p. exist, with closed trajectories around them. A closed separatrix,
passing through 0, encloses them For amplitudes sulliciently large, cyclical orbits about
0 do exist. Ti indicates only that 0 is a saddle. V,(J) gives &A,, A) < 0 (a saddle) for
small amplitudes, but as Ai, A, + co, c$(A,, A) become >O, indicating a center.
Example. f(x) = - x + x”, n an odd integer >l. Thus, two other isolated singular
points exist at the real roots of x”- ’ - 1 = 0, i.e. at & 1. Sincef(x) is odd, the trajectories
are symmetric about the y axis, so A, = A, = A, and

c3(A,R) = -1 + A”-’ (4.9)

Forany.&> -2,~~ = 0 for some value of A, indicating a separatrix. Again, for R = -4,
c3 = 0 yields A = &-1, so that the singular points away from 0 are given correctly. (However,
quantitative results for the period have been quite poor for rZ= -4 [4].)
316 M. L. J&ELBBRG and H. H. PENMAN

5. QUA.DRATIC FlUC-FiON

As an example of a weak focus, we consider a quadratic friction problem, and compare

our results with the lowest order Krylov-Bogoliubov-Mitropolsky (K&M.) approxima-
tion. The exact equations are
g-=: Y (5Sa)
Y-_ -x - 3lYlY9 (5.lb)
which have the first integral, for y > 0,

b” = 1 - x - (1 + aJe-“*-X,
where y = 0 when x = -a i < 0. The next x-axis crossing, a2 > 0, is given by

(1 - a,)ea2 = (1 + a,)e-“I. (5.2)

T, replaces (5.1) by the simple harmonic oscillator equations, so that 0 is predicted a
center instead of a focus. If w(x, A) = (1 - x2)“-*, the V,(J) approximation to (5.1) is,
for y. > 0,
g=y (5.3a)
3 = - x - 2yy, (5.3b)
where
r = A‘$@ + 2)/2&I-(A + #). (5.4)
From the solution of the linearly-damped oscillator (y2 -=z1) equations (5.3), we have
A, = ai = A,exp(Btan-‘ fl-‘), (5.5a)
A#) = a1 exp (-N-J), (5.5b)

where fi = y/(1 - r”)*, and A* is the maximum velocity for this half-cycle, ai can be found
as a function of y from (5.4) and (5Sa), after which we can determine Az(ai), using (55b).
In the K.B.M. first approximation [S, p. 611, the approximate amplitude a2 where the
trajectory crosses the positive x axis is related to a1 by
a2 = a,/(1 + &)- (5.6)

In Fig. 1, the errors in the ~plit~~ AZ@),a2 as functions of a1 are plotted (several values
of 1 are used). We note that, on the range considered, E.= 6 gives the best agreement with
the exact solution, while I = 0 and K.B.M. yield comparable errors.
For small initial amplitudes a,, A, x al, and y and fi are also small, so that (5.5b) be-
comes, to first order in al,
A2(#QkQST= R/I % 1 - JX r(J + 2)a,/2I-(A + 3). (5.7)
From (5.6), we obtain, in this limit,
a2 = ai0 - k). (5.8)

Thus, when U,(n) is ultraspherical linearization, it agrees with the K.B.M. first approxima-
tion for a2 (and hence with the exact result) for 1= 0, as a1 -, 0.
 Phase plane analysis of non-linear systems wing weighted linearization 317

FIG. 1. A#) - a2 and a2 - a2 vs q.

6. THE LlENARD PROBLEM

It has been shown, by Lienard and others [7], that the second-order differential equation
jt + f(x) i + g(x) = 0 (6.1)
will have a unique stable periodic solution if:

(1) f(x) is continuous and even, andf(0) < 0;

(2) g(x) is continuous, satisfies the Lipshitz condition, and xg(x) > 0, for x # 0; and
(3) F(x) + f cc as x + + co, respectively, where F(x) = Xf(u) du, and F(x) has a single
d
positive zero at x = a, while for x & a, F increases monotonically.
Lety= ZZ+ F(x). Then the differential equation (6.1) can be written
i=y- F(x) K-1

j = -g(x). (6.2b)
Sincef(x) is even, F(x) is odd, and thus 0 is a s.p. Then T1 yields the matrix (assuming g is
differentiable at 0)

(6.3)

so that J = g’(0) 2 0, I = -f(O)/2 > 0. Thus, T1 indicates that the s.p. is an unstable
node or focus in the x,y plane, and does not predict the limit cycle. (The exact solution
will, however, behave like au unstable node or focus very near 0.)
U,(A) linearization yields the matrix
c,(& 2) ckt4,~)
I . (6.4)
C c,(&J) 0
318 M. L. ADELEERG
and H. H. DMbk4N

Here I = c,/2, where

1
ci = - F(A,x) x w(x, 2) dx/A, 1 x2 w(x, 1) dx.
1 d
For A2 sufficiently small, cr > 0, since -F > 0 near 0. But for A2 large, i.e. A, & u,
-F -=c0 over much of the range of integration, so that cl becomes negative. Thus there is
always a value of A, such that ci = 0. Similarly, c,(A,, 2) is always negative and c,(A,, J)
positive, so that J > 0. Thus, for some value of A,, U,(;1) will indicate the s.p. is a center,
with the trajectory passing through (A,,O). These two requirements can be shown to
combine to indicate that there is a stable limit cycle.
However, Ur(13)might indicate more than one limit cycle. But iff(x) crosses the positive
x axis once only, and if w(x, 2) is non-decreasing for x > 0, then it can be shown that the
Vi@) limit cycle is unique.
As an example, we consider the well-known van der Pol equation
R - ,ff(f -x2)2+x=0, (6.5)
which satisfies the above conditions for p> 0.9 If ultraspherical polynomials are used for
the U,(J) linearization, the matrix (6.4) becomes

(6.6)

For A, very small, the s.p. 0 appears to be an unstable node or focus, as for 7;. But, for
A, = ,/&W. + 2)], f = 0 and we obtain a center, with the trajectory passing through
(A,, 0). For R,= 0, the usual first approximation for the amplitude of oscillation for the
van der Pol equation, A = 2, results. For trajectories with smaller x amplitudes, A <
,,/[2(2 + A)], U,(d) predicts a trajectory of increasing amplitude, while if A > ,/[2(2 + J.)],
the trajectories are decreasing in amplitude. Thus, U,(i) predicts a trajectory pattern
similar to that of the correct unique stable limit cycle.

7. KEIL’S PROBLEMS

We now consider isolated s.p.‘s, where J = 0 but not all of the c, vanish so that Ti
incorrectIy predicts a line of s.p.‘s. In the first problem of Keil, I # 0, so that the exact
equations in normal form are
g= -x+X&,y) (7. la)

i = yF&, Yh (7.lb)
It is known [l, pp. 256-2661 that on each side of the y-axis, a single trajectory is tangent
(as t + co) to the x-axis at 0. The equation, x = XNt, defines a curve (denoted by M)
tangent to the y-axis at 0. For y > 0, the following rules apply:
A. If & > 0, a single trajectory is tangent (as t -+ - 03) to the y-axis at 0; all others
in the half-plane remain near 0 for a finite time; i.e. 0 is a saddle.
t p < 0 is equivalent to changing the sign oft in (6.5). Then the limit cyck is unstable, and is still so indicated
bY v,ca
 Phase plane analysis of non-linear systems usingweighted linearization 3l9

B. If &, < 0, all trajectories in the half-plane are tangent (as t + co) to the y-axis at
0 ; i.e. 0 is a (two-tangent) node.
For y c 0, the rules on the sign of Y),,, are reversed
In the second problem of Keil, I = 0, so that the equations in normal form are
3 = Y + X,(X,Y) (7.2a)
3 = YNL(X,Y). (7.2b)
It is known [ 1, pp. 267-2691 that curve M (defined by y = - X,) is tangent to the x-axis
at 0. For x > 0, the following rules apply:
C. If YIM> 0 on each side of the y-axis, a finite number of trajectories are tangent to
the x-axis at 0. Inside the two adjacent to the x-axis, the phase plane has the ap-
pearance of a saddle.
D. If )‘lu < 0, either no trajectories or infinitely many are tangent to the x-axis at 0.
The former possibility corresponds to a center or a focus, the latter to a (one-tangent)
node.
If x < 0, the rules on the sign of )ilM are reversed.
At present, we can apply U,(n) to Keil’s problems only when they are in normal form,
and X, and Y,,_ are sums of functions of x or y. Thus, Keil’s first and second problems
become, respectively,
1= -x+p(x)+hCy) (7.3a)

3 =_I-(.4+ st_Y), (7.3b)

and

3 = y + p(y) + h(x) (7.4a)

3 = fW + g(x). (7.4b)

We further assume that for sufficiently small arguments none of the above functions
vanish except at 0. With this assumption,

jr(Az)zwdz . b Cr(W4~41 s(& zw dz

= hll,. (7.5)
!i!$/tz,,dz =!z jas(Az) mdz
Z

Thus, if lim r(z)/s(z) = Q, lim c,(A,, J)/cJAb 2) = Q, where the subscript refers to the
2-O A,+0

function for which c(A,, 2) is the coefficient. In the following, we make constant use of the
above relation and of the fact that, if s(z) > 0 for z > 0 (z < 0), c, > 0 (c, < 0), where
henceforth we shall omit the arguments of the c’s.
In Keil’s first problem, we see that, as A, + 0, I(A, 2) + -4, and J(A, 1) + 0, so that
for small A,, J(A, 12)< I’(A, 2). Thus U,(A) can indicate only a node or a saddle-in
agreement with the exact analysis. In further agreement, U,(L) indicates isolated tangents
320 M.L. ADELBERG andH.H. DENMAN

near the x-axis asymptotic (as t -+ co) to 0. These conclusions follow immediately, since,
for x, y small and y 4 X, 2 % -xandd=y/x(wheree = tan- ’ y/x), so that for x > 0,
8 > 0 (here y > 0) and oice versa. Thus, the nature of the s.p. in the U,(2) approximation
depends only on the behavior of the tangent near the y-axis ; U,(A) indicates a saddle
(node) if the trajectory of this tangent approaches 0 as t --* - co (as t --) co). Setting x = rcy,
where rc is small, and substituting into the approximate equations

1= -(I- c&J
x+ ChY

3= CfX + c,y,

we obtain, for small Ai: K FZc,,, and y 1r = (c~c,, + c,)y, where y 1T denotes evaluation of
yalongx=rcy.Ify[r>O@(r < 0), U,(n) indicates a saddle (node).
Thus, the U,(n) analysis will agree with the exact results provided y(r has the same sign
as 31M The latter is given by :

3lM= fCwA + dY)l+ sot), (7.6)

where M is given by x = h(y) + a(y), with lim ~~(y)//i(y)= 0. If limf(z) h(z)/z g(z) = 0, it
y-0 z-r0

follows from (7.6), sincefis of higher order than linear, that y (M has the sign of g(y). Provided
we take A, 2 max (Al, A,), we can show t 31T also has the sign of g(y). (Here y > 0.) If,
however, lim g(z)lf[h(z)] is finite and > - 1, ylM has the sign off[h(y)], while )jlT has the
z-r0
sign of cfch [again A, z max (Al, A&J. It is easy to see that, in this case, the sign of 31T
agrees with that of JJtM; e.g. for y > 0 and small, if h(y) c 0, c,, < 0 and, since the argumest
offis ~0 and small, cI has sign opposite that off[h(y)], so that c,c, has the same sign.
In Keil’s second problem, only a tangent to the x axis is possible. Writing y = rcx and
proceeding as before, we write

i = (1 + cp)y + c,,x

j= CJY + C&

from which

2(1 + c& = cf - c,, + ,,[(c/ - c,,)2 + 451.

Along y = rcx,

’ = +{‘J + CLk J[(c, - c,,)2 + 4c,]} X.

We consider only the case

;2 [l.fCz>l+ l&)l12/ze(4 = 0. (7.7)

t From a formula similar to (7.Q it follows that, for A, 2 maxti,, A,), fimO c~ Jc, = 0.
I-
 Phase plane malysif of mm-linear sysiem us&g weighted limwrization 321

Thus, for A, small and A,, A, 3 As, Ah, we obtain

K 5%f Jc, (7.8a)

3 5?: * (JCJX. (7.8b)

Therefore, when x > 0, U,(n) indicates a saddle (with both tangents tending to the x-axis)
if g(x) > 0, and a focus or center if&c) < 0. This can be put in agreement with an exact
analysis if (7.7) is strengthened to

lim Ctf(z)J + l~z~lJ/g(z) = 0. (7.9)

r-4
With this assumption, it can be shown that, for x > 0, g(x) > 0, the finite number referred
to in Ruie C is one and the lower (upper) trajectory approaches 0 as t -+ co (t + - co) as
is indicated by (7.8b). Furthermore, for x > 0, g(x) < 0, it can be shown that no trajectory
can approach 0 along any tangent, i.e. 0 is a focus or a center.
As an example, consider the case where, for X,y # 0 but small, yf(y) < 0, x/z(x) < 0, and
xg(x) < 0. We define

P(y) = ‘p(z) dz and G(x) = ig(z) dz.

d
Thus, the positive definite form (for small x and y)

V= z+p(Y)-G(x)

has a negative definite time derivative :

3 = 6J + PtY))fCy) - &)h(~)~
so that 0 is asymptotically stable (in this case a focus, due to (79)). U,(J) also indicates a
stable focus since Z(Ab A) = &zf -t c,,) and is strictly less than zero for A, small.

8. NON4SOLAlED SINGULAR POINT

We consider here an example (a slight generalization of one discussed by Pars [7])

of a non-isolated singular point,
gt=y-. x”, n>l, x>o (8.la)
y = 0. (8.lb)
The solution of (8.lb) is y = y,; we take y, > 0. Then the solution to (&la) will have the
limit, as t --, co,
x + y;ln, (8.2)
so that the motion is stable. T,, however, discards x” in (8.la), and thus gives
x = y*t + x0, (8.3)
which is unstable.
322 M. L. ADELLIERG
and H. H. DENMAN

Using V,(n) in [0, AJ, we replace x” by

c,(n, A)/I;- l x. (8.4)
Substituting (8.4) in (8.la). we obtain
In [(yO - c,A;-’ x)/y,] = - c,A”,-’ t, (8.5)
with the initial conditions x = 0, y = y,, at t = 0. Then,
lim x = y,/c,A;- ‘. (8.6)
t-m
However, since x has been assumed to vary in [0, A,], for consistency and so that A, is
taken as small as possible, we also have x + A,. Thus
A, = y&A;-‘, (8.7)
or
x + A, = (y&)““, (8.8)
which is stable, and agrees with the exact result in y:“‘.

9. NON<’ SINGULAR POINT

As an example of a case for which X and Y are not C’ at the s-p. 0, and thus 7’r can
not be applied, we consider
I=y (9.la)
y= -x-H(y), (9.lb)

where for y # 0, If is C2, with yZ-ZCy)

> 0, H’(y) > 0, and yH”(y) < 0. Also, the integrals
-1

WH, WH (9.2)
d s

exist. Under these conditions, Malgarini [l, p. 3431 has shown that all trajectories are
tangent to the x-axis at 0 (as t + co) except one in the second quadrant and another in
the fourth which are tangent to the y-axis at 0 (as t -P co).
The Lr,@) equations are
I=y (9.3a)
y= -x-&y, (9.3b)
where ci can be written

c; = (l/N)i[fH(+A*z)/A,z]z’wdz.

Since yHCy) > 0, ci > 0, and because H- ’ is integrable, cl: diverges as A + 0. In (9.3).
J(A,II) = 1, Z(A,A) = - &cj$, and thus as A + 0, 0 < J < Z2, so that V,(A) correctly
describes the s.p. as a two-tangent stable node. Moreover, techniques similar to those
used in Sect. 7 for the analysis of Keil’s problems show that the behavior of the tangents,
and the trajectories near them, as given by u,(n), is correct.
 Phase plane analyslFof non-linear systevw using weighted linearization 323

10. DISCUSSION

We have developed herein a weighted linearization procedure U,(A) for analyzing

non-linear oscillating systems governed by equations of the form (IA), and compared
the results with the usual s.p. analysis based on Taylor series linearization (T1), where the
latter was applicable. We lirst showed that, wherever T, indicates correctly the nature of
the s.p., U,(A) does also. But, as shown in Sect. 3, in some cases where T1 indicates a center
and the actual trajectory in the phase plane is a focus, U,(A) correctly predicts the focus.
Further, where Tl gives a center correctly, as in certain free oscillation problems, it in-
correctly predicts a period of oscillation independent of amplitude, while U,(A) indicates
this dependence. In the soft~~g-having case, Tl indicates only a saddle near 0, while
U,(A) correctly predicts the shift to closed cycles as the amplitude is increased.
For the quadratic friction example of a weak focus (Sect. 5), T1 ignores the friction and
predicts a center. If U,(A) is the ultraspherical polynomial linearization, it gives the stable
focus for small and intermediate amplitudes, and, for I = 4, gives good quantitative
results for a2(a1). In the Lienard problem of Sect. 6, Tl correctly indicates the s.p. is an
unstable focus, but can not predict the existence of thz stable limit cycle, while U,(k) can
do both.
In Keil’s problem (Sect. 7), TXindicates a line of s.p.‘s where the system actually has an
isolated sp., so that 7” should not be used. U,(A), however, can be applied, and is shown
to give the correct qualitative features of the phase plane portrait under some conditions.
In Sect. 8, we examined a problem in which the s.p. is not isolated. Here Tl indicates an
unstable solution, while U,(A) indicates correctly the stability of the solution, and predicts
the correct qualitative features of the solution as well. Finally, in Sect. 9, we examined a
class of problems in which Y is not Cl, and thus Tt can not be used. In this case also,
U,(A) can be applied and successfully predicts certain important qualitative features of the
trajectories.
Thus it seems clear that the U,(A) analysis offers substantial advantages over the con-
ventional T, analysis. It is always at least as accurate as T,, but in a number of cases where
I; is incorrect, or is correct only near the s.p. but the system behaves quite differently at
some distance from the s.p., or where Tl is not applicable, U,(A) not only can be applied,
but gives a good qualitative picture of the phase plane trajectories, and sometimes useful
quantitative results.

Note oddcd in proof: J. E. Howard [SIAM J. @pi. hfatb. 16,747 (I%@] has wed a two-lim Chebyahev fora
approximationin tk a@&8 of fnt osci%tjons.

[l] 0. SATQXINE and R. CONIT,‘Non-hear Differential Equations. Pergamon Press (1964).

[2] C. LANCZDS, Tables of the Chebyshev Polynomiak N.B.S. Appl. Math. Series 9, U.S. Govt. Printing Office
(1962).
[3] M. L. ADELBERO, Application of Uitraspherical Poiynomiak ro Singular Point Analysis. M. S. Essay. Wayne
State University (1966).
[4] H. H. DENMANand Y. K. LKJ, Q. up@. Math. 22,273 (1965).
(Sj H. H. BOCJOLIUBOV and U. A, MITROF~XSN.Asymptotic Methods in the Theory of Non-linea? Osciiiafions,
2nd edition (in Russian), pp. 6&68. Govt. Publishing House, Pbysico-Mathematical Literature, Moskow
(1958).
[6] N. M-Y, Non-linear Oscilkzttons, p. 104. Van Nostrand (1962).
[7j L. A. PARS,A TIeotise on Analytical Dynamics, p. 378. Wiley (1965).

(Received 21 February 1968)

324 M. L. Aorrinna~ and H. H. t%WAN

RWOn dbloppe une analyse bask SWdes approximations hneaires pond&es pour certaines equations
~~r~t~~~ o&mires non ha&&es et on ~app~que a ia description des points et des traje&oires singuliers
dam fe plan de phase. On compare cotte technique a i’analyse de Poincari; bas& sur la linearisation en s&ies
de Taylor, lorsque cette derni&e eat applicable. On obtient les r&hats num&iques pour un probl&ne de friction
quadratique et on les compare aver le r&&tat du premier ordre de ~lov-~go~iu~v-Mi~opoI~y. La nouvelle
technique peut au& &re appliqute aver suc&e aux probl&nes de Lienard, Keil et Malgarini ainsi qu’a un cas
special de point singuiier non isol& ou les bnearisations en series de Taylor sont inapphcabls ou conduisent a
resultats qualitativement incorrects.

~~~E~e Analysisfiirsties nichthneare, ~w~~ic~ ~e~tial~eieh~~ wird ent-

wickelt, die auf gewogenen linearen Gina basiert und auf die IBM van singuhlren Punkten
und Rahen in der Pbasenebene angewendet wird. Die Methode wird mit der Analysis von Poincare verglichen,
die auf einer Linearisieruug mit Hilfe van Taylor&hen basiert. IGmerische Ergebnisse werden fttr ein quadra-
tischea R~b~~obi~ abgeleitet und mit dem Resultat emter Ores nach Krylov, Rogoliubov und Mitro-
polsky verghchen Die neue Metbode enveist sic% als ebenfalh erfolgreich an Problemen von Lienard, Keil und
MaQarini, und in einem speziellea Fall einea nicht isoberten, singuKren Pun&es, wenn Linearisiemngen durch
Taylor&her8 ungiWg sind oder qualitativ falsche Rrgebnisse hefem.

A~OT~IWF-AHWII~BocaoBaHabXt% Ha Becomx artHe&mx npa6mHceHHffx paapa6aTbrBaeTm ~JIR:

BeKoTopbIx HemirreZ1Bbfx 06VKHoBeHHbIx miQtQlepemviaJIbmx ypaBaeaii8 II npmtemeTcrr J&JIB
OIIUCaHHf3oco6urx TOW3HuTpaeKTOpSi~B ~asOBO~nJIOCHOCTH.3TOTMBTO~CPaBHHB8eTCRC anax?iaoY
llyamaps ~OH~B~H~~ Ha ~KBeap~ea~n~ pima TeftJropa KorEa orsa ~p~~eB~~a. %meaHlre
peaynbTaTbs nonysem am3 KB~pa~~~o~ npo6Jrew Tperim EIcpaBHem c pesyzbTaTom nepBor0
np~BmH(emR:: Kpkmosa-~orom06oBa-bfMTpoIroJIbcKoro. HOB~I& nfeToa ycwermio npnmewreTcK
To~eKnpo6Jrenaa~~beaap~a,KetrJra~ManbrapnHnu~c~esaanbnoanyc~yvaHl~e-~sonnpoBa~~ofl
oco6oiogTOSKK, ~~eap~ea4~~ pwa TaUopa ~e~p~~eB~M~ mm me AamT KariecTBeKzwe nenpa-