mathematics of computation, volume 25, NUMBER115, JULY, 1971

The Method of Christopherson for Solving Free

Boundary Problems for Infinite Journal Bearings
by Means of Finite Differences
By Colin W. Cryer*

Abstract. A method for solving free boundary problems for journal bearings by means
of finite differences has been proposed by Christopherson. We analyse Christopherson's
method in detail for the case of an infinite journal bearing where the free boundary problem
is as follows: Given T > 0 and h(t) find r E (0, T] and p(t) such that (i) [ffipj = h' for
t G (0, r), (Ü)/>(0) = 0, (iii) p(t) = 0 for t S [r, r], and (iv) p'(T - 0) = 0.
First, it is shown that the discrete approximation is accurate to 0([At]2) where Ai is
the step size. Next, it is shown that the discrete problem is equivalent to a quadratic pro-
gramming problem. Then, the iterative method for computing the discrete approximation
is analysed. Finally, some numerical results are given.

1. Introduction. A journal bearing consists of a rotating cylinder which is

separated from a "bearing surface" by a thin film of lubricating fluid (see Fig. 1).
The fluid is fed in at A and flows out at B. The width of the film is smallest at C,
and we set / = d/dc where 0 is as shown in Fig. 1.
Between C and B, the width of the film increases so that the pressure in the
lubricating fluid may be expected to decrease. We assume that for t = r the pressure
becomes so low that the fluid vaporizes. The point t = t, the interface between the
two phases of the fluid, is called the free boundary.
The mathematical problem can now be formulated (see Pinkus and Sternlicht
[7, p. 41 and p. 46]):
Problem 1. Find a functionp(t) and a constant t such that/» G C[0, T]C\ e<2,(0, t),
and

(1.2) p(t) =0, ráííT,

(1.3) piO) = 0,

(1.4) jfpÍT)=0.
In Problem 1, p(t) is proportional to the fluid pressure, while Eq. (1.1) is Reynolds'
equation for the pressure in a lubricating film.
Received February 13, 1970, revised January 19, 1971.
AMS 1970subject classifications.Primary 65L10; Secondary76D99, 76T05,90C20.
Key words and phrases. Christopherson's method, free boundary problems, finite differences,
journal bearings, quadratic programming, lubrication theory.
* Sponsored by the Mathematics Research Center, United States Army, Madison, Wisconsin,
under Contract No. DA-31-124-ARO-D-462,and the Office of Naval Research, under Contract
No. N00014-47-A-0128-0004.
Copyright © 1971, American Mathematical Society
435

License or copyright restrictions may apply to redistribution; see http://www.ams.org/journal-terms-of-use

 436 COLIN W. CRYER

t=T

CYLINDER

BEARINGSURFACE
Figure 1. Cross-section of a journal bearing

In order that Problem 1 be well defined, it is necessary that h(t), the width of the
film, satisfy certain conditions. Throughout this paper, we will assume that
h G ea)[0, T] and that
(1.5) h(t) > 0, / G [O, T],

t G (0, 1),
(1.6)
dh
> 0, tEH.T),
dt
(1.7) h(T) ^ h(0).
It will be shown in Section 2 that conditions (1.5) through (1.7) ensure that there
exists a unique solution to Problem 1.
Conditions (1.5) and (1.6) are always satisfied in practice, but this is not true of
(1.7). However, as we shall see in Section 2, condition (1.7) can be imposed without
any loss of generality.
In 1941, Christopherson [3] proposed a method for solving journal-bearing
problems numerically. A partial analysis of the method was given by Gnanadoss and
Osborne [6]. In the present paper, we present a detailed analysis of Christopherson's
method as applied to Problem 1.

License or copyright restrictions may apply to redistribution; see http://www.ams.org/journal-terms-of-use

 INFINITE JOURNAL BEARINGS 437

This paper is based upon a technical report (Cryer [4]) in which further details,
in particular all proofs, will be found.
Acknowledgement. The author is indebted to T. Ladner who wrote the program
used to obtain the numerical results in Section 7.

2. The Analytic Problem. In this section, we first analyse Problem 1 and then
formulate an equivalent problem, Problem 2.
Let

(2.1) â2(t) = f h'2(s) ds, t G [0, T],

Jo

(2.2) ¿3(0 = f h-\s) ds, t G [0, T],

Jo

(2.3) <p(t)= ê2(t)/M0, t G (0, T],

(2.4) W) = hiO&sit)- ê2(0, t G [0, 71.
Lemma 2.1. There is a unique constant <r, 1 < <r < T, such that (i) \}/(t) < 0,for
t G (0, 0%(ii) ft,) = 0; and (iii) *(t) > 0,for t G (<r,T}.
For a G (0, T], let q(t; a) be the function such that (i) £q(t; a) = 0, t G (0, a),
(ii) q(0; a) = 0; and (iii) q(t; a) = 0, t G [a, T¡; where £ is as in (1.1). It is easily
proved that q(t; a) is unique and is given by
q(t; a) = â2(t) - <p(a)ô3(0, t G [0, a],

= 0, t G la, T].
Theorem 2.2. There is exactly one solution [p(t), t) of Problem 1. If o- is as in
Lemma 2.1, then t = a, and p(t) = q(t; <r).
Proof. The proof is a straightforward generalization of previous results (Birkhoff
and Hays [1, p. 132], Pinkus and Sternlicht [7, p. 46], and Gnanadoss and Osborne [6]).
Problem 2. Find {p(t), r} such that
(2.5) t - sup{a G (0, T]: q(t; a) ^ 0 for t G LO,T]},
(2.6) p(t) = q(t; t).
That is, find the largest interval in which a nonnegative solution of Reynolds'
equation exists.
Problem 2 was first suggested by Gnanadoss and Osborne [6], and the next
theorem is a generalization of their results:
Theorem 2.3. Problems 1 and 2 are equivalent.
We conclude this section with a discussion of condition (1.7). First, we note that
it is the role of condition (1.7) to ensure that the lubricating fluid occurs in both the
liquid and gaseous phases. If (1.7) is not satisfied, then it is possible, for example if
B is close to C (see Fig. 1), for the fluid to occur only in the liquid phase.
Secondly, we note that there is no loss of generality in assuming (1.7). For suppose
that h\t) and f are suchthat (i) h G e(1)[0,f\, (ii)h(t) > 0, t G [0,f ];(iii)dh/dt < 0,
t G (0, 1); and (iv) dh/dt > 0, t G (1, f); but that h(f) < h(0). Let T > T and
h G e(1)[0, T] be such that (1.5) through (1.7) are satisfied and h(t) = h(t) for

License or copyright restrictions may apply to redistribution; see http://www.ams.org/journal-terms-of-use

 438 COLIN W. CRYER

t G [0, f]; clearly, h and 7"can be chosen in many ways. Let Problem 1 and Problem Ï
be the problems corresponding to h and h, respectively. Clearly, if Problem Ï has a
solution {p(t), f}, then Problem 1 has a solution {p(t), t) where r = f, p(t) = p(t)
for t G [0, T], and p(t) = 0 for t G (f, T\. Hence, to solve Problem Ï, we may first
find the solution {p(t), r} of Problem 1; by Theorem 2.2, this solution exists. If
r > T, Problem ï has no solution. On the other hand, if t ^ t, then Problem Î has
the solution \p(t), f} where f = r and p(t) = p(0 for t G [0, f].

3. The Discrete Approximation. Let {p(t), t } be the solution of Problems 1

and 2. We subdivide the interval [0, T] into N subintervals, each of length At, so that
N = T/At. We seek an integer m and an (N + 1)-vector P = {P¡\,j = 0,1, • ■• , N,
such that m At œ t and P, œ p(jAt).
We approximate the Reynolds' equation (1.1) by the finite-difference equation

(3.1) (LP),. = ^-5 AihU/tVPi) ~ ¿ AÄf_1/a= 0, 0 < i < TV,

where A and V denote the forward and backward difference operators respectively,
hi = h(iAt), and A,_1/2= h([i - %]At).
In order to avoid certain trivial possibilities, we assume that At á f, and that
7V«¡£3. We also make an additional assumption about h(t), namely, that
(3.2) h(T - (At)/2) è h((At)/2).
Assumption (3.2) is trivially satisfied by, if necessary, slightly increasing T and
modifying the definition of h(t) appropriately (see Section 2 where a similar device
is used).
Let
i
(3.3) 72(0 = £ (¿i-1/2)"2, 0 g / á N,

(3.4) 78(f) = ¿ (À,_1/2r3, 0 g i ^ TV,

1-1

(3.5) *,. = hiO/hiO, 1 á i á N,

(3.6) ¥,. = hi +i/2I3(i) - I2(i), 0 £ f £ 7f - 1,
with the convention that
o o
E (A--1/2)"2= E (ä,--^)"3 - 0.
¿-i ¿-1

Noting the analogy between Eqs. (1.1) and (3.1) and between Eqs. (2.1) through
(2.4) and Eqs. (3.3) through (3.6), we are led to
Lemma 3.1. There is a unique integer n, 1 — 3(At)/2 g nAt Ú T — 2(Ai), such
that (i) % < 0,for 1 ^ i < n; (ii) *n á 0; and (iii) % > 0,for n < i £ N - 1.
For any integer /, 1 ^ / ^ N, let Q(I) = {ß,(/)} be the (N + l)-vector such that
[LQiDh= 0, 1 á / á /- L
ßo(0 = o,
ß,(0 =0, l ^ i é N,

License or copyright restrictions may apply to redistribution; see http://www.ams.org/journal-terms-of-use

 INFINITE JOURNAL BEARINGS 439

where L is as in (3.1). It is easily proved that Q(/) is unique and is given by

ß,(0 = At{I2(i)- $«73(0}, 0£ /á /,
= 0, l á i á N.
We can now formulate the discrete analog of Problem 2:
Problem 2D. Find {P, m\ such that,
(3.7) m = max {/: 1 g / g TV;0,(0 ^ 0 for 0 ^ ï g TV},
(3.8) P = Q(m).
Theorem 3.2. 77zere ¿s exactly one solution {V, m\ of Problem 2D. If n is as in
Lemma 3.1, then m = n + 1, a«G?P = Q(m).

4. Error Analysis. In Theorems 2.2 and 3.2, explicit expressions for the solution
{p(t), t} of Problem 2 and the solution {P, m\ of Problem 2D are given. By comparing
these expressions, error estimates can be obtained. Under the assumption that
(4.1) h G e<2)[0, T],
lengthy, but straightforward, computations yield
Theorem 4.1. There are positive constants (AOo and K such that if
(4.2) Ai ^ (AOo,
then
(4.3) \t - mAt\ è 5(AO/8,
and
(A.A) \p(jAt) - P,\â K(At)2, 0 ^ j g TV.
It follows from Theorem 4.1 that {P, m\ is as accurate an approximation to
{PÍO, Tí as could be hoped for. For we can at best have that
(4.5) \t - mAt\ g (AO/2,
and (4.3) is almost as good as (4.5). Since we can at best have (4.5), we might expect
that \p(jAt) — P¡\ = O(At). Instead, the gods have smiled, and we have (4.4). More
prosaically, the "reason" why (4.4) holds is that the condition p'(r) = 0 makes p
"relatively insensitive to errors in r."

5. Another Discrete Approximation. In this section, we formulate a second

discrete approximation to Problem 2, Problem 3D, which can be shown to be
equivalent to Problem 2D. The reason for introducing Problem 3D is that Chris-
topherson's method is best understood if it is regarded as an algorithm for solving
Problem 3D.
Set M = N — 1 and denote by A the M X M matrix with components
Au = -(hi+i/2f, if 7= i+ 1,
(5 1} = l(hi+i/2? + (A,-V2)8]. if 7= i,
= -(hi-i/2f, if; = / - 1,
= 0, otherwise,

License or copyright restrictions may apply to redistribution; see http://www.ams.org/journal-terms-of-use

 440 COLIN W. CRYER

for 1 í£ i, j ^ M. Noting (1.5), we see that A is a symmetric irreducibly diagonally

dominant matrix with positive diagonal entries, so that A is positive definite (Varga
[9, p. 23]).
Denote by B the M-vector with components

(5.2) Bi = -(At)[hi +i/2 - ft,_x/*L láiáM.

Problem 3D. Find M-vectors X and Y such that
(5.3) AX - Y = B,
(5.4) XrY = 0,

(5.5) XâO, Y à 0.
Using the theory of quadratic programming and the detailed information available
concerning the solution of Problem 2D, we obtain
Theorem 5.1. Problem 3D has a unique solution. Problems 2D and 3D are equiva-
lent. If {P, m\ and {X, Y) are the solutions of Problems 2D and 3D, respectively, then
(5.6) Xi = Pi, \ Û i ú M,
(5.7) Y = AX - B,
(5.8) 7>0= PN = 0,
(5.9) m = inf {i; Y¡ > 0},
(5.10) Y¡ = 0, 1 g i < m,
(5.11) Yi > 0, m g i ^ M,
(5.12) Xi > 0, 1 á i á m - 2,
(5.13) Xm-t = -(AtWm-i/{(hm-i/2)3I3(m)} ^ 0,
(5.14) Xi = 0, m ^ i ^ M.

6. The Iterative Solution of the Discrete Approximation. In this section, we

analyse the algorithm used by Christopherson to compute the solution of the discrete
approximations to Problem 1.
Algorithm 6.1. Choose an M-vector X(0) = {X^} where X(0' ^ 0. Choose a
relaxation parameter o, where 0 < w < 2.
Generate a sequence of M-vectors Xe" = {*<"}, R<" = {R^}, and YU) =
{F,w}, k = 1, 2, • ■• , using the equations,

(6.1) R?+" = B, - E AttX,Mi - E ¿„Af.

j-i ,-i
(6.2) Jt?+I) = max {0, X[k) + aR?+1)/A«},
(6.3) if+1) = -*<*+1) + ^,U,(*+1) - X(ik)).

The reader will have observed that Algorithm 6.1 consists of applying S.O.R.
(systematic overrelaxation) to the equations AX = B with the proviso that the iterates
X"' be nonnegative. This was the way in which Algorithm 6.1 was viewed by

License or copyright restrictions may apply to redistribution; see http://www.ams.org/journal-terms-of-use

 INFINITE JOURNAL BEARINGS 441

Christopherson except that, since he worked by hand, he used relaxation rather

than S.O.R. The condition that the vectors X(t) be nonnegative arises naturally from
the physical restraint that the lubricating fluid cannot support negative pressures.
Christopherson used Algorithm 6.1 without explicitly formulating the discrete
problem that he was solving. Of the two formulations of the discrete problem that
we have developed, Problems 2D and 3D, it seems to us that Problem 3D lies closest
in spirit to Christopherson's ideas.
In computations for Cameron and Wood [2], Fox (working by hand) used
Algorithm 6.1 with relaxation instead of S.O.R.; Raimondi and Boyd [8] (using an
IBM 704) used the Liebmann or Gauss-Seidel method instead of S.O.R.; finally, the
use of S.O.R. was suggested by Gnanadoss and Osborne [6].
Throughout the remainder of this section, we denote the solutions of Problems
2D and 3D by {P, m\ and {X, Y}, respectively, and assume that X"° and Y(h) are
generated using Algorithm 6.1.
Using the results of Cryer [5], we obtain
Theorem 6.1. For any X(0) ^ 0, Xik) -►X and YM -^Yasfc^œ.
Next, we consider the speed with which Algorithm 6.1 converges. We define the
asymptotic rate of convergence of Algorithm 6.1 to be

(6.4) R(A, B,co) = -log { sup lim sup ||X(i) - X\\1/k\ ,

Ix<°'ä0 *-.» J
where 11• 11denotes any vector norm.
We need certain concepts from the theory of S.O.R. (see Varga [9]). Let Ä be a
p X p positive definite matrix. Let Ä = D — Ê — F, where ß is a diagonal matrix
while Ë and F are, respectively, strictly upper and strictly lower triangular matrices.
Then the point successive relaxation matrix corresponding to Ä is given by
(6.5) £.(Ä) = (D - WE)"1¡(I - «)ß + o>f\.
The point Jacobi matrix J is given by
(6.6) J = (ßy-'tE + F].
The asymptotic rate of convergence for £„(Ä) is given by
(6.7) Ä.[£B(Ä)] = -log {p[£.(Ä)]},
where p[£„(Ä)] is the spectral radius of £„(Ä). Finally, the optimum relaxation
parameter wb = wb(A) satisfies
(6.8) 2U£„6(Ä)] = maxTU£„(Ä)].

Theorem 6.2. Assume that ^m_! > 0. Then

R(A,B,w) = ZUJE.rA'"-"]),
where A<™~1) is the (m — 1) X (m — 1) matrix consisting of the first (m — 1) rows and
columns of A. If wopt = co¡,(A<m_1)),
then wopt ^ ù>b(A)andR(A,B, w) g R(A,B, uOBt),
0 < £0< 2.

7. Numerical Results. In this section, we present numerical results for an

infinitely long full journal bearing to illustrate the theoretical results of the preceding
sections.

License or copyright restrictions may apply to redistribution; see http://www.ams.org/journal-terms-of-use

 442 COLIN W. CRYER

The equations for an infinitely long full journal bearing are (Pinkus and Sternlicht
[7, p. 42 and p. 46])
<'•» íV'<Á =dJT- •<»<*•
(7.2) p(6) = 0, 9,gH 2k,

(7.3) p(0) = 0,

(7.4) ^p(0) = 0, 0 = e2,

(7.5) w(6) = 1 + £ cos 6.

Here, e is the eccentricity ratio and satisfies 0 5í € < 1.
Introducing the variable
(7.6) t = 8/k,
it is found that;? satisfies Eqs. (1.1) through (1.4) with T = 2, r = 02/7r,and
(7.7) A(0 = (1 + € cos-n-0/vV
It is easily verified that h satisfies (1.5), (1.6), (1.7), (3.2), and (4.1), so that all the
results of the preceding sections are valid.
For h given by (7.7), the solution of Problem 1 can be given analytically (Pinkus
and Sternlicht [7, p. 47]); the expressions are complex and will not be given here.
Numerical results were obtained for the case e = .8. The computations were
performed on the UNIV AC 1108 computer at the University of Wisconsin; this
computer uses eight-decimal floating-point arithmetic. The analytic solution \p(t), t}
was computed directly. The discrete approximation {P, m} was computed by using
Algorithm 6.1; the iterations were terminated when
(7.8) l|R(*+"ll- = max \R¡k+v\ g 10"7.
The initial approximation X<0) was always taken to be identically zero.
Two experiments were carried out. In the first experiment, N was taken equal to
64 while cowas varied. We were primarily interested in determining the number of
iterations required to converge, that is the number of iterations required before (7.8)
was satisfied. The results (see Table 7.1) showed that a reasonable strategy was to
set co = co* where
(7.9) co* = 2/{l + sin(7r/TV)}.
(co* is the optimum overrelaxation parameter for the discrete two-point boundary-
value problem VAu¡ = b¡, 1 Ú j è N — 1; u0 = uK = 0.)
In the second experiment, cowas taken equal to co* while N was varied. We were
primarily interested in the difference between {p, r\ and {P, m}. Setting
(7.10) \\p - P|U = max \p(jAt) - P,\,
i
the dependence of \\p — P||„ upon At is shown in Table 7.2.
Bearing in mind that the UNIV AC 1108 works to only eight decimal places, it is
clear that the results are in agreement with the assertion of Theorem 4.1 that

License or copyright restrictions may apply to redistribution; see http://www.ams.org/journal-terms-of-use

 INFINITE JOURNAL BEARINGS 443

Table 7.1. Number of iterations to converge (N — 64)

No. of iterations to
co converge
1.0 811
1.1 670
1.2 558
1.3 453
1.4 362
1.5 288
1.6 216
1.7 146
1.8 70
1.9 136
co*= 1.90645 146

Table 7.2. Dependence of\\p — P||„ upon At

No. of iterations
TV At \\p — P||„ to converge
64 .03125 .016017 146
128 .015625 .002725 268
256 .0078125 .000756 513
512 .00390625 .000170 923
1024 .001953125 .000073 1640

\\p — P||» Ú K(At)2. The other assertion of Theorem 4.1, namely that |r — mAt\ g
5(A0/8, was always satisfied.

Computer Sciences Department

The University of Wisconsin
Madison, Wisconsin 53706

1. G. Birkhoff & D. F. Hays, "Free boundaries in partial lubrication," /. Mathematical

Phys., v. 42, 1963, pp. 126-138. MR 27 #3168.
2. A. Cameron & W. L. Wood, "The full journal bearing," Inst. Mech. Engrs. J. Proc,
v. 161, 1949,pp. 59-64.
3. D. G. Christopherson, "A new mathematical method for the solution of film lubrica-
tion problems," Inst. Mech. Engrs. J. Proc, v. 146, 1941, pp. 126-135. MR 3, 285.
4. C. W. Cryer, The Method of Christopherson for Solving Free Boundary Problems for
Infinite Journal Bearings by Means of Finite Differences, Technical Report #72, Computer
Sciences Dept., University of Wisconsin, Madison, Wisconsin, 1969.
5. C. W. Cryer, "The solution of a quadratic programming problem using systematic
overrelaxation," SI AM J. Control (To appear.)
6. A. A. Gnanadoss & M. R. Osborne, "The numerical solution of Reynolds' equation
for a journal bearing," Quart. J. Mech. Appl. Math., v. 17, 1964, pp. 241-246.
7. O. Pinkus & B. Sternlicht, Theory of Hydrodynamic Lubrication, McGraw-Hill,
New York, 1961.
8. A. A. Raimondi & J. Boyd, "A solution for the finite journal bearing and its applica-
tion to analysis and design. Ill, Trans. Amer. Soc. Lubrication Engrs., v. 1, 1958, pp. 194-209.
9. R. S. Varga, Matrix Iterative Analysis, Prentice-Hall, Englewood Cliffs, N. I., 1962.
MR 28 #1725.

License or copyright restrictions may apply to redistribution; see http://www.ams.org/journal-terms-of-use