C. W.

Wampler Complete Solution of the Nine-

A. P. Morgan Point Path Synthesis Problem for
Mathematics Department,
General Motors Research Laboratories,
Four-Bar Linkages
Warren, Ml 48090
The problem of finding all four-bar linkages whose coupler curve passes through
nine prescribed points has been a longstanding unsolved problem in kinematics.
A. J. Sommese Using a combination of classical elimination, multihomogeneous variables, and
Mathematics Department, numerical polynomial continuation, we show that there are generically 1442 non-
University of Notre Dame, degenerate solutions along with their Roberts cognates, for a total of 4326 distinct
Notre Dame, IN 46556 solutions. Moreover, a computer algorithm that computes all solutions for any given
nine points has been developed.

Introduction
The approximate synthesis of a given path by use of four- of degenerate solutions. By the theory of "parameter poly-
bar linkages has been studied extensively. Formulations in nomial continuation" (Morgan and Sommese, 1989), we may
terms of four or five precision points along with specifications ignore all the degenerate solutions and use only the nonde-
on crank angles or the position of the hinges of the mechanism generate ones as start points in subsequent continuations to
have been solved (Freudenstein and Sandor, 1959; Shigley and find all nondegenerate solutions to any other problem of the
Uicker, 1980; Erdman and Sandor, 1984; Morgan and Wamp- class. Thus, we have not only established the generic number
ler, 1989; Subbian and Flugrad, 1989). However, the problem of nondegenerate solutions to the problem, but also have de-
of finding four-bar linkages whose coupler curve passes through veloped an efficient computer algorithm for finding them.
nine precision points, which was formulated as early as 1923 In any particular example, not all of the 1442x3 = 4326
(Alt), has until now defied complete solution. Since nine gen- solutions are useful. Most give linkages with complex link
eral precision points is the largest number that can be pre- lengths, whereas others give real linkages that exhibit branch
scribed, this formulation gives a designer maximum control or order defects, or that have poor transmission angles, etc.
over the shape of the coupler curve. We discuss these issues in the context of several test problems.
The first serious attempt to solve the nine-point problem The papers (Wampler, Morgan and Sommese, 1990; Wamp-
appears to have been conducted by Roth and Freudenstein ler and Morgan, 1990) contain tutorial material on the math-
(1963), as a special case of their treatment of geared five-bar ematical techniques we have used. In particular, the reader
mechanisms. They employed a type of numerical continuation, may wish to consult these papers for discussions of the multi-
which they called the "bootstrap procedure," to find some homogeneous Bezout number and the method of numerical
solutions. More recently Tsai and Lu (1989) applied a new reduction (via parameter continuation; also referred to as "the
continuation method, called the "cheater's homotopy," to method of the generic case").
increase the reliability of the procedure, but they also did not
attempt to find all solutions. Since the problem has many
solutions, most of which are either degenerate or have branch Problem Formulation and Reduction
or order defects, it is often difficult to find an acceptable
solution by trial-and-error procedures. Only by finding all The most concise formulation of the problem is obtained
nondegenerate solutions can one be sure to find the best mech- by representing the links as vectors in the complex plane. Our
anism or, in some cases, verify that no acceptable solution derivation follows that of Roth and Freudenstein (1963), but
exists. with a change of variables that allows for subsequent reduction.
We solve the problem using a combination of analytical and Referring to Fig. 1, let P0 be the first precision point, at which
numerical tools. First, we reformulate the problem to analyt- the four-bar is given by quadrilateral ABCD with coupler tri-
ically reduce the polynomial system of equations that describe angle CP0D. Summing the vectors around the left-hand loop
the problem. We then use numerical polynomial continuation gives
in multihomogeneous variables (Morgan and Sommese, 1987) u=x-a. (1)
to solve a problem with randomly generated precision points,
thereby determining the generic structure of the solution set Figure 2 shows the four-bar after a displacement of 5, to pre-
for these problems. In addition to 1442 sets of Robert's cognate cision point Pj, with corresponding angular displacements 0,-,
triples, the solution set includes several higher-dimensional sets \j, fij of the coupler triangle, link AD and link BC, respectively.
Now, the vector loop equation becomes
Contributed by the Mechanisms Committee for publication in the JOURNAL
OF MECHANICAL DESIGN. Manuscript received November 1990. ue'y=xe"'j+8i-a. (2)

Journal of Mechanical Design MARCH 1992, Vol. 114/153

Copyright © 1992 by ASME
Downloaded From: http://mechanicaldesign.asmedigitalcollection.asme.org/ on 05/19/2014 Terms of Use: http://asme.org/terms
 Fig. 2 Displacement of the four-bar to a new precision point P}

y* with x, a, y, b, yj, respectively, in equations (6, 7, 8) to

Fig. 1 Four-bar ABCD with coupler triangle CP0D at the initial precision make them polynomial. Moreover, we can reduce the degrees
point P0 of the equations by introducing the new variables n, n, m,
m defined as follows
Substituting equation (1) and performing the same derivations n = ax, n = ax, m = by, rfi = By. (9)
for the right-hand side of the mechanism, we obtain Finally, we eliminate some solutions at infinity by partitioning
(x-a)eiXJ = xeieJ + SJ-a (3) the variables into 10 groups:
[x, x, a, a, n, n], [y,y, b, b, m, m\, [yj, yj},j = 1.....8
(y-b)ei"J=yeieJ + SJ-b (4)
and introducing homogeneous coordinates for each group,
Multiplying each side of these equations by its complex con-
respectively, as x°, y°, yj, y = 1 8. With all of these modi-
jugate and letting
fications, we may rewrite equations (6-9) as
yj = eieJ-l, (5)
nx°-ax, nx° = dx, my° = by, my° = Sy (10)
we get
(n-SjX)yj+ (n-bjX)yj+ [bj(a-x)
(a* -6/)xyj+ia-bj)x*yf + bj(a*-x*)
+ bj{a -x)- bjbjX0]yj = 0 (11)
+ 8f(a-x)-8j8f=0 (6)
(m - Sjy)yj+ (m - bjy)yj+ [bj(6-y)
(b* - 6/ )yyJ+ (b- 8j)y*y? + 8j(b* -y*)
+ 8j(b-y)-bjSJy°]y° =0 (12)
+ 6f(b-y)-6j&f =0 (7)
where the asterisk indicates complex conjugation. From equa- y/Yj + Yy7° + yjyj = 0 (13)
tion (5), we also have the identity
Equations (10-13) are a set of 28 equations in 28 unknowns.
yjy! + yj+y}=o. (8) To reduce the number of variables, we solve equations (11,
Since conjugation is not algebraic, some modification of the 12) using Cramer's rule to obtain for j= 1,...,8
proceding equations is necessary to make them polynomial, n~SjX bj(d-x)+Sj(a-x)-8j§jX
which is a prerequisite to finding all solutions. The usual ap- (14)
yr- m - bjy bj(b-y) +bj(b-y)- bjbjy0
proach is to explicitly write each variable in terms of its real
and imaginary parts, for example, we may replace x with bj(d^x)+bj(a-x)- bjbjX0 n - SjX
X\ + ix2, and so on. Of course, x* is then replaced by x\ - ix2. yj= (15)
If we then treat x{ and x2 as complex variables and solve the bj(b-y)+Sj{b-y)-bjbjy0 m-bjy
polynomial system, only those solutions for which X\ and x2
are both real have physical meaning. Note that X\ + ix2 and n-SjX n — bjX
yj= rh - bjy m - bjy (16)
X\ - ix2 are complex conjugates of each other if, and only if,
X\ and x2 are real. A simpler treatment is to retain x as a
complex variable and replace x* by an independent complex Substituting these expressions into equation (13) and retaining
variable, say x, and so on. Now, the condition for a solution equations (10), we have a system of 12 equations in 12 un-
to have physical meaning is that x* =x, etc. knowns.
Although Roth and Freudenstein used the real and imaginary The system consists of 4 quadrics and 8 quartics for a total
parts of x, a, y, b as variables, a set of synthesis equations degree of 2 4 4 8 = 1,048,576. However, the 2-homogeneous Be-
essentially the same as theirs is obtained as follows. For each zout number is only 286,720. This number is the result of a
precision point Pj, j= 1 8, treat equations (6, 7) as a pair certain combinatorial calculation that can be mechanized as
follows. First construct Table 1, listing the degree of each
of linear equations in yj and y*, solve these using Cramer's
• equation with respect to each of the two groups of variables,
rule, and substitute into equation (8). This gives eight equations
arbitrarily named the a-group and the /3-group. Then, form a
in eight unknowns. Solutions to this set of equations corre-
single term for each equation, being the sum of the degrees
spond to four-bar linkages whose coupler curve passes through
times either a or /3, as appropriate, and multiply all these terms
the nine precision points Pj, j = 0,... ,8. Each of the eight equa-
together to get the polynomial (2a)2(2/3)2(2a + 2/3)8. Now, since
tions is a seventh degree polynomial, yielding a total degree
there are 6 variables in each group, we want the coefficient of
for the system of 78 = 5,764,801.
a6/?6 in this polynomial, which turns out to be 2128!/(4!4!) =
Reduction. We reduce the polynomial system by intro- 286,720. (The ratcoef command in Macsyma and similar com-
ducing some new variables and using multihomogeneous co- mands in other symbolic manipulation programs can sperform
ordinates. As discussed above, we first replace x*, a*, y*, b*, this calculation.) Thus, from the original figure of 7 , our re-

1 5 4 / V o l . 114, MARCH 1992 Transactions of the AS ME

Downloaded From: http://mechanicaldesign.asmedigitalcollection.asme.org/ on 05/19/2014 Terms of Use: http://asme.org/terms

 Table 1 Degree table for 2-homogeneous Bezout count h=m = Q, which makes yj = 7°= 0. This case is purely a math-
Equation a-group /J-group Terni ematical figment, since the coupler triangle becomes a point
{x, xyatayn^h\ {y,y,6,6, m,m} and the linkage does not move. In either case, we are not
nx° = ax 2 0 2a interested in such linkages.
nz° — ax 2 0 2a In fact, there are several other sets of degenerate solutions,
my° = by 0 2 2/3 but they all obey one of the following conditions:
my" = by 0 2 20
.Eq. (13) 2 2 (2a + 2/3)8 x = 0 o r 7 = 0 o r f = 0 or y = 0, (19)
j=l,...,8
x=y or x=y. (20)
duction has cut the number of possible solutions by a factor Again, we are not interested in linkages satisfying these con-
of slightly more than 20. ditions.
We remark as an aside that the 10-homogeneous Bezout If we try to solve the problem by applying Newton's method
number for equations (10-13) is also 286,720, but we prefer from an initial guess, there is a high likelihood that it will
the formulation with a fewer number of variables. converge to one of these degenerate cases. This is one of the
Roberts Cognates. A well-known result due to Roberts motivations for developing a more sophisticated numerical
(1875) is that every four-bar linkage has two four-bar cognates; approach.
that is, there are two other four-bar linkages that produce the Precision Points. It is known (Roth and Freudenstein, 1963;
same coupler curve. Clearly, if a four-bar linkage solves our Primrose and Freudenstein, 1963) that a four-bar coupler curve
nine-point problem, then so will both of its cognates. Referring can intersect a circle or a line at most 6 times. Therefore, of
to equations (3, 4), suppose that a linkage and its angular the nine precision points, no more than 6 can lie on a circle
displacement are given by z = (x, a, y, b, dj, X,-, nj) for a dis- or line. This conclusion can easily be derived from our for-
placement of the coupler point of <5/. Then, the corresponding mulation. After substituting from equations (14-16) into (13),
position of one of its cognates is given by the formula we note that for a given four-bar linkage, equation (13) can
r(x, a, y, b, dj, X,-, ^ ) be viewed as a polynomial equation in dj, 5,-. We then ask how
many solutions this equation shares with an arbitrary circle in
/(x-a)y bx-ay ^ \ ,„„, the plane, which has the form
= {- -, *-,a-x,a,\jHj,6j). 17
y x-y x-y ) /MySy-H 026,- +1838,+04 = 0. (21)
This may be confirmed by noting that if we rewrite equations Treating equations (13, 21) as 2-homogeneous, we note that
(3, 4) a s / ( z ) = 0 and g(z)=0, respectively, then f(r(z)) equation (13) is bi-cubic and equation (21) is bilinear, so that
= (xg(z)-yf(z))/(x-y) and g(r(z))=f(z), so that if z the Bezout number is the coefficient of era in the polynomial
satisfies the equations then so does r(z) . The fact that cognates (3a + 3a)(a + a), which is 6. For a line, we have the special
come in triples can be confirmed by showing that case 0i = 0 and the argument still holds.
r(r(r(z)))=z.
In our polynomial formulation of the problem, the angles Summary. Using only analytical means, we have shown
have been eliminated and variables (x, a, y, b) have been in- that the nine-point problem has at most 286,720 isolated
troduced. To compute a cognate of a solution in these vari- solutions and, because of the two-fold S symmetry, these
ables, we drop the angular entries from the formula of equation represent at most 143,360 distinct mechanisms. Moreover,
(17) and apply it to both (x, a, y, b) and (x, a, y, B) inde- these mechanisms must appear in groups of three according to
pendently. The corresponding values of n, n, m, m are then Roberts theorem. The fact that 143,360 is not a multiple of 3
calculated from their definitions in equation (9). For brevity, is not a problem because this Bezout calculation includes many
we do not write out the complete formula, but simply refer to degenerate solutions and for most of these the Roberts cognates
it hereafter as R{z). are not well-defined. Finally, while we can solve the problem
Symmetry. In addition to the cognate groups, our equa- for nine generally placed precision points, no more than six
tions contain a two-fold symmetry. If we simply exchange the of the points can lie on a circle or a line.
variables in the two homogeneous groups, that is, if we replace
z by S(z) where
Numerical Reduction by Continuation
S{x, x, a, a, n, n, x°, y, y, b, 6, m, m, y°)
If the nine-point problem actually had 143,360 solutions,
= (y> y> b, 6, m, m, y°, x, x, a, a, n, n, x°), (18) it would seem impractical to compute all the solutions as a
the new point also satisfies the equations. This operation simply matter of routine in real design problems. However, it turns
exchanges the rows in each determinant in equations (14-16), out that the vast majority of this Bezout number is due to the
thereby changing the sign of each of yjt yj, yj, which leaves degenerate solution sets, which can be eliminated numerically.
equation (13) unchanged. The physical interpretation of this The method of reduction requires a one-time computation of
symmetry is that we may exchange the labels on the left and the full set of 143,360 solutions for a generic set of precision
right sides of the four-bar linkages, thereby exchanging the points. These solutions are then sorted as to whether or not
values of x with y, etc., while the linkage itself remains the they satisfy the side conditions, equations (19, 20), defining
same. degeneracy. Only the nondegenerate solutions can lead to non-
For a general point z, the Roberts cognate formula and the degenerate solutions for other sets of precision points, so the
symmetry formula give a group of six points as follows: z, number of these solutions determines the cost of solving sub-
sequent problems. Thus, we can reduce the problem to man-
R(z),R(R(z)),S(z),S(R(z)),S{R(R(z))).Becauseofthe
ageable proportions by a single, albeit expensive, numerical
identity R(S(z)) =S(R(R(z))), the group is closed.
computation. This approach is justified by the theory of pa-
Degenerate Solutions. Our set of equations admits several rameter continuation (Morgan and Sommese, 1989) and is
types of degenerate solutions. One of these occurs when developed in the context of a robotics problem in Wampler
z = S(z), that is x=y, a=b, etc. In that case, the mechanism and Morgan (1990).
degenerates into a two-degree-of-freedom linkage that can reach We proceed as follows. L e t / ( p , z) = 0 represent our poly-
any point inside an annulus centered on the fixed pivots, which nomial formulation of the nine-point problem, equations (10,
are coincident. Another degenerate case occurs when x=y = 13), where z is the set of 12 variables and/? is the set of path

Journal of Mechanical Design MARCH 1992, Vol. 1 1 4 / 1 5 5

Downloaded From: http://mechanicaldesign.asmedigitalcollection.asme.org/ on 05/19/2014 Terms of Use: http://asme.org/terms

 sequently, the end result of the Cheater's Homotopy would We would also like to acknowledge the work of Scott Town-
be a program that tracked at least 90,000 continuation paths, send in developing computer codes.
rather than our 1442. This work represents a cooperative interdisciplinary effort.
The order of the names given on this report has no significance.
There is no primary author.
Summary and Conclusion
We have established that the nine point path synthesis References
problem for four-bar linkages generically has 4326 distinct Alt, H., 1923, "Uber die Erzeugung gegebener ebener Kurven mit Hilfe des
nondegenerate linkages occurring in 1442 cognate triples, and Gelenkvierseits," Zeitschrift fur angewandte Mathematik und Mechanik, Vol.
we have developed a computer program to find all of them by 3, pp. 13-19.
tracking 1442 continuation paths. This is a reduction by a Erdman, A. G., and Sandor, G. N., 1984, Mechanism Design: Analysis and
factor of nearly 4000 from the best total degree (78) previously Synthesis, Vols. 1 & 2, Prentice-Hall, Englewood Cliffs, NJ.
Freudenstein, F., and Sandor, G. N., 1959, "Synthesis of Path-Generating
reported for the problem. The reduction is accomplished in Mechanisms by Means of a Programmed Digital Computer," ASME Journal
two stages: an analytical reduction based on both classical of Engineering for Industry, Series B, Vol. 81, pp. 159-168.
elimination and a 2-homogeneous formulation, and a numer- Li, T. Y., Sauer, T., and Yorke, J. A., 1989, "The Cheater's Homotopy:
ical reduction based on parameter polynomial continuation. An Efficient Procedure for Solving Systems of Polynomial Equations," SIAM
J. of Numerical Analysis, Vol. 26, pp. 1241-1251.
The numerical reduction step tracked 143,360 solution paths Morgan, A. P., 1986, " A Transformation to Avoid Solutions at Infinity for
and established that only 4326 of the endpoints were nonde- Polynomial Systems," Applied Mathematics and Computation, Vol. 18, pp.
generate. Numerical checks for Roberts cognates and complex 77-86.
Morgan, A. P., 1987, Solving Polynomial Systems Using Continuation for
conjugates provided strong evidence that the solution list is Scientific and Engineering Problems, Prentice-Hall, Englewood Cliffs, NJ.
complete. Morgan, A. P., and Sommese, A. J., 1987, " A Homotopy for Solving General
Of the 1442 solution triples for a set of physical precision Polynomial Systems that Respects m-Homogeneous Structures," Appl. Math.
points, the number of physically meaningful linkages (i.e., real andComput., Vol. 24, pp. 101-113.
Morgan, A. P., and Sommese, A. J., 1989, "Coefficient Parameter Poly-
link lengths) varies. In the examples, this number ranged from nomial Continuation," Appl. Math, and Comput., Vol. 29, pp. 123-160.
21 to 120 cognate triples. Morgan, A. P., and Wampler, C. W., 1989, "Solving a Planar Four-Bar
While the numerical reduction was quite expensive (332 CPU Design Problem Using Continuation," Advances in Design Automation—1989:
hours, IBM 3081), the problem can now be solved in a man- Mechanical Systems Analysis, Design and Simulation, B. Ravani, ed., ASME
DE-Vol. 19-3, pp. 409-416.
ageable time (0.5-3 hours by Transputer, 1-6 CPU hours by Primrose, E. J. F., and Freudenstein, F., 1963, "Geared Five-Bar Motion,
IBM 3090). Although these times might be reduced for these Part 2—Arbitrary Commensurate Gear Ratio," ASME /. of Applied Mechanics,
particular problems by fine tuning the code's run parameters, June, pp. 170-175.
we believe it is more representative of the expected behavior theRoberts, S., 1875, "On Three-Bar Motion in Plane Space," Proceedings of
London Mathematical Society, III (1869-71), pp. 286-319.
not to do so. Roth, B., and Freudenstein, F., 1963, "Synthesis of Path-Generating Mech-
The rapid advances in recent years in techniques for poly- anisms by Numerical Methods," ASME Journal of Engineering for Industry,
nomial continuation and the fact that mainframe speeds can Vol. 85, Series B, No. 3, pp. 298-306.
now be exceeded by a desktop computer assembled from com- Shigley, J. E., and Uicker, J. J., 1980, Theory of Machines and Mechanisms,
McGraw-Hill, New York.
merically available parallel technology have brought about a Subbian, T., and Flugrad, D. R., 1989, "Four-Bar Path Generation Synthesis
great leap in the feasibility and affordability of solving this by a Continuation Method," Advances in Design Automation—1989: Mechan-
and other similarly difficult mechanism synthesis problems. ical Systems Analysis, Design and Simulation, B. Ravani, ed., ASME DE-Vol.
19-3, pp. 425-432.
Tsai, L.-W., and Lu, J.-J., 1989, "Coupler-Point-Curve Synthesis Using
Homotopy Methods," Advances in Design Automation—1989: Mechanical Sys-
Acknowledgments tems Analysis, Design and Simulation, B. Ravani, ed., ASME DE-Vol. 19-3,
pp. 417-424.
We would like to acknowledge the University of Notre Dame Wampler, C. W., and Morgan, A. P., 1991, "Solving the 6R Inverse Position
for providing computer time on their IBM 3081. The third Problem Using a Generic-Case Solution Methodology," Mechanism and Ma-
author would like to thank the National Science Foundation chine Theory, Vol. 26, No. 1, pp. 99-106.
Wampler, C. W., Morgan, A. P . , and Sommese, A. J., 1990, "Numerical
(DMS 87-22330) and the University of Notre Dame for their Continuation Methods for Solving Polynomial Systems Arising in Kinematics,"
support. ASME JOURNAL OF MECHANICAL DESIGN, Vol. 112, No. 1, pp. 59-68.

Journal of Mechanical Design MARCH 1992, Vol. 114/159

Downloaded From: http://mechanicaldesign.asmedigitalcollection.asme.org/ on 05/19/2014 Terms of Use: http://asme.org/terms
 Table 2 Precision points for the example problems
;' Problem 1 Probl :m 2 Piob em 3 Problem 4
0 0.8961867 -0.09802917 0.00 0.0 0.25 0.00 1.000 0.00000000
l 1.2156535 -1.18749100 0.30 -0.1 0.52 0.10 0.875 0.96824583
2 1.5151435 -0.85449808 0.70 0.0 0.80 0.70 0.750 1.32287565
3 1.6754775 -0.48768058 0.82 0.2 1.20 1.00 0.625 1.56124949
4 1.7138690 -0.30099232 0.90 0.4 1.40 1.30 0.500 1.73205080
5 1.7215236 0.03269953 0.90 0.7 1.10 1.48 0.375 1.85404962
6 1.6642029 0.33241088 0.60 0.7 0.70 1.40 0.250 1.93649167
7 1.4984171 0.74435576 0.10 0.5 0.20 1.00 0.125 1.98431348
8 1.3011834 0.92153806 0.00 0.3 0.02 0.40 0.000 2.00000000

Table 3 Program statistics for the example problems

Problem CPU (min.) CPU (min.) Function No. Real
Number IBM 3090 Transputer* Evaluations Mechanisms
1 123 46 2.8 x 106 21
2 74 28 1.8 X 106 45
3 69 25 1.7 x 106 64
4 321 181 9.5 x 106 120
*T800, 48-nodes

at the nonsingular nondegenerate solutions of f(p0, z)=0 (p0 Fig. 3 A new solution to Roth and Freudenstein's problem. (See the
random, complex) and following the continuation paths of discussion of Problem 1.)
f((l-t)p0 + tpuz)=0 (25)
from / = 0 to t= 1. Since, as discussed in the previous section,
every nondegenerate solution off(p0, z) = 0 turned out to also
be nonsingular, we will by this method find every nonde-
generate solution for any set of generally placed precision
points. For some special cases, such as Problem 4 below, we
will have nondegenerate solutions that lie on positive-dimen-
sional solution sets. The method will then find sample solutions
in such sets.
By the Roberts cognate and symmetry relations, iff(p, z)=0
then f(p, R(z))=0 and f(p, S(z))=0, that is, the solution
paths of equation (25) will appear in groups of 6 according to
the cognate and symmetry relations. Therefore, by tracking
one of each of the 1442 cognate triples, we will find all the
solutions of our target problem. By applying the cognate and
symmetry formulas, these 1442 representative solutions yield
4326 distinct mechanisms and 8652 actual solutions. (Recall
that z and S(z) give the same mechanism.)
Our parameter continuation program used the same path
tracking method as described above. Typically, about 1 percent
to 5 percent of the paths presents some numerical difficulty
(depending, for example, on the singularity of the endpoint
set) in which case the program automatically tries one of the
path's cognates. This strategy has proven to be very effective.
Real Problems. Given nine precision points, we compute Fig. 4 A solution to Problem 2. All solutions had branch or order errors.
the 8 path increment vectors 5y and let §,- = &*. Among the
solutions we compute, only those with x=x* a = a* etc., are yet to discover a missing conjugate. Since the pairings change
physically meaningful. (This is analogous to choosing only the from problem to problem, this is further strong evidence that
real solutions in more typical formulations where the coeffi- we have the complete nondegenerate solution list for problem
cients of the polynomial system are real.) Furthermore, by Po-
taking the complex conjugate of equations (10-13), one may
verify that if S,- = 6/ then paired with any solution (x, x, Examples. The precision points for four example problems
a, &,...) is another solution (x* x* a*, a*,...), which we call its are listed in Table 2. The results of our program are sum-
conjugate. Note that physical solutions are self-conjugate in marized in Table 3, which lists the number of physical cognate
this sense, while all other solutions must appear in conjugate triples found for each problem, the CPU times for two ma-
pairs. chines, and the number of function evaluations. This last figure
We note that our original problem p0 was nonphysical in is equal to the total number of prediction and correction steps
that Sj^&f. Accordingly, its solutions do not appear in con- used in tracking all 1442 paths. The two CPU times are for
jugate pairs nor do the continuation paths of equation (25), an IBM 3090 (mainframe) and a 48-node Transputer network
but for a physical problem px the conjugate of each path (desktop). Because each path is tracked independently, the
endpoint should appear. Just as the number of physical (i.e., multiple nodes of the Transputer network operate very effi-
self-conjugate) solutions changes from problem to problem, ciently in parallel to exceed the speed of the mainframe.
the pairing between the nonphysical solutions also changes. Our Problem 1 is Example 2 from Roth and Freudenstein
For physical problems, our program checks for the presence (1963). The same problem was studied by Tsai and Lu (1989).
of the conjugate of each solution in the fully expanded list of Each of these papers presented one linkage for this problem.
8652 solutions. After solving several such problems, we have Our complete solution yielded 21 physical cognate triples (i.e.,

Journal of Mechanical Design MARCH 1992, Vol. 114/157

Downloaded From: http://mechanicaldesign.asmedigitalcollection.asme.org/ on 05/19/2014 Terms of Use: http://asme.org/terms

 three problems, this one generated singular path endpoints,
which accounts for the large CPU time: 181 min. Among the
singular endpoints were various Cardan (elliptic) mechanisms
of both crank-slider and slider-slider types. Neglecting these,
we found 120 physical cognate triples, one of which is shown
in Fig. 6.
Comparison to Alternative Methods
We wish to briefly contrast our method to two alternatives
that have been proposed, both based on continuation. The
"bootstrap" method (Roth and Freudenstein, 1963) used real
variables. This had the advantage that it would never yield a
non-physical complex answer. However, a continuation path
between two real solutions to the nine point problem can have
segments that branch into the complex domain, and thus the
real portions of the path do not connect. Roth and Freudenstein
Fig. 5 This mechanism and its cognates are the only viable solutions
proposed two heuristics to circumvent such difficulties: inter-
to Problem 3. changing which precision point is used as P0, and redirecting
the continuation path according to a quality index. Another
aspect of their method is that they used a fixed step in the
interpolation parameter /, which makes the procedure prone
to path-crossing unless the step is very small. In fact, they
reported obtaining different linkages by changing only the step
size. Of course, the intent of their effort was only to find some
linkages, not all of them, and the scope of the work included
geared five-bars, not just the four-bar equivalent case. In this
context, the work was very successful.
Tsai and Lu (1989) studied only the four-bar case, as we
also have done. Their principal improvement was to work in
complex space. To solve a nine point problem, they proposed
first selecting a subset of five points, solving a synthesis with
prescribed crank angles, and picking four other points on the
resulting coupler curves. This gave several starting linkages,
all satisfying five of the desired nine precision points. Then,
the precision points on the starting coupler curve were given
a small random, complex perturbation to move the problem
out of the reals. As justified by the "Cheater's Homotopy"
(Li, Sauer, and Yorke, 1989), this guaranteed that the contin-
Fig. 6 One of many viable solutions to Problem 4. The long link ex- uation path from start to target would be nonsingular and
tending off the page is about 5.5 times longer than the crank. would yield a nonsingular final solution. However, the final
solution could be complex and therefore nonphysical. As in
63 mechanisms), although most of them have branch or order the bootstrap method, this method can find several different
errors. The problem is difficult because eight of the nine pre- linkages, but because each starting linkage solves a different
cision points lie near a circle. This means that it is nearly nine-point problem, the approach does not provide an orga-
degenerate, so it is not surprising that this problem took a nized way to find all solutions. Also, in either method, the
greater amount of CPU time: 46 min. As expected, the two final linkage often bears little resemblance to the initial one,
previously published solutions were found, along with several and so it may have branch or order errors, bad transmission
others, such as the one shown in Fig. 3, that pass through the angle, etc. with no recourse except to try another initial guess.
precision points in the proper order. However, assuming the Tsai and Lu used the multiple solutions to their five-point
intention is a coupler curve with a smooth arc through Px,... ,P8 problem, along with their Roberts cognates, to generate new
and a simple return through P0, none of the mechanisms would initial guesses.
be acceptable. If this problem was part of an actual design, By finding all solutions, we can pick the one with the most
some reconsideration of the placement of the precision points desirable characteristics. Any numerical method for finding
would be in order. all nondegenerate solutions will have to solve the entire prob-
Problem 2 has points in an oval shape. It took 28 min and lem, including many degenerate solutions, at least once. An
yielded 45 physical cognate triples. In this case, every coupler alternative to our parameter continuation approach is the
curve had either a branch or order defect. Figure 4 shows one closely related Cheater's Homotopy (Li, Sauer, and Yorke,
solution with a branch defect. Having found all solutions, we 1989), which could also be used to find all solutions. However,
know that an adjustment of the precision points is necessary. the numerical reduction proposed there is to solve f(p *,
Otherwise, if we were using trial-and-error, we might attempt z) + b * =0, where both the parameters p * and the additional
many trials before giving up, without knowing even then constants b* are random, complex. Then, for a subsequent
whether the problem was feasible. problem pu use the homotopy
Problem 3 has three points on one circle and four on another f((l-t)p*+tpuz) + (.l-t)b*=0, (26)
so that by the addition of a suitably located dyad connected
to the coupler point one would have a double-dwell six-bar using as start points only the finite solutions from the original
linkage. This problem required 25 min and gave 64 physical calculation. This approach eliminates only the solutions at
cognate triples. Only the mechanism shown in Fig. 5 (and its
cognates) give a simple curve through the points. infinity and does not eliminate degenerate finite solutions. In
Problem 4 has 9 points on the ellipse x?+y2/4= 1 taken at the case of the nine-point problem, only about one-third of
evenly spaced values of xin the first quadrant. Unlike the other the degenerate solutions lie at infinity. Also, the additional
constants b * destroy the Roberts cognate simplification. Con-
158/Vol. 114, MARCH 1992 Transactions of the ASME
Downloaded From: http://mechanicaldesign.asmedigitalcollection.asme.org/ on 05/19/2014 Terms of Use: http://asme.org/terms
 sequently, the end result of the Cheater's Homotopy would We would also like to acknowledge the work of Scott Town-
be a program that tracked at least 90,000 continuation paths, send in developing computer codes.
rather than our 1442. This work represents a cooperative interdisciplinary effort.
The order of the names given on this report has no significance.
There is no primary author.
Summary and Conclusion
We have established that the nine point path synthesis References
problem for four-bar linkages generically has 4326 distinct Alt, H., 1923, "Uber die Erzeugung gegebener ebener Kurven mit Hilfe des
nondegenerate linkages occurring in 1442 cognate triples, and Gelenkvierseits," Zeitschrift fur angewandte Mathematik und Mechanik, Vol.
we have developed a computer program to find all of them by 3, pp. 13-19.
tracking 1442 continuation paths. This is a reduction by a Erdman, A. G., and Sandor, G. N., 1984, Mechanism Design: Analysis and
factor of nearly 4000 from the best total degree (78) previously Synthesis, Vols. 1 & 2, Prentice-Hall, Englewood Cliffs, NJ.
Freudenstein, F., and Sandor, G. N., 1959, "Synthesis of Path-Generating
reported for the problem. The reduction is accomplished in Mechanisms by Means of a Programmed Digital Computer," ASME Journal
two stages: an analytical reduction based on both classical of Engineering for Industry, Series B, Vol. 81, pp. 159-168.
elimination and a 2-homogeneous formulation, and a numer- Li, T. Y., Sauer, T., and Yorke, J. A., 1989, "The Cheater's Homotopy:
ical reduction based on parameter polynomial continuation. An Efficient Procedure for Solving Systems of Polynomial Equations," SIAM
J. of Numerical Analysis, Vol. 26, pp. 1241-1251.
The numerical reduction step tracked 143,360 solution paths Morgan, A. P., 1986, " A Transformation to Avoid Solutions at Infinity for
and established that only 4326 of the endpoints were nonde- Polynomial Systems," Applied Mathematics and Computation, Vol. 18, pp.
generate. Numerical checks for Roberts cognates and complex 77-86.
Morgan, A. P., 1987, Solving Polynomial Systems Using Continuation for
conjugates provided strong evidence that the solution list is Scientific and Engineering Problems, Prentice-Hall, Englewood Cliffs, NJ.
complete. Morgan, A. P., and Sommese, A. J., 1987, " A Homotopy for Solving General
Of the 1442 solution triples for a set of physical precision Polynomial Systems that Respects m-Homogeneous Structures," Appl. Math.
points, the number of physically meaningful linkages (i.e., real andComput., Vol. 24, pp. 101-113.
Morgan, A. P., and Sommese, A. J., 1989, "Coefficient Parameter Poly-
link lengths) varies. In the examples, this number ranged from nomial Continuation," Appl. Math, and Comput., Vol. 29, pp. 123-160.
21 to 120 cognate triples. Morgan, A. P., and Wampler, C. W., 1989, "Solving a Planar Four-Bar
While the numerical reduction was quite expensive (332 CPU Design Problem Using Continuation," Advances in Design Automation—1989:
hours, IBM 3081), the problem can now be solved in a man- Mechanical Systems Analysis, Design and Simulation, B. Ravani, ed., ASME
DE-Vol. 19-3, pp. 409-416.
ageable time (0.5-3 hours by Transputer, 1-6 CPU hours by Primrose, E. J. F., and Freudenstein, F., 1963, "Geared Five-Bar Motion,
IBM 3090). Although these times might be reduced for these Part 2—Arbitrary Commensurate Gear Ratio," ASME /. of Applied Mechanics,
particular problems by fine tuning the code's run parameters, June, pp. 170-175.
we believe it is more representative of the expected behavior theRoberts, S., 1875, "On Three-Bar Motion in Plane Space," Proceedings of
London Mathematical Society, III (1869-71), pp. 286-319.
not to do so. Roth, B., and Freudenstein, F., 1963, "Synthesis of Path-Generating Mech-
The rapid advances in recent years in techniques for poly- anisms by Numerical Methods," ASME Journal of Engineering for Industry,
nomial continuation and the fact that mainframe speeds can Vol. 85, Series B, No. 3, pp. 298-306.
now be exceeded by a desktop computer assembled from com- Shigley, J. E., and Uicker, J. J., 1980, Theory of Machines and Mechanisms,
McGraw-Hill, New York.
merically available parallel technology have brought about a Subbian, T., and Flugrad, D. R., 1989, "Four-Bar Path Generation Synthesis
great leap in the feasibility and affordability of solving this by a Continuation Method," Advances in Design Automation—1989: Mechan-
and other similarly difficult mechanism synthesis problems. ical Systems Analysis, Design and Simulation, B. Ravani, ed., ASME DE-Vol.
19-3, pp. 425-432.
Tsai, L.-W., and Lu, J.-J., 1989, "Coupler-Point-Curve Synthesis Using
Homotopy Methods," Advances in Design Automation—1989: Mechanical Sys-
Acknowledgments tems Analysis, Design and Simulation, B. Ravani, ed., ASME DE-Vol. 19-3,
pp. 417-424.
We would like to acknowledge the University of Notre Dame Wampler, C. W., and Morgan, A. P., 1991, "Solving the 6R Inverse Position
for providing computer time on their IBM 3081. The third Problem Using a Generic-Case Solution Methodology," Mechanism and Ma-
author would like to thank the National Science Foundation chine Theory, Vol. 26, No. 1, pp. 99-106.
Wampler, C. W., Morgan, A. P . , and Sommese, A. J., 1990, "Numerical
(DMS 87-22330) and the University of Notre Dame for their Continuation Methods for Solving Polynomial Systems Arising in Kinematics,"
support. ASME JOURNAL OF MECHANICAL DESIGN, Vol. 112, No. 1, pp. 59-68.

Journal of Mechanical Design MARCH 1992, Vol. 114/159

Downloaded From: http://mechanicaldesign.asmedigitalcollection.asme.org/ on 05/19/2014 Terms of Use: http://asme.org/terms