heuristic methods for computing the minimal multi-homogeneous bézout number
TRANSCRIPT
![Page 1: Heuristic methods for computing the minimal multi-homogeneous Bézout number](https://reader035.vdocuments.net/reader035/viewer/2022081212/575021e11a28ab877ea1fd2c/html5/thumbnails/1.jpg)
Applied Mathematics and Computation 146 (2003) 237–256
www.elsevier.com/locate/amc
Heuristic methods for computing theminimal multi-homogeneous B�eezout number
Ting Li, Zhenjiang Lin, Fengshan Bai *,1
Department of Mathematical Sciences, Tsinghua University, Beijing 100084, PR China
Abstract
The multi-homogeneous B�eezout number of a polynomial system is the number of
paths that one has to follow in computing all its isolated solutions with continuation
method. Each partition of variables corresponds to a multi-homogeneous B�eezoutnumber. It is a challenging problem to find a partition with minimal multi-homogeneous
B�eezout number. Two heuristic numerical methods for computing the minimal multi-
homogeneous B�eezout number are presented in this paper. Some analysis of computa-
tional complexity are given. Numerical examples show the efficiency of these two
methods.
� 2002 Elsevier Inc. All rights reserved.
Keywords: Multi-homogeneous B�eezout number; Continuation method; System of polynomial
equations
1. Introduction
For a system of polynomial equations one may wish to determine the
number of isolated solutions and then compute all of them. Garcia and
Zangwill [3] and Drexler [1] suggest that homotopy continuation could be used
to find the full set of isolated solutions of polynomial systems. During the last
* Corresponding author.
E-mail address: [email protected] (F. Bai).1 Supported by National Science Foundation of China G19871047 and National Key Basic
Research Special Fund G1998020306.
0096-3003/$ - see front matter � 2002 Elsevier Inc. All rights reserved.
doi:10.1016/S0096-3003(02)00540-4
![Page 2: Heuristic methods for computing the minimal multi-homogeneous Bézout number](https://reader035.vdocuments.net/reader035/viewer/2022081212/575021e11a28ab877ea1fd2c/html5/thumbnails/2.jpg)
238 T. Li et al. / Appl. Math. Comput. 146 (2003) 237–256
two decades, this method has been developed to be a reliable and efficient
numerical algorithm.Consider a polynomial system of equations,
P1ðx1; x2; . . . ; xnÞ ¼ 0;
P2ðx1; x2; . . . ; xnÞ ¼ 0;
..
.
Pnðx1; x2; . . . ; xnÞ ¼ 0;
ð1:1Þ
where X ¼ ðx1; x2; . . . ; xnÞ 2 Cn. Denote P ¼ ðP1; P2; . . . ; PnÞT . The classical ho-
motopy method for polynomial system is based on B�eezout theory. The number
of isolated solutions of (1.1), and hence the number of curves one has to follow
in the continuation, is bounded above by the total degree (i.e., B�eezout number)
TD ¼Qn
i¼1 di, where di is the degree of the ith equation Pi. However, TD is
often far larger than the number of isolated solutions that the system (1.1)
really has. Hence it may waste too much time by following those unnecessarycurves.
Morgan and Sommese [10] propose the multi-homogeneous B�eezout theory.It is shown that the multi-homogeneous B�eezout number also gives an upper
bound for the number of isolated solutions of a polynomial system. Different
way of partitioning the variables produces different multi-homogeneous B�eezoutnumber. It is desired to find a partition whose multi-homogeneous B�eezoutnumber is the smallest among all possible variable partitions. In fact the
minimal multi-homogeneous B�eezout number is usually smaller (sometime evenfar smaller) than B�eezout number TD. Thus smaller number of paths is followed
in the multi-homogeneous homotopy method. Wampler [12] presents an ex-
haustive search method on finding the optimal partition of variables. However,
it only works for small problems, since the computational complexity of the
algorithm grows exponentially as a function of n, the number of variables.
In this paper, two heuristic methods (fission method and assembly method)
are presented. The idea of our methods is aroused by evolution methods [4,9].
All possible partitions of variables in our algorithms can be regarded as apopulation of individuals which undergoes fission or assembly transforma-
tions. These individuals strive for survival: a selection scheme, biased towards
fitter individuals, selects the next generation. After some generations, it gives
an individual that hopefully represents the optimal partition or a reasonable
approximation of it. The computational cost of these two methods are much
lower than that of Wampler�s exhaustive search method. Large amount of
numerical computation shows that the optimal partition can be obtained for
most of the problems by fission method and/or assembly method.The plan of this paper is as follows. In the next section, general ideas and
basic concepts for the minimal multi-homogeneous are introduced. Basic
![Page 3: Heuristic methods for computing the minimal multi-homogeneous Bézout number](https://reader035.vdocuments.net/reader035/viewer/2022081212/575021e11a28ab877ea1fd2c/html5/thumbnails/3.jpg)
T. Li et al. / Appl. Math. Comput. 146 (2003) 237–256 239
problems and the method by Wampler [12] are also given. The fission method
is presented in Section 3, and the assembly method is presented in Section 4. InSection 5 a comparison of these two methods with Wampler�s method is given
in several aspects. Numerical results are given in Section 6. Finally, conclusions
and some discussions are given in Section 7.
2. The minimal multi-homogeneous B�eezout number
Consider the polynomial system (1.1). Let X ¼ ðx1; x2; � � � ; xnÞ 2 Cn, and a
partition of it be T ¼ ðxð1Þ; xð2Þ; . . . ; xðmÞÞ, where
xðjÞ ¼ ðxj1 ; xj2 ; . . . ; xjkj Þ; j ¼ 1; 2; . . . ;m:
K ¼ ½k1; k2; . . . ; km� is called the partition vector of T above. The subscripts
clearly should satisfyPm
j¼1 kj ¼ n;m6 n.Assume that the degree of the polynomial PiðxÞ with respect to xðkÞ in (1.1) is
dik, k ¼ 1; 2; . . . ;m. Then the corresponding degree matrix under the partition T
is given by the following:
D ¼
d11 d12 � � � d1md21 d22 � � � d2m... ..
. ...
dn1 dn2 � � � dnm
26664
37775:
The multi-homogeneous B�eezout number of the partition T, denoting by Bm, is
defined as the coefficient of ak11 ak22 � � � akmm in the following polynomial:
/ða1; a2; . . . ; amÞ ¼ ðd11a1 þ d12a2 þ � � � þ d1mamÞ � � � ðdn1a1 þ dn2a2 þ � � �þ dnmamÞ:
It is equivalent to have
Bm ¼X
ði1;i2;...;inÞd1i1d2i2 � � � dnin ; ð2:1Þ
where i1; i2; . . . ; in 2 f1; 2; . . . ;mg, and the integer j ðj ¼ 1; 2; . . . ;mÞ appears inthe index i1; i2; . . . ; in exactly kj times.
Among all possible partitions of variables of a polynomial system, the
minimal multi-homogeneous B�eezout number is denoted by mB. The following
well known example shows the significance of the difference between the
minimal multi-homogeneous B�eezout number and the classical B�eezout number.
Example 2.1. Consider the matrix eigenvalue problem:
Ax ¼ kx; x 2 Cn; k 2 C: ð2:2Þ
![Page 4: Heuristic methods for computing the minimal multi-homogeneous Bézout number](https://reader035.vdocuments.net/reader035/viewer/2022081212/575021e11a28ab877ea1fd2c/html5/thumbnails/4.jpg)
240 T. Li et al. / Appl. Math. Comput. 146 (2003) 237–256
One can view it as a polynomial system of x 2 Cn, k 2 C:
Ax ¼ kx;gTx ¼ 1;
�ð2:3Þ
where g 2 Cn is a randomly chosen vector. Clearly, the classical B�eezout number
of system (2.3) is TD ¼ 2n. But we all know that the eigenvalue problem (2.2)
only has n solutions. Let the partition of variables in (2.3) be fx1; . . . ; xng; fkg.The corresponding Bm is exactly n. Hence mB ¼ n for the system (2.3). The
difference between the minimal multi-homogeneous B�eezout number mB ¼ nand the classical B�eezout number TD ¼ 2n is very dramatic, particularly as n is alittle big.
In order to obtain the minimal over all possible Bm, Wampler�s method has
to visit all possible partitions of variables. Hence the following two sub-
problems has to be considered.Go through all possible partitions. The total number of all possible partitions
is Bell number, denoted by BðnÞ. The combinatorics model for Bell number is:
all possible ways of putting n distinct balls into m identical boxes, where
16m6 n. The following recursive relationship is hold for Bell numbers.
BðnÞ ¼Xn�1k¼0
n� 1
k
� �BðkÞ; Bð0Þ ¼ 1:
An estimation is given by [7]
ðn=2Þðn=2Þ < BðnÞ < n!:
This means that Bell number is exponentially increasing as n grows. For ex-
ample, Bð5Þ ¼ 32, Bð10Þ ¼ 115; 975, Bð15Þ ¼ 1; 382; 958; 545. When n is big, it
is impossible to go through so many partitions. That is the reason whyWampler�s method only works for small problems.
Compute the multi-homogeneous B�eezout number for any given partition. Arecursive algorithm is established by Wampler [12]:
bðD;K; iÞ ¼ 1 if i ¼ nþ 1;Pmj¼1;kj 6¼0 dij � bðD;K � ej; iþ 1Þ otherwise
�
and Bm ¼ bðD;K; 1Þ. Here K ¼ ½k1; k2; . . . ; km� is the partitioning vector and D is
the degree matrix, ej ¼ ð0; 0; . . . ; 0; 1; 0; . . . ; 0Þ 2Zn. The computational cost
of such an algorithm is also exponentially increasing as n grows.
In the following sections, we present two efficient heuristic methods to
finding out the partition which corresponds to mB, the minimal multi-homo-geneous B�eezout number, or an approximation of it, without exhaustive
searching for all partitions. To compare the results properly, Wampler�s re-
![Page 5: Heuristic methods for computing the minimal multi-homogeneous Bézout number](https://reader035.vdocuments.net/reader035/viewer/2022081212/575021e11a28ab877ea1fd2c/html5/thumbnails/5.jpg)
T. Li et al. / Appl. Math. Comput. 146 (2003) 237–256 241
cursive algorithm mentioned above is adopted to computing Bm for any given
partition.
3. Fission method
Let X ¼ ðx1; x2; . . . ; xnÞ be the unknown vector in system (1.1), and a par-
tition of it be
T ¼ ðxð1Þ; xð2Þ; . . . ; xðmÞÞ;
where xðjÞ ¼ ðxj1 ; xj2 ; . . . ; xjkj Þ, j ¼ 1; 2; . . . ;m. For simplicity, we often denote
the partition T above as fG1;G2; . . . ;Gmg, where Gj ¼ fj1; j2; . . . ; jkjg,j ¼ 1; 2; . . . ;m.
Example 3.1. Suppose X ¼ ðx1; x2; . . . ; x5Þ and T ¼ fxð1Þ; xð2Þ; xð3Þg, where
xð1Þ ¼ fx1; x2g, xð2Þ ¼ fx3; x5g, xð3Þ ¼ fx4g. Then T is denoted by f1; 2g;f3; 5g; f4g.
To eliminate the non-uniqueness, we always assume that
ij < il for all j < l;i1 < j1 for all i < j:
This means that variables inside a group are arranged in an increasing order oftheir subscripts, and the groups in a partition are also placed in an increasing
order of the subscripts of their first variables.
Definition 3.1. T1 ¼ ðxð1Þ1 ; xð2Þ1 ; . . . ; xðmþ1Þ1 Þ is called a partition fissured from the
partition T0 ¼ ðxð1Þ0 ; xð2Þ0 ; . . . ; xðmÞ0 Þ if there exists an index i, ð16 i6mÞ such that
xðjÞ1 ¼ xðjÞ0 8j < i;
xðjþ1Þ1 ¼ xðjÞ0 8j > i;
(j ¼ 1; 2; . . . ;m
and
xðiÞ1 [ xðiþ1Þ1 ¼ xðiÞ0 :
Example 3.2. Consider the following
f1; 2; 3g f4; 5; 6g f7; 8g+ fissure
f1; 2; 3g f4gf5; 6g f7; 8g
Denote T1 ¼ ff1; 2; 3g; f4g; f5; 6g; f7; 8gg, T0 ¼ ff1; 2; 3g; f4; 5; 6g; f7; 8gg. ByDefinition 3.1, T1 is a partition fissured from the partition T0.
![Page 6: Heuristic methods for computing the minimal multi-homogeneous Bézout number](https://reader035.vdocuments.net/reader035/viewer/2022081212/575021e11a28ab877ea1fd2c/html5/thumbnails/6.jpg)
242 T. Li et al. / Appl. Math. Comput. 146 (2003) 237–256
Definition 3.2. If T1 is fissured from T0, then T0 is called the parent of T1, and T1is called a child of T0.
It is easy to see that a partition T may have more than one child. We call all
of T�s children its offspring.
Example 3.3. Consider a partition T0 ¼ ff1; 2; 3g; f4; 5gg. Its offspring is the
following set fT11; T12; T13; T14g, where
T11 ¼ f1gf2; 3gf4; 5g; T12 ¼ f1; 2gf3gf4; 5g;T13 ¼ f1; 3gf2gf4; 5g; T14 ¼ f1; 2; 3gf4gf5g:
Assume the partition of variables chosen at step t is denoted by P ðtÞ. One of
the groups in the current partition fissures to generate new partitions. The setof all possible partitions generated from P ðtÞ in this way is called the offspring
of PðtÞ, and is denoted by CðtÞ. The fitness of each element in CðtÞ is evaluatedby computing its multi-homogeneous B�eezout number. A general structure of
the fission algorithm is as follows.
Algorithm 3.1. Fission method
begin
t 1;
initialize P ðtÞ;whileðt < nÞdobegin
fissure P ðtÞ to yield CðtÞ;evaluate CðtÞ;select P ðt þ 1Þ from CðtÞ;t t þ 1;
end
end
The fission method can be thought as a simplified evolution process. Regard
P ðtÞ as a population of individuals at the generation t. Each individual rep-
resents a partition of variables, and it is evaluated by measuring its fitness. One
of the individuals fissures to form new individuals called offspring CðtÞ, arethen evaluated. A new population is formed by selecting the most fit individual
from the offspring population. After several generations, the algorithm con-
verges to one individual, which hopefully represents an optimal or suboptimal
partition of variables.
![Page 7: Heuristic methods for computing the minimal multi-homogeneous Bézout number](https://reader035.vdocuments.net/reader035/viewer/2022081212/575021e11a28ab877ea1fd2c/html5/thumbnails/7.jpg)
T. Li et al. / Appl. Math. Comput. 146 (2003) 237–256 243
Remark 3.1. For the initial population of individuals, one natural idea is to use
T ¼ f1; 2; . . . ; ng since T cannot be produced by any partitions through fis-suring operation. Namely, T has no parents. The corresponding Bm under such
partition T is the classical B�eezout number TD.
Remark 3.2. The evaluation function. It is the criterion to determining indi-
viduals� fitness. The most fit individual can be selected from the offspring
population by using the evaluation function. Bm, the multi-homogeneous
B�eezout number of the variable partition, is regarded as the evaluation function.
Remark 3.3. The Criterion to stop. The process of evolution from the initial
population T ¼ f1; 2; . . . ; ng continues until the nth population Tn ¼ f1g;f2g; . . . ; fng, which cannot be fissured any further. Then the process of evo-
lution ends.
Here we adopt the strategy of storing the optimal individual during theprocess of calculation. The optimal partition obtained by fission method is
denoted by fission partition (FP), and the corresponding multi-homogeneous
B�eezout number Bm is denoted by fission B�eezout (FB). Let us survey the pro-
cedure of Fission method calculation through a simple example.
Example 3.4. Consider a polynomial system:
x21 þ x2 þ 1 ¼ 0;
x1x3 þ x2 þ x4 ¼ 0;
x1 þ x2x3 þ x4 ¼ 0;
x1 þ x3x4 þ 4 ¼ 0:
In the following, we denote the present optimal individual by POI. The Fissionmethod runs as follows:
Pð1Þ : T ¼ f1 2 3 4g;
D ¼
2
2
2
2
2664
3775;
Bm ¼ TD ¼ 16;
POI : T ¼ f1; 2; 3; 4g; Bm ¼ 16:
![Page 8: Heuristic methods for computing the minimal multi-homogeneous Bézout number](https://reader035.vdocuments.net/reader035/viewer/2022081212/575021e11a28ab877ea1fd2c/html5/thumbnails/8.jpg)
244 T. Li et al. / Appl. Math. Comput. 146 (2003) 237–256
By looking for all possible partitions fissured from T ¼ f1; 2; 3; 4g, it gives
Cð1Þ :T1 ¼ f1; 2; 3gf4g T2 ¼ f1; 2; 4gf3g T3 ¼ f1; 2gf3; 4g T4 ¼ f1; 3; 4gf2g
D ¼
2 0
2 1
2 1
1 1
2664
3775 D ¼
2 0
1 1
1 1
1 1
2664
3775 D ¼
2 0
1 1
1 1
1 2
2664
3775 D ¼
2 1
2 1
1 1
2 0
2664
3775
Bm ¼ 16 Bm ¼ 6 Bm ¼ 10 Bm ¼ 16
T5 ¼ f1; 3gf2; 4g T6 ¼ f1; 4gf2; 3g T7 ¼ f1gf2; 3; 4g
D ¼
2 1
2 1
1 11 1
2664
3775 D ¼
2 1
1 1
1 21 1
2664
3775 D ¼
2 1
1 1
1 21 2
2664
3775
Bm ¼ 13 Bm ¼ 14 Bm ¼ 16
The set Cð1Þ contains seven elements. By taking one from Cð1Þ whose corre-
sponding Bm is the smallest, the present optimal individual is obtained as fol-
lows:
POI : T ¼ f1; 2; 4gf3g; Bm ¼ 6:
P ð2Þ should be taken from the set C(1), and here it is identical to the present
optimal individual at this step.
Pð2Þ : T ¼ T2 ¼ f1; 2; 4gf3g:
Similarly, by looking for all possible partitions fissured from T ¼ f1; 2; 4gf3g,it gives
Cð2Þ :
T1 ¼ f1gf2; 4gf3g T2 ¼ f1; 2gf3gf4g T3 ¼ f1; 4gf2gf3g
D ¼
2 1 0
1 1 1
1 1 1
1 1 1
2664
3775 D ¼
2 0 0
1 1 1
1 1 1
1 1 1
2664
3775 D ¼
2 1 0
1 1 1
1 1 1
1 0 1
2664
3775
Bm ¼ 12 Bm ¼ 12 Bm ¼ 11
![Page 9: Heuristic methods for computing the minimal multi-homogeneous Bézout number](https://reader035.vdocuments.net/reader035/viewer/2022081212/575021e11a28ab877ea1fd2c/html5/thumbnails/9.jpg)
T. Li et al. / Appl. Math. Comput. 146 (2003) 237–256 245
The POI, the present optimal individual at this step is the same as last step,
since there is no element in Cð2Þ whose corresponding Bm could be any smallerthan 6.
POI : T ¼ f1; 2; 4gf3g; Bm ¼ 6:
However P ð3Þ has to be an element from Cð2Þ, whose corresponding Bm is the
smallest over all elements in Cð2Þ. Hence
Pð3Þ : T ¼ T3 ¼ f1; 4gf2gf3g;
Cð3Þ :
T1 ¼ f1gf2gf3gf4g
D ¼
2 1 0 0
1 1 1 1
1 1 1 1
1 0 1 1
2664
3775
Bm ¼ 14
POI : T ¼ f1; 2; 4gf3g; Bm ¼ 6;
Pð4Þ : T ¼ T1 ¼ f1gf2gf3gf4g:
Hence the optimal partition of variables obtained by fission method is
FP ¼ f1; 2; 4gf3g, and the corresponding FB ¼ 6.
For n ¼ 4, the total number of all possible partitions is Bð4Þ ¼ 15. The
minimal multi-homogeneous B�eezout number mB ¼ 6. Adopting fission
method, we calculate 12 partitions to obtain the value of the optimum indi-
vidual which equal to mB. As the number of variables n grows, the percentageof partitions calculated among all possible ones will decrease.
4. Assembly method
An analogous algorithm which basically reverses the procedure of fission
method is presented in this section. It is called assembly method.
Definition 4.1. We say that a partition T1 ¼ ðxð1Þ1 ; xð2Þ1 ; . . . ; xðmÞ1 Þ is assembledfrom
T0 ¼ ðxð1Þ0 ; xð2Þ0 ; . . . ; xðmþ1Þ0 Þ;
![Page 10: Heuristic methods for computing the minimal multi-homogeneous Bézout number](https://reader035.vdocuments.net/reader035/viewer/2022081212/575021e11a28ab877ea1fd2c/html5/thumbnails/10.jpg)
246 T. Li et al. / Appl. Math. Comput. 146 (2003) 237–256
if there exists i, j (without loss of generality, we assume that i < j), such
that
xðiÞ1 ¼ xðiÞ0 [ xðjÞ0 ;
xðkÞ1 ¼ xðkÞ0 8k < j and k 6¼ i;
xðkÞ1 ¼ xðkþ1Þ0 8kP j:
8>><>>:
Example 4.1. Consider
f1; 2; 3g f4gf5; 6g f7g+ assemble
f1; 2; 3g f4; 5; 6g f7g
Analogous to fission method, if T1 is assembled from T0, then T0 is called
parent of T1; and T1 is called a child of T0.A general structure of the assembly algorithm is as follows:
Algorithm 4.1. Assembly method
begint 1;
initialize P ðtÞ;whileðt < nÞdobegin
assemble P ðtÞ to yield CðtÞ;evaluate CðtÞ;select P ðt þ 1Þ from CðtÞ;t t þ 1;
end
end
Remark 4.1. In the assembly method, we take T ¼ f1gf2g � � � fng as initial
population since T cannot be generated from any partitions through assem-
bling. The process of evolution from T continues until the nth population
Tn ¼ f1; 2; . . . ; ng, which cannot be assembled. Then the algorithm ends.
Here again, we adopt the strategy of storing the optimal individual in theprocess of calculation. The partition obtained by assembly method is denoted
by assembly partition (AP), and the corresponding Bm is denoted by assembly
B�eezout number (AB).
![Page 11: Heuristic methods for computing the minimal multi-homogeneous Bézout number](https://reader035.vdocuments.net/reader035/viewer/2022081212/575021e11a28ab877ea1fd2c/html5/thumbnails/11.jpg)
T. Li et al. / Appl. Math. Comput. 146 (2003) 237–256 247
5. Complexity of algorithms
In Section 2 we point out that in order to obtain the minimal Bm, all possible
partitions of variables have to be gone through by Wampler�s method. The
total number of all possible partitions is the Bell number which grows expo-
nentially.
In the following, we consider the numbers of partitions which have to be
visited by the fission method and assembly method. This is certainly related to
the complexity of our methods.
Lemma 5.1. There are 2n�1 � 1 different possible cases to partitioning a set withn elements into two subsets.
Proof. Deleting all multiplicities, the number of different possible cases to
partitioning a set with n elements into two subsets should always be
1
2fC1
n þ C2n þ � � � þ Cn�1
n g:
Notice that
C0n þ C1
n þ � � � þ Cnn ¼ ð1þ 1Þn ¼ 2n
and thus the fact
C1n þ � � � þ Cn�1
n ¼ 2n � 2:
Hence the conclusion holds. �
Lemma 5.2. Suppose x; y 2N [ f0g, then 2x þ 2y 6 2xþy þ 1.
Proof. When x ¼ 0 or y ¼ 0, the equality holds. Hence the result holds. When
xy 6¼ 0, one can assume xP y, without loss of generality. Hence
2x þ 2y 6 2� 2x 6 2xþy < 2xþy þ 1:
Thus the conclusion holds. �
Lemma 5.3. Suppose n 2N, m 2N and 16m6 n; xi 2N, i ¼ 1; . . . ;m. IfPmi¼1 xi ¼ n, then ð2x1�1 � 1Þ þ � � � þ ð2xm�1 � 1Þ6 2n�m � 1.
Proof. It is proved by induction on the number m.
(1) Let m ¼ 1, then x1 ¼ n. Thus 2x1�1 � 1 ¼ 2n�1 � 1, which proves the result.
(2) Assume that the result of the lemma holds for m ¼ k, that is,
2x1�1 � 1þ � � � þ 2xk�1 � 16 2n�k � 1:
![Page 12: Heuristic methods for computing the minimal multi-homogeneous Bézout number](https://reader035.vdocuments.net/reader035/viewer/2022081212/575021e11a28ab877ea1fd2c/html5/thumbnails/12.jpg)
248 T. Li et al. / Appl. Math. Comput. 146 (2003) 237–256
Now consider the case of m ¼ k þ 1. By Lemma 5.2, one has
2x1�1 � 1þ 2x2�1 � 16 2x1þx2�2 � 1 ¼ 2ðx1þx2�1Þ�1 � 1:
By the assumption of induction and the inequality above, one gets
2x1�1 � 1þ � � � þ 2xkþ1�1 � 16 2ðx1þx2�1Þ�1 � 1þ � � � þ 2xkþ1�1 � 1
6 2ðx1þx2�1þx3þ���þxkþ1Þ�k � 1 ¼ 2n�1�k � 1
¼ 2n�ðkþ1Þ � 1:
Hence the result is proved. �
Now we consider the number of partitions used by fission method, referring
Algorithm 3.1 for details.
P(1): T1 ¼ f1; 2; . . . ; ng, the partition that the algorithm starts.
C(1): It is obtained by partitioning T1 into all possible two subsets.
By Lemma 5.1, the number of partitions generated here is 2n�1 � 1.
P(2): T2 ¼ fG1;G2g is chosen from Cð1Þ. Suppose jG1j ¼ x1, jG2j ¼ x2.C(2): Case 1: fG11;G12;G2g, where G11 [ G12 ¼ G1.
By Lemma 5.1, the number of partitions generated in this way is
2x1�1 � 1.
Case 2: fG1;G21;G22g, where G21 [ G22 ¼ G2.
By Lemma 5.1, the number of partitions generated in this way is
2x2�1 � 1.
Hence the number of elements in C(2) is bounded above by 2n�2 � 1, by
Lemma 5.3. Analogously, the number of elements in CðkÞ is bounded above by2n�k � 1.
Thus the total number of partitions used by fission method is
Sumf 6 1þ ð2n�1 � 1Þ þ ð2n�2 � 1Þ þ � � � þ ð2n�ðn�1Þ � 1Þ¼ ð2n�1 þ 2n�2 þ � � � þ 21Þ � nþ 2 ¼ 2n � 2� nþ 2 ¼ 2n � n:
This proves the following result.
Theorem 5.1. Let the number of partitions used in fission method is Sumf . Thenone has Sumf ¼ Oð2nÞ.
The result above shows that the number of partitions visited by fission
method is much less than that of Wampler�s method, but it still grows expo-
nentially as n grows.
![Page 13: Heuristic methods for computing the minimal multi-homogeneous Bézout number](https://reader035.vdocuments.net/reader035/viewer/2022081212/575021e11a28ab877ea1fd2c/html5/thumbnails/13.jpg)
T. Li et al. / Appl. Math. Comput. 146 (2003) 237–256 249
Remark 5.1. When each population has a single set fissured from its parent,
the equality above holds. This is the worst case in the computation.
Consider an example as follows.
P(1): T1 ¼ f1; 2; 3; 4g and the number of Cð1Þ is 24�1 � 1;
P(2): T2 ¼ f1; 3; 4gf2g and the number of Cð2Þ is 23�1 � 1 ¼ 24�2 � 1;
P(3): T3 ¼ f1; 3gf2gf4g and the number of Cð3Þ is 22�1 � 1 ¼ 24�3 � 1;
P(4): T4 ¼ f1gf2gf3gf4g.Sumf ¼ 1þ 24�1 � 1þ 24�2 � 1þ 24�3 � 1 ¼ 24 � 4.
For the assembly method, one can easily get the number of partitions used
as follows:
Table
Compa
N
PNm
PNFB
PNA
Suma ¼ 1þ C2n þ C2
n�1 þ � � � þ C22
¼ 1þ nðn� 1Þ2
þ ðn� 1Þðn� 2Þ2
þ � � � þ 1 ¼ Oðn3Þ:
Hence the following result is proved.
Theorem 5.2. The number of partitions that assembly method visits is a poly-nomial function of n.
Notice that the numbers of partitions used by fission method and assembly
method are much less than that of Wampler�s method. In Table 1, N denotes
the number of the unknowns and PNmB, PNFB and PNAB denote the number of
partitions visited by Wampler�s method, fission method and assembly method,
respectively.
The numbers of partitions required in those three methods clearly satisfy
PNmB > PNFB > PNAB. Fission method and assembly method both use less
partitions than Wampler�s method and save time for calculation through suchmodification. The number of partitions computed by assembly algorithm even
grows as a polynomial function of N. But what we care about most is whether
the destination function can reach the optimal value. In the next section, nu-
merical experiments are presented for special examples and also for randomly
generated systems of sparse polynomial equations, respectively.
1
rison of the numbers of partitions
3 4 5 6 7 8 10 12 15
B 5 15 52 203 877 4140 115975 4213597 1382958545
<5 <12 <27 <58 <121 <248 <1014 <4084 <32753
B 5 11 21 36 57 85 166 287 561
![Page 14: Heuristic methods for computing the minimal multi-homogeneous Bézout number](https://reader035.vdocuments.net/reader035/viewer/2022081212/575021e11a28ab877ea1fd2c/html5/thumbnails/14.jpg)
250 T. Li et al. / Appl. Math. Comput. 146 (2003) 237–256
6. Numerical results
A large amount of numerical computation is done with the fission method
and the assembly method, which shows the efficiency of these two methods. All
of our computations in this paper are performed on PIII/700, with C as the
programming language.
6.1. Numerical examples
The fission method and the assembly method are first tested with 20 poly-
nomial systems, which are all concrete problems from science and engineering.
We refer the details of those 20 examples to the appendix of [7]. Computational
results are given. Some comparisons are also made with Wampler�s method.
Notations used are as follows:
N the number of unknown
TD total degree, i.e. the classical B�eezout numbermB the minimal multi-homogeneous B�eezout number
FB the fission B�eezout number, the optimal Bm obtained by fission method
AB the assembly B�eezout number, the optimal Bm obtained by assembly
method
Time time spent on calculations
PN the number of partitions calculated in algorithms
Note that ‘‘<0.0001’’ shown in Table 2 means that the time spent for suchcomputation cannot be measured by the computer we use.
Results of Table 2 clearly show that the computing time and the number of
partitions of the fission method and the assembly method are far less than
those of Wampler�s method, as the scale of problems becomes a little big.
All the computational results by our fission method, reach the optimal so-
lution. The fission B�eezout numbers computed for those 20 examples are all
equals to their minimal multi-homogeneous B�eezout numbers. The assembly
B�eezout numbers, which are obtained by the assembly method, differ fromminimal multi-homogeneous B�eezout numbers only for one problem among 20.
For this one, the assembly B�eezout number is not far away from the minimal
multi-homogeneous B�eezout number.
From Table 2, we can see that the number of partitions which has to be
calculated by the assembly method is the smallest. That is also clear from the
analysis in Section 5. However, Table 2 clearly show that the computing time it
consumes is much more than that of the fission method. By carefully com-
paring the calculation procedures of these two methods, one can easily find thereason for this. The fission method uses more partitions. But about half of
them, the size of the corresponding degree matrix is n� 2. The assembly
![Page 15: Heuristic methods for computing the minimal multi-homogeneous Bézout number](https://reader035.vdocuments.net/reader035/viewer/2022081212/575021e11a28ab877ea1fd2c/html5/thumbnails/15.jpg)
Table 2
Computation results for 20 examples
N TD mB FB AB
Eq. (1) 2 16 10 10 10
Eq. (2) 4 625 384 384 384
Eq. (3) 4 256 96 96 96
Eq. (4) 4 144 62 62 62
Eq. (5) 4 900 450 450 450
Eq. (6) 5 16 16 16 16
Eq. (7) 6 8 8 8 8
Eq. (8) 8 5764801 645120 645120 645120
Eq. (9) 8 128 16 16 16
Eq. (10) 6 64 20 20 20
Eq. (11) 11 2048 320 320 320
Eq. (12) 8 576 193 193 193
Eq. (13) 7 4608 1361 1361 1361
Eq. (14) 10 3628800 3628800 3628800 3628800
Eq. (15) 9 362880 362880 362880 362880
Eq. (16) 6 1024 216 216 216
Eq. (17) 5 108 56 56 56
Eq. (18) 4 1344 368 368 368
Eq. (19) 10 96 44 44 48*
Eq. (20) 8 256 96 96 96
Time-mB Time-FB Time-AB PNmB PNFB PNAB
Eq. (1) 2 < 0.0001 <0.0001 <0.0001 2 2 2
Eq. (2) 4 <0.0001 <0.0001 <0.0001 15 10 11
Eq. (3) 4 <0.0001 <0.0001 <0.0001 15 10 11
Eq. (4) 4 <0.0001 <0.0001 <0.0001 15 10 11
Eq. (5) 4 <0.0001 <0.0001 <0.0001 15 11 11
Eq. (6) 5 <0.0001 <0.0001 <0.0001 52 26 21
Eq. (7) 6 0.0100 <0.0001 <0.0001 203 46 36
Eq. (8) 8 5.5480 0.1100 0.5800 4140 160 85
Eq. (9) 8 1.8620 0.0710 0.6000 4140 181 85
Eq. (10) 6 0.0100 <0.0001 <0.0001 203 45 36
Eq. (11) 11 1490.53 0.4010 1221.95 678570 1187 221
Eq. (12) 8 0.2200 0.0100 0.1700 4140 169 85
Eq. (13) 7 0.2400 0.0100 0.0500 877 108 57
Eq. (14) 10 6113.00 4.3670 209.86 115975 607 166
Eq. (15) 9 192.37 0.9110 15.022 21147 305 121
Eq. (16) 6 0.0100 <0.0001 <0.0001 203 45 36
Eq. (17) 5 <0.0001 <0.0001 <0.0001 52 26 21
Eq. (18) 4 <0.0001 <0.0001 <0.0001 15 11 11
Eq. (19) 10 2967.01 6.4400 89.8420 115975 693 166
Eq. (20) 8 0.8610 0.0100 0.5700 4140 165 85
T. Li et al. / Appl. Math. Comput. 146 (2003) 237–256 251
method requires less partitions, comparing with the fission method. But for
more than half of all partitions, the size of the corresponding degree matrix is
![Page 16: Heuristic methods for computing the minimal multi-homogeneous Bézout number](https://reader035.vdocuments.net/reader035/viewer/2022081212/575021e11a28ab877ea1fd2c/html5/thumbnails/16.jpg)
252 T. Li et al. / Appl. Math. Comput. 146 (2003) 237–256
n� ðn� 1Þ. It is much more expensive to computing Bm for an n� ðn� 1Þdegree matrix. Hence a lot of time has been spent on computing Bm in assemblymethod.
6.2. Further numerical results
In order to carefully compare the efficiency of the fission method, the as-sembly method with Wampler�s method, numerical experiments are done for
sparse polynomial systems which are generated randomly in this section.
Let n denote the number of unknowns, d and s denote the degree and the
sparsity of a polynomial system, respectively. Note that numbers n and d give
number of all possible monomials in a polynomial. Hence the sparsity is the
proportion that monomials appear in the polynomial.
For example, let n ¼ 3, d ¼ 2. Then the set M of all possible monomials is
given by
M ¼ f1; x1; x2; x3; x21; x1x2; x1x3; x22; x2x3; x23g;
where T ¼ 10 gives the number of elements of the set M. Let n ¼ 6, d ¼ 5,
s ¼ 5%, then T ¼ 462, N ¼ 23. That is, each polynomial of the system can
contain not more than 23 monomials.
Some notations used in this section are given as follows.
n the number of unknowns of a polynomial systemd the maximum degree of the polynomial system
s the sparsity of the polynomial system
PS the percentage of success in reaching mB
EB relative error between computing results r and mB, namely
ðr �mBÞ=mB
ET relative error between computing results r and TD, namely
ðTD� rÞ=TDTime-mean the mean of the time spent on calculationPN-mean the mean of the number of partitions used
ETB relative error between total degree TD and multi-homogeneous B�eezoutnumber mB, namely ðTD�mBÞ=TD
Take n ¼ 6, d ¼ 4, s ¼ 5%, and n ¼ 8, d ¼ 3, s ¼ 5%. 400 instances are
computed for each case. Computational results are given in Table 3.
In Table 3, numerical examples indicate that one can obtain the optimal
partition for almost all cases by these two methods. It shows that our twomethods are extraordinary efficient. Note that the assembly method requires
less partitions than the fission method, by the analysis in Section 5 and also
numerical results in Table 3. However the assembly method takes more com-
![Page 17: Heuristic methods for computing the minimal multi-homogeneous Bézout number](https://reader035.vdocuments.net/reader035/viewer/2022081212/575021e11a28ab877ea1fd2c/html5/thumbnails/17.jpg)
Table 3
Computation results for random examples
PS (%) Time-mean PN-mean
n ¼ 6, d ¼ 4, s ¼ 5%
Wampler�s method 100 0.0178 203
Fission method 97.25 0.0028 49
Assembly method 96.75 0.0053 36
n ¼ 8, d ¼ 3, s ¼ 5%
Wampler�s method 100 6.4485 4140
Fission method 96.25 0.0692 177
Assembly method 96.75 0.2967 85
T. Li et al. / Appl. Math. Comput. 146 (2003) 237–256 253
puter time, shown by the numerical results. That is because computing the
multi-homogeneous B�eezout number for a partition with more groups is much
more expansive than that for a partition with less groups.
More computation is done for different sparsity. The ratio of success is
plotted in the Fig. 1.
It is shown by Fig. 1 that the percentage that the computational results bythe fission or assembly method reach mB, decreases as the sparsity s decrease.
Fig. 1. The percentage of success to reaching mB.
![Page 18: Heuristic methods for computing the minimal multi-homogeneous Bézout number](https://reader035.vdocuments.net/reader035/viewer/2022081212/575021e11a28ab877ea1fd2c/html5/thumbnails/18.jpg)
Table 4
Difference between TD and mB (n ¼ 6, d ¼ 3)
ETB 0 >0 and <0.3 P0.3 and <0.5 P0.5 and <0.7 P0.7
s ¼ 8% 53% 22.5% 10.5% 10% 4%
s ¼ 4% 1% 6.5% 19% 30% 43.5%
Table 5
Distribution of the difference between AB and mB (n ¼ 6, d ¼ 3, s ¼ 4%)
<0.02 P0.02 and
<0.05
P0.05 and
<0.1
P0.1 and
<0.2
P0.2 and
<0.5
P0.5
EB 29.11% 32.91% 15.19% 11.4% 3.8% 7.59%
Table 6
Difference between AB and mB (n ¼ 6, d ¼ 3, s ¼ 4%)
<0.02 <0.05 <0.1 <0.2 <0.5
EB 29.11% 62.02% 77.21% 88.61% 92.41%
254 T. Li et al. / Appl. Math. Comput. 146 (2003) 237–256
It sounds unreasonable. The results shown by Table 4 give a solid answer for
that. As s ¼ 8%, among all 400 polynomial system we computed, 53% of them
make mB ¼ TD, and only 4% of them make mB and TD dramatically differ-
ent. However as the sparsity s ¼ 4%, only 1% of them make mB ¼ TD, while
43.5% of them make mB and TD dramatically different. Hence at the cases of slarger, although one get mB through computation for many cases, it may only
give a trivial, since mB ¼ TD holds at a very high probability. As the sparsity sgoes smaller, one can hope to get more non-trivial results.
As a heuristic method, it is acceptable to have an approximate optimal
solution through the computation. Then the distance between the approximate
solution and the exact solution is what we concern. Some statistics of the non-
success cases for n ¼ 6, d ¼ 3, s ¼ 4% by assembly method is given in Tables 5
and 6.
Statistics in Tables 5 and 6 clearly shows that those computational results
obtained by assembly method which do not equal to mB, are mostly good
approximations of mB. Similar conclusion hold for other case of n and d, and itis also true for fission method.
7. Conclusions and discussions
The fission method and the assembly method are proposed and analyzed in
this paper. These methods, supporting by a large amount of numerical com-
![Page 19: Heuristic methods for computing the minimal multi-homogeneous Bézout number](https://reader035.vdocuments.net/reader035/viewer/2022081212/575021e11a28ab877ea1fd2c/html5/thumbnails/19.jpg)
T. Li et al. / Appl. Math. Comput. 146 (2003) 237–256 255
putation, are fairly efficient. Comparing with Wampler�s method [12] and Li
and Bai�s method [7], the number of partitions that our methods use, reducesdramatically. That of the assembly method is even reduced to n3, which is
substantial as n increases.
Hence we are able to make the size of the problems larger in our
computations, but still limited. The reason is that we have to compute the
multi-homogeneous B�eezout number for each given partition of variables.
Wampler�s algorithm with row expansion recursion [12], which is the best
method so far to our knowledge, is used. This makes the whole computation
still too costly for large systems at present. Thus an efficient algorithm ofcomputing multi-homogeneous B�eezout number for a given partition remains
an open question.
Combinatorial geometry has come to play an important role in homotopy
methods during the last several years [2,5,6,8,11]. With some combinatorial
geometric techniques, one can get another bound, which is called the stable
mixed volume, for the number of the isolated solutions. It may be tighter than
the minimal multi-homogeneous B�eezout number. Huber and Sturmfels [5]
constructed the starting system with random lifting and convex hull techniques.Such a procedure is called polyhedral homotopy. However it is very expensive
to obtain the starting system for the current polyhedral homotopy method.
Therefore, we may obtain a partition with a satisfactory multi-homogeneous
B�eezout number within a reasonable amount of time by our methods. It may be
larger than the stable mixed volume. But the total computational cost for the
multi-homogeneous homotopy method may still be less.
References
[1] F.J. Drexler, Eine Methode zur Berechnung samtlicher Losungen von Polynomgleichungs-
systemen, Numer. Math. 29 (1977) 45–58.
[2] T.A. Gao, T.Y. Li, X.S. Wang, Finding isolated zeros of polynomial systems in Cn with stable
mixed volumes, J. Symbolic Comput. 28 (1999) 187–211.
[3] C.B. Garica, W.I. Zangwill, Finding all solutions to polynomial systems and other systems of
equations, Math. Programming 16 (1979) 159–176.
[4] M. Gen, R. Cheng, Genetic Algorithms and Engineering Optimization, Wiley-Interscience,
New York, 2000.
[5] B. Huber, B. Sturmfels, A polyhedral method for solving sparse polynomial systems, Math.
Comput. 64 (1995) 1541–1555.
[6] B. Huber, B. Sturmfels, Bernstein�s theorem in affine space, Discrete Comput. Geom. 17 (1997)
137–141.
[7] T.J. Li, F. Bai, Minimizing multi-homogeneous B�eezout number by a local search method,
Math. Comput. 70 (2000) 767–787.
[8] T.Y. Li, Numerical solution of multivariate polynomial systems by homotopy continuation
methods, Acta Numer. (1997) 399–436.
[9] Z. Michalewicz, Genetic AlgorithmsþData Structures ¼ Evolution Programs, Springer-
Verlag, 1992.
![Page 20: Heuristic methods for computing the minimal multi-homogeneous Bézout number](https://reader035.vdocuments.net/reader035/viewer/2022081212/575021e11a28ab877ea1fd2c/html5/thumbnails/20.jpg)
256 T. Li et al. / Appl. Math. Comput. 146 (2003) 237–256
[10] A.P. Morgan, A.J. Sommese, A homotopy for solving general polynomial systems that respect
m-homogeneous structures, Appl. Math. Comput. 24 (1987) 101–113.
[11] J. Verschelde, K. Gatermann, R. Cools, Mixed-volume computation by dynamic lifting
applied to polynomial systems solving, Discrete Comput. Geom. 16 (1996) 69–112.
[12] C.W. Wampler, B�eezout number calculations for multi-homogeneous polynomial systems,
Appl. Math. Comput. 51 (1992) 143–157.