Available online at www.sciencedirect.

com

Applied Mathematics and Computation 204 (2008) 658–670

www.elsevier.com/locate/amc

Necessary and sufficient conditions

for a function to be separable
C.P. Viazminsky
Department of Physics, University of Aleppo, Syria

Abstract

The necessary and sufficient conditions for a function F(x1, . . . , xn) to be totally or partially separable are derived. It is
shown that a function is totally separable if and only if each component of the gradient vector of LnjFj depends only on the
corresponding variable. The conditions of separability are expressed neatly in terms of the matrix FF,ij  F,iF,j which has to
be diagonal if the function is to be totally separable, and has to assume a diagonal block form in order that the function is
partially separable. The conditions of separability are also given without using derivatives. For polynomials in two vari-
ables, the conditions of separability are shown to hold if and only if the product of the first column and the first row of the
coefficients matrix is equal to the matrix itself. This promotes an easy computational scheme for checking, and actually
carrying out, variable separation. The latter results are extended to polynomials in n variables.
Ó 2008 Elsevier Inc. All rights reserved.

Keywords: Functions factorization; Polynomials separation

1. Introduction

Separation of variables in a function is a common practice in special types of ordinary and partial differ-
ential equations [1]. Separation of variables plays an essential role is statistical distributions, where indepen-
dence of random variables is equivalent to the possibility of factorizing the common distribution function to
products of distributions each of which depends on one random variable [2]. The process of factorization of
polynomials was studied in [3], where a separation scheme was designed. The latter scheme however, requires
function’s derivatives for its implementation.
Before getting involved in the technicalities it may be helpful to give some basic general facts underlying the
subject of this work. A function F(x1, . . . , xn), defined on a region D  Rn , is said to be separable with respect
to the variable x1 (or in the variable x1) if it can be written as a product of two functions, defined on D, with
one function depending only on x1, whereas the other is independent of x1. The function F(x1, . . . , xn) is said
totally separable (or just separable) if it is separable in all its variables. When separable, the function
F(x1, . . . , xn) can be factorized into a product of n functions fi(xi) that depend on distinct variables. A function
F(x1, . . . , xn) is called partially separable if it is separable with some, but not all, of its variables. The process in

E-mail address: [email protected]

0096-3003/$ - see front matter Ó 2008 Elsevier Inc. All rights reserved.
doi:10.1016/j.amc.2008.02.007
 C.P. Viazminsky / Applied Mathematics and Computation 204 (2008) 658–670 659

which the function F is factorized to products that depend on distinct variables is called variable separation.
As mentioned earlier, variable separation could be partial or total. It is clear that the factors fi(xi) appearing in
variable separation of a function F are unique up to arbitrary constant multipliers ai, such that Pni¼1 ai ¼ 1.
We finally point out to the close link between additive separability and multiplicative separability. A func-
tion F(x1, . . . , xn) is additively separable, or separable with respect to addition, if it can be written in the form
F ðx1 ; . . . ; xn Þ ¼ g1 ðx1 Þ þ    þ gn ðxn Þ:
It is clear that F is separable with respect to addition if and only if exp(F) is separable with respect to multi-
plication. Or equivalently, F is separable with respect to multiplication if and only if LnjFj is separable with
respect to addition.

2. Separation of one variable

We assume throughout this section that the function F : D  Rn ! R is differentiable with respect to x1
and its partial derivative oF/ox1  F,1 is continuous on D.
Theorem 1. The function F(x1, . . . , xn) is separable in its first variable if and only if
F ;1 =F ð2:1Þ
depends only on x1.

Proof. ()) If F is separable with respect to x1, it can be written in the form
F ¼ f1 ðx1 ÞR1 ðx2 ; . . . ; xn Þ; ð2:2Þ
and hence
F ;1 =F ¼ f1;1 =f1 : ð2:3Þ
As the right hand-side of Eq. (2.3) is a function of x1 alone, so must be its left hand-side.
(Ü) Assume that F,1/F is a function X1of x1 alone:
F ;1 =F ¼ X 1 ðx1 Þ: ð2:4Þ
Integrating both sides of this equation with respect to x1 yields
Z
LnjF =R1 ðx2 ; . . . ; xn Þj ¼ X 1 ðx1 Þdx1 ; ð2:5Þ

where R1 is an arbitrary function which does not depends on x1. Solving (2.5) for F yields
F ¼ f1 ðx1 ÞR1 ðx2 ; . . . ; xn Þ; ð2:6Þ
where
Z 
jf1 ðx1 Þj ¼ exp X 1 ðx1 Þdx1 : ð2:7Þ

If R1(x2, . . . , xn) is left arbitrary, then (2.6) yields all functions that satisfy (2.4). For our objective function
F(x1, . . . , xn), the function R1 is determined by
R1 ðx2 ; . . . ; xn Þ ¼ F ðx1 ; . . . ; xn Þ=f1 ðx1 Þ:  ð2:8Þ
Note that we could have solved Eq. (2.4), using the standard method of integrating a linear partial differ-
ential equation, to obtain the same general solution (2.6).

3. Conditions of total separability

Theorem 2. Assume that the function F(x1, . . . , xn) is of class C1 on its domain of definition D  Rn . In order that
F is totally separable, it is necessary and sufficient that
660 C.P. Viazminsky / Applied Mathematics and Computation 204 (2008) 658–670

F ;i =F ð3:1Þ
depends on xi alone, for i = 1, . . . , n.

Proof. ()) Assume that F is separable in all its variables, i.e. it can be factorized in the form
F ¼ f1 ðx1 Þ    fn ðxn Þ: ð3:2Þ
It is evident that
F ;i =F ¼ fi;i =fi ð3:3Þ
depends on xi alone. This is true for i = 1, . . . , n.
(Ü) Assume that F,i/F is dependent on xi only, where i = 1, . . . , n, and set
F ;i =F ¼ X i ðxi Þ: ð3:4Þ
It is easy to check that the simultaneous system of linear partial differential equation (3.4) has the following
integrals:
F ðx1 ; . . . ; xn Þ ¼ f1 ðx1 ÞR1 ðx2 ; . . . ; xn Þ
¼ f2 ðx2 ÞR2 ðx1 ; x3 ; . . . :; xn Þ
.. ð3:5Þ
.
¼ fn ðxn ÞRn ðx1 ; . . . :; xn1 Þ:
From the last equations it follows that, for each value of i, the ratio
F =fi ðxi Þ ¼ Ri ðx1 ; . . . ; xi1 ; xiþ1 ; . . . ; xn Þ ð3:6Þ
is independent of xi. Multiplying the function F/f1, which is independent of x1, by 1/f2, yields F/f1f2 which is
certainly still independent of x1. Moreover, and if F/f1f2 depends on x2, then its product by f1(x1), namely, f1
(F/f1f2) = F/f2 must depends on x2, which is in contradiction with (3.5). Thus F/f1f2 is independent of both x1
and x2. The proof proceeds by induction to conclude that F/f1f2    fn = const, because this ratio is independent
of all the variables x1, . . . , xn. Thus
F ¼ const:f1  f2    fn ; ð3:7Þ
and the proof is complete. h

Remarks

1. Had we left the constant multiplier in (3.7) arbitrary we would have obtained all functions that satisfy the
system of differential equations (3.4).
2. Since separability of a function F(x1, . . . , xn) with respect to its first (n  1) variables implies its total sep-
arability, a second version of Theorem 2 is:

Theorem 2a
A function F(x1, . . . , xn) of class C1(D) is totally separable if and only if F,i/F is a function of xi alone, for any
set of n  1 chosen variables.

3. Theorem 2 may be deduced as a direct generalization of Theorem 1. Indeed, the function F = f1R1, which is
already separable with respect to x1, is also separable with respect to x2, i.e. it is has the form
F = f1f2R12(x3, . . . , xn), if and only if F,2/F depends on x2 alone. The proof proceeds to conclude that F
is separable with respect to x1, . . . , xn1, and thus having the form:
F ¼ f1    fn1 R12n1 ðxn Þ ð3:8Þ
if and only if the conditions (3.1) are valid for i = 1, . . . , n  1. The last form however shows that F is sep-
arable in all its variables. We may thus take fn  R12  n1(xn), so that the constant in (3.7) is 1.
 C.P. Viazminsky / Applied Mathematics and Computation 204 (2008) 658–670 661

4. The conditions of separability (3.4) may be expressed compactly as

rLnjF j ¼ ðX 1 ðx1 Þ; . . . ; X n ðxn ÞÞ; ð3:9Þ
which is equivalent to
Xn
dLnjF j ¼ X i ðxi Þdxi ; ð3:10Þ
i¼1

which in turn is equivalent to

n 
PR
jF j ¼ c  exp X i ðxi Þdxi ;
i¼1
ð3:11Þ
n
i:e: F ¼ c  P fi ðxi Þ:
i¼1

The content of this remark constitutes another proof of Theorem 2 which may be rephrased in the following
way.

Theorem 2b. A function F of class C1(D) is totally separable if and only if each component of $LnjFj depends
only on the corresponding variable.
If F is separable, the discussion above together with Eq. (3.4), promotes a scheme for its separation by the
formula
Z
n F ;i
F ¼ const: P exp dxi : ð3:12Þ
i¼1 F
This scheme however, is not practical, since it requires performing a number of order 3n operations of partial
differentiation to calculate F,i, divisions to calculate F/F,i and integrations. We shall find much easier separa-
tion schemes in the forthcoming sections.
Theorem 3. A function F(x1, . . . , xn) of class C2(D) is totally separable if and only if the following conditions hold:
FF ;ij  F i F j ¼ 0 ði–jÞ: ð3:13Þ

Proof. By Theorem 2 F is separable if and only if, for each i, (F,i/F) depends on xi alone. These conditions are
equivalent to the following:
ðF ;i =F Þ;j ¼ 0 ði–jÞ:
The theorem follows on noting that
 
F ;i F ;ij F  F ;i F ;j
¼ :  ð3:14Þ
F ;j F2
On account of the symmetry of Eq. (3.13) with respect to its indices, the number of independent conditions that
are to be satisfied by a function F, in order to be separable, reduces to 12 nðn  1Þ conditions corresponding to

FF ;ij  F ;i F ;j ¼ 0 ði < jÞ: ð3:15Þ

The latter conditions have to be proved all valid if the function F is to be separable. For n = 2, we obtain
Scott’s theorem [1,3] which states that a function F(x, y) is separable if and only if
FF ;xy  F ;x F ;y ¼ 0:

4. Separation with respect to sets of variables

Let I be a proper subset of the set {1, 2, . . . n}, and J = {1, . . . , n}  I. The variables xi (i 2 I) will be denoted
collectively by xI, so that a function G that depends only on the variables xi(i 2 I) will be denoted by G(xI).
Similarly, H(xJ) will denote a function which depends only on xj (j 2 J).
662 C.P. Viazminsky / Applied Mathematics and Computation 204 (2008) 658–670

Theorem 4. In order that a function F(x1, . . . , xn) of class C1(D) is separable in the set of variables xI, and hence
capable of assuming the form,
F ¼ GðxI Þ  H ðxJ Þ; ð4:1Þ
it is necessary and sufficient that the functions
F ;i
 gi ðxI Þ ð8i 2 IÞ: ð4:2Þ
F
depend only on xI.

Proof. First, it is clear that if F is separable in the form (4.1), then conditions (4.2) are satisfied, and hence are
necessary. In proving the converse we suppose that the given function F satisfies the conditions (4.2). Let us
look temporarily on the variables xj (j 2 J) in F as parameters, and thus denote the differential of LnjFj, under
this assumption, by dILnjF j. Now
X F ;i
dI LnjF j ¼ dxi
i2I
F
By hypothesis, the right hand-side does not involve the variables xJ, and hence the left hand-side, when xJ are
considered parameters, also does not involve xJ. Thus the right hand-side is a total differential, which we des-
ignate by dg(xI), because the left hand-side is so. Therefore
X F ;j
dLnjF j ¼ dgðxI Þ þ dxj :
j2J
F

Since dLnjFj  dg is a total differential, the sum on the right hand-side must be a total differential, which is
designated by dh(xJ). Therefore LnjFj = g(xI) + h(xj), and F = G(xI)H(xJ), where g(xI) = LnjG(xI)j, and
h(xI) = LnjH(xJ)j. h

Corollary 1. It is clear that a function F is separable in xI if and only if it is separable in xJ, and hence, iff
F ;j
 hj ðxJ Þ ðj 2 J Þ: ð4:3Þ
F
In practice our choice of either set I or J for checking separability will be due to the expected simplicity gained by
such a choice.

Theorem 5. A function F(x1, . . . , xn) of class C2(D) is separable in xI if and only if

FF ;ij  F ;i F ;j ¼ 0 ði 2 I; j 2 J Þ: ð4:4Þ

Proof. If F is separable in xI, then for each (i 2 I) F,i/F depends only on xI and hence (F,i/F),j = 0 for all
(i 2 I,j 2 J). These conditions are obviously equivalent to (4.4). Conversely, if the conditions (4.4) hold, then
(F,i/F),j = 0(i 2 I,j 2 J), which imply that F,i/Fare functions of xI only. By Theorem 4 F is separable in the set
xI.
Let P = {I1, . . . , Ir} be a partition of {1, . . . ,n}. Generalization of the last two theorems to treat separation
with respect to the latter subsets is straightforward. h

Theorem 4g. The function F, which is of class C1(D), is separable in accordance with the partition P iff
F is =F  gis ðI s Þ ðis 2 I s ; s ¼ 1; . . . ; rÞ:

Theorem 5g. The function F, which is of class C2(D), is separable in accordance with the partition P iff
FF ;is it  F ;is F ;it ¼ 0 ðis 2 I s ; it 2 I t ; s < t; s; t ¼ 1; . . . ; rÞ
Note that Theorems 1–3 correspond to particular cases of the last two theorems. Theorem 1 corresponds to
I = {1}, Theorems 2 and 3 corresponds to Is = {xs}, (s = 1, . . . , n).
 C.P. Viazminsky / Applied Mathematics and Computation 204 (2008) 658–670 663

A good tactic to explore whether a function F is separable or not, and if separable, to determine its type of
separability, is to form the matrix
M ij ¼ FF ;ij  F ;i F ;j ði; j ¼ 1; . . . ; nÞ
The function F will be

(i) totally separable if Mij is diagonal,

(ii) separable in the set of variables xI if Mij has a diagonal block form, when the indices {1,. . .,n} are reor-
dered so that I comes first,
(iii) non-separable otherwise.

In the latter results the matrix o2LnjFj/oxioxj can clearly replace Mij.

5. Separation using margins

Definition. Let (x1, . . . , xn) = (xI, xJ). The function F(xI, aJ) will be called the J-margin of F at aJ.
The following theorem shows that a function F is separable in the set of variables I, if and only if its product
by its value at any point (aI, aJ) is equal to the product of its J-margin at aJ by its I-margin at aI.

Theorem 6. A function F(x1, . . . , xn) is separable with respect to the set of variables xI if and only if
F ðaI ; aJ ÞF ðxI ; xJ Þ ¼ F ðxI ; aJ ÞF ðaI ; xJ Þ; ð5:1Þ
where (aI, aJ) is a any point in the domain of F at which F(aI, aJ) – 0. Moreover, if the latter condition holds, then
F can be written as
F ¼ GðxI ÞH ðxJ Þ; ð5:2Þ
where
GðxI Þ ¼ F ðaI ; aJ Þ1 F ðxI ; aJ Þ; H ðxj Þ ¼ F ðaI ; xJ Þ: ð5:3Þ

Proof. If F is separable in the form (5.2), then

F ðaI ; xJ Þ ¼ GðaI ÞH ðxJ Þ;

ð5:4Þ
F ðxI ; aJ Þ ¼ GðxI ÞH ðaJ Þ:
Multiplying the last equations side to side yields
F ðxI ; aJ ÞF ðaI ; xJ Þ ¼ GðaI ÞH ðaJ ÞGðxI ÞH ðxJ Þ ¼ F ðaI ; aJ ÞF ðxI ; xJ Þ:
To prove the converse we assume that the conditions (5.1) hold. Dividing both sides of Eq. (5.1) by F(aI, aJ)
yields the separable form (5.2) of F in which G and H are given by (5.4). h
Even if F is separable in a set of variables I, we cannot a priori specify this set. In fact we have to check the
validity of the conditions (5.1) for all sets that contain one element, two elements, and so on. The number of
n1
conditions that have to be checked is of order 12 fC 1n þ C 2n þ    þ C n1
n g, which is 2  1 conditions. Owing to
the large magnitude of this number when n is large, it may be better to appeal to the criterion (4.4), which
requires less numbers of checking operations. For a small value of n the current criterion (5.1) is indisputably
the easiest.

6. Some properties of separable functions

The following facts are almost direct consequences of the definition of separable functions:

(i) The product of two separable functions is a separable function. (The product function is defined on the
intersection domain.)
664 C.P. Viazminsky / Applied Mathematics and Computation 204 (2008) 658–670

(ii) The reciprocal of a separable function is separable. (If F is defined on D, then F is defined on D  O,
where O is the set of zeros of F. The domain D of 1/F may be extended to points at which F itself is
not defined. Example: F ¼ sin x= cos y; D ¼ R  fð2k þ 1Þ p2 : k 2 Zg; D ¼ R  fkp : k 2 Zg.)
(iii) The set of separable functions S defined on D and having no zeros in D, form a group with respect to
function multiplication. This follows because S is closed under product and taking reciprocals, the sep-
arable function 1 acts as identity, and product is associative.
(iv) If F is separable, then so are F n ðn 2 ZÞ; jF j; jF ja ða 2 RÞ:We proceed now to consider the partial deriva-
tives (mixed and unmixed) of a separable function F. It is assumed that the function F is differentiable on
its domain of definition at least of the order required for these expressions to be meaningful.
(v) It is easy to show that
ok F ok fi F ok fi
k
¼ k Ri ¼ :
oxi oxi fi oxki
In particular F,i = F(fi,i/fi).
r r
orþs F F o fi o fj
(vi) ox r oxs ¼ f f oxr oxs ði–jÞ:
i j i j i j

(vii) From (v) and (vi)

orþs F or F os F
F ¼ ði–jÞ:
oxri oxsj oxri oxsj
For r = s = 1,we get the conditions of separability (4.4).
s þþsr s sr
(viii) F r1 oox1s1 oxsFr ¼ oox1sF1    ooxsFr ;where {i1, . . . , ir} is a subset of {1, 2, . . . ,n}. In particular
i1 ir i1 ir

F n1 F ;12n ¼ F ;1 F ;2    F ;n :
(ix) (F,i/F),j = (F,j/F),i = 0.
(x) Each partial derivative of a separable function of any rank (mixed and unmixed) is separable. This can
be seen from the fact that differentiation with respect to a variable xi affects only the factor fi. This can
also be seen from (viii).

The last fact (x) constitutes a proof of the following obvious theorem.
Theorem 7. A function F that has a certain partial derivative (of any order) which is not separable, is itself not
separable.
The latter theorem can be used as a tool for refuting separability.
Example: F = x3y + x2y2 + xy + y2. As we recognize F,xx = (6x + 2y)y as non-separable, F is also non-
separable.
It is evident that separability of all partial derivatives of a function are not a sufficient condition for its sep-
arability. This best can be seen through a counter example. The function F = x2 + y2 is non-separable, and
yet all its partial derivatives are separable.
If F,1 is not separable then so are all functions of the type F + R1(x2, . . . , xn). If F,12 is not separable then so are
all functions of the form F þ x1 R12 þ x2 R012 , where R12 and R0 12 are arbitrary functions independent of x1 and x2.

7. Separability and linear transformations

We introduce here a new set of variables xi defined by the general linear transformations
Xn
xi ¼ T ij xj þ bi ði ¼ 1; . . . ; nÞ; ð7:1Þ
j¼1

in which Tij is an (n  n) matrix and bi is an n-vector. A function F(x1, . . . , xn) transforms under (7.1) as an
invariant
F ðx1 ; . . . ; xn Þ ¼ F ðx1 ðxÞ; . . . ; xn ðxÞÞ; ð7:2Þ
where the coordinates x1 ; . . . ; xn are denoted collectively by x.
 C.P. Viazminsky / Applied Mathematics and Computation 204 (2008) 658–670 665

The question posed here is that: does separability persist under the transformation (7.1)? If not, is it possible
to impose additional conditions so that a separable function F remains separable? And to what extent do these
additional conditions narrow the class of functions fulfilling this property? The answer to the first part of this
question is intuitively negative. As a counter example, we mention that the separable function F = xy trans-
forms under ðx ¼ x þ y ; y ¼ x  y Þ to the non-separable form F ¼ x2  y 2 : To answer the second and third
parts we revert to Eq. (4.4), which give the necessary and sufficient conditions of separability, and consider
the transform of its left hand-side under (7.1):
X
n X
n
F F ;ij  F ;i F ;j ¼ T ik T jl ðFF ;kl  F ;k F ;l Þ ¼ T ik T jk ðFF ;kk  F ;k F ;k Þ ði–jÞ;
k;l¼1 k¼1

since FF,kl  F,klF,kl = 0 if (k – l). The left hand side vanishes if and only if FF,kk  F,kF,k = 0 for all k. Thus a
function F has to fulfill the equations
FF ;ij  F ;i F ;j ¼ 0 ð7:3Þ
for all i and j, including i = j, in order to remain separable under the transformation (7.1). In the latter case the
set of Eq. (7.3) form a tensor equation (F is an invariant, F,i is a vector, and F,ij is a tensor of second rank) with
respect to the transformation (7.1). This implies that if this equation is valid with respect to one set of vari-
ables, it is valid with respect to all set of variables that result from the original set by a general linear trans-
formation (7.1). The vanishing of the quantities (7.3) for i = j is equivalent to vanishing of (F,i/ F),i = (fi,i/fi),i,
which are generally functions of xi, and need not to vanish. Therefore, a separable function remains separable
under the transformation (7.1) if and only if (fi,i/fi),i = 0 for each i. This is equivalent to fi = exp(ai xi + ci),
where ai and ci are arbitrary constants. Hence a function that retains its separability under the transformation
(7.1) is of the form
" #
Xn
F ¼ exp ai x i þ c ;
i¼1

where ai (i = 1, . . . , n) and c are arbitrary constants. This result might have been guessed from start.
In the following two sections, we determine the conditions imposed on the coefficients of a polynomial
P(x1, . . . , xn) in order to be separable.

8. Variable separation for polynomials

As a starting point we consider the product of two polynomials Xn (x) and Ym(y) of degrees n and m,
respectively:
X
n X
m n X
X m n X
X m
X n ðxÞY m ðyÞ ¼ ai x i  bj y j ¼ ai bj xi y j  cij xi y j : ð8:1Þ
i¼0 j¼0 i¼0 j¼0 i¼0 j¼0

The degree of the leading term anbmxnym, namely n + m, is called the degree of the polynomial. The other
terms have degrees that descend from (n + m  1) to 0. The total number of terms in the product is
(n + 1)  (m + 1) which gives rise to the same number of coefficients cij = aibj. These coefficients however,
are determined by the n + m + 2 values ai and bj. It is clear therefore that severe limitations are imposed
on the coefficients of a polynomial P(x, y) so that it admits variable separation. Note that the condition
cnm – 0, which guarantees that the polynomial is of degree n + m, discards some forms of polynomials as ut-
terly non-separable. To single out such types of anomalous polynomials, we take the term in which i + j is
maximum, say m + n. If there exists any other term xkyl in which k + l = n + m, or k > n, or l > m, then
the polynomial P(x, y) is non-separable. For example, the polynomials

P 1 ¼ x2 y 4 þ x3 y 3 ; P 2 ¼ x3 y 3 þ xy 4 ; P 3 ¼ x2 y 3 þ xy 4 ; P 4 ¼ x3 y 2 þ xy 4

are certainly not separable. A polynomial P(x, y) of degree n + m is denoted by Pnm(x, y). The factors Xn(x)
and Ym(y) appearing in the separation of Pnm(x, y) are unique up to constant multipliers of the form a and
666 C.P. Viazminsky / Applied Mathematics and Computation 204 (2008) 658–670

1/a. With no loss of generality, we shall take the coefficients of the leading terms in the given polynomial and
in its factors equal to 1. Thus we set
cnm ¼ 1; an ¼ 1; bm ¼ 1: ð8:2Þ
This choice is possible since cnm = anbm. Now, the polynomial Pnm(x, y) is separable if and only if there exists
polynomials Xn(x) and Ym(y) such that the following equality:
Xn X m n X
X m
cij xi y j ¼ ai bj x i y j ð8:3Þ
i¼0 j¼0 i¼0 j¼0

holds identically in x and y. This identity holds if and if

cij ¼ ai bj ði ¼ 0; 1; . . . ; n; j ¼ 0; 1; . . . ; mÞ: ð8:4Þ
Eq. (8.4) thus constitute necessary and sufficient conditions for the separability of Pnm(x, y). The latter condi-
tions together with the conventions (8.2) determine, if Pnm was separable, the coefficients of the factor poly-
nomials Xn(x) and Ym(y), as
cim ¼ ai bm ¼ ai ; cnj ¼ an bj ¼ bj : ð8:5Þ
Substituting for ai and bj from the latter equations in (8.4) we obtain the necessary and sufficient conditions
for the separability of Pnm(x, y) in terms of its coefficients
cij ¼ cim cnj ði ¼ 0; 1; . . . ; n; j ¼ 0; 1; . . . ; mÞ: ð8:6Þ
We sum up the results we have obtained above in the following theorem:
Theorem 8

(i) A polynomial Pnm(x, y) is separable if and only if the (product of the first column and first row of its coef-
ficient matrix C is equal to C itself:
0 1 0 1
cnm ¼ 1 cnm1    cn0 cnm ¼ 1
B c cn1m1    cn10 C B C
B n1m C B cn1m C
C¼B B .. .. C
.. C B¼ B .. Cð cnm ¼ 1 cnm1    cn0 Þ:
C ð8:7Þ
@ . .  . A @ . A
c0m c0m1    c00 c0m
(ii) If Pnm(x, y) is separable then the coefficients of its factors Xn(x) and Ym(y) are given by the first column and
first row of C, respectively:
Xn X
m
X n ðxÞ ¼ cim xi ; Y m ðyÞ ¼ cnj y j : ð8:8Þ
i¼0 j¼0
Proof

(i) Follows from (8.6) on noting that (8.7) is just the matrix form of the latter equations.
(ii) Follows from (8.5). h

Denoting the ith row and the jth column of the matrix C by Ri and Cj, respectively, there follows from (8.7)
that each of the following relations is necessary and sufficient for Pnm (x, y) to be separable:
ðiÞ C ¼ ð C1cnm1 C 1  cn0 C 1 Þ:
0 1
R1
B C
B cn1m R1 C
B C
ðiiÞ C ¼ B
B ..
C:
C
B . C
@ A
c0m R1
 C.P. Viazminsky / Applied Mathematics and Computation 204 (2008) 658–670 667

Thus we have:
Theorem 9. Each of the following statements is necessary and sufficient for the polynomial Pnm(x, y)to be
separable:

(i) Each column of the coefficients’ matrix C is a multiple of its first column by the first element of the column
concerned.
(ii) Each row of the coefficients’ matrix is a multiple of the first element of this row by the first row of C.

An algorithm for checking whether Pnm(x, y) is separable, and if it was, to carry out variable separation is
the following:

(i) Write down the coefficients cij as the (n + 1)  (m + 1) matrix C appearing in (8.7).
(ii) Form matrix product Dijof the first column and first row of the matrix C.
(iii) If Dij – Cij, the separation is impossible.
(iv) If Dij = Cij, form the factors Xn(x) and Ym(y) by (8.8).

It is noted that having determined the coefficients of the factors polynomials by (8.5), it suffices to fulfill the
remaining relations of (8.6), namely those corresponding to (i = 0, 1, . . . , n  1; j = 0, 1, . . . , m  1). Thus we
have:
Theorem 10. The polynomial Pnm(x, y) is separable in the form Pnm = Xn(x)Ym(y), where Xn(x) and Ym(y) are
given by (8.8) if and only if its coefficients cij satisfy the matrix equation
0 1 0 1
cn1m1    cn11 cn10 cn1m
B .. .. .. C B . C
B . .  . C B . C
B C ¼ B . Cð cnm1    cn1 cn0 Þ: ð8:9Þ
B C B C
@ c1m1    c11 c10 A @ c1m A
c0m1    c01 c00 c0m

Example 1. Consider the polynomial

P 43 ðx; yÞ ¼ x4 y 3 þ 2x4 y 2  x4 y þ 3x4  3x3 y 3  6x3 y 2 þ 3x3 y  9x3 þ 5x2 y 3 þ 10x2 y 2  5x2 y þ 15x2 þ 2xy 3
þ 4xy 2  2xy þ 6x þ 7y 3 þ 14y 2  7y þ 21:

Since the coefficients matrix is equal to the product of its first column by its first row:
0 1 0 1
1 2 1 3 1
B C B C
B 3 6 3 9 C B 3 C
B C B C
B 5 10 5 15 C B C
B C ¼ B 5 Cð 1 2 1 3 Þ;
B C B C
@ 2 4 2 6 A @ 2 A
7 14 7 21 7

the given polynomial is separable, and

P 43 ðx; yÞ ¼ ðx4  3x3 þ 5x2 þ 2x þ 7Þðy 3 þ 2y 2  y þ 3Þ:
Note that the coefficients of X4(x) are the elements of the first column and those of Y3(y) are the elements of
the first row in the coefficients matrix.
Finally it is noted that one may make use of margins to set up a computational scheme in which Pnm(x, y) is
separated after verifying the possibility of its separation. Indeed, and if c00 – 0,
Xn X
m
P nm ðx; 0Þ ¼ ci0 xi ; P nm ð0; yÞ ¼ c0j y j : ð8:10Þ
i¼0 j¼0
668 C.P. Viazminsky / Applied Mathematics and Computation 204 (2008) 658–670

The factors Xn(x) and Ym(y) are proportional to the latter corresponding margins. Since the coefficients of the
leading terms in these factors are equal to 1, it is evident that these two margins must be divided by cn0 and
c0m, respectively. The latter numbers are not zero if and only if c00 – 0, since c00 = cn0c0m. Thus we state on:
Theorem 11. If c00 – 0, the factors of the separable polynomial Pnm(x, y) are

X n ðxÞ ¼ c1
n0 P nm ðx; 0Þ; Y m ðyÞ ¼ c1
0m P nm ð0; yÞ ð8:11Þ
It is evident that any computational scheme based on the last relations must, either verify first that Pnm(x, y)
is separable, and hence go through the necessary requirements of Theorem 9 or Theorem 10, or alternatively,
perform the product of the factors (8.11), to ensure that this product is identical to Pnm(x, y).

Example 2. We revert to the separable polynomial given in Example 1, and calculate the ingredients of (8.11).

cn0  c40 ¼ 3; c0m  c03 ¼ 7;

P 43 ðx; 0Þ ¼ 3ðx  3x3 þ 5x2 þ 2x þ 7Þ;
4

P 43 ð0; yÞ ¼ 7ðy 3 þ 2y 2  y þ 3Þ:

It is clear that

X n ðxÞ ¼ c1
40 P 43 ðx; 0Þ; Y m ðyÞ ¼ c1
03 P ð0; yÞ

as specified by Theorem 11.

In comment on the latter theorem we note that if Pnm(0, 0) = c00 = 0, then we may use any other point, and
apply Theorem 6. If Pnm(1, 1) – 0, then the un-normalized form of the factors are
!1
X
X n ðxÞ ¼ cij P n ðx; 1Þ; Y m ðyÞ ¼ P nm ð1; yÞ:
ij

9. Polynomials in n variables

In this section, we discuss total separation of polynomials of degree N1 +    + Nn in n variables

X
N1 X
Nn
P N 1 N n ðx1 ; . . . ; xn Þ ¼ ... ci1 in xi11    xinn ; ð9:1Þ
i1 ¼0 in ¼0

where cN 1 N 2 –0. The factors of P N 1 N n , when it is totally separable, are denoted by X1(x1), . . . ,Xn(xn), where

X
Nr
X r ðxr Þ ¼ ari xir ðr ¼ 1; . . . ; nÞ: ð9:2Þ
i¼0

It is evident that the coefficients of P N 1 N n must relate to the coefficients of its factors by the equations
ci1 i2 in ¼ a1i1 a2i2    anin ðir ¼ 0; 1; . . . ; N r ; r ¼ 1; . . . ; nÞ; ð9:3Þ
which constitute necessary and sufficient conditions for its total separability. We choose to take the coefficients
of the leading terms in P and its factors equal to 1:

cN 1 N n ¼ 1; a1N 1 ¼    ¼ anN n ¼ 1: ð9:4Þ

From (9.3) and (9.4) the coefficients of the factors are given by
a1i1 ¼ ci1 N 2 N n ; a2i2 ¼ cN 1 i2 N 3 N n ; . . . ; anin ¼ cN 1 N 2 N n1 in : ð9:5Þ
 C.P. Viazminsky / Applied Mathematics and Computation 204 (2008) 658–670 669

The necessary and sufficient conditions for separability, according to (9.3) and (9.5), are reduced to
ci1 in ¼ ci1 N 2 N n  cN 1 i2 N 3 N n    cN 1 N n1 in : ð9:6Þ

We summarize the latter results by the following theorem:

Theorem 12. Consider the polynomial P N 1 N n ðx1 ; . . . ; xn Þ.

(i) P is totally separable if and only if its coefficients fulfill the relations (9.6).
(ii) If P is separable then the separation form is
X
N1 X
N2 X
Nn
P¼ ci1 N 2 N n xi11  cN 1 i2 N 3 N n xi22    cN 1 N n1 in xinn : ð9:7Þ
i1 ¼0 i2 ¼0 in ¼0

It is noted that the relations (9.5) and (9.6), and no matter how large n is, may be programmed easily, giving
rise to a computational scheme that checks and carries out separation. For hand manipulations the following
obvious theorem, which is given for three variables, proves useful.
Theorem 13. The polynomial PIJK(x, y, z) is totally separable iff

(i) the matrix [cijK] is equal to the product of its first column and first row, and
(ii) the matrices [cijk], (k = K  1, . . . , 1,0) are multiples of [cijK] by cIJk.

Generalization of the latter theorem to n variables should have become clear. In order that a polynomial
P N 1 ...N n ðx1 ; . . . ; xn Þ is separable the following must hold:

– The product of the column ½ci1 N 2 N n  by the row ½cN 1 i2 N 3 N n  must equal to the matrix ½ci1 i2 N 3 N n .
– The matrices ½ci1 i2 i3 N 4 N n ; ði3 ¼ N 3  1; . . . ; 1; 0Þ are multiples of ½ci1 i2 N 3 N n  by cN 1 N 2 i3 N 4 N n .
– The matrices ½ci1 i2 i3 i4 N 5 N n ; ði3 ¼ N 3  1; . . . ; 1; 0; i4 ¼ N 4  1; . . . ; 1; 0Þ are multiples of ½ci1 i2 i3 N 4 N n  by
cN 1 N 2 N 3 i4 N 5 N n , and so on.

If the previous conditions hold, P is factorized as given by (9.7).

Manipulations based on the last generalization become more clumsy as n increases, and it is recommended
to use the computational scheme based on (9.6).
Example 3
P 234 ðx; y; zÞ ¼ x2 y 3 z4 þ 2x2 y 3 z þ x2 yz4 þ 2x2 yz þ 2xy 3 z4 þ 4xy 3 z þ 2xyz4
þ 4xyz þ 3y 3 z4 þ 6y 3 z þ 3yz4 þ 6yz: ð9:8Þ
According to Theorem 12, P234(x, y, z) is separable iff
cijk ¼ ci34 c2j4 c23k :
Utilizing Theorem 13, we fix k temporarily at its maximum value 4, make the substitution c234 = 1 in the last
equations, then check that cij4 = ci34c2j4 benefiting of Theorem 9. Indeed
0 1 0 1
1 0 1 0 1
B C B C
½cij4  ¼ @ 2 0 2 0 A ¼ @ 2 Að 1 0 1 0 Þ: ð9:9Þ
3 0 3 0 3

To carry on the checking process, it suffices to show that the coefficients matrices corresponding to the remain-
ing values of k must appear as multiples of the matrix [cij4] by c23k. Since c233 = c232 = c230 = 0, there must
appear in P the term in z only, with coefficients given by the matrix c231 [cij4] = 2[cij4]. A glance at the given
polynomial shows that this is the case, and hence P is separable. Since only two coefficients c23k, do not vanish,
670 C.P. Viazminsky / Applied Mathematics and Computation 204 (2008) 658–670

with that corresponding to k = 1 being equal to 2 (of course c234 can not vanish, and it is certainly equal to
one) we have by (9.9)
P 234 ðx; y; zÞ ¼ ðx2 þ 2x þ 3Þðy 3 þ yÞðz4 þ 2zÞ:

References

[1] D. Scott, When is an ordinary differential equation separable, American Mathematics Monthly 92 (1985) 22–423.
[2] A.M. Mood, F.A. Graybill, Introduction to the Theory of Statistics, McGraw-Hill Book Company, Inc., New York, 1963 (Chapter 9).
[3] R. Ali, G. Mosa, Factorization of separable variable polynomials, Al Manarah 4 (1999) 43–49.