Journal of Nonlinear and Convex Analysis, vol. 10, pp. 429, 435, 2009
A note on convexity of two signomial functions
Jein-Shan Chen 1 Department of Mathematics National Taiwan Normal University
Taipei 11677, Taiwan Chia-Hui Huang 2
Department of Information Management Kainan University
Taoyuan 33857, Taiwan March 24, 2008
(revised on January 12, 2009)
Abstract. In this note, we provide correct proofs for showing the convexity of two signomial functions which are frequently used in some recent papers [4, 6, 7, 8, 9] by Tsai et al.. Their arguments contain repeated flaws that motivate our work of this note.
Key words. Convexity, positive definite, Hessian matrix, signomial function.
1 Motivation and Basic Concepts
In this note, we consider two signomial functions whose convexity play important roles in some recent papers [4, 6, 7, 8, 9] dealing with geometric programming problems. How- ever, the verifications therein contain some certain flaws and those incorrect arguments are repeatedly appeared and cited. From point of scientific research’s view, we hereby provide correct proofs for them.
First, we recall what signomial function is. A function f : IRn++ → IR defined as f (x) = cxα11xα22· · · xαnn,
where c > 0 and αi ∈ IR for all i, is called a monomial function or simply a monomial, see [2]. Note that the exponents αi of a monomial can be any real numbers, but the
1Member of Mathematics Division, National Center for Theoretical Sciences, Taipei Office. E-mail:
jschen@math.ntnu.edu.tw.
2E-mail:leohuang@mail.knn.edu.tw
coefficient c must be nonnegative. A sum of monomials, namely, a function of the form
f (x) =
∑N k=1
ckxα11kxα22k· · · xαnnk,
where ck > 0 and cik ∈ IR, is called a posynomial function with N terms or simply a posynomial. A signomial is a linear combination of monomials of some positive variables x1, . . . , xn. Generally speaking, signomials are more general than posynomials.
Next, we review some basic concepts and properties of symmetric matrices which will be used in subsequent analysis. These materials can be found in regular textbooks regarding matrix analysis and convex functions, e.g., [1, 3]. Let f be defined on an open convex set D ⊆ IRn and be twice differentiable, it is known that (i) f is convex on D if and only if the Hessian matrix ∇2f (x) is positive semidefinite (p.s.d. for short) at each x ∈ D; (ii) if ∇2f (x) is positive definite (p.d. for short) at each x ∈ D, then f is strictly convex. The converse of (ii) is false, see the counterexample f (x) = x4. Another important criterion for positive definiteness of a symmetric matrix A is via its leading principal minors as below. For convenience, we denote△kas the leading principal minors of A.
Lemma 1.1 Let A be an n× n nonzero symmetric matrix.
(a) If A is positive semidefinite, then all its leading principal minors are nonnegative with not all of them being zero, i.e., △k ≥ 0, k = 1, 2, . . . , n and not all △k = 0.
(b) A is positive definite if and only if all its leading principal minors are positive, i.e.,
△k> 0, for all k = 1, 2, . . . , n.
The positive definiteness of a symmetric matrix can be described not only by its leading principal minors, but also by all principal minors. More specifically, the positivity of any nested sequence of n principal minors of A (not just the leading principal minors) is necessary and sufficient for A to be positive definite (see [3, Theorem 7.2.5]). On the other hand, if all principal minors of A are nonnegative, then A is positive semidefinite (see [3, page 405]).
The converse of Lemma 1.1(a) is false. For example, let A =
1 0 0 0 0 0 0 0 −1
, we have
⟨x, Ax⟩ = x21 − x23 which is not always nonnegative for all x ∈ IR3. But △1 = 1 ≥ 0,
△2 = 0≥ 0, △3 = 0≥ 0. In fact, the converse of Lemma 1.1(a) is true only for n = 2, see [1, page 112]. From the aforementioned discussion, we know that we can not tell the positive semidefiniteness of a symmetric matrix by its leading principal minors whereas we can do it for positive definiteness. Nonetheless, we still can reach the conclusion of the
positive semidefiniteness of a symmetric matrix by the nonnegativeness of its eigenvalues.
This can be seen as below.
Lemma 1.2 Let A be an n× n nonzero symmetric matrix. Then, the followings hold.
(a) A is p.s.d. if and only if all of its eigenvalues are nonnegative with at least one eigenvalue being zero.
(b) A is p.d. if and only if all of its eigenvalues are positive.
To close this section, we state another important relation between ln f (x) and f (x) on their convexity that will be needed for proving our main results, i.e., suppose f is defined on a convex set D⊆ IRnand f (x) > 0 for all x∈ D, then the convexity of ln f(x) implies f (x) being convex. Note that the converse is false, for instance, f (x) = x2 is convex but ln f (x) = 2 ln|x| is not convex.
2 Main Results
Now we are ready to present our main results which show that the following two signo- mial functions are convex functions. As mentioned earlier, signomial functions play an important role in geometric programming. In particular, the convexity of such functions will help in designing solution methods for it which is the main motivation for this note.
Proposition 2.1 Let f1 : IRn++ → IR be defined as f1(x) = c1
∏n i=1
xαii, where c1 > 0 and αi ≤ 0 for all i = 1, 2, . . . , n. Then f1 is a convex function.
Proof. Since c1 > 0, it is enough to show that ef1(x) =
∏n i=1
xαii is convex.
Let g(x)=ln ef1(x)=
∑n i=1
ln xαii=
∑n i=1
αiln xi. Then, we have
∇g(x) = [α1
x1 α2
x2 · · · αn xn
]T
and ∇2g(x) =
−α1
x21 0 · · · 0 0 −α2
x22 · · · 0 ... ... . .. ... 0 0 · · · −αn
x2n
Due to αi ≤ 0 for all i = 1, 2, . . . , n, we know that all eigenvalues of ∇2g(x) are non-
negative which implies (by Lemma 1.2(a)) that ∇2g(x) is positive semidefinite. Thus, g(x)=ln ef (x) is a convex function which yields ef1(x) being a convex function. 2
Proposition 2.2 Let f2 : IRn++ → IR be defined as f2(x) = c2
∏n i=1
xαii, where c2 < 0 and
αi > 0 for all i = 1, 2, . . . , n with 1−
∑n i=1
αi ≥ 0. Then f2 is a convex function.
Proof. It is not hard to compute that [∇f2(x)]i = c2αixαii−1
∏n j=1,j̸=i
xαjj. In other words,
∇f2(x) =
c2α1xα11−1xα22· · · xαnn c2α2xα11xα22−1· · · xαnn
...
c2αnxα11xα22· · · xαnn−1
.
In addition, it can be verified that [∇2f2(x)]
ij = ∂2f2(x)
∂xi∂xj
=
αiαj
xixjf2(x), if i ̸= j, αi(αi− 1)
x2i f2(x), if i = j, namely,
∇2f2(x)
=
c2α1(α1− 1)x−21
∏n i=1
xαii c2α1α2x−11 x−12
∏n i=1
xαii · · · c2α1αnx−11 x−1n
∏n i=1
xαii c2α2α1x−12 x−11
∏n i=1
xαii c2α2(α2− 1)x−22
∏n i=1
xαii · · · c2α2αnx−12 x−1n
∏n i=1
xαii
... ... . .. ...
c2αnα1x−1n x−11
∏n i=1
xαii c2αnα2x−1n x−12
∏n i=1
xαii · · · c2αn(αn− 1)x−2n
∏n i=1
xαii
Moreover, the determinant of ∇2f2(x) can be computed and be shown by induction as det[
∇2f2(x)]
= (−c2)n ( n
∏
i=1
αixnαi i−2 ) (
1−
∑n i=1
αi )
. (1)
Now, we will complete the proof by discussing the following two cases.
Case (i): If 1−
∑n i=1
αi = 0, we will show that yT∇2f2(x) y ≥ 0 for any y ∈ IRn which
says ∇2f2(x) is a positive semidefinite matrix by definition, and hence f2(x) is a convex function under this case. To see this, we first write out the expression of yT∇2f2(x) y as below
yT∇2f2(x) y
= c2
∏n i=1
xαii
α1(α1− 1)x−21 y12+ α1α2x−11 x−12 y1y2 +· · · + α1αnx−11 x−1n y1yn + α2α1x−12 x−11 y1y2+ α2(α2− 1)x−22 y22 +· · · + α2αnx−12 x−1n y2yn
+ ... ... ...
+ αnα1x−1n x−11 y1yn+ αnα2x−1n x−12 y2yn+· · · + αn(αn− 1)x−2n yn2
= c2
∏n i=1
xαii
α1x−11 y1[
(α1− 1)x−11 y1 + α2x−12 y2+· · · + αnx−1n yn] + α2x−12 y2[
α1x−11 y1+ (α2− 1)x−12 y2+· · · + αnx−1n yn]
+ ... ... ...
+ αnx−1n yn[
α1x−11 y1+ α2x−12 y2+· · · + (αn− 1)x−1n yn]
= c2
∏n i=1
xαii
α1x−11 y1[
α1x−11 y1+ α2x−12 y2+· · · + αnx−1n yn− x−11 y1] + α2x−12 y2[
α1x−11 y1+ α2x−12 y2+· · · + αnx−1n yn− x−12 y2]
+ ... ... ...
+ αnx−1n yn[
α1x−11 y1+ α2x−12 y2+· · · + αnx−1n yn− x−1n yn]
= c2
∏n i=1
xαii
{ (
α1x−11 y1+ α2x−12 y2+· · · + αnx−1n yn)2
− (
α1x−21 y21 + α2x−22 y22+· · · + αnx−2n yn2) }
. (2)
Next, we will argue that the whole thing inside the big parenthesis of (2) is nonpositive by applying Cauchy-Schwarz inequality. In order to apply Cauchy-Schwarz inequality, we make the following arrangement:
[(√α1x−11 y1)2
+(√
α2x−12 y2)2
+· · · +(√
αnx−1n yn)2] [ (√
α1)2+ (√
α2)2+· · · + (√ αn)2
]
≥[
α1x−11 y1+ α2x−12 y2+· · · + αnx−1n yn]2
. (3)
Since [(√α1)2
+(√
α2)2
+· · · +(√
αn)2]
= 1, inequality (3) is equivalent to (α1x−11 y1 + α2x−12 y2+· · · + αnx−1n yn
)2
−(
α1x−21 y12+ α2x−22 y22+· · · + αnx−2n yn2)
≤ 0.
This together with c2 < 0 implies that yT∇2f2(x) y ≥ 0 for any y ∈ IRn. Thus, we complete the proof of case (i).
Case(ii): If 1−
∑n i=1
αi > 0, then we know from (1) that
△i = (−c2)i ( i
∏
j=1
αj xiαj j−2 ) (
1−
∑i j=1
αj )
, (4)
where △i denotes the i-th leading principal minor of the Hessian matrix of f2(x). Note that c2 < 0, αi > 0 for all i = 1, 2,· · · , n, and 1 −
∑n i=1
αi > 0. Therefore, it can be seen that △i > 0 for all i = 1, 2,· · · , n, which implies (by Lemma 1.1(b)) that ∇2f2(x) is a positive definite matrix. This says that f2(x) is strictly convex under this case. 2
For Proposition 2.1, Tsai et al. claimed that (e.g. [4, Prop. 5(i)], [6, Prop. 1] and [9, Prop. 2]) all principal minors△k≥ 0 and concluded directly that f1 is a convex function.
As mentioned earlier, this property holds only for n = 2 and is not satisfied for general n≥ 3. For Proposition 2.2, Tsai et al. made the same mistakes again and did not notice that the case 1−∑n
i=1αi = 0 will cause the error therein (e.g. [4, Prop. 5(ii)], [6, Prop.
2] and [9, Prop. 3]).
We want to point out that our results also provide an alternative proof for the main result (Theorem 7) of [5]. Indeed, Maranas and Floudas in [5, Theorem 7] further discuss another condition as below
∃j such that αj ≥ 1 −
∑n i̸=j
αi, and αi ≤ 0, ∀i ̸= j, i = 1, 2, · · · n. (5) to guarantee that f1 defined as in Prop. 2.1 is a convex function. Our approach can be also employed to verify this fact. To see this, we arrange all powers αi in decreasing order. In other words, without loss of generality, we assume
α1 > α2 ≥ · · · ≥ αn. (6)
Notice that condition (5) implies that α1 is positive and all the other α2,· · · , αn are nonpositive with α1 ≥ 1 −∑n
i=2αi. As mentioned in Prop. 2.1, we only need to show that the function ef1(x) =
∏n i=1
xαii is convex. By similar arguments as in the proof of Prop.
2.2, we know that
△bi = (−1)i ( i
∏
j=1
αj xiαj j−2 ) (
1−
∑i j=1
αj )
,
where b△i denotes the i-th leading principal minor of the Hessian matrix of ef1(x). From conditions (5) and (6), it is easily verified that
( 1−
∑i j=1
αj )
< 0 for each i. It is also
not hard to observe that
∏i j=1
αj is positive if i is odd, and is negative if i is even. In other words, for each i there holds
(−1)i ( i
∏
j=1
αjxiαj j−2 )
< 0.
In addition, we observe that b△n= 0 when α1 = 1−∑n
i=2αi. Thus, from all the above, we have either
△b1 > 0,· · · , b△n−1 > 0, b△n > 0 if α1 > 1−
∑n i=2
αi (7)
or
△b1 > 0,· · · , b△n−1 > 0, b△n= 0 if α1 = 1−
∑n i=2
αi. (8)
Then, Lemma 1.1(b) says that ∇2fe1(x) is positive definite for case (7) whereas following the similar arguments as in Prop. 2.2 implies that ∇2fe1(x) is positive semidefinite for case (8). Thus, we conclude that ef1 is also a convex function under condition (5).
Acknowledgement. The authors would like to thank anonymous referees for their com- ments and suggestions which help improve this paper a lot.
References
[1] Berkovitz, L.D. 2002, Convexity and Optimization in IRn, John Wiley & Sons, Inc..
[2] Boyd, S., Vandenberghe, L. 2004, Convex Optimization, Cambridge University Press.
[3] Horn, R.A., Johnson, C.R. 1985, Matrix Analysis, Cambridge University Press, New York.
[4] Li, H.-L., Tsai, J.-F. 2005, Treating free variables in generalized geometric global optimization programs, Journal of Global Optimization, vol. 33, pp. 1–13.
[5] Maranas, C.D., Floudas, C.A. 1995, Finding all solutions of nonlinearly con- strained systems of equations, Journal of Global Optimization, vol. 7, pp. 143–182.
[6] Tsai, J.-F. 2005, Global optimization for nonlinear fractional programming problems in engineering design, Engineering Optimization, vol. 37, pp. 399–409.
[7] Tsai, J.-F., Lin, M.-H. 2006, An optimization approach for solving signomial discrete programming problems with free variables, Computers and Chemical Engi- neering, vol. 30, pp. 1256–1263.
[8] Tsai, J.-F., Li, H.-L., Hu, N.-Z. 2002, Global optimization for signomial discrete programming problems in engineering design, Engineering Optimization, vol. 34, pp.
613–622.
[9] Tsai, J.-F., Lin, M.-H., Hu, Y.-C. 2007, On generalized geometric programming problems with non-positive variables, European Journal of Operational Research, vol.
178, pp. 10–19.