Formal Power Series and Algebraic Combinatorics

Nankai University, Tianjin, China, 2007

Inequalities for Symmetric Means

Allison Cuttler, Curtis Greene, and Mark Skandera

Abstract. We study Muirhead-type generalizations of families of inequalities due to Newton, Maclaurin

and others. Each family is defined in terms of a commonly used basis of the ring of symmetric functions in n
variables. Inequalities corresponding to elementary symmetric functions and power sum symmetric functions
are characterized by the same simple poset which generalizes the majorization order. Some analogous results
are also obtained for the Schur, homogeneous, and monomial cases.

Résumé. Nous étudions les inégalités de Muirhead et ses analogues qui géneralise les inégalités de Newton,
Maclaurin, et des autres. Chaque famille d’inégalités correspond a une base de l’anneau de polinômes
symmétriques avec n variables. Avec un seule poset qui géneralise l’ordre “majorization”, nous caracterisons
les familles qui d’inégalités qui correspondent aux bases {eλ } et {pλ }. Nous avons des résultats analougues
pour les bases {hλ }, {sλ }, {mλ }.

1. Introduction
Commonly used bases for the vector space Λrn of homogeneous of degree r symmetric functions in
n variables x = (x1 , . . . , xn ) are the monomial symmetric functions {mλ (x) | λ ⊢ r}, elementary symmetric
functions {eλ (x) | λ ⊢ r}, (complete) homogeneous symmetric functions {hλ (x) | λ ⊢ r}, power sum symmetric
functions {pλ (x) | λ ⊢ r}, and Schur functions {sλ (x) | λ ⊢ r}. (See [Sta99, Ch. 7] for definitions.) When
r = 0 we adopt the conventions m0 (x) = e0 (x) = h0 (x) = s0 (x) = 1 and p0 (x) = n.
To each element gλ (x) of these bases, we will associate a term-normalized symmetric function Gλ (x)
and a mean Gλ (x) by
gλ (x) p
d
(1.1) Gλ (x) = , Gλ (x) = Gλ (x).
gλ (1n )
Note that {Gλ (x) | λ ⊢ r} forms a basis of Λrn , and that the functions {Gλ (x) | λ ⊢ r}, while symmetric, are
not polynomials in x and therefore do not belong to the ring of symmetric functions Λn . In the definition
of Gλ (x), we assume r > 0.
The term symmetric mean is often used to describe a function G(x) : Rn≥0 → R≥0 which is symmetric
in x1 , . . . , xn and which for all a, b ∈ Rn≥0 , c ∈ R≥0 satisfies
(1) min(a) ≤ G(a) ≤ max(a),
(2) a ≤ b (componentwise) implies G(a) ≤ G(b),
(3) limb→0 G(a + b) = G(a).
(4) G(ca) = cG(a).
(See, e.g., [Bul03, p. 62].) For fixed n and two means F, G, we will write F(x) ≤ G(x) or G(x) − F(x) ≥ 0
if we have F(a) ≤ G(a) for all a ∈ Rn≥0 . We define the inequality F (x) ≤ G(x) analogously. Note that if the
degrees of F (x) and G(x) are equal, then we have F (x) ≤ G(x) if and only if we have F(x) ≤ G(x).

2000 Mathematics Subject Classification. Primary 05E05; Secondary 26E60.

Key words and phrases. symmetric functions, means, partitions.

1
2 Allison Cuttler, Curtis Greene, and Mark Skandera

The study of inequalities in symmetric means has quite a long history. (See, e.g., [Bul03], [HLP34].)
Perhaps the best known such inequality is that of the arithmetic and geometric means,
E1 (x) ≥ En (x).
Other symmetric function inequalities which may be stated in terms of means are due to
(1) Muirhead [Mui03]: For λ, µ ⊢ r,
Mλ (x) ≤ Mµ (x) if and only if µ majorizes λ, equivalently,
Mλ (x) ≤ Mµ (x) if and only if µ majorizes λ,
(2) Maclaurin [Mac29]: For 1 ≤ i ≤ j ≤ n,
Ei (x) ≥ Ej (x),
(3) Newton [New, p. 173]: For 1 ≤ k ≤ n − 1,
Ek,k (x) ≥ Ek+1,k−1 (x), equivalently,
Ek,k (x) ≥ Ek+1,k−1 (x),
(4) Schlömilch [Sch58]: For 1 ≤ i ≤ j,
Pi (x) ≤ Pj (x),
(5) Gantmacher [Gan59, p. 203]: For k ≥ 1,
pk,k (x) ≤ pk+1,k−1 (x), equivalently,
Pk,k (x) ≤ Pk+1,k−1 (x), equivalently,
Pk,k (x) ≤ Pk+1,k−1 (x),
(6) Popoviciu [Pop34]: For 1 ≤ i ≤ j,
Hi (x) ≤ Hj (x),
(7) Schur [HLP34, p. 164]: For k ≥ 1,
Hk,k (x) ≤ Hk+1,k−1 (x), equivalently,
Hk,k (x) ≤ Hk+1,k−1 (x).
It is easy to use the inequalities of Newton, Gantmacher, and Schur to derive the inequalities of Maclaurin,
Schlömilch, and Popoviciu, respectively. Furthermore, we will show in Section 4 that Muirhead’s inequalities
imply those of Newton and Gantmacher.
Note that the term-normalized symmetric functions and mean which we have associated to a symmetric
function in (1.1) are defined only for a finite number n of variables. Nevertheless, we may essentially
eliminate dependence upon n from the inequalities enumerated above by considering them to be inequalities
in sequences of functions,
G = (G(x1 ), G(x1 , x2 ), G(x1 , x2 , x3 ), . . . ),
G = (G(x1 ), G(x1 , x2 ), G(x1 , x2 , x3 ), . . . ).
More generally, we will define partial orders on such sequences by declaring F ≤ G if we have F (x) ≤ G(x)
for all n > 0, and F ≤ G if we have F(x) ≤ G(x) for all n > 0. The principal goal of this paper is to
characterize the infinite partial orders on the sequences {Gλ | λ ⊢ 1, 2, . . . } corresponding to the common
bases of the ring of symmetric functions.
Note that Muirhead’s inequalities are indexed by pairs of arbitrary partitions of a fixed integer, while
other inequalities are indexed by one- or two-part partitions, not necessarily of a fixed integer. Using the
common bases of Λrn , we will state and prove inequalities in pairs (Gλ , Gµ ) of means for which λ and µ are
arbitrary and do not necessarily partition the same integer. We consider elementary means and power sum
means in Sections 2-3, and other means and inequalities in Section 4.
 INEQUALITIES FOR SYMMETRIC MEANS 3

2. Elementary Means
Generalizing Maclaurin’s and Newton’s inequalities are inequalities of the form Eλ ≤ Eµ , which we
characterize in Theorem 2.3 in terms of sequence majorization. Let α = (α1 , α2 , . . . ) be a weakly decreasing
nonnegative sequence. If the components of α are integers (respectively, rationals) which sum to r, we say
that α is an integer partition (respectively, a rational partition) of r and write α ⊢ r or |α| = r.
Given two sequences α, β, with |α| = |β| < ∞, we write α  β and say that β majorizes α if for each
index i the corresponding initial partial sums satisfy
α1 + · · · + αi ≤ β1 + · · · + βi .
(If necessary, append zeros to the end of either sequence so that the partial sums are defined.) For c ∈ R≥0 ,
d ∈ N, we define operations of scalar multiplication α 7→ cα, and repetition α 7→ αd by
cα = (cα1 , cα2 , . . . ),
αd = (α1 , . . . , α1 , α2 , . . . , α2 , , . . . ).
| {z } | {z }
d d

Given an integer partition λ = (λ1 , λ2 , . . . ), we will define the transpose partition λ⊤ = (λ⊤1 , λ⊤2 , . . . ) by
λ⊤j = max{i | λi ≥ j}.
The following facts are either well-known or easy to see.
Observation 2.1. Let λ, µ be integer partitions with |λ| = |µ|, let c be a positive real number, and let
d be a positive integer.
(1) (λ⊤)⊤ = λ.
(2) (dλ⊤)⊤ = λd .
(3) (((λ⊤)d )⊤ = dλ.
(4) λ  µ ⇐⇒ cλ  cµ ⇐⇒ λd  µd ⇐⇒ λ⊤  µ⊤.
For λ = (λ1 , . . . , λℓ ) ⊢ r, the term-normalized elementary symmetric function Eλ (x) is given by
e (x)
Eλ (x) = n

λ
,
λ1 · · · λnℓ
p
and the corresponding elementary mean is the rth root of this, Eλ (x) = r Eλ (x). Since λd is a partition of
dr, we have the following stability property of elementary means under the repetition operation.
Observation 2.2. For any partition λ and integer d ≥ 1 we have Eλ = Eλd .
The following theorem shows that inequalities of the form Eλ ≤ Eµ are characterized by the majorization
order on rational partitions of 1.
Theorem 2.3. Given integer partitions λ and µ, we have
λ⊤ µ⊤
Eλ ≤ Eµ if and only if  .
|λ| |µ|
Proof. First let us consider the case that |λ| = |µ|. This case, asserting the equivalence of Eλ ≤ Eµ
and λ⊤  µ⊤, was first proved in [Cut06, Thm. 5.7].
(⇒) Suppose that λ⊤  µ⊤. Then for some index i we have
λ⊤1 + · · · + λ⊤i > µ⊤1 + · · · + µ⊤i .
Choosing a number n ≥ max(λ1 , µ1 ) and specializing the symmetric functions Eλ (x), Eµ (x) at
x1 = · · · = xi = t,
xi+1 = · · · = xn = 1,
we obtain polynomials in N[t] of degrees λ⊤1 + · · · + λ⊤i and µ⊤1 + · · · + µ⊤i , respectively. Thus we have
lim [Eλ (t, . . . , t, 1, . . . , 1) − Eµ (t, . . . , t, 1, . . . , 1)] = ∞,
t→∞

which implies that Eλ  Eµ .

4 Allison Cuttler, Curtis Greene, and Mark Skandera

(⇐) Now suppose that λ⊤  µ⊤, or equivalently that λ  µ, and write λ = (λ1 , . . . , λℓ ). If λ covers µ in
the majorization order, then there exist indices 1 ≤ j < k ≤ ℓ for which we have
µ = (λ1 , . . . , λj−1 , λj − 1, λj+1 , . . . , λk−1 , λk + 1, λk+1 , . . . , λℓ ).
Thus Eµ (x) − Eλ (x) is equal to
Eλ (x)
(Eλj −1 (x)Eλk +1 (x) − Eλj (x)Eλk (x)).
Eλj (x)Eλk (x)
Rewriting Newton’s inequalities as
E1 (x) E2 (x) E3 (x)
≥ ≥ ≥ ··· ,
E0 (x) E1 (x) E2 (x)
we see that Eλj (x)Eλk (x) ≤ Eλj −1 (x)Eλk +1 (x), which implies that Eλ ≤ Eµ . If λ does not cover µ in the
majorization order, then there exists a sequence of partitions
µ = ν (0) ≤ ν (1) ≤ · · · ≤ ν (m) = λ,
in which each comparison of consecutive partitions is a covering relation, such that we have
m−1
X
Eµ (x) − Eλ (x) = (Eν (i) (x) − Eν (i+1) (x)) ≥ 0,
i=0

and consequently Eλ ≤ Eµ .
Now consider the case that |λ| and |µ| are not equal. By Observation 2.2, we have
Eλ (x) = Eλ|µ| (x), Eµ (x) = Eµ|λ| (x).
Since λ|µ| and µ|λ| are both partitions of |λ| · |µ|, we have Eλ ≤ Eµ if and only if (λ|µ| )⊤  (µ|λ| )⊤. By
λ⊤ µ⊤
Observation 2.1 this is equivalent to the condition |λ|  |µ| . 

From Theorem 2.3, we see that the poset describing inequalities Eλ ≤ Eµ of elementary means is
isomorphic to the dual of the majorization order on rational partitions of 1. For all r ≥ 1, this poset
contains the (dual of the) majorization order on partitions of r. Figure 1 shows the restriction of this poset
to elementary means Eλ corresponding to integer partitions of 1, . . . , 6. Note that some integer partitions
do not appear in this poset because of the equalities implied by Observation 2.2. Partitions of integers
dividing 6 have been emphasized to show the embedding in this poset of the dual of the majorization order
on partitions of 6.
From Theorem 2.3, we see that this poset is not finitary: between any two partitions, there must be
another. In particular, if |λ| = |µ| = r and λ is covered by µ in the majorization order on partitions of r,
then λ2 is not covered by µ2 in the majorization order of parititions of 2r.
Define the semiring N to be the set of all nonnegative linear combinations of products of symmetric
functions of the forms
{Ei,j (x) − Ei+1,j−1 (x) | 1 ≤ j ≤ i ≤ n − 1} ∪ {Ei (x) | 1 ≤ i ≤ n}.
The proof of Theorem 2.3 shows that each difference Eλ − Eµ with µ majorizing λ belongs to N.

3. Power Sum Means

We now generalize Schlömilch’s and Gantmacher’s inequalities by characterizing inequalities of the form
Pλ ≤ Pµ . Note that for λ = (λ1 , . . . , λℓ ) ⊢ r, the term-normalized power sum symmetric function Pλ (x) is
given by
pλ (x)
Pλ (x) = ,
nℓ
p
and the corresponding power sum mean is the rth root of this, Pλ (x) = r Pλ (x). Like the elementary basis
{eλ (x) | λ ⊢ r} of Λrn , the power sum basis {pλ (x) | λ ⊢ r} is multiplicative. We therefore have the following
equalities.
Observation 3.1. For any partition λ and integer d ≥ 1 we have Pλ = Pλd .
 INEQUALITIES FOR SYMMETRIC MEANS 5

21111

2111

211

21

221

2 3111

321 311

32 31

3 411

42 41

51 4

Figure 1. Equivalence classes of partitions of 1, . . . , 6 ordered by inequalities in elementary means.

To prove a power sum analog of Theorem 2.3, we whall use the following alternative characterization of
the majorization order on integer partitions.
Lemma 3.2. Given a partition λ = (λ1 , . . . , λℓ ), define a function ψλ : N → N by
ψλ (u) = max {λ1 + · · · + λk − ku}.
1≤k≤ℓ

Then for two partitions λ, µ ⊢ r, we have that λ  µ if and only if ψλ (u) ≤ ψλ (u) for u = 0, . . . , r.
Proof. Note that ψλ (u) is equal to the number of boxes in columns u + 1, . . . , r of the Young diagram
of λ,
ψλ (u) = r − (λ⊤1 + · · · + λ⊤u ).
Thus the condition ψλ (u) ≤ ψµ (u) for u = 0, . . . , r is equivalent to the condition λ⊤  µ⊤, which in turn is
equivalent to λ  µ. 
The following result shows that inequalities of the form Pλ ≤ Pµ , are characterized by the majorization
order on of rational partitions of 1. In fact, we have Pλ ≤ Pµ if and only if we have Eλ ≥ Eµ .
Theorem 3.3. Given integer partitions λ and µ, we have
λ⊤ µ⊤
Pλ ≤ Pµ if and only if  .
|λ| |µ|
6 Allison Cuttler, Curtis Greene, and Mark Skandera

Proof. Let λ = (λ1 , . . . , λm ℓ), µ = (µ1 , . . . , µm ). Let us first consider the case that |λ| = |µ|. In this
case the theorem asserts the equivalence of Pλ ≤ Pµ and λ  µ.
(⇒) Suppose that λ  µ. Using Lemma 3.2, choose an index j such that ψλ (j) > ψµ (j), and consider
the functions φλ (t), φµ (t) defined by
ℓ ℓ
1 Y λi X X
φλ (t) = Pλ (t, 1, . . . , 1) = jℓ (t + tj ) = tλi1 +···+λik −kj ,
| {z } t i=1
k=0 {i1 ,...,ik }
tj
m m
1 Y µi X X
φµ (t) = Pµ (t, 1, . . . , 1) = jm (t + tj ) = tµi1 +···+µik −kj .
| {z } t i=1
k=0 {i1 ,...,ik }
tj

These are polynomials in t of degrees ψλ (j) and ψµ (j), respectively. Thus we have
lim [φλ (t) − φµ (t)] = ∞,
t→∞

which implies Pλ  Pµ .
(⇐) Suppose that λ  µ. Then we proceed as in the proof of Theorem 2.3 with Gantmacher’s inequalities
replacing those of Newton and conclude that Pλ ≤ Pµ .
In the case that λ and µ are not equal, we again proceed as in the proof of Theorem 2.3. Applying
Observation 3.1, we see that
Pλ (x) = Pλ|µ| (x), Pµ (x) = Pµ|λ| (x).
λ⊤ µ⊤
Thus we have Pλ ≤ Pµ if and only if λ|µ|  µ|λ| , or equivalently, if and only if |λ|  |µ| . 

The similarity of Theorems 2.3 and 3.3 is somewhat curious. Apparently we have that Pλ ≤ Pµ if and
only if Eλ ≥ Eµ . Thus the poset of power sum means is dual to the poset of elementary means.
Define the semiring G to be the set of all nonnegative linear combinations of products of symmetric
functions of the forms
{Pi+1,j−1 (x) − Pi,j (x) | 1 ≤ j ≤ i ≤ n − 1} ∪ {Pi (x) | 1 ≤ i ≤ n}.
The proof of Theorem 3.3 shows that each difference Pµ (x) − Pλ (x) with µ majorizing λ belongs to G.

4. More Inequalities
The authors have obtained some results and have formulated some conjectures concerning characteriza-
tion of inequalities of the forms Sλ ≤ Sµ , Mλ ≤ Mµ , Hλ ≤ Hµ , for λ and µ not necessarily partitions of the
same integer.
Let us consider the generalization of Muirhead’s inequalities which would characterize all inequalities of
the form Mλ ≤ Mµ . Note that a formula for the term-normalized monomial symmetric function is given by
mλ (x)
Mλ (x) = n

,
α1 ,...,αr ,n−ℓ

where αj is equal to the number of parts of λ which are equal to j. Unlike the elementary and power sum
bases, the monomial basis {mλ (x) | λ ⊢ r} of Λrn , is not multiplicative. Nevertheless, rational partitions of 1
seem to characterize monomial means as follows.
Conjecture 4.1. Given integer partitions λ and µ, we have
λ µ λ⊤ µ⊤
Mλ ≤ Mµ if and only if  and  .
|λ| |µ| |λ| |µ|
Equivalently, Mλ ≤ Mµ if and only if Eλ⊤ ≤ Eµ⊤ and Pλ ≤ Pµ .
λ µ
This conjecture suggests defining a double majorization order on integer partitions by λ E µ if |λ|  |µ|
⊤ ⊤
λ µ
and |λ|  |µ| . It is easy to see that when |λ| = |µ|, double majorization becomes ordinary majorization.
Thus for each r ∈ N, the majorization order of partitions of r is a subposet of the double majorization order.
Figure 2 shows the restriction of this poset to integer partitions of of 1, . . . , 5.
 INEQUALITIES FOR SYMMETRIC MEANS 7

Figure 2. Double majorization of partitions of 1, . . . , 5.

It is well-known that Muirhead’s inequalities imply many other inequalities. For instance, a straightfor-
ward computation shows that Gantmacher’s inqualities are essentially special cases of those of Muirhead,
n−1
Pk+1,k (x) − Pk,k (x) = (Mk+1,k−1 (x) − Mk,k (x)).
n
A lengthier computation shows that one can obtain Newton’s inequalities by adding various inequalities of
Muirhead.
Theorem 4.2. We have
k−1
X
Ek,k (x) − Ek+1,k−1 (x) = dj (M2k−j 1i+2j−k (x) − M2k−j−1 12j+2 (x)),
j=0

where


k−1 n−k
j j+1
dj =
.
(n − k) nk
Proof. Omitted. 

Thus by the comments at the ends of Sections 2 and 3, the inequalities in these sections can be derived
from those of Muirhead. It would therefore be interesting to express homogeneous and Schur differences
Hλ (x)−Hµ (x) and Sλ (x)−Sµ (x) as nonnegative linear combinations of Muirhead differences Mλ (x)−Mµ (x).
The authors have obtained partial results suggesting that this is possible in many cases.
Generalizing all of the inequalities considered so far are those of the form F ≤ G, where the means are
constructed from homogeneous elements f (x) ∈ Λk and g(x) ∈ Λℓ which need not be commonly used basis
elements of these spaces,
X X
f (x) = cκ m κ , g(x) = dλ mλ .
κ⊢k λ⊢ℓ
It would be interesting to characterize these inequalities in terms of the coordinate sequences (cκ )κ⊢k . By
the homogeneity of means, it would be sufficient to restrict one’s attention to coordinate sequences which
sum to 1.
To ensure that the means are well defined, one would have to require that the symmetric functions F (x)
and G(x) evaluate nonnegatively on all nonnegative vectors (of finite support). This condition does not
imply F and G to be monomial nonnegative, or even to be nonnegative linear combinations of Muirhead
differences. This suggests the following question.
8 Allison Cuttler, Curtis Greene, and Mark Skandera

Question 4.3. Which symmetric functions f (x) ∈ Λ (F (x) ∈ Λrn ) satisfy f (a) ≥ 0 (F (a) ≥ 0) for all
real nonnegative vectors a of finite support?

5. Acknowledgments
The authors are grateful to Lynne Butler for helpful conversations.

References
[Bul03] P. S. Bullen, Handbook of means and their inequalities, Mathematics and its Applications, vol. 560, Kluwer Academic
Publishers Group, Dordrecht, 2003, Revised from the 1988 original [P. S. Bullen, D. S. Mitrinović and P. M. Vasić,
Means and their inequalities, Reidel, Dordrecht; MR0947142]. MR MR2024343 (2005a:26001)
[Cut06] A. Cuttler, Symmetric mean inequalities, Haverford College undergraduate thesis, May 2006.
[Gan59] F. R. Gantmacher, The Theory of Matrices, vol. 2, Chelsea, New York, 1959.
[HLP34] G. H. Hardy, J. E. Littlewood, and G. Pólya, Inequalities, Cambridge University Press, Cambridge, 1934.
[Mac29] C. Maclaurin, A second letter to Martin Foulkes, Esq., concerning the roots of equations with the demonstrations of
other rules in algebra, Phil. Trans. 36 (1729), 59–96.
[Mui03] R. F. Muirhead, Some methods applicable to identities of symmetric algebraic functions of n letters, Proc. Edinburgh
Math. Soc. 21 (1902/03), 144–157.
[New] I. Newton, Arithmetica universalis: sive de compositione et resolutione arithmetica liber, Opera, I.
[Pop34] T. Popoviciu, Sur un théorème de Laguerre, Bull. Soc. Sci. Cluj 8 (1934), 1–4.
[Sch58] O. Schlömilch, Über Mittel grössen verschiedener Ordnung, Z. Math. Phys. 3 (1858), 301–308.
[Sta99] R. Stanley, Enumerative Combinatorics, vol. 2, Cambridge University Press, Cambridge, 1999.

Department of Mathematics, University of California, San Diego

E-mail address: [email protected]

Department of Mathematics, Haverford College

E-mail address: [email protected]

Department of Mathematics, Lehigh University

E-mail address: [email protected]