Network Utility Maximization With Nonconcave Utilities Using

Sum-of-Squares Method
Maryam Fazel Mung Chiang
Control and Dynamical Systems, Caltech Electrical Engineering, Princeton University

Abstract—Network Utility Maximization problem has re- behavior models and network efficiency-fairness charac-
cently been used extensively to analyze and design dis- terization. By allowing nonlinear concave utility objec-
tributed rate allocation in networks such as the Internet. tive functions, NUM substantially expands the scope of
A major limitation in the state-of-the-art is that user utility the classical LP-based Network Flow Problems.
functions are assumed to be strictly concave functions, mod-
eling elastic flows. Many applications require inelastic flow Consider a communication network with L links, each
models where nonconcave utility functions need to be maxi- with a fixed capacity of cl bps, and S sources (i.e., end
mized. It has been an open problem to find the globally op- users), each transmitting at a source rate of xs bps. Each
timal rate allocation that solves nonconcave network utility source s emits one flow, using a fixed set L(s) of links in
maximization, which is a difficult nonconvex optimization its path, and has a utility function Us (xs ). Each link l is
problem.
shared by a set S(l) of sources. Network Utility Max-
We provide a centralized algorithm for off-line analysis imization (NUM), in its basic version, is the following
and establishment of performance benchmark for noncon- problem of maximizing the total utility of the network
cave utility maximization. Based on the semi-algebraic ap- P
s Us (xs ),P
over the source rates x, subject to linear flow
proach to nonlinear optimization, we employ convex sum- constraints s:l∈L(s) xs ≤ cl for all links l:
of-squares relaxations for polynomial optimization, solved
by a sequence of semidefinite programs, to obtain increas- P
maximize U (x )
ingly tighter upper bounds on total achievable utility. Sur- Ps s s
prisingly, in all our experiments, a very low order and subject to s∈S(l) xs ≤ cl , ∀l, (1)
often a minimal order relaxation yields not just a bound xº0
on attainable network utility, but also the globally maxi-
mized network utility. Using a sufficiency test, whenever where the variables are x.
the bound can be proven to be exact, a globally optimal There are many nice properties of the basic NUM
rate allocation is recovered. In addition to polynomial utili- model due to several simplifying assumptions of the util-
ties, sigmoidal utilities can be transformed into polynomials ity functions and flow constraints, which provide the
and are handled by our algorithm. Furthermore, using two mathematical tractability of problem (1) but also limit its
alternative representation theorems for positive polynomi- applicability. In particular, the utility functions {Us } are
als, we present price interpretations in economics terms for often assumed to be increasing and strictly concave func-
these relaxations, extending the classical interpretation of tions. In this paper, we investigate the extension of the
independent congestion pricing on each link to pricing for basic NUM to maximization of nonconcave utilities.
the simultaneous usage of multiple links. Assuming that Us (xs ) becomes concave for large
Keywords: Nonconvex optimization, network util- enough xs is reasonable, because the law of diminishing
ity, rate allocation, algebraic geometry, sum of squares marginal utility eventually will be effective. However, Us
method may not be concave throughout its domain. In his semi-
nal paper published a decade ago, Shenker [17] differen-
I. I NTRODUCTION tiated inelastic network traffic from elastic traffic. Utility
functions for elastic traffic were modeled as strictly con-
A. Background: Basic network utility maximization cave functions. While inelastic flows with nonconcave
Since the publication of the seminal paper [6] by Kelly, utility functions represent important applications in prac-
Maulloo, and Tan in 1998, the framework of Network tice, they have received little attention and have scarcely
Utility Maximization (NUM) has found many applica- any mathematical foundation, except very recent publica-
tions in network rate allocation algorithms and Internet tions [9], [3], because of their intrinsic intractability in the
congestion control protocols (e.g., [10]), as well as user utility maximization framework.
 P
B. Review: Canonical distributed algorithm l∈L(s) λl is available to source s, where source s maxi-
A reason that the the assumption of utility function’s mizes a strictly concave function Ls (xs , λs ) over xs for a
concavity is upheld in almost all papers on NUM is that it given λs :
leads to three highly desirable mathematical properties of x∗s (λs ) = argmax [Us (xs ) − λs xs ] , ∀s. (2)
the basic NUM:
• It is a convex optimization problem, therefore the The Lagrange dual problem is
global minimum can be computed (at least in central-
minimize g(λ) = L(x∗ (λ), λ)
ized algorithms) in worst-case polynomial-time com- (3)
subject to λ º 0
plexity [2].
• Strong duality holds for (1) and its Lagrange dual where the optimization variable is λ. Any algorithms that
problem, i.e., the difference between the optimized find a pair of primal-dual variables (x, λ) that satisfy the
value of (1) and that of its dual problem (the optimal KKT optimality condition would solve (1) and its dual
duality gap) is zero [1], [2]. Zero duality gap enables problem (3). One possibility is a distributed, iterative sub-
a dual approach to solve (1). gradient method, which updates the dual variables λ to
• Minimization of a separable objective function over solve the dual problem (3):
linear constraints can be conducted by distributed al-   +
gorithms. X
Indeed, the basic NUM (1) is such a ‘nice’ optimiza- λl (t+1) = λl (t) − α(t) cl − xs (λs (t)) , ∀l
s:l∈L(s)
tion problem that its theoretical and computational prop-
(4)
erties have been well studied since the 1950s in the field
where t is the iteration number and α(t) > 0 are step
of monotropic programming, e.g., as summarized in [15].
sizes. Certain choices of step sizes, such as α(t) =
For network rate allocation problems, a dual-based dis-
β/t, β > 0, guarantee that the sequence of dual variables
tributed algorithm has been widely studied (e.g., in [6],
λ(t) will converge to the dual optimal λ∗ as t → ∞. The
[10]), and is summarized below.
primal variable x(λ(t)) will also converge to the primal
Zero duality gap for (1) states that the solving the La- optimal variable x∗ . For a primal problem that is a con-
grange dual problem is equivalent to solving the primal vex optimization, the convergence is towards the global
problem (1). The Lagrange dual problem is readily de- optimum.
rived. We first form the Lagrangian of (1): The sequence of the pair of algorithmic steps (2,4)
  forms a canonical distributed algorithm that globally
X X X
L(x, λ) = Us (xs ) + λl cl − xs  solves network utility optimization problem (1) and the
s l s∈S(l) dual (3) and computes the optimal rates x∗ and link prices
λ∗ .
where λl ≥ 0 is the Lagrange multiplier (link price) as-
sociated with the linear flow constraint on link l. Additiv- C. Summary of results
ity of total utility and linearity of flow constraints lead to It is known that for many multimedia applications, user
a Lagrangian dual decomposition into individual source satisfaction may assume non-concave shape as a function
terms: of the allocated rate. For example, the utility for streaming
   
applications is better described by a sigmoidal function:
X X X
L(x, λ) = Us (xs ) −  λl  xs  + cl λl with a convex part at low rate and a concave part at high
s l∈L(s) l rate, and a single inflexion point x0 (with Us00 (x0 ) = 0)
X X separating the two parts. The concavity assumption on Us
s
= Ls (xs , λ ) + cl λl
s l
is also related to the elasticity assumption on rate demands
P
by users. When demands for xs are not perfectly elastic,
where λs = l∈L(s) λl . For each source s, Ls (xs , λs ) = Us (xs ) may not be concave.
Us (xs )−λs xs only depends on local xs and the link prices Suppose we remove the critical assumption that {Us }
λl on those links used by source s. are concave functions, and allow them to be any nonlin-
The Lagrange dual function g(λ) is defined as the ear functions. The resulting NUM becomes nonconvex
maximized L(x, λ) over x. This ‘net utility’ maximiza- optimization and significantly harder to be analyzed and
tion obviously can be conducted distributively by the solved, even by centralized computational methods. In
each source, as long as the aggregate link price λs = particular, a local optimum may not be a global optimum
 3
Theorem in real algebraic geometry, we provide a cen-
tralized computational method to bound the total network
2.5
utility in polynomial-time. A surprising result is that for
2
the examples we have tried, wherever we could verify the
result, the tightest possible bound (i.e., the globally opti-
mal solution) of NUM with nonconcave utilities is com-
U(x)

1.5

puted in polynomial-time with a very low order relaxation.

1
This efficient numerical method for off-line analysis also
provides the benchmark for distributed heuristics. We also
0.5
examine two forms of sigmoidal utilities, and use a change
0
of variables to transform the original problem into one that
0 2 4 6 8 10 12
x
involves only polynomials. The sum-of-squares approach
Fig. 1. Some examples of utility functions Us (xs ): it can be con- mentioned above can then be applied.
cave or sigmoidal as shown in the graph, or any general nonconcave Our focus has been not only on calculating numerical
function. If the bottleneck link capacity used by the source is small bounds for the problem, but also on understanding the in-
enough, i.e., if the dotted vertical line is pushed to the left, a sigmoidal ner workings of the relaxations, and the mechanism be-
utility function effectively becomes a convex utility function.
hind the tightening of the upper bound, in the context of
NUM problems. In this regard, we have examined two
and the duality gap can be strictly positive. The standard polynomial representations that are particularly suited for
distributive algorithms that solve the dual problem may an economics/price interpretation of NUM. One result is
produce infeasible or suboptimal rate allocation. Global that the classical pricing of congestion on a link is ex-
maximization of nonconcave functions is an intrinsically tended to pricing of the usage of multiple links.
difficult problem of nonconvex optimization. Indeed, over These three different approaches: proposing distributed
the last two decades, it has been widely recognized that but suboptimal heuristics (for sigmoidal utilities) in [9],
“in fact the great watershed in optimization isn’t between determining optimality conditions for the canonical dis-
linearity and nonlinearity, but convexity and nonconvex- tributed algorithm to converge globally (for all nonlin-
ity” (Quote from Rockafellar [16]). ear utilities) in [3], and proposing efficient but centralized
Despite such difficulties, there have been two very re- method to compute the global optimum (for a wide class
cent publications on distributed algorithm for nonconcave of utilities that can be transformed into polynomial utili-
utility maximization. In [9], it is shown that, in general, ties) in this paper, are complementary in the study of dis-
the canonical distributive algorithm that solves the dual tributed rate allocation by nonconcave NUM, a difficult
problem may produce suboptimal, or even infeasible, rate class of nonlinear optimization.
allocation, and a ‘self-regulation’ heuristic is proposed to
avoid the resulting oscillation in rate allocation. How- II. G LOBAL MAXIMIZATION OF NONCONCAVE
ever, the heuristic converges only to a suboptimal rate al- NETWORK UTILITY
location. In [3], a set of sufficient conditions and nec- A. Sum-of-squares method
essary conditions is presented under which the canonical
First consider a NUM with convex polynomial utili-
distributed algorithm still converges to the globally opti-
ties, such as Us (xs ) = x2s . Sigmoidal utilities will be
mal solution. However, these conditions may not hold in
considered in subsection III.B. For notational simplicity,
many cases. In summary, currently there is no theoreti-
we assume the domain of definition of the Us (xs ) implies
cally polynomial-time and practically efficient algorithm
xs ≥ 0.
(distributed or centralized) known for nonconcave utility
maximization. P
maximize U (x )
Ps s s (5)
In this paper, we removes the concavity assumption on subject to s∈S(l) xs ≤ cl , ∀l.
utility functions, thus turning NUM into a nonlinear, non-
convex optimization problem with a strictly positive du- We would like to bound the maximum network util-
ality gap. Such problems in general are NP hard, thus ity by γ in polynomial time and search for a tight bound.
extremely unlikely to be polynomial-time solvable even Had there been no link capacity constraints, maximizing
by centralized computations. Using a family of convex a polynomial is already an NP hard problem, but can be
semidefinite programming (SDP) relaxations based on the relaxed into a SDP [18]. This is because testing if the fol-
sum-of-squares (SOS) method and the Positivestellensatz lowing bounding inequality holds γ ≥ p(x), where p(x)
is a polynomial of degree d in n variables, is equivalent The optimization variables are γ and all of the coefficients
to testing the positivity of γ − p(x), which can be re- in polynomials σ0 , σl , σjk . Note that x is not an opti-
laxed into testing if γ − p(x) can be written as a sum of mization variable; the constraints hold for all x, therefore
P
squares: p(x) = ri=1 qi (x)2 for some polynomials qi , imposing constraints on the coefficients of the polynomial
where the degree of qi is less than or equal to d/2. If a terms in x. This formulation uses Schmudgen’s represen-
polynomial can be written as a sum of squares, it must be tation of positive polynomials. Two alternative represen-
non-negative, but not vice versa. Conditions under which tations are discussed in section IV.
this relaxation is tight were studied by Hilbert. Quadratic Let D be the degree of the expression in the first con-
polynomials are one such case. Determining if a sum of straint in (6). For a fixed order D, the problem can be
squares decomposition exists can be formulated as an SDP solved via SDP. As D is increased, the expression includes
feasibility problem, thus polynomial-time solvable. more terms, the corresponding SDP becomes larger, and
Constrained nonconcave NUM can be relaxed by a gen- the relaxation gives tighter bounds. An important prop-
eralization of the Lagrange duality theory, where the La- erty of this nested family of relaxations is the following:
grange multiplier terms are nonlinear combinations of the by the Positivstellensatz, for D large enough but finite,
constraints instead of linear combinations in the standard the relaxation becomes exact. In general, such D can be
duality theory. The key result is the Positivstellensatz [19] exponentially large in the number of variables n.
in real algebraic geometry, which states that, for a system To show how this relaxation includes the dual-based so-
of polynomial inequalities, either there exists a solution lution to the basic NUM with strictly concave utilities as a
in Rn or there exists a polynomial which is a certificate special case, let the right hand side of the expression in (6)
that no solution exists. This infeasibility certificate is re- be first order. Then the relaxation is simply
cently shown to be also computable by SDP of sufficient
minimize γ
size [12], [11], a process that is referred to sum-of-squares
subject to
method and automated by the software SOSTOOLS [13]. P P P
γ − s Us (xs ) = σ0 + l σl (cl − s∈S(l) xs ), ∀x
Furthermore, the bound γ itself can become an opti-
σ0 ≥ 0, σl ≥ 0, ∀l.
mization variable in the SDP and can be directly mini-
mized. A nested family of SDP relaxations, each indexed In this case, σ0 and σi are simply non-negative constants.
by the degree of the infeasibility certificate polynomial It is easy to show that the above problem is equivalent to
and polynomial-time solvable, is guaranteed to produce the following one, which is precisely the dual:
the exact bound. Of course, given the problem is NP hard, X X X
minimize max{ Us (xs ) + σl (cl − xs ) + σ0 }
it is not surprising that the worst-case number of SDP x
s l s∈S(l)
relaxations needed is exponential in the number of vari-
subject to σl ≥ 0, ∀l.
ables. 1 What is surprising is that in applying SOSTOOLS (7)
to nonconcave utility maximization, only a very low or- There is a standard price interpretation for this simple
der, sometimes the minimum order relaxation already pro- first-order relaxation, which we shall extend to higher-
duces the globally optimal solution for nonconcave NUM. order relaxation later in this paper. For the case of con-
cave utilities, the dual variables σ can be interpreted as
B. Application of SOS method to nonconcave NUM link prices, and the bound γ is exact. In the non-concave
utility case, the duality gap is in general not zero, and the
Using sum-of-squares and the Positivstellensatz, the γ is only an upper bound; however σ can still be inter-
following relaxation for problem (5) can be derived. It preted as link prices. If the lth capacity constraint is vio-
yields an upper bound on the objective value of the prob- lated, we pay an extra charge proportional to the amount
P
lem. of violation, with price σl (since cl − s∈S(l) xs is nega-
tive and subtracts from the total utility). Similarly, we are
minimize γ
rewarded proportional to the amount of under-used capac-
subject to
P P P ity. In sharp contrast to the concave utility case, for non-
γ − s Us (xs ) = σ0 (x) + l σl (x)(cl − s∈S(l) xs )+
P P P concave utilities, these are not equilibrium prices and do
j,k σjk (x)(cj − s∈S(j) xs )(ck − s∈S(k) xs ) + . . . , ∀x
not result in optimal or even feasible rate allocation, un-
σ0 (x), σl (x), σjk (x), . . . are SOS. less the relaxation is exact. In a sense, this pricing scheme
(6)
fails to provide the right incentives for the users to respect
1
For more details and many application of SOS methods, see [12] the constraints. In section IV, we discuss this interpreta-
and references therein. tion for higher order relaxations.
 Higher order relaxations can improve the upper bound. 1) Formulate the relaxed problem (6) for a given de-
For example, let the right hand side expression in (6) be gree D.
second order. We have 2) Use SDP to solve the Dth order relaxation, which
can be conducted using SOSTOOLS [13].
minimize γ 3) If the resulting dual SDP solution satisfies the suffi-
subject to cient rank condition, the Dth order optimizer γ ∗ (D)
P P P
γ − s Us (xs ) = σ0 (x) + l σl (cl − s∈S(l) xs )+ is the globally optimal network utility, and a cor-
P P P
j,k σjk (cj − s∈S(j) xs )(ck − s∈S(k) xs ), ∀x responding x∗ can be obtained. Otherwise, γ ∗ (D)
σl ≥ 0, σjk ≥ 0, ∀l, j, k may still be the globally optimal network utility but
σ0 is SOS. is only provably an upper bound.
(8) 4) Increase D to D + 2, i.e., the next higher order re-
Note that we now have terms that involve the product of laxation, and repeat.
the constraints. This problem is in fact the Lagrange dual
for problem (5) with some added redundant constraints; In the following section, we give examples of the ap-
namely, the pairwise product of every two non-negative plication of SOS relaxation to the nonconcave NUM. We
P P also apply the a version of the above sufficient test (as im-
terms (cj − s∈S(j) xs )(ck − s∈S(k) xs ). As mentioned
before, this problem can be solved via SDP, and yields a plemented in SOSTOOLS) to check if the bound is exact,
bound that is at least as strong as the first-order case. The and if so, we recover the optimum rate allocation x∗ that
expression in (8) can also be written as achieve this tightest bound.
P P P
γ− s Us (xs )P
− l σl (cl − s∈S(l) xs )−
P P
j,k σjk (cj − s∈S(j) xs )(ck − s∈S(k) xs ) is SOS. III. N UMERICAL E XAMPLES
(9)
So we seek γ and polynomials σi , σjk that make this A. Polynomial utility examples
polynomial SOS. Clearly a polynomial of odd degree can-
not be SOS, so we need to consider only the cases where First, consider the case of quadratic utilities, i.e.,
the expression has even degree. Therefore, the degree of Ui (xi ) = x2i as the simplest case. We can also handle
the first non-trivial relaxation is the smallest even number weights on the utilities, cubic or higher order polynomi-
P
greater than or equal to degree of s Us (xs ). als as utilities, or Ui of different orders for different users,
A key question is the following: How do we find out, in a similar fashion. We present a few examples that are
after solving an SOS relaxation, if the bound happens to typical, in our experience, of the performance of the re-
be exact? Fortunately, there is a sufficient test that can re- laxations.
veal this, using the properties of the SDP and its dual so- Example 1. This first example is a small and illustrative
lution. In [5], [7], a parallel set of relaxations, equivalent one. Consider the simple 2 link, 3 user network shown in
to the SOS ones, is developed in the dual framework. The Figure 2, with c = [1, 2].
dual of checking the nonnegativity of a polynomial over
a semi-algebraic set turns out to be finding a sequence of x2 x3
moments that represent a probability measure with sup-
port in that set. To be a valid set of moments, the sequence c1 c2
should form a positive semidefinite moment matrix. Then,
each level of relaxation fixes the size of this matrix, i.e.,
considers moments up a certain order, and therefore solves
an SDP. This is equivalent to fixing the order of the poly- x1
nomials appearing in SOS relaxations. The sufficient rank Fig. 2. Network topology for example 1.
test checks a rank condition on this moment matrix, as
discussed in [5].
In summary, we have the following Algorithm 1 for The optimization problem is
centralized computation of a globally optimal rate alloca-
P 2
tion to nonconcave utility maximization, where the util- maximize s xs
ity functions can be written as or converted into polyno- subject to x1 + x2 ≤ 1
(10)
mials (details about such conversions are in the next sec- x1 + x3 ≤ 2
tion): x1 , x2 , x3 ≥ 0.
The first level relaxation with D = 2 is this simple network, either by using the sufficient test de-
maximize γ scribed above or by inspection, we find that the rate vector
subject to x0 = [1, 0, 0, 1] achieves this bound, hence the bound
γ − (x21 + x22 + x23 ) − σ1 (−x1 − x2 + 1) − σ2 (−x1 is exact. As another example, with c = [2, 3, 4, 1], we
−x3 + 2) − σ3 x1 − σ4 x2 − σ5 x3 − σ6 (−x1 − x2 + 1) obtain γ = 10. Again, we find that x0 = [0, 0, 3, 1]
(−x1 − x3 + 2) − σ7 x1 (−x1 − x2 + 1) − σ8 x2 (−x1 achieves this upper bound, which is therefore exact. The
−x2 + 1) − σ9 x3 (−x1 − x2 + 1) − σ10 x1 (−x1 − x3 + 2)− globally optimal solution to the NUM problem is again
σ11 x2 (−x1 − x3 + 2) − σ12 x3 (−x1 − x3 + 2)− efficiently computed.
σ13 x1 x2 − σ14 x1 x3 − σ15 x2 x3 is SOS, Example 3. We now consider a larger example. Con-
σi ≥ 0, i = 1, . . . , 15. sider the network shown in Figure 4 with 7 links. We
(11) allow 9 users, with the following routing table that lists
The first constraint above can be written as xT Qx for x = the links on each user’s path.
[1, x1 , x2 , x3 ]T and an appropriate Q. For example, the x1 x2 x3 x4 x5 x6 x7 x8 x9
(1,1) entry which is the constant term reads γ −σ1 −2σ2 − 1,2 1,2,4 2,3 4,5 2,4 6,5,7 5,6 7 5
2σ6 , the (2,1) entry, coefficient of x1 , reads σ1 +σ2 −σ3 +
3σ6 − σ7 − 2σ10 , and so on. The expression is SOS if and Let c = [5, 10, 4, 3, 7, 3, 5], and we obtain the bound
only if Q ≥ 0. γ = 116 with D = 2, which turns out to be globally op-
However, since the polynomial is a quadratic in this timal. The sufficient condition for exactness of the bound
specific example, being SOS and being nonnegative are holds here, and allows us to recover the globally optimal
the same. Thus the problem simplifies to minimizing γ rate vector directly: x0 = [5, 0, 4, 0, 1, 0, 0, 5, 7].
subject to γ = −x1 (σ1 + σ2 − σ3 + 3σ6 − σ7 − 2σ10 ) − In this example, exhaustive search is too computationally
. . . + (σ1 + 2σ2 + 2σ6 ), ∀x. Equating the coefficients intensive, and the sufficient condition test plays an impor-
results in linear constraints in σ and γ, and the problem tant role in proving the bound was exact and in recovering
can be solved using a simple linear program with 8 con- x0 .
straints and 15 variables (this coincides with the LP relax-
ation discussed in the next section). The optimal γ is 5, c3
c6
which is achieved by, e.g., σ1 = 1, σ2 = 2, σ3 = 1, σ8 =
c5
1, σ10 = 1, σ12 = 1, σ13 = 1, σ14 = 2 and the rest of c1 c2
c4
the σi equal to zero.
In this example, many of the σi could be chosen to be
c7
zero. This means not all product terms appearing in 11
are needed in constructing the SOS polynomial. Such in-
Fig. 4. Network topology for example 3.
formation is valuable from the decentralization point of
view, and can help determine to what extent our bound
can be calculated in a distributed manner. This is a topic
for future work. B. Sigmoidal utility examples
Example 2. Consider the 4 link, 4 user network shown Now consider sigmoidal utilities in a standard form:
in Figure 3. 1
Us (xs ) = ,
x3
1 + e s xs +bs )
−(a

x1 c3 where {as , bs } are constant integers. Even though these

sigmoidal functions are not polynomials, we show the
c1 c2
x4 problem can be cast as one with polynomial cost and con-
straints, with a change of variables.
Example 4. Consider the simple 2 link, 3 user example
x2 c4
shown in Figure 2 for as = 1 and bs = −5.
The NUM problem is to
P 1
Fig. 3. Network topology for example 2. maximize s 1+e−(xs −5)
subject to x1 + x2 ≤ c1
(12)
If we set all link capacities {cl } to 1, using an SOS re- x1 + x3 ≤ c2
laxation with D = 2, we obtain the upper bound γ = 2. In x ≥ 0.
 1
Let ys = 1+e−(x s −5)
, then xs = − log( y1s − 1) + 5. alent to
Substituting for x1 , x2 in the first constraint, arranging P
terms and taking exponentials, then multiplying the sides
maximize s ys
subject to xns − ys xns − ays = 0
by y1 y2 (note that y1 , y2 > 0), we get P (14)
s∈S(l) xs ≤ cl , ∀l
(1 − y1 )(1 − y2 ) ≥ e(10−c1 ) y1 y2 , x≥0

which again can be accommodated in the SOS method and

which is polynomial in the new variables y. This applies
be solved by Algorithm 1.
to all capacity constraints, and the non-negativity con-
1 The benefit of this choice of utility function is that the
straints for xs translate to ys ≥ 1+e 5 . Therefore the whole
largest degree of the polynomials in the problem is n + 1,
problem can be written in polynomial form, and SOS
so it grows linearly with n. The disadvantage compared
methods apply. This transformation renders the problem
to the exponential form for sigmoidal functions is that the
polynomial for general sigmoidal utility functions, with
location of the inflection point and the slope at this point
any as and bs .
cannot be set independently.
We present some numerical results. These example is
small enough to be solved by hand, and we could verify
that the bounds obtained in the following examples are IV. A LTERNATIVE REPRESENTATIONS FOR CONVEX
exact. Here SOS relaxations of order 4 (D = 4) were RELAXATIONS TO NONCONCAVE NUM
used. For c1 = 4, c2 = 8, we find γ = 1.228, which The SOS relaxation we used in the last two sections is
turns out to be a global optimum, with x0 = [0, 4, 8] based on Schmudgen’s representation for positive polyno-
as the optimal rate vector. For c1 = 9, c2 = 10, we find mials over compact sets described by other polynomials.
γ = 1.982 and x0 = [0, 9, 10]. Now place a weight of In this section, we briefly discuss two other representa-
2 on y1 , while the other ys have weight one, we obtain tions of relevance to the NUM, that are interesting from
γ = 1.982 and x0 = [9, 0, 1]. both theoretical (e.g., interpretation) and computational
In general, if as 6= 1 for some s, however, the degree of (e.g., efficiency) points of view.
the polynomials in the transformed problem may be very
high. If we write the general problem as A. LP relaxation
P1 Exploiting linearity of the constraints in NUM and with
maximize
Ps 1+e−(as xs +bs ) the additional assumption of nonempty interior for the
subject to s∈S(l) xs ≤ cl , ∀l, (13)
feasible set, we can use Handelman’s representation [4]
x ≥ 0, and refine the Positivstellensatz condition to obtain the
following convex relaxation of nonconcave NUM prob-
each capacity constraint after transformation will be
lem:
Q
s (1 − ys )rls Πk6=s ak ≥ maximize γ
Q
Q P Q rls a
k6=s k subject to
exp(− s as (cl + s rls /as bs )) s ys , L
P X Y P
γ− αl
s Us (xs ) = σα (cl − s∈S(l) xs ) , ∀x
where rls = 1 if l ∈ L(s) and equals 0 otherwise. Since α∈N L l=1
the product of the as appears in the exponents, as > 1 σα ≥ 0, ∀α,
significantly increases the degree of the polynomials ap- (15)
pearing in the problem and hence the dimension of the where the optimization variables are γ and σα , and α de-
SDP in the SOS method. notes an ordered set of integers {αl }.
P
Another utility function that can be used to represent Fixing D where l αl ≤ D, and equating the coeffi-
sigmoidal functions is the following rational function: cients on the two sides of the equality in (15), yields a
linear program (LP). (Note that there are no SOS terms,
xns therefore no semidefiniteness conditions.) As before, in-
Us (xs ) = ,
a + xns creasing the degree D gives higher order relaxations and
a tighter bound.
where the inflection point is x0 = ( a(n−1)
n+1 )
1/n and the
We provide a (partial) price interpretation for prob-
0 n−1 n+1 1/n
slope at the inflection point is Us (x ) = 4n ( a(n−1) ) . lem (15). First, normalize each capacity constraint as
P
Let ys = Us (xs ), the NUM problem in this case is equiv- 1 − ul (x) ≥ 0, where ul (x) = s∈S(l) xs /cl . We can
interpret ul (x) as link usage, or the probability that link l depends on the rates. The physical meaning of such
is used at any given point in time. Then, in (15), we have prices, and the computational aspects of this relaxation
terms linear in u such as σl (1 − ul (x)), in which σl has remain to be explored.
a similar interpretation as in concave NUM, as the price
of using link l (at full capacity, due to the normalization). V. C ONCLUSIONS AND FURTHER EXTENSIONS
We also have product terms such as σjk (1 − uj (x))(1 −
We consider the NUM problem in the presence of in-
uk (x)), that indicate the probability of simultaneous us-
elastic flows, i.e., flows with nonconcave utilities. Despite
age of links j and k, for links whose usage probabilities
its practical importance, this problem has not been stud-
are independent (e.g., they do not share any flows). Prod-
ied widely, mainly due to the fact it is a nonconvex, NP-
ucts of more terms can be interpreted similarly.
hard problem. There has been no effective mechanism,
While the above price interpretation is not complete and
centralized or distributed, to compute the globally optimal
does not justify all the terms appearing in (15) (e.g., pow-
rate allocation for nonconcave utility maximization prob-
ers of the constraints; product terms for links with shared
lems in networks. This limitation has made performance
flows), it does provide some useful intuition: this relax-
assessment and design of networks that include inelastic
ation results in a pricing scheme that provides better in-
flows very difficult.
centives for the users to observe the constraints, by putting
To address this problem, we employed convex SOS
additional reward (since the corresponding term adds pos-
relaxations, solved by a sequence of SDPs, to obtain
itively to the utility) for simultaneously keeping two links
high quality, increasingly tighter upper bounds on total
free. Such incentive helps tighten the upper bound and
achievable utility. In practice, the performance of our
eventually achieve a feasible (and optimal) allocation.
SOSTOOLS-based algorithm was surprisingly good, and
This relaxation is computationally attractive since we
bounds obtained using a polynomial-time (and indeed a
need to solve an LPs instead of the previous SDPs at each
low-order and often minimal order) relaxation were found
level. However, significantly more levels may be required.
2 to be exact, achieving the global optimum of nonconcave
NUM problems. Furthermore, a dual-based sufficient test,
if successful, detects the exactness of the bound, in which
B. Relaxation with no product terms case the optimal rate allocation can also be recovered.
Putinar [14] showed that a polynomial positive over a This surprisingly good performance of the proposed algo-
compact set (with an extra assumption that always holds rithm brings up the question whether there is any particu-
for linear constraints as in NUM problems) can be repre- lar property or structure in nonconcave NUM that makes
sented as an SOS-combination of the constraints. This it especially suitable for SOS relaxations.
yields the following convex relaxation for nonconcave We further examined the use of two more specialized
NUM problem: polynomial representations, one that uses products of con-
straints with constant multipliers, resulting in LP relax-
maximize γ ations; and at the other end of spectrum, one that uses a
subject to ‘linear’ combination of constraints with SOS multipliers.
P P P
γ − s Us (xs ) = L l=1 σl (x)(cl − s∈S(l) xs ), ∀x We expect these relaxations to give higher order certifi-
σ(x) is SOS, cates, thus their potential computational benefits need to
(16) be examined further. We also show they admit useful eco-
where the optimization variables are the coefficients in nomics interpretations (e.g., prices, incentives) that pro-
σl (x). Similar to the SOS relaxation, fixing the order D vide valuable insight on how the SOS relaxations work in
of the expression in (16) results in an SDP. This relax- the framework of link congestion pricing for the simulta-
ation has the nice property that no product terms appear, neous usage of multiple links.
that is, the relaxation becomes exact with a high enough Another research issue of interest in nonconcave NUM
D without the need of product terms. However, this de- problems is decentralization methods. Algorithm 1 is not
gree might be much higher than what the previous SOS easy to decentralize, given the products of the constraints
method requires. or polynomial multipliers that destroy the separable struc-
We note yet another price interpretation: this time the ture of the problem. However, when relaxations become
link price is given by an SOS polynomial multiplier that exact, the sparsity pattern of the coefficients can provide
2
From a numerical point of view, the recent paper [8] reports conver-
information about partially decentralized computation of
gence problems with this type of relaxation, but the problematic cases optimal rates. For example, if after solving the NUM off-
do not include our problem. line, we obtain an exact bound, then if the coefficient of
the cross-term xi xj turns out to be zero, it means users i [16] R. T. Rockafellar, “Lagrange multipliers and optimality,” SIAM
and j do not need to communicate to each other to find Review, vol. 35, pp. 183-283, 1993.
[17] S. Shenker, “Fundamental design issues for the future Internet,”
their optimal rates. An interesting next step in this area of
IEEE J. Sel. Area Comm., vol. 13, no. 7, pp. 1176-1188, Sept.
research is to investigate distributed version of Algorithm 1995.
1 through message passing among clusters of network [18] N. Z. Shor, “Quadratic optimization problems,” Soviet J. Com-
nodes and links. It is also worth continuing to explore put. Systems Sci., vol 25, pp. 1-11, 1987.
[19] G. Stengle, “A Nullstellensatz and a Positivstellensatz in semial-
other types of nonconcave functions that can be trans-
gebraic geometry,” Math. Ann., vol. 207, pp.87-97, 1974.
formed into polynomials and handled by SOS methods,
in addition to the two sigmoidal forms we already exam-
ined in this paper.

ACKNOWLEDGEMENT
We would like to thank very helpful discussions with
Pablo Parrilo, Steven Low, and John Doyle.

R EFERENCES
[1] D. P. Bertsekas, Nonlinear Programming, Athena Scientific, 1999.
[2] S. Boyd and L. Vandenberghe, Convex Optimization, Cambridge
University Press, 2004.
[3] M. Chiang, S. Zhang, and P. Hande, “Distributed rate allocation
for inelastic flows: Optimization framework, optimality condi-
tions, and optimal algorithms,” Proc. IEEE Infocom, Miami, FL,
March 2005.
[4] D. Handelman, “Representing polynomials by positive linear
functions on compact convex polyhedra,” Pacific J. Math., vol.
132, pp. 35-62, 1988.
[5] D. Henrion, J.B. Lasserre, “Detecting global optimality and ex-
tracting solutions in GloptiPoly,” Research report, LAAS-CNRS,
2003.
[6] F. P. Kelly, A. Maulloo, and D. Tan, “Rate control for communica-
tion networks: shadow prices, proportional fairness and stability,”
Journal of Operations Research Society, vol. 49, no. 3, pp.237-
252, March 1998.
[7] J.B. Lasserre, “Global optimization with polynomials and the
problem of moments,” SIAM J. Optim., vol. 11, no. 3, pp. 796-
817, 2001.
[8] J.B. Lasserre, “Polynomial programming: LP-relaxations also
converge,” SIAM J. Optimization, vol. 15, no. 2, pp. 383-393,
2004.
[9] J. W. Lee, R. R. Mazumdar, and N. Shroff, “Non-convex optimiza-
tion and rate control for multi-class services in the Internet,” Proc.
IEEE Infocom, Hong Kong, China, March 2004.
[10] S. H. Low, “A duality model of TCP and queue management
algorithms,” IEEE/ACM Tran. Networking, vol. 11, no. 4, pp. 525-
536, Aug. 2003.
[11] P. A. Parrilo, Structured semidefinite programs and semi-
algebraic geometry methods in robustness and optimization,” PhD
thesis, Caltech, May 2002.
[12] P. A. Parrilo, “Semidefinite programming relaxations for semi-
algebraic problems,” Math. Program., vol. 96, pp.293-320, 2003.
[13] S. Prajna, A. Papachristodoulou, P. A. Parrilo, “SOSTOOLS:
Sum of squares optimization toolbox for Matlab, available from
http://www.cds.caltech.edu/sostools, 2002-04.
[14] M. Putinar, “Positive polynomials on compact semi-algebraic
sets,” Indiana University Mathematics Journal, vol. 42, no. 3, pp.
969-984, 1993.
[15] R. T. Rockafellar, Network Flows and Monotropic Programming,
Athena Scientific, 1998.