Available online at www.sciencedirect.

com

Operations
Research
Letters

Operations Research Letters 31 (2003) 167 175

www.elsevier.com/locate/dsw

Mirror descent and nonlinear projected subgradient methods

for convex optimization
Amir Beck, Marc Teboulle
School of Mathematical Sciences, Tel-Aviv University, Ramat-Aviv 69978, Israel
Received 20 August 2002; received in revised form 28 October 2002; accepted 31 October 2002

Abstract
The mirror descent algorithm (MDA) was introduced by Nemirovsky and Yudin for solving convex optimization problems.
This method exhibits an e3ciency estimate that is mildly dependent in the decision variables dimension, and thus suitable
for solving very large scale optimization problems. We present a new derivation and analysis of this algorithm. We show
that the MDA can be viewed as a nonlinear projected-subgradient type method, derived from using a general distance-like
function instead of the usual Euclidean squared distance. Within this interpretation, we derive in a simple way convergence
and e3ciency estimates. We then propose an Entropic mirror descent algorithm for convex minimization over the unit
simplex, with a global e3ciency estimate proven to be mildly dependent in the dimension of the problem.
c 2003 Elsevier Science B.V. All rights reserved.

Keywords: Nonsmooth convex minimization; Projected subgradient methods; Nonlinear projections; Mirror descent algorithms; Relative
entropy; Complexity analysis; Global rate of convergence

1. Introduction
Consider the following nonsmooth convex minimization problem,
(P)

minimize f(x) s:t: x X Rn :

Throughout the paper we make the following assumptions on problem (P):

Assumption A.
(a) X is a closed convex subset in Rn with nonempty
interior.

(b) The objective function f : X R is a convex

Lipschitz continuous function with Lipschitz constant Lf with respect to a ?xed given norm  ,
i.e., |f(x) f(y)| 6 Lf x y, x; y X .
(c) The optimal set of (P) denoted by X is nonempty.
(d) A subgradient of f at x X is computable. An
element of the subdi@erential 9f(x) is denoted by
f (x).
We are interested in ?nding an approximate solution
to problem (P), within
0, i.e., to ?nd x X such
that
f(x) f := f(x) minf(x) 6
:
xX

Corresponding author.
E-mail addresses: [email protected] (A. Beck),
[email protected] (M. Teboulle).

A standard method to solve (P) is the subgradient projection algorithm, (see e.g. [2] and references

c 2003 Elsevier Science B.V. All rights reserved.

0167-6377/03/$ - see front matter

PII: S 0 1 6 7 - 6 3 7 7 ( 0 2 ) 0 0 2 3 1 - 6

168

A. Beck, M. Teboulle / Operations Research Letters 31 (2003) 167 175

therein), which generates iteratively the sequence {xk }

via
xk+1 = X (xk tk f (xk ));
tk 0 (a stepsize);

(1.1)

where X (x) = argmin{x y | y X } is the

Euclidean projection onto X .
The key advantage of the subgradient algorithm is
its simplicity, provided that projections can be easily
computed, which is the case when the constraints
set X is described by simple sets, e.g., hyperplanes,
balls, bound constraints, etc. Its main drawback is
that it has a very slow rate of convergence. Indeed,
consider the convex problem (P) with X convex
compact, and denote by Diam(X ) the diameter of X ,
i.e., Diam(X ) := maxx; yX x y . Then, the
optimal e3ciency estimate for the subgradient
method with stepsizes tk = Diam(X )k 1=2 , k = 1; : : : ;
is (see [10]):
min f(xs ) min f(x)

16s6k

xX

6 O(1)Lf Diam(X )k 1=2 ;

(1.2)

where O(1) stands for a positive absolute constant.

Thus, like all gradient based methods, one can obtain
in a very small number of iterations a low accuracy
optimal value, (say one or two digits) but then within
further iterations no more progress in accuracy can
be achieved and the method is essentially jamming.
However, a key feature of gradient methods is also
the fact that while their rate of convergence is very
slow, the rate is almost independent of the dimension of the problem. In contrast to this, more e3cient
sophisticated algorithms, such as for example interior
point-based methods, which require for example at
each iteration Newton-type computations, i.e., the
solution of a linear system, are often defeated even for
problems with a few thousands of variables, and a
fortiori for very large-scale nonsmooth problems.
Therefore, for constrained problems where low accurate solutions is su3cient and the dimension is huge,
gradient type methods appear as natural candidates
for developing potential practical algorithms. The
recent paper of Ben-Tal et al. [1] on computerized
tomography demonstrates very well this situation
through an algorithm based on the mirror descent algorithm (MDA for short) introduced by Nemirovski

and Yudin [10]. It is shown there that it is possible to

solve e3ciently a convex minimization problem over
the unit simplex, with millions of variables.
Motivated by the recent work of Ben-Tal et al. [1],
in this paper we concentrate on the analysis of the
basic steps of the (MDA) which is recalled in Section 2. We show in Section 3, that the (MDA) can be
viewed as a simple nonlinear subgradient projection
method, where the usual Euclidean projection operator
is replaced by a nonlinear/nonorthogonal type projection operator based on a Bregman-like distance function (see e.g. [3,4,14] and references therein). With
this new interpretation of the (MDA), we derive in
a simple and systematic way convergence proofs and
e3ciency estimates, see Section 4. In Section 5 we
concentrate on optimization problems over the unit
simplex and propose a new algorithm called the entropic mirror descent algorithm (EMDA). The EMDA
is proven to exhibit an e3ciency estimate which is almost independent in the dimension n of the problem
and in fact shares the same properties of an algorithm
proposed in [1] for the same class of problems, but is
given explicitly by a simple formula. Finally, in the
last section we outline some potential applications and
extensions for further work.
2. The mirror descent algorithm (MDA)
The idea of the algorithm is based on dealing with
the structure of the Euclidean norm rather than with
local behavior of the objective function in problem
(P). Roughly speaking, the method originated from
functional analytic arguments arising within the in?nite dimensional setting, between primal and dual
spaces. The mathematical objects associated with f
and x are not vectors from a vector space E, but elements of the dual vector space to E, which consists
of linear forms on E. The Euclidean structure is not
the only way to identify the primal-dual spaces, and
it is possible to identify the primal and dual spaces
within a wider family which includes as particular
case, the classical Euclidean structures. This idea and
approach was introduced by Nemirovsky and Yudin
[10], and the reader is refereed to their book for a more
detailed motivation and explanations. We will show
below, that there is a much simpler and easy way to
motivate, explain, and construct the MDA. For now,

A. Beck, M. Teboulle / Operations Research Letters 31 (2003) 167 175

169

let us consider the basic steps involve in the original

MDA.
Consider the problem (P) satisfying Assumption A.
The (MDA) further assumes the following objects,
which can be freely chosen as long as they satisfy the
following hypothesis:

closed convex set X , using the well known relations

(I + NX )1 = X and (9 )1 = 9 , (see [11]), one
thus has

Fix any norm   in Rn (which will play a role in

the choice of the stepsize).
Let : X R be a continuously di@erentiable and
strongly convex function on X with strong convexity parameter  0.
The conjugate of , de?ned by (y)=maxxX { x;
y (x)} is assumed to be easily computable.

Therefore, the (MDA) yields

The basic steps of the MDA can be described as

follows, see [10,1] (for comparison the set Y there is
set to be equal to X in [1, p. 84]).
MDA: Start with y1 dom and generate the
sequence {xk } X via the iterations
xk =

(yk );

yk+1 = (xk ) tk f (xk );

xk+1 =

(yk+1 )

( (xk ) tk f (xk ));

(2.3)
(2.4)

(2.5)

where tk 0 are appropriate step sizes.

The method looks at this stage somewhat hard to
understand or even to motivate (besides the very rough
explanation given above). In the next section we will
give a very simple interpretation which will explain
and reveal the structure of this algorithm. In the mean
time, let us consider a basic example which clearly
indicates that the MDA appears to be as a natural
generalization of the subgradient algorithm.
Example 1. Let   be the usual l2 norm in Rn and
let (x) := 12 x2 for x X and + for x  X . The
function is clearly proper, lsc and strongly convex
with parameter  = 1, and continuously di@erentiable
on X . A straightforward computation shows that the
conjugate of is given by : Rn R

(z) = 12 (z2 z X (z)2 )

with (z) = X (z). Indeed, since 9 (x) = (I +

NX )(x), where NX denotes the normal cone of the

z 9 (x) x = (I + NX )1 (z)
= X (z) =

(z):

xk = X (yk );

(2.6)

yk+1 = xk tk f (xk );

(2.7)

xk+1 = X (xk tk f (xk ));

(2.8)

i.e., we have recovered the subgradient projection

algorithm.
3. Nonlinear projection methods
It is well known (see e.g. [2]) that the subgradient algorithm can be viewed as linearization of
the so-called proximal algorithm [12], (or as an explicit scheme of the corresponding subdi@erential
inclusion). Indeed, it is immediate to verify that the
projected subgradient iteration (1.1) can be rewritten
equivalently as



1
k+1
 k
k 2
x argmin x; f (x ) +
x x  :
2tk
xX
In [14], it has been shown that more general proximal maps can be considered by replacing the usual
Euclidean quadratic norms with some sort of more
general distance-like functions. As explained there,
the principal motivation for such kind of distances is
to be able to use one which reKects the geometry of
the given constraints set X , so that in particular with
such an appropriate choice, the constraints can often
be automatically eliminated. In a similar way, we can
thus construct nonlinear projection subgradient methods, by considering iteration schemes of the form



1
k+1
 k
k
x argmin x; f (x ) + D(x; x ) ;
(3.9)
tk
xX
where D(u; v) replaces 21 u v2 , and should verify
the property D(u; v) is nonnegative, and D(u; v) = 0
if and only if u = v. We prove below, that the MDA

170

A. Beck, M. Teboulle / Operations Research Letters 31 (2003) 167 175

is nothing else, but the nonlinear subgradient projection method (3.9), with a particular choice of D based
on a Bregman-like distance generated by a function
. Note, that the hypothesis on D will be somewhat
di@erent from the usual Bregman based distances assumed in the literature (see e.g. [8,14], and references
therein).
Let : X R be strongly convex and continuously
di@erentiable on int X . The distance-like function is
de?ned by B : X int(X ) R given by
B (x; y) = (x) (y) x y; (y):

(3.10)

The basic subgradient algorithm based on B is as

follows.
Subgradient algorithm with nonlinear projections
(SANP): Given B as de?ned in (3.10) with
as
above, start with x1 int X , and generate the sequence
{xk } via the iteration



1
k+1
 k
k
= argmin x; f (x ) + B (x; x ) ;
x
tk
xX
tk 0:

(3.11)

When can be continuously extended on X , (e.g.,

as in Example 1), then we can consider the function B
de?ned on X X . Note that in this case one needs not
to start with x1 int X and SANP can start with any
arbitrary point x1 Rn . With X =Rn and (x)= 12 x2
one obtains B (x; y) = 12 x y2 thus recovering the
classical squared Euclidean distance and SANP is just
the classical subgradient algorithm.
We now turn to the question of having SANP a
well de?ned algorithm. When is continuously di@erentiable on X , then the strong convexity assumption
immediately implies that the algorithm which starts
with x1 Rn is well de?ned and produces a sequence
xk X , k. When is only assumed to be di@erentiable on int X , we clearly need to guarantee that the
next iterate stays in the interior of X , so that B can
be de?ned on X int X . For that, it su3ces to make
the following assumption:
 (xt ) + as t ; {xt } int X
with xt x bd X;

(3.12)

where bd X denotes the boundary of X . Note that

(3.12) is just to say that is essentially smooth, (see

[11]). With this additional assumption on together

with the strong convexity, it follows that the sequence
{xk } is well de?ned i.e., xk int X , k. An interesting
choice for satisfying (3.12) will be considered in
Section 5.
It is interesting to note the di@erences between the
two classes of algorithms which then emerged from
(SANP). The ?rst class with continuously di@erentiable on X leads to noninterior methods with iterates
xk X . This is exactly the setting of the MDA. Typical examples of
in that case will involve power
of norms on X , see Example 1 and [1]. On the other
hand, the second class, with satisfying (3.12) will be
an interior type subgradient algorithm producing sequences xk int X . Note that the analysis we develop
in the rest of this paper will hold for both classes of
algorithms with the additional assumption (3.14) on
when needed.
We ?rst recall some useful facts regarding strongly
convex functions, and their relations with conjugates
and subdi@erentials. These results can be found in [13,
Section 12H].
Proposition 3.1. Let : Rn R{+} be a proper
convex and lsc function and let  0. Consider the
following statements:
(a) is strongly convex with parameter .
(b) u v; x y x y2 , whenever u 9(x);
v 9(y); i.e., the map 9 is strongly monotone.
(c) The inverse map (9)1 is everywhere single valued and Lipschitz continuous with constant 1 .
(d) is ;nite everywhere and di<erentiable.
Then, (a) (b) (c) (d).
As written above in (3.11), the resemblance between MDA and SANP is still not obvious. However,
we ?rst note that the main step of SANP can be written in a more explicit way. Writing down formally the
optimality conditions for (3.11), we obtain the following equivalent forms for SANP:
0 tk f (xk ) + (xk+1 ) (xk ) + NX (xk+1 );
( + NX )(xk+1 ) (xk ) tk f (xk );
xk+1 ( + NX )1 ( (xk ) tk f (xk )):

(3.13)

A. Beck, M. Teboulle / Operations Research Letters 31 (2003) 167 175

Proposition 3.2. The sequence {xk } X generated

by MDA corresponds exactly to the sequence generated by SANP.
Proof. By de?nition of the conjugate function, one
has (z) = maxxX { x; z (x)}. Writing the optimality conditions for the later we obtain 0 z
(x) NX (x), which is the same as x ( +
NX )1 (z). But, since is strongly convex on X , then
using Proposition 3.1, is ?nite everywhere and differentiable and one has: =(9 )1 . Thus, the later
inclusion is just the equation
x = ( + NX )

(z) =

(z) = (9 )

Using these relations, SANP can be written as follows. Let yk+1 := (xk ) tk f (xk ) and set xk =
(yk ). Then, SANP given by (3.13) reduces to
xk+1 = (yk+1 ), which are exactly the iterations
generated by MDA.
Note that when satis?es (3.12), then SANP reduces
to: xk+1 = ( )1 ( (xk ) tk f (xk )).

Theorem 4.1. Suppose that assumption A is satis;ed

for the convex optimization problem (P). Let {xk } be
the sequence generated by SANP with starting point
x1 int(X ). Then, for every k 1 one has
(a)

min f(xs ) min f(x)

16s6k

xX

B (x ; x1 ) + 21
6
k

k

2
 s 2
s=1 ts f (x )

s=1 ts

(4.15)

(b) In particular, the method converges, i.e.,

min16s6k f(xs ) minxX f(x) 0 provided that

ts = ; tk 0; k :
s

Proof. Let x be an optimal solution of (P). Optimality for (3.11) implies:

x xk+1 ; tk f (xk ) + (xk+1 ) (xk ) 0;
x X
and thus in particular for x = x we obtain

4. Convergence analysis
With this interpretation of the MDA, viewed as
SANP, its convergence analysis can be derived in a
simple way. The key of the analysis, relies essentially
on the following simple identity which appears to be
a natural generalization of the quadratic identity valid
for the Euclidean norm.
Lemma 4.1 (Chen and Teboulle [5]). Let S Rn
be an open set with closure SM and let : SM R
be continuously di<erentiable on S. Then for any
three points a; b S and c SM the following identity
holds true
B (c; a) + B (a; b) B (c; b)
= (b) (a); c a:

171

(4.14)

We will need some further notations. Let z =

max { x; z | x Rn ; x 6 1} be the (dual) conjugate
norm. The convergence results for the SANP (and
hence MDA) are given in the following theorem. We
assume that the sequence xk produced by SANP is
well de?ned (see Section for the appropriate condition
on ).

x xk+1 ; (xk ) (xk+1 ) tk f (xk )

0:

(4.16)

Using the subgradient inequality for the convex function f one obtains
0 6 tk (f(xk ) f(x ))
6 tk xk x ; f (xk )
= x xk+1 ; (xk )
(xk+1 ) tk f (xk )

(4.17)

+ x xk+1 ; (xk+1 ) (xk )

(4.18)

+ xk xk+1 ; tk f (xk )

(4.19)

:= s1 + s2 + s3 ;

(4.20)

where s1 ; s2 ; s3 denotes the three right-hand side terms

(4.17)(4.19). Now, we have
s1 6 0; [by (4:16)];

172

A. Beck, M. Teboulle / Operations Research Letters 31 (2003) 167 175

Proof. Writing the KKT optimality conditions for the

(equivalent) convex problem



c + (2)1 z T Dz T
: b z=u ;
inf
z;u0
u

s2 = B (x ; xk ) B (x ; xk+1 ) B (xk+1 ; xk )
(by Lemma 4:1);
s3 6 (2)1 tk2 f (xk )2 + 21 xk xk+1 2 ;
the later inequality following from a; b 6 (2)1 a2
+ 21 b2 , a; b Rn . Therefore, recalling that
B (; ) is -strongly convex, i.e., B (xk+1 ; xk ) +
21 xk xk+1 2 6 0, it follows that
tk (f(xk ) f(x )) = s1 + s2 + s3
6 B (x ; xk ) B (x ; xk+1 )
+ (2)1 tk2 f (xk )2 :

(4.21)

Summing (4.21) over k = 1; : : : ; s one obtains,

s

tk (f(xk ) f(x )) 6 B (x ; x1 ) B (x ; xs+1 )
k=1

+ (2)

s


Theorem 4.2. Suppose that assumption A is satis;ed

for the convex optimization problem (P). Let {xk } be
the sequence generated by SANP with starting point
x1 int X . Then, with the stepsizes chosen as

2B (x ; x1 ) 1
;
(4.23)
tk :=
Lf
k

min f(xs ) min f(x)

16s6k

Since B (; ) 0, it follows from the last inequality

that
min f(xs ) min f(x)
xX

B (x ; x1 ) + (2)1
6
k

We can now derive the e3ciency estimate for

SANP.

one has the following e?ciency estimate

tk2 f (xk )2 :

k=1

16s6k

yields the desired result.

k

2
 s 2
s=1 ts f (x )

s=1 ts

;
(4.22)

proving (a). Assuming that tk 0 and

tk = as
k , it thus follows from (4.22) that min16s6k
f(xs ) minxX f(x) 0 as k , proving (b).
The above convergence result allows for deriving
the best e3ciency estimate of the method, by choosing an appropriate step size. The best stepsize is
obviously obtained by minimizing the right-hand side
of the inequality (4.22), with respect to t Rk++ . We
need the following technical result.
Proposition 4.1. Given c 0, b Rd++ and D a symmetric positive de;nite matrix one has

c + (2)1 z T Dz
2c
inf
=
Tz
T D1 b
d
b
b
zR++

with optimal solution z = (2c=bT D1 b)D1 b.

6 Lf

xX

2B (x ; x1 ) 1
:

k

(4.24)

Proof. The right-hand side of (4.22) is upper bounded

by
k
B (x ; x1 ) + (21 )L2f s=1 ts2
;
(4.25)
k
s=1 ts
Minimizing (4.25) with respect to t1 ; t2 ; : : : ; tk 0,
and invoking Proposition 4.1 with c := B (x ; x1 ); b :=
e = (1; 1; : : : ; 1)T and D = L2f I where I is the k k
identity matrix, one gets the desired step size and
e3ciency estimate.
Clearly, in order to make this result practical, one
has to be able to upper bound the quantity B (x ; x1 ),
which depends on the (obviously unknown) optimal
solution x , so that the step size and the e3ciency estimate can be computed. This can be done by de?ning
for any y int X %[ ; y] := maxxX B (x; y). Thus,
one can replace in the estimate (4.15), the quantity
B(x ; x1 ) by %[ ; x1 ], provided the later quantity is
?nite. This is particularly true whenever X is assumed
compact, as done in [1] for MDA in which case the
results of Ben-Tal et al. [1] are recovered through

A. Beck, M. Teboulle / Operations Research Letters 31 (2003) 167 175

Theorem 4.2. Also, note that when we replace the

di@erentiability assumption for on int X and (3.12)
holds, then one obtains an interior subgradient
algorithm to minimize f over X , where in (4.23)
and (4.24) the quantity B(x ; x1 ) is replaced by
%[ ; x1 ] for any x1 int X .

Proposition 5.1. Let e : & R be the entropy function de;ned in (5.27). Then,
(a) e is 1-strongly convex over int & with respect
to the  1 norm, i.e.,
e (x) e (y); x y
=

5. Application: minimization over the unit simplex

As explained before, the key elements needed to
implement the MDA and analyze its e3ciency rely on
our ability to compute the conjugate function of
e3ciently; to evaluate the strong convexity constant
of , and to upperbound the quantity B (x ; x1 ). In
this section, we begin by recalling brieKy the results of
Ben-Tal et al. [1], where the authors analyze the problem (P) of minimizing a convex
n function f over the
unit simplex & := {x Rn : j=1 xj = 1; x 0}, and
we introduce a new method for this class of problems.
The MDA1 (Ben-Tal et al. [1]): Let (x)= 1 (x) :=
21 x2p with p := 1 + (ln n)1 . It was proved in [1]
that the following e3ciency estimate for MDA1 holds:
xX

16s6k

6 O(1)

(ln n)1=2 max16s6k f (xs )

(5.26)

and thus, the MDA1 with 1 can outperformed the

usual gradient method (obtained with 2 on &) by
a factor of (n=ln n)1=2 , which for large n, can make
a huge di@erence. This method was considered as a
nearly optimal algorithm for the class of problems
under consideration. For further details, see [1].
We now propose a di@erent choice for to solve
the minimization problem (P) over the unit simplex &,
which shares the same e3ciency estimate. The function appears to be quite natural due to the simplex
constraints, and is the so-called entropy function
de?ned by
e (x)

n


xj ln xj if x &;

n


(xj yj ) ln

j=1

x y21 ;

xj
yj

x; y int &:

(b) The conjugate of e is the functione : Rn R

n
with e C (Rn ) given by e (z) = ln j=1 ezj , and
 e (x) as x xM &.
(c) For the choice x1 = n1 e, and = e one has
B (x ; x1 ) 6 ln n, x &.
Proof. (a) The strong convexity of e follows from
a fundamental inequality in information Theory [7].
For completeness we give here a di@erent and simple
proof. Let : R++ R be de?ned by
(t) = (t 1) ln t 2

min f(xs ) min f(x)

+ otherwise;(5.27)

j=1

where we adopt the convention 0 ln 0 0.

The entropy function de?ned on & possesses some
remarkable properties collected below.

173

(t 1)2
;
t+1

t 0:

It is easy to verify that (1) =  (1) = 0 and that

 (t) 0 t 0. Therefore is convex on (0; )
and it follows that (t) 0, t 0. Therefore, with
t := xj =yj it follows that x; y int &:
n

j=1

(xj yj ) ln

 (xj yj )2
xj

2
yj
xj + y j
j=1

n

xj + yj (xj yj )2
x +y
2
( j 2 j )2
j=1

2
n

x
+
y

y
|
|x
j
j
j
j

xj +yj
2
2

()

j=1

= x y21 ;
where the inequality (*) follows from the convexity of
the quadratic function and the fact that (x + y)=2 &.
(b) Using the de?nition of the conjugate and simple
calculus gives the desired results, see also [11].

174

A. Beck, M. Teboulle / Operations Research Letters 31 (2003) 167 175

n
(c) Substituting = e (x) = j=1 xj ln xj in the
de?nition of B we obtain with xj1 = n1 , j,

n

xj
xj ln
B (x ; x1 ) =
xj1
j=1
=

n


xj ln xj + ln n 6 ln n;

x &;

j=1

the last inequality being true since the entropy function

is always nonpositive on &.
Remark 5.1. If for some j one has yj = 0; xj 0, the
left-hand side of the strong convexity inequality in (a)
is + and there is nothing to prove. Likewise, when
we reverse x with y. Thus, recalling that 0 ln 0 0, it
follows that the strong convexity inequality given in
(a) remains true for all x; y &.
Using the entropy function e in (3.12), we thus obtain
a very simple algorithm for minimizing the convex
function f over &, which is given explicitly by
The entropic descent algorithm (EDA)
Start with x1 int & and generate for k = 1; : : : ; the
sequence {xk } via:

k
xjk etk fj (x )
2 ln n 1
k+1
;
xj = n
; tk =
k tk fj
(xk )
Lf
k
j=1 xj e
where f (x) = (f1 (x) ; : : : ; fn (x))T 9f(x).
Applying Theorem 4.2 and Proposition 5.1 we immediately obtain the following e3ciency estimate for
the EMDA.
Theorem 5.1. Let {xk } be the sequence generated by
EMDA with starting point x1 = n1 e. Then, for all
k 1 one has

which still requires the solution of one-dimensional

nonlinear equation at each step of the algorithm to
compute 1 .
6. Concluding remarks and further applications
We have presented a new derivation and analysis of
mirror descent type algorithms. In its current state, the
proposed approach has given rise to new insights on
the properties of Mirror descent methods, bringing it
in line of subgradient projection algorithms based on
Bregman-based distance-like functions. This has led
us to provide simple proofs for its convergence analysis and to introduce the new algorithm (EMDA) for
solving convex problems over the unit simplex, with
e3ciency estimate mildly dependent on the problems
dimension. Many issues for potential extensions and
further analysis include:
m
Extension to the cases where f(x) = l=1 fl (x)
which can be derived along the analysis of incremental subgradients techniques [9,1] and numerical
implementations for the corresponding EMDA.
The choice of other functions can be considered
in SANP, (see for example [14,8]) to produce other
interior subgradient (gradient) methods.
Extension to semide?nite programs, in particular for
problems with constraints of the type
Z Sn ;

tr(Z) = 1; Z  0

and which often arise in relaxations of combinatorial optimization problems. This can be analyzed
within the use of a corresponding entropic function
de?ned over the space of positive semide?nite symmetric matrices (see for example [6] and references
therein).

min f(xs ) min f(x)

xX

16s6k

2ln n

f (xs )

:
k

(5.28)

Thus, the EMDA appears as another useful candidate algorithm for solving large scale convex minimization problems over the unit simplex. Indeed,
EMDA shares the same e3ciency estimate than the
(MDA1 ) obtained with 1 , but has the advantage of
being completely explicit, as opposed to the (MDA1 )

References
[1] A. Ben-Tal, T. Margalit, A. Nemirovski, The ordered subsets
mirror descent optimization method with applications to
tomography, SIAM J. Optim. 12 (2001) 79108.
[2] D. Bertsekas, Nonlinear Programming, 2nd Edition, Athena
Scienti?c, Belmont, MA, 1999.
[3] L.M. Bregman, A relaxation method of ?nding a common
point of convex sets and its application to the solution
of problems in convex programming, USSR Computational
Mathematics and Mathematical Physics 7 (1967) 200217.

A. Beck, M. Teboulle / Operations Research Letters 31 (2003) 167 175

[4] Y. Censor, A. Lent, An iterative row-action method for
interval convex programming, J. Optim. Theory Appl. 34 (3)
(1981) 321353.
[5] G. Chen, M. Teboulle, Convergence analysis of a proximallike minimization algorithm using Bregman functions, SIAM
J. Optim. 3 (1993) 538543.
[6] M. Doljanski, M. Teboulle, An interior proximal algorithm
and the exponential multiplier method for semi-de?nite
programming, SIAM J. Optim. 9 (1998) 113.
[7] J.H.B. Kemperman, On the optimum rate of transmitting
information, Ann. Math. Statist. 40 (1969) 21562177.
[8] K.C. Kiwiel, Proximal minimization methods with generalized Bregman functions, SIAM J. Control Optim. 35 (1997)
11421168.

175

[9] A. Nedic, D. Bertsekas, Incremental subgradient methods for

nondi@erentiable optimization, SIAM J. Optim. 12 (2001)
109138.
[10] A. Nemirovski, D. Yudin, Problem complexity and Method
E3ciency in Optimization, Wiley, New York, 1983.
[11] R.T. Rockafellar, Convex Analysis, Princeton University
Press, Princeton, NJ, 1970.
[12] R.T. Rockafellar, Monotone operators and the proximal point
algorithm, SIAM J. Control Optim. 14 (1976) 877898.
[13] R.T. Rockafellar, R.J.B Wets, Variational Analysis, Springer,
New York, 1998.
[14] M. Teboulle, Entropic proximal mappings with application
to nonlinear programming, Math. Oper. Res. 17 (1992)
670690.