Network Tomography: Estimating Souree-Destination Traffie Intensities

from Link Data

Y. Vardi

Journal ofthe American StatisticalAssociation, Volume 91, Issue 433 (Mar., 1996),
365-377.

Stable URL:
http://links.jstor.org/sici?sici=0162-1459%28199603%2991%3A433%3C365%3ANTESTI%3E2.0.CO%3B2-2

Your use of the JSTOR archive indicates your acceptance of JSTOR' s Terms and Conditions of Use, available at
http://www.jstor.org/about/terms.html. JSTOR's Terms and Conditions ofUse provides, in part, that unless you
have obtained prior permission, you may not download an entire issue of ajoumal or multiple copies of articles, and
you may use content in the JSTOR archive only for your personal, non-commercial use.

Each copy of any part of a JSTOR transmission must contain the same copyright notice that appears on the screen or
printed page of such transmission.

Journal ofthe American Statistical Association is published by American Statistical Association. Please contact
the publisher for further permissions regarding the use of this work. Publisher contact information may be obtained
at http://www.jstor.org/joumals/astata.html.

Journal ofthe American Statistical Association

©1996 American Statistical Association

JSTOR and the JSTOR logo are trademarks of JSTOR, and are Registered in the U.S. Patent and Trademark Office.
For more information on JSTOR contactjstor-info@umich.edu.

©2003 JSTOR

http://www.jstor.org/
WedMar 1214:12:412003
Network Tomography: Estimating Source-Destination
Traffic Intensities From Link Data
y. VAROI

The problem of estimating the node-to-node traffic intensity from repeated measurements of traffic on the links of a network is
formulated and discussed under Poisson assumptions and two types of traffic-routing regimens: deterministic (a fixed known path
between each directed pair of nodes) and Markovian (a random path between each directed pair of nodes, determined according
to a known Markov chain fixed for that pair). Maximum likelihood estimation and related approximations are discussed, and
computational difficulties are pointed out. A detailed methodology is presented for estimates based on the method of moments.
The estimates are derived algorithmically, taking advantage of the fact that the first and second moment equations give rise to a
linear inverse problem with positivity restrictions that can be approached by an EM algorithm, resulting in a particularly simple
solution to a hard problem. A small simulation study is carried out.
KEY WORDS: Communication network; Computer network; EM algorithm; Linear inverse problems; Maximum-likelihood;
Moment method; Origin-destination tables; Poisson traffic; Positivity constraints; Trip matrix.

1. INTRODUCTION independent. We denote vector s and matrices by boldface

We consider the problem of estimating the traffic intensity
between all source-destination (directed) pairs (SD pairs
heretofore) of nodes of a network from repeated measure-
ments of traffic flow along the directed links of the net-
work. For concreteness and simplicity of presentation, we
discuss the problem in terms of "packets" or "messages"
transmitted over a communication network although the
development is applicable to any network for which our
formal assumptions (described later) hold. We assume that
the SD pairs transmit messages over a strongly connected
directed network (Le., there exists a direct path between any
two nodes). The networks are divided into two categories:
fixed routing (deterministic)andrandom (Markovianlrout-
ing networks. In fixed routing networks, the directed path
between any two nodes is fixed and known and remains the
same for all messages traveling between these two nodes.
In random routing networks, the directed path taken by a
message traveling from the source node j¡ to the destinatioil
nodej, is determined according to a fixed known Markov
chain specific to the SD pair (]1,]2). From a purely tech-
nical standpoint, fixed routing is a special case of random
routing, much as a sequence of iid random variables is a
special case of a Markovian sequence. But there are' many
advantages to treating the two cases separately. and so, for
and transpose by a prime. X(k) == (xi k ) , ... ,X~k)), is the
SD 'transmission vector at period k, k == 1, ... , K. Let r
be the number of directed links in the network, and let
A denote.the r x e routing matrix for a fixed-routing net-
work. A is a zero-one matrix; its rows corresponds to the
directed links, its columns correspond to the SD pairs, and
its entries, aij' s, are "1" or "O"according to whether link
i does or does not belong to the directed path of the SD
pair ]. (For simplicity, we index the SD pairs and the di-
rected links with one dimensional indices, typically j and
i. This may occasionally lead to a slight ambiguity as we
may refer to the third SD pair (say) as "] == 3" and also as
] == (jI, ]2) == (a, b) (say). The alternative, however, would
be overly cluttered expressions with too many subscripts
and superscripts, throughout the article.) The measured data
are the traffic on all the links of the network for K time
. ds.. y(k) -_("l".(k)
perlo L 1 , ... , "l".(k)),
L r
k· - 1
, - , ... "
K ,where
k == 1, ... , K. (1)
Our goal is to estimate A == (Al, ... , Ac ) ' from y(l),
within the framework of this model. Because of
... , y(k}

now, we proceed with the case of fixed routing, postpon-
ing the discussion and treatmerit of random routing until
Section 5.
Let c denote the number of SD pairs (c == n(n - 1) in a
network of n nodes), and let Xjk) be the number of trans-
mitted messages for SD pair ] at measurement period k. We
assume that Xjk} rv Poisson(Aj),j == 1, ... , c, k == 1, ... , K,
Y. Vardi is Professor of Statistics, Department of Statistics, Rutgers Uni-
versity, Busch Campus, Piscataway, NJ 08855. The problem originated in
a consulting project to DARPA (Becker and Wilks 1993) regarding the
statistical analysis of computer data networks, and was brought to his
attention by D. Pregibon, C. L. Mallows and P. McCullagh contributed
useful comments.Teanne Yue helped with prograrnming, and National Sci-
ence Foundation provided financial support. The works of Sherali et al.,
and Bell, were brought to his attention,at the proofreading stage of this
article, by R. Becker.
the network interpretation of the model, e (== n(n - 1))
is typically larger than r (== O(n)), and this raises interest-
ing questions regarding identifiability of theparameters and
consistency of the estimates. Note that although the model
is linear, it is neither a "typical" linear regression nor a
random-effects model, because of the zero-one structure of
A, the nonnegativity constraints on the .parameters, andthe
Poisson assumption.
ExampleI, Consider the network of Figure 1. There
are (4 x 3 ==) 12 SD pairs and 7 directed links. The routing
matrix A is given by
© 1996 American Statistical Association
~ournal ofthe American Statistical Association
March 1996, Vol. 91, No. 433, Theory and Methods

365
366 Journal 01 the American StatisticalAssociation, March 1996

simple and will typically be satisfied by any real-life rout-

ing matrix A. This is somewhat unexpected, because in the
simplest example, when r == 1, the Y's are sums of indepen-
dent Poisson variables and so the intensity vector A is not
identifiable (e.g., A == (1,1,0,1), so that Y == Xl +X2+X4
and hence only the sum Al + A2 + A4 is identifiable but not
the individual components Al, A2'A3' A4.)
The following holds.
Lemma. If the columns of the routing matrix A are all
distinct, and each column has at least one nonzero entry,
then A is identifiable. '

Note. A zero column means that the corresponding SD

pair is not connected by a path. Likewise, a zero row means
Figure 1. A Sketch ot a Four-Node Directed Network. (The actual that the corresponding link is not a part of the network.
routs between any two nodes are separately specified by the matrix A Thus any real-life routing matrix A would not have either
of Example 1.1.)
columns or rows of all zeros.
1 2 3 4 5 6 7 8 9 10 11 12
ab ae ad ba be bd ea C;b ed da, db de
Proof. See the Appendix.
1 (a--+b) 1 O O O O O O O O O O O Corollary. If e > 2T - 1, then sorne of the individual
2 (a--+e) O 1 1 O O 1 O O O O O O
. 3 (b --+ a) O O O 1 O 1 1 O O 1 O O rates Al, ... , Ac cannot be estimated separately.
4 (b --+ e) O O O O 1 O O O O O O O
5 (e--+b) O O O O O O 1 1 O 1 1 O
Proofi , 2T
- 1 is the maximal number of distinct zero-
6 (e --+ d) O O 1 O O 1 O O 1 O O O
7 (d--+e) O O O O O O O O O 1 1 1 one vectors with at least one "1" in each vector.

Note: The routing is deterministic and is .prespecified for 3. E8TIMATION (FIXED ROUTING)
each SD pairo (This assumption is relaxed.in Seco 5). For
We start with a discussion of the maximum likelihood
instance, although in Figure 1 both paths, b ,'---+ e ---+ d and
estimator (MLE) and an iterative procedure,' based on the
b ---+ a ---+ e ---+ d, connect the SD pair (bd), the routing
EM algorithm..The trouble with this approach, however,
matrix specifies that messages from b· tod always travel on
is .that the iteration formula is very hard .to compute. In
the path b ---+ a ---+' e ---+. d. Traffic from d to .b, on the other
the course of the discussion, .to demonstrate the difficulty
hand, always travels on the path d ---+ e ---+ b.
of the problem, wegive a simple but constructive example
In Section 2 we show that if the columns of A are all
where the likelihood equation has a. unique solution that
nonzero and different from each other, then A is identifi-
differs .from theMLE, which is also unique. Alternative
able. In' Section 3 we discuss estimation via the methods
methods that are potentially more computationally tractable
of maximum likelihood (ML) and moments. ML estimation
arediscussed next. We then propose and experiment with
is numerically complicated for the problem at hand. But
a particularly simple method motivated by the method of
the method of moments is simple, because the moment-
moments. Given that we are estimating parameters within
equations make up a LININPOS (LINear INverse POSitive)
the exponential family, it is reasonable to, expect that this
problem and so can be solved using a simple EM/ML algo-
method is relatively efficient. Our simulation experiment
rithm (Snyder, Schultz, and O'Sullivan 1992; Vardi and Lee
(discussed in the next section) indicates that the method
1993) to produce a method-of-moments estimate. "Wethen
produces .approximately unbiased estimates of A already
simulate a small network to check the performance of such
for moderate values of K.
an estimate. The simulation resultssuggest that the estimate
tends toward unbiasedness already fo.r a moderate number 3.1 Maximum Likelihood Estimation
of repeated link-traffic measurements. In Section 4 we of-
fer sorne possible modificationsand discuss an estimation We start with an interesting "toy" example that demon-
strategy when the Poissonassumption can be,viewed only as strates the difficulty of the problem due to boundary solu-
a crude approximation. The proposed method of weighted tions.
moment equations thenamounts to estimationbased on the Example 3.1: (r == 2, e == 3, K == 1). Let X,
first moments (which do not depend on the Poisson assump- rv Poissoru.x.), i == 1,2,3 independent, and suppose that
tions) regularized by the second moments (which do depend we are given
on the Poisson assumption). Finally, in Section 5 we derive
method-of-moments estimates for networks with Marko-
vian routing.

2. IDE·NTIFIABILITY
Our goal is to derive the ML estimate of A. In the notation
The conditions for identifiability of A are surprisingly of the earlier sections, we have
Vardi: Network Tomography 367

and, theoretically, the EM algorithm (Dempster, Laird, and

The following holds: The unique solution of the likeli-
hood equations is A == (0,1,2)'. But the unique MLE is
~ == (1,0,1)'. To verify this, we first derive the likelihood
of the data: (2) has only two possible solutions in natural
numbers for the X's:
Rubin 1977) can be used to search for a solution. But the
implementation is a different story. If we treat the X(k), k
== 1, , K, as the complete (unobserved) data and y(k), k
= 1, , K, as the incomplete (observed) data, then the EM
algorithm, in vector notation, is
A (n+l) = E[Xly(I), ... , y(K), A (n)],
n= 0,1, ... ,

x == (1,0,1)' and X == (0,1,2)',

(4) (.x(O) > O, arbitrary). Because of linearity oí E and inde-
pendence across k's, (8) becomes
and so the likelihood of the data is K
.x (n+l) = ~ 2: E[X(k) Iy(k), .x (n)], n = 0, 1, . . . . (9)
L(A) - p~ {Y == (1, 2)'} k=l

p~ {X == (1, 0, 1)'} + p~ {X == (O, 1, 2)'} The trouble with this iteration formula is that the summands

== (AIA3 + A2A~/2) exp( -Al - A2 - A3). (5) E[XIY,A] (10)

Let l(A) == log L(A). Setting the partial derivatives of l equal
to zero, we get the likelihood equations
(superscripts ignored for simplicity) are extremely hard to
calculate, as they require first finding all the solutions in
natural numbers ofAX == Y. To demonstrate the details of
the procedure, consider again our "toy" example, where A
and Y are given by (3). From (4), we derive

(e) (2AIA3 + A2A~)-12(AI + A2A3) == 1. (6)

[XIY, A] r-;»

Equations (óa) and (ób) imply that

( ~) with probo 2,Atf(2,Al + ,A2,A3),
A3 == 2.
(11)
Substituting this back into (6) results in the equations
and so (9) leads to the iteration

°
which have a unique solution Al == and A2 == 1. Thus
A == (O, 1,2)' is the unique solution of the likelihood equa-
tions. But the MLE is ~ == (1,0,1)'. To prove this, first
note that the MLE must be a boundary point of the region
Ai 2:: 0, i = 1,2,3, because no interior point satisfies the
likelihood equations (as we just proved). From (5), it is clear
that any candidate for a MLE must have A3 > and one of °
Al or A2 strictlypositive (but not both, because otherwise it
will be an interior point), If Al = 0, then we need to max-
imize (A2A~/2) exp( ~A2 - A3) subject to A2' A3 > 0, and
this maximum is attained at A2 = 1 and A3 = 2. If A2 = 0,
then we need to maximize AIA3 exp(-AI - A3) subject to
AI,A3 > 0, and this maximum is attained at Al = A3 = 1.
Thus there are two possible candidates for MLE: (O, 1,2)'
and (1,0, 1)'. Comparing the likelihoods of these candi-
dates, one verifies that L((O, 1,2)') < L((l, 0,1)'), and so
the unique MLE is A == (1,0, 1)'.
3.2 An EM Derivation of the MLE
The likelihood equations (O == 8l/8Aj,j 1, ... ,c,l
- log L) in vector notation can be expressed as
(12)
The computational difficulties in replicating this process
and in calculating (10) for anything but a "toy problem'tare
apparent. Unconnected to this computational difficulty, note
here that both (0,1,2)' and (1,0,1)' are stationary points of
the iteration (12), and so, depending on the starting point,
the algorithm may converge to a non-MLE point. .
Another interesting question is whether the log-
likelihood isconcave, which would have simplified the
likelihood optimization problema After sorne algebra, it is
possible to express the matrix of second derivatives of the
log-likelihood function l as
2l
H == ( 8 )
- 8Ai8Aj
K
== 2: cov[x(k) / Aly(k) , A]
k=l

0=
K
~ 2:E.\[X(k)ly(k) = AX(k)] ~.x,
k=l
(8) - diag (t, E[X(k) /.x 2Iy(k),.x]) , (13)
368 Journal of the American Statistical Association, March 1996

where the quotient vector s X(k) / A and X(k) / A2 are the large K. We start 'with the first. Assume that
vectors of coordinate-wise ratios. It follows that
A == diag(A), (17)

and consider the joint distribution of (X', Y')' when Y

AX. Then

It can then be shown that this quantity is not necessarily

::; O for all A, and so l is not necessarily concave. Note, and
however, that when the true A, say A*, is an interior point,
then for large K's, l is concave in the neighborhood of A*. XIY r-;» Nc{A + AA'(AAA)-l(y - AA),
To see this, apply the law of large numbers to (14) to get
A - AA'(AAA')-l AA}. (19)

~ ~zizjHij
"',J
----+ E"*var,, [t ~: XiIY]
",=1
In particular, (10) is approximated as A + AA' (AAA -1)
(Y - AA). Applying this to (9) results in the iteration for-

- E"*E,, [t ~} XiI Y ] , (15)

mula
K '
,X(n+l) = ~L [,X(n) +A(n)A/(AA(n)A/)-l
where Y = AX. Setting A = A* in (15), we can apply the k=l
general formula

var[U] = E(var[UIV]) + var(E[UIV]) We have not experimented with this method, but we suspect
to get that it may perform badly for sorne cases because the nor-
mal approximations (17), on which the iteration is based,
(right side of (15) with A = A*) is a poor approximation unless we ··know a priori that all
of the Ai' s are large. Other potential difficulties with this
= varj, (t ~i. Xi)
i=l '"
approach are the cumulative approximation error in each
step of the iteration (because we apply the normal approx-
imation to each term in (9) separately), potential numerical

- varj, ( E" [t ~: XiI Y]) instability from the matrix inversion in each iteration, and
the possibility of converging to negative values of Ai'S.
In the second normal approach,we approximate the dis-

- E" [t (~:r -. tribution of Y when K is large witha normal distribution:

K
Y = ~ Ly(k) d~. Nr(A'x,K-1AAA/), A = diagjX).
(A = A*), k=l

(16)
where the last equation follows from theequality of the
mean and the variance for Poisson random variables and
from the independence of the Xi's. Thus if K is large and
the algorithm is iterated in the vicinity of the true A (as-
sumed interior), then the algorithm converges to anMLE
near the true A. In fact, irrespective of the numerical dif-
ficulties in calculating the MLE, for large K the standard
asymptotic properties of the MLE should hold for the case
at hand.
3.3 MLE and Normal Approximations
Two types of normal approximations can be considered
"natural" here. The first is a normal approximation of each
of the summands in (9), which amounts to approximating
(10) using multivariate normal theory. The second is based
on the centrallimit theory (CLT) approximation of Y for
(21)
Treating the right side of (21),as the exact distribution of Y,
the log-likelihood of Y (modulo multiplicative andadditive
constants) is
l(A) = -lag IAAA'I- K(Y.- AA).'(AAA')-l(Y - AA).
(22)
A Ml.E'based on this approximation would seekto max-
imize l(A) of (22) subject to the constraints Ai 2:: O, i
1, ... ,c. When K is large, the second term is the
dominant term in (22), and this suggests argminA>oK(Y
- AA)'(AAA)-l(y - AA) as a reasonable large-sample
substitute for the MLE. Note that this is a weighted least
square with positivity constraints and with weights depend-
ing on A, which can be estimated by the sample covariance
matrix of the Y's.
Vardi: Network Tomography 369

3.4 Estimation Based on Sample Moments gebraically inconsistent system of equations (with proba-
3.4.1 First and Second Moment Equations. The approx- bility ~ 1). Furthermore, due to sampling variability, we
imate normal distribution of V is completely determined by can also expect to observe sorne negative sample covari-
the mean vector, AA, and covariance matrix, AAA', of Y. ance values, Sii' 's, whereas the theoretical covariance val-
Thus we can equate the sample's first and second moments ues, ¿l ailai'lAl, are always nonnegative. Finally, when
to their theoretical values to obtain a linear (in A) system ¿l auaei = O and Sii' =1= O, we get a degenerate false iden-
of estimating equations: tity Sii' = O, which, as before, is attributable to sampling
variability and should be deleted. These algebraically in-
e
consistent equations are to be expected, and discussion on
(a) E(Yi) = Yi = L ailAl, i = 1, ... ,r how to handle them follows the simulation experiment dis-
l=l cussed latero For now, suppose that all the constants on the
and left side of (24) are strictly positive (Le., Yi, Sii' > O) and
that B has no rows of zeros. (Clearly, A has no such rows.)
(b) -("l.l'" x> ) _ ~ ~ y, (k)y(k) _ "l7."l7.
Then, because all of the entries of A and B are nonnegative
COY .li, .li' - K L...J- i i' .I.i.li'
k and because (V', S')' >. O and A is constrained to be 2:: O,
e (24) is of the general form of a LININPOS problem, and we
= L~ilai'lAl' 1 < i < i' < r. (23) propose using an EM algorithm to"solve" it. Formally, this
l=l solution is the minimizer of the Kullback-Leibler-directed
distance between the observed sample moments and their
These are r(r + 3)/2 linear equations that can be written, theoretical values (modulo a certain natural scaling), and it
in vector notation, as can also be viewed as an MLE of the algebraic solution un-
der a multinomial model. (Details have been given in Snyder
(24) et al. 1992 and in Vardi and Lee 1993.)
3.4.3 On the Relation to Maximum Likelihood Estima-
where S is the sample covariance matrix .stretched out as
tion. Left-multiplying the likelihood equation (8) by A, we .
a vector of length r(r + 1)/2 (which, for simplicity of
see that any stationary point of the log-likelihood function,
presentation here, we index by a double subscript (i, i') 1
::; i ::; i' ::; r, ordered Iexicographically, say),
l, also satisfies the first-moment equations V = AA. The
reverse implication is not true,but a partial converse holds.
_ 1. ~ (k) (k) -- Specifically, if the columns of A are linearly independent
Sii' - K L...J ~ ~, - YiYi', (which is not the case in our typical application where e
k
> r), and A* solves the first-moment equations, V = AA*,
and B is an (r(r +~1)/2) x e matrix (rows indexed by then A* solves (8) and hence is a stationary point of l. To
(i, i'), 1 ::; i ::; i' ::; r, to match the indexing of S) with prove this, denote the right side of (8) by 1](A) and note that
the (i, i')th row of B being the element-wise product of V = AA* implies A1](A*) = O, which implies 1](A*) = O
row i and row i' of the matrix A. for A with linearly independent columns, so that A* is a
solution of (8).
Example 3.2. .Suppose that It is important tonote, however, that because of the non-
negativity constraints on A, the MLE may be a boundary
A=
(1O O 1
1 1
1
O
O 1 O 1
¡). point and so is not necessarily a stationary point of l, as
demonstrated in Example 3.1.
3.4.4 Details of the Method: Implementation for Matri-
Then
ces in Block Form. The canonical form of the EM iteration
1 O 1 1 1 SIl for solving the LININPOS problem V = AA is
O O 1 O 1 S12
O O O 1 1 S13
B= S=
O 1 1 O 1 S22
O 1 O O 1 j = 1,.... ,c. (26)
S23
O 1 O 1 1 S33

3.4.2 A Closer Look at the Moment Equations (24). If the linear system is given in a block form as
First we recall that because the X, 's are independent and
the aij' s are zeros or 1s, the marginal distributions of the
Yi's are also Poisson and so
(27)
i = 1, ... .r. (25)

On the other hand, sampling variability will typically re-

sult in Yi =1= Su, i = 1, ... , r, and so (24) will be an al- where A is r x c, V is r x 1, and S is m x 1 (indexed as r
370 Journal of the American Statistical Assoclatíon, March 1996

+ 1, ... , r + m) and B is m x e (rows indexed as r + 1, and Lee 1993),

... , r + m), then (26) becomes the following iteration:
~J.(A , Y , A) ==
-
Aj ' "
L..-J~
aij
A'Yi j == 1, ... , c, (30)
A·J a.j i LJk aik k
Aj ~
~r
LJi= 1 aij + ~r+m b
LJi=r+ 1 ij for any matrix A and vectors Y and A, with their respective
positivity constraints.
3.4.5 A Simulation Experiment. To simulate the net-
work described in Example 1.1, we took A == (Al,
j == 1, ... , c. (28) A2, ... ,A12) == (1,2, ... , 12) to be the "daily" transmission
rate between all 12 pairs and generated daily data on the
A similar block decomposition was first utilized by lusem seven links for 100 days, y(l) (= (y}l), ... , Y7(1)),), ... ,
andSvaiter (1994). Equation (28) can now be written as y(lOO). We then calculated the sample means, Y, and
covariances, S, and applied the algorithm (29). This
led to an estimate ~ == (~l' ~2, ... ' ~12)" We repeated
this whole process 50 times to produce 50 such esti-
A (1) A (50)
mates, A , ... , A . The mean vector and covariance
with "." denoting surnmation over the subscript that it re- matrix based on these 50 estimates were X == (1.01,
places (e.g., a.j == Ei aij) .. Here ~j is defined by the canoni- 2.37,2.68,4.72,5.06,5.79,6.84,7.92,9.25,9.87, 11.33, 12.14)
cal EM iteration for LININPOS problems (as noted in Vardi and

.01 -.01 -.01 -.10 -O .03 -.01 -.09 -.01 -.07 .08 -.02
1.41 -.62 .44 .02 -1.22 -.40 .15 .89 -.10, -.81 -.07
3.00 .88 -.27 -1.48 -.18 -.04 -.51 .10 .69 -.62
8.62 .15 -1.55 -2.15 6.56 .46 -7.05 -1.88 3.87
.22 .20 -.24 .33 .10 .08 -.27 .05
OOV(~) == * * reflect * * 3.78 -.51 -.27 -.97 -.01 .67 .55
6.96 -5.03 .46 -2.84 2.46 -1.1'5
10.79 -.13 -3.48 -7.38 5.76
1.78 .24 -.87 .07
12.69 -1.60 -4.52
11.50 -5.93
8.54

The simulation results suggest that the method leads to
unbiased estimation of A and that the variance of the esti-
mated Aj increases monotonically with Aj.
3.4.6 On Algebraic Inconsistency and Bad Equations.
Algebraic inconsistency of (24) of the type described in
connection with Equation (25) is not a real problem. In fact,
this inconsistency often implies that the ML/EM solution
of the LININPOS problem (24) has a unique solution. (This
follows from Byrne 1993 and from Roeder, Devlin, and
Lindsay 1989.) Intuitively, the point is that both equations
Yi == EYi and Su == EYi are 'believable and valid as mo-
ment equations under the Poisson model, and so we let the
"geometryof Kullback-Leibler distances" determine the
best compromise between all these inconsistent equations.
But the situation is different when we have infeasible equa-
tions (e.g., when Sii' < O but El ailai'l > O) or false iden-
tities (e.g., when Sii' =1= O but El ailai'l == O). These are
known to be impossible moment equations under the Pois-
son model and should be given no weight by comparison
to feasible moment equations. But assigning no weight to
these equations amounts to simply deleting them. Because
there is no shortage of moment equations, deleting the bad
ones seems a reasonable strategy. It is also related to our
later discussion of reducing weights to sorne moment equa-
tions.
4. REGULARIZ~TION:WEIGHTED
MOMENT-EQUATIONS
The first-moment equations Y == AA follow from the
linearity of the system (1) and is independent of the Poisson
assumption. The second-moment equations depend strongly
on the Poisson model, and so in. applications where this
model is only a crude approximation, one may want to give
these equations smaller weight relative to the first-moment
equations. This is achieved by replacing S and B in (27)
with eS and eB, for O < e < 1. This in turn changes (29)
into
Note that when e -+ O, the solution is based entirely on the
first moments, and as such it is independent of any proba-
Vardi: Network Tomography 371

bilistic assumption (but most likely is not a good estimate
Of'A because of shortage of estimating equations). On the
other hand, when e -+ 1, the solution is based on the first-
and second-moment equations for the Poisson model. Thus
e can be interpreted as a regularization parameter, when the
regularization in thiscase attempts to pull the model toward
a Poisson model. This approach, in a different context and
with a goal similar in spirit to Silverman, Jones, Wilson,
and Nychka's (1990) EMS and Green's (1990a,b) one-step-
late EM, was suggested by Iusem and Svaiter (1994), who
discussed convergence properties of (31).
5. RANDOM (MARKOVIAN) ROUTING
Here we consider the same problem under the same model
assumptions as in Sections 1-4, except that the assump-
tion of fixed routing is replaced by Markovian routing. For
specificity and ease of description, we use the term SD ad-
dress (and sometimes just address) to mean the combined
sender and receiver addresses, so that a packet traveling
from source i, to destination j2 is said to have an SD ad-
dress j (= (jI, j2) ). (This is like a standard letter, which
usually includes both sender and receiver addresses.) As
before, the routing regimen is specified by an r x e matrix
A, but now the entries are conditional probabilities:
The 12 columns of A correspond to the 12 Markovian net-
works depicts in Figure 2.
As in the earlier sections, we observe the traffic on the
links and are interested in estimating thetransmission rates
for all .SD addresses,· based on these data and the matrix
A. The problem at hand resembles the estimation problem
in PET (Shepp and Vardi 1982; Vardi, Shepp, and Kauf-
man 1985), where the link data (also Poisson, as sums
of "thinned" Poisson variables) are the equivalent of the
"detector tubes" data in the PET problem. But there is a cru-
cial difference that makes the current problem much harder.
Although the detector tubes data in PET are mutually in-
dependent, the link data are highly dependent. In fact, the
current problem is similar to a PET problem in which the
detector tubes overlap. Thus ML estimation appears to be
very complicated. Because of this, we focus again on the
method of moments.
Despite the significantly more complicated nature of the
random routing regimen, it turns out that the first and sec-
ond moments of the observed link data are linear in A, and,
as in the case of fixed routing, the moment equations con-
stitute a LININPOS problem that can be "solved" using an
EM/ML algorithm. The hard part is then the calculation of
EYi and cov(Yi, Yi,), 1 ~i, i' ~ r, which are needed to set
up the moment equations, These are derived latero

aij is the conditional probability that a packet with SD- 5.1 Estimation: First and Second Moments
address j passes through link i = (i l , i 2), conditioned on the 5.1.1 Network Thinning of Poisson Variables. In gen-
packet being at node i l . (It is assumed that this probability eral, "thinning" of Poisson variables is a reference to
is independent of how the packet arrived at i l but depends the following general property: IfX rv Poissoní.x)
only on its SD address j.) . and [YI , ... , YkIX] rv multinomial(X;7rI,'" ,7rk), then
For any fixed SD pair, the Markovian nature of the routing YI, ... , Yk are mutually independent andr, rv Poissonf.or.),
is a consequence of the assumption that the routing of any i ~ 1, ... , k. (The index k here should not be confused with
packet at each particular node does not depend on the path the index k for themeasurement period in earliersections,
through which it arrived at the node, but depends only on as the derivation of the moments in this section requires
it being there and on its SD address. Thus each column consideration of only a single measurement period.)
of the matrix A corresponds to a directed network with e We can use this property as follows: Let X j be the num-
ber of packets with SD address j; X j rv Poissont.xj). Let
nodes and r links, with probabilities attached to each link.
Typically, if the probability of passing through a given link Y/ be the number of packets with address j passing through
is zero, then we do not include the link in the diagram of link i (= (i l , i 2)). It follows from the "thinning property"
the network. Wefurther assume that the routing matrixA is that
such that a packet cannot revisit a node, and that for each Y/ rv Poisson(AjPl), (32)
SD address j = (jl,j2), the corresponding column of A
specifies a network of paths allleading from i, (source) to where
j2 (destination).
pI probability that a packet with address j
Example 5.1. Consider a network of 4 nodes (and hence passes through link i.
12 SD addresses) and 9 possible links given by the follow-
ing matrix A (blank entries are equal to zero): Furthermore, because of the independence of the Xj's, the
1 2 3 4 5 6 7 8 9 10 11 12 Y/ 's are independent across j for fixed i, and so
ab ae ad ba be bd ca eb ed da db de
1 a ~ b .8 .2 .2
2a ~ e .2 .8 .8 1 1 i = 1, ... , r. (33)
3b~ a 1 .2 .1 1 1
A= 4b~ e .8 .8 .1 1
d
b~
5
6 e~ b . 8
.2 1
.2
.8
.8 .8
1
.2 1 1 Next, we express pi in terms of the routing matrix A.
t e-s « .2 .8 1 .2 .2 .8 We define a chain of connected links as a path, and we say
8b~d 1 1 1 .8 .8 .2 that a path is possible if it has positive probability. Specif-
9 d~e 1 .2 .2 .8
ically, a path (i l, ... , ik) is possible for SD address j if it
372 Journal of the American Statistical Association, March 1996

ah ad
Q . Q
!.2 ¡.s
~~ i
G)E 1.0 @
s
Q
GJ
i 2
CE)
~~
1.0
~)'\j)

ha
a

1.0

't~~
ea eh ed
. G G G
O
Q CE)
G)E .@ 7~-
1.0 GJ 1.0~'\j)

da dh de
Q Q Q
·1.0/ Q
@<
iY'Z@ .8

Figure 2. The Four-Node Network of Example 5.1 With Its Markovian Routing Scheme (as Specified by the Matrix A) Gives Rise to 12 Markov
Networks Corresponding to the 12 Possible SD Addresses, as Depicted Here.

is possible to travel in sequence through i l , i 2 , ... , ik un- For example,

der the network probabilities for SD address j.We say
that a path is complete for SD address j if it is possi- pi = P~d = .2 X 1.0 + .8 X .2 X 1.0
ble and leads from the source (JI) to the destination (j2)'
Let B{ denote the set ofall complete paths for address j = .36 (= aI3 X a53 + a23 X a63 X a53)'
that pass through link i. Then the chain rule for probabil-
ities (P(D 3D2DI) = P(D3ID2DI)P(D2IDI)P(DI)) ando 5.1.2 Calculating the First and Second Moments of Y.
the Markovian nature of the routing imply that the proba- As it .stands, formula (34) is somewhat hard to calculate
bility for a packet to travel on a possible path (i l , ... , ik) because of the complexity of constructing the set of paths
in B{ is the product ai1j ... aikj' Hence B{ . Here we give a relatively simple method for deriving
P from A, using Markov chains techniques, and demon-
(34) strateit onthe foregoing example. Expressions (33) and (34)
give us the necessary ingredients for the moment equations
based on the mean and variance, because
(i = 1, ... ,r,j = 1, ... ,c).
EY=PA, Var y = PA, (35)
Example 5.1 (continued). We get the following p!'s,
i = 1, ... ,9 and j = 1, ... ,12 (blank entries are equal to where var Y = (varYI , ... , varYr ) ' and P is given in (34).
zero): To construct additional second-moment equations corre-
sponding to cov(Yi, Yi,), we need the following. Let
1 2 3 4 5 6 7 8 9 10 11 12
ab ae ad ba be bd ca eb ed da db de
1
2
a
a
-+
-+
b
e
.8
.2
.2
.8
.2
.8 .2 .1
~1, = the number of packets with SD address j
p=
3
4
5
b
b
b
-+
-+
-+
a
e
d
.16
.04 .36
1 .2
.8
.1
.1
.8
1

.2
1
.2 passing through both link i and link r.
6 e -+ b .16 .16 .8 .8 .2 .2 .2
7 e d .04 .64 .2 .2 .2 .8
8 d
-+
-+ b .04 .2 .2 .8 .2 .2 Then, again from the "thinning property" of Poisson vari-
9 d e .04 .2 .2 .8
-+
ables, we have
Vardi: Network Tomography 373

Example 5.2. Consider the network depicted in Figure

3. We have

,Bg,s == {(1, 5" 8, 9,13), (2,8,9,13), (3,6,8,9,13),

(1,5,8,9,11,14),(2,8,9,11,14),(3,6,8,9,11,14)},

B4,s == 1, B~,6 == 1,

B~,s == B~ == {(1, 5,8,9,13), (1,5,8,9,11,14), (1,5,8,12),

(2,8,9,13),(2,8,9,11,14),(2,8,12),(3,6,8,9,13),
(3,6,8,9, 11,t4), (3,6,8,12)},

and
Figure 3. A Oiagram ot a Network Corresponding to SO Address j
(= 01, le) With 14 Links Labeled 1, ... , 14.
Bi,lO == {(3, 6,10,13), (3,6, 10, 11, 14)}.
~~, rv Poisson(AjPh'), 1 ~ i, i' ~ T, j == 1, ... , c,
Note: Because the sum of the probabilities over all(par-
(36) tial) paths originating at any given node and terminating at
where
the destination is 1,we can modify the definition of i , toBl
Bl
disregard the tails of all paths in i , beyond the point at
Ph' == probability that a packet with SD address j which they pass both link i and link i', (This is so because
passes through both link i and link i', the absence of loops implies that all such paths either pass
firstthrough i and then through i', or the other way around
Now we can decompose ~j and ~~ as follows: ~j == ~~, for all such paths. Otherwise, there would be a loop be-
+ U1.,
'l:¿
and y~ == y~, + U!,., where U1., is the number of
'l, 'l,'l, 'l, 'l, 'l,'l, tween i and i', contradicting our assumption. This permits
packets with address j passing through link i but not i', Bl
modifying the definition of i , to be the set of all possi-
and Uti is the number of packets with address j passing ble paths for address j starting at j 1 and passing through
through link i' but not i. Because of the thinning property, both link i and i' and terminating at either link i or link i'
U/i" Uti' and ~~, are statistically independent, and thus we (whichever is furthest from j 1).) An analogous statement is
get of course true for Bl. With this modified definition, we get
in Example 5.2

Bg,s == {(1, 5, 8, 9), (2; 8, 9), (3,6,8,9)},

== L COV(~~, + U~" ~~, + Uti)
j
B~,6 == 1,
o L COV(~{/ ~{, )
j

B~ ,s == B~ == {(1, 5,8), (2,8), (3,6,8)},

(37)
and
(Here (1) follows from the independence of ~j,s across i.
(2) follows from the independence of ~~"U~" and as ci; Bi ,10 == ,{3, 6, ID}.
explained earlier; and (3) follows from (36).) ,
As in the derivation of (34), the chain rule for probabili- This simplifies the calculation of (34) and (38),considerably,
ties and the Markovian nature of the routing lead to as it reduces the number of summand and the length of the
products. Putting (37) and (38) together, we get
1 ~ i, i' ~ T,

j == 1, ... , c, (38)
where
set of all complete paths for address i. which is of the same general form as (35) combined
which pass through both link i and link i', with (34).
374 Journal 01 the American Statistical Association,. March 1996

@ (M + 1) X (M + 1) transition probability matrix between

®
i~~
2 ¡.s
~
1.0 )@
Figure 4. The Markov Network ot SO Address (a, d) ot Example 5.1.
5.1.3 A Simple Method for Constructing P of (34).
For the following discussion, we fix the address j, and with-
out loss of genera1ity assume that the "active" links (Le.,
links with positive probabilities) are labeled 1, ... , M. Con-
sider a Markov chain with M + 1· states, where the states
are the M links plus an additional dummy link that is an
absorbing state labeled M + 1 (or "end"), corresponding to
an imaginary state that follows the last link of the path. (Al1
paths are eventually absorbed at this state.) Let Q(j) be the
these links. To reduce the notational clutter, we suppress
the superscript j and use Q for Q(j). Q. is derived directly
from the matrix A as follows: For two links, i == (i l , i 2)
and i' == (i~, i~), if a direct transition from i to i' is
possible, so that i 2 == i~, then we set qi,i' == a{. If
i 2 =1= i~ (Le., i is not directly connected to i'), then
qi,i' == o. If i 2 == j2 (Le., == the destination node), then
the next state is necessarily M + 1, so that qi,M+I ==
1 if i 2 == j2. Also qM+I,M+I == 1. This is somewhat
hard to see formally but easier to see via an example.
Example 5.1 (continued). Consider the address j
== (ad) (== "3") corresponding to the third column of A
in Example 5.1. The corresponding network is depicted in
Figure 4..There are five active links, and the matrix Q is
the following 6 x 6 transition probability matrix:

1 2 3 4 5 6
(a ~ b) (a ~ e) (b~ d) (e ~ b) (e ~ d) (end)
1 (a ~ b) O O 1 O O O
Q _ 2 (a -+ e) O O O .2 .8 O
(40)
- 3 (b ~ d) O O O O o. 1
4 (e ~ b) O O 1 O O O
5 (e ~ d) O O O O O 1
6 (end) O O O O O 1

We shall also need the two-step transition probability matrix chain ever visits state i is then
O O O O O ·1
O O .2 O O .8 i== 1, ... ,M, (QO == 1).
O O O O O 1
Q2== (41)
O O O O O 1 This is seen as follows: Because there are no loops in the
O O O O O 1 chain, the event [chain visits state i] is a union of mutually
O O O O O 1 exclusive events, u~o [chain visits state i at step k], and
and the three-step transition probability matrix so we get(cf. (34»

O O O O O 1 P[chain visits state i]

O O O O O 1 M

Q3== O O O O O 1
(42)
L P[chain visits state i at step k]
O O O O O 1 k=O
O O O O O 1
O O O O O 1 (44)

Note that because there are no loops (by assumption), such

a chain is always absorbed at state M + 1 after at most
M steps, and so QM is a matrix with columns of all zeros
except the last column, which is all 1s.
The initial probabilities of the chain, say 1r, are also read
off the matrix A. These probabilities are positive only for
links connected to the source (JI) of address j, and in this
case 1ri == a~. In Example 5.1, for j == (a,d) == "3", we get
Bringing back into (44) the suppressed dependency of Q
and 1r on the SD address j, and noting that the event "chain
visits state i" is equivalent to a "packet with SD address j
passing through link i," we get
(45)

== (.2, .8, O, O, O, O) (1re nd == O). (43) Example 5.1 (continued). For the address j == (ad), we
get
The probability that the chain is in "state i," i == 1, ... , M,
after k steps is then {-¡r'Qk)i' and the probability that the
Vardi: Network Tomography
This is obtained as the first five (= M) coordinates of
-¡r(j)' [Qo + QI + Q2]:
(.2, .8,0,0,0,0)[1 + QI + Q2]
= (.2, .8,0,0,0, O) + (0,0, .2, .16, .64, O)
+ (0,0, .16,0,0, .84) = (.2'-.8, .36, .16, .64, .84).
Compare the vector (.2, .8, .36, .16, .64) with the third col-
umn of the matrix P in Example 5.1 and note that they are
the same, except that we deleted the inactive links of the
network in the current calculation, because for these links
pid = O.
5.1.4 A Simple Method for Constructing P of (38).
The absence of loops induces a partial order on the net-
work with respect to each address j. For any pair of links,
i and i', they may be unconnected, in which case they are
unordered, or they may be connected by a possible path,
in which case they are' ordered relative to their distance
from JI. We say that i <j i' if i and i' are connected and i
precedes i' relative to the address j. If i and i' are uncon-
nected, then P{i' = O. If they are connected, and supposing
(without loss of generality) that i <ji', then a conditional
probability argument gives the following factorization:
(46)
where i = (i l , i 2), i' = (i~, i~), and j = (jl,j2). Here is pi
as defined before (33), and Pi~i2,j2) is the probability that a
packet with address (i 2 , j2) (starting at node iz and ending at
node j2) passes through link i', under the transition proba-
bilities of thenetwork with the original address j = (jI, j2).
Both pi and Pi~i2,j2) can be calculated using the method
given in (45).
5.1.5 Putting it Together. Expressions (45) and (35)
together give the mean and variance of the .Yi 's, and ex-
pressions (46) and (37) together give the covariances of the
Yi's. For estimation, we compare these quantities to the
empirical moments, based on repeated measurements at the
links (exactly as in Seco 3),and treat the resulting system of
linear equations as a LININPOS problem. Much of the dis-
cussion in Sections 1-4, which relates to implementation
of the method in connection with bad equations, regular-
ization toward the Poisson model via reduced weights to
second-moment equations, and so on, is aH relevant to the
random-routing case.
6. CONCLUSIONS ANO FINAL REMARKS
6.1 On the Problem (Formulation and Methodology)
The problem of estimating the node-to-node traffic in-
tensity from repeated measurements of traffic on the links
of a network is formulated and discussed under Poisson
assumptions. The MLE and related approximations are dis-
cussed, and numerical difficulties are pointed out. A detailed
methodology is presented for the method of moments. For
the Poisson model, estimates based on the method of mo-
ments are consistent and asymptotically normal. The esti-
375
mates are derived algorithmically, taking advantage of the
fact that the first- and second-moment equations give rise to
a linear inverse problem with positivity restrictions, which
can be approached with an EM algorithm. This is done
for two different types of network routing schemes: fixed
(deterministic) and random (Markovian). A small simula-
tion study is presented for a network with fixed routing,
and the results indicate tendency toward unbiasedness al-
ready for moderate sample sizes.
6.2 On the Poisson Assumption
Without any probabilitymodel, one is left only with the
linear equations (1), in which the X(k) are from unspecified
distribution(s). This cannot lead to anything useful. Thus a
model is needed to tie all these equations together and relate
the data to sorne underlying SD traffic pattern. A Poisson
model is the most natural starting point for various reasons
including the popularity of Poisson-related models for time-
dependent processes such as conventional telephone and
packet traffic (Heffes and Lucantoni, 1986), as well as ve-
hicular traffic; the simplicity of transition from inference for
Poisson variables (as in this article) to inference for Poisson
processes; and the linearity and positivity in A of the first
two moments, which facilitate a simple estimation proce-
dure. The method of weighted moment equations of Section
4 addresses estimation under deviations from the Poisson
assumption and can also be viewed as a regularization tech-
nique for the Poisson modelestimation. Clearly, however,
there are situations of network data where Poisson-related
models are totally inappropriate (see, for example, Leland et
al. 1993). Such cases are governed by different probability
models, and, accordingly, different estimation procedures,
tailored to the specific models, would need to be derived.
6.3 On Applications to Road Planning
Knowing the node-to-node traffic intensity is a great de-
sign tool for traffic engineers and urban planners, as it is
essential for deciding which nodes need to be linked di-
rectly to reduce network congestiono Historically, road side
interviews and other types of surveys have been used for es-
timáting the SD traffic intensities in transportation studies.
In recent years more algebraic and statistical methods based
on link data have been proposed. See Sherali, Sivanandan,
and Hobeika (1994), Bell (1991), and the references therein,
for important recent work and current practices. A forth-
coming paper will discuss our methodology in the context
of these applications, and its relation to the methods pro-
posed in this field.
6.4 On an Early Interest in Networks
Finally, Wf; would like to include earlier references (sug-
gested by W. Whitt and M. Segal) to a related problem.
Kruithof (1937) posed the problem of projecting from mea-
sured node-to-node teletraffic data to sorne future values,
based on estimates of total originating and terminating traf-
fic only. Krupp (1979) presented Kruithof's method, treated
it with modern mathematical tools, and related it to more
recent mathematical concepts, such as Kullback-Leibler in-
376
formation divergence. Kruithof's method can be consid-
ered a dual of the EM algorithm for LININPOS problems,
as it also minimizes a Kullback-Leibler criterion function
for such problems, but it does the minimization with re-
spect to the first argument of the Kullback-Leibler func-
tion, .whereas the EM does it with respect to the second
argument (see also Byrne 1993 for a related analysis.)
APPENDIX: PROOF OF LEMMA OF SECTION 2
Let
r
Sj = L A~j) = number of l's in column A (j). (A.1)
i=l
Without loss of generality, assume that the columns of A are ar-
ranged monotonically in Sj, 1 :::; SI :::; ... :::; s-: To prove identifi-
ability, we need to show that if
Pv(yIA) = Pv(yIX) for a set of y's of measure 1(A), (A.2)
then necessarily aA = X. The proof has three steps. In Step 1 we
show that (A.2) implies E Ai = E .xi, in Step 2 we show that
it implies Al = .xl, and in Step 3 we use induction to complete
the proof. In all threesteps we calculate the probabilities of the
possible solution vectors, x, to the equations y = Ax under the
Poisson model,
Step 1: Take y = O (vector of all zeros). Then, because the
components of x mustbe natural numbers (because the x/s are
Poisson variables) and because A has no zero columns, the only
feasible solution of y = Ax is x = o. Combining this with (A.2),
we have e-}JAi = Pv(OIA) = Pv(OIX) = e-}J'xi, so that
(A.3)
Step 2: Now take y = A (1) and consider again the solution,
in x, to y = Ax. Clearly, x = lil == (1, O, ... ,O)' is a solution.
"Io see that this is the unique solution, suppose (by negation) that
there is another solution. Because A (1) is a vector of zeros and
1s, so must be any feasible solution x. If the first coordinate of
x is 1, then the remaining must all be zeros, because we have
A (1) = E~=l xiA (i), so that SI = E~=l xis, = SI + E~=2 xis«,
and so X2 = ... = Xc = O. If the first coordinate of x is zero,
then we get SI = E~=2 xis«, and because SI :::; Si, i = 2, ... ,c,
we can have at most one xi(2 :::; i :::; c) equal to 1, but because x
is a solution to A (1) = Ax, it would imply A (1) = A (i) for sorne
i E {2, ... ,c}, contradicting the assumption that all the columns
of A are different. Thus we have
Py(A(1) lA) = Px(8 l IA) = e-}JAiAl
and, similarly, Pv(A (1) IX) = e-}J'xi.xl. It then follows from (A.2)
and (A.3) that
Al = .xi. (A.4)
Step 3 (induction step): Suppose that
Ai=.xi, i=1, ... ,j-1, forany jE{2, ... ,c}. (A.5)
We want to show that Aj = .xj. Take y = A (j) and consider again
the equation y = Ax. Clearly, x = lij == (O, ... , O, 1, O, ... ,O)'
(1 at position i. and zero elsewhere) is a solution. Also, because
A (j) is a vector of zeros and ls, and because the x/s are Poisson
variables, any feasible solution must be a vector of zeros and 1s.
For similar reasons as in Step 2, any solution must have x j + 1
Journal 01 the American Statistical Association, March 1996
= Xc = O. Furthermore, if a solution has Xj = 1, then
it must also have Xl = ... = Xj-l = O, because for such a
solution, Sj = Sj + E~:: xis«, and hence Xl = ... = Xj-l = O
too. In other words, the only feasible solution of A (j) = Ax with
Xj = 1 is x = lij • Therefore, if A (j) = Ax has any solution
in addition to x = lij , then it must be a vector of zeros and ls
and have a tail of zeros starting at position j or earlier, so that
Xj = Xj+l = ... = Xc = o. Let xi , ... ,Xk be all such distinct
feasible solutions to A (j) = Ax, and let b, ~ {I, ... ,j - 1} index
the coordinates with value 1 of x., i = 1, ... ,k. (For example, if
x¡ = (1, 0,1, O... O) and X2 = (1,1, O... O), then b: = {1, 3} and
ba = {1, 2}.) We then have .
Pv(A(j) lA) Px(lij or x¡ or X2 ... or
xkl A) = e-}JAi(Aj + rr Ai + rr Ai + ... + rr Ai) .
iEbl iEb2 iEbk ~
(Note that lij and xi , ... ,Xk are all distinct.) Similarly,
Pv(A(j) IX) = e-}J'xi(Aj + rr + rrAi
iEb2
Ai + ... + rr
iEbk
Ai) .
iEbl
It follows from (A.2), (A.3), and the induction assumption (A.5)
that
. (A.6)
where the last equality follows from the induction assumption
because
bl ~ {1, ... ,j - 1}, l = 1, ... , k.
This completes the proof of the induction step and of the lernma.
[Received January 1994. Revised May 1995.}
REFERENCES
Becker, R. A., and Wilks, A. R. (1993), "Network Traffic Analysis,"
AT&T Bell Labs technical report.
Bell, M. G. (1991), "The Estimation of Origin-Destination Matrices by
Constrained Generalised Least Squares," Transportation Research B,
25B, pp. 13-22.
Byrne, C. L. (1993), "Iterative Image Reconstruction AIgorithms Based
on Cross-Entropy Minimization," IEEE Transactions on Imaging Pro-
cesses, 2, 96-103.
Dempster, A. P., Laird, N. M., and Rubin,D. B. (1977)~ "Maximum Like-
lihood from Incomplete Data via the EM AIgorithm" (with discussion),
Journal of the Royal Statistical Society, Ser. B, 39, 1-38.
Green, P. J. (1990a), "Bayesian Reconstructions From Emission Tomog-
raphy Data Using a ModifiedEM AIgorithm," IEEE Transactions in
MedicalImaging, 9,84-93.
- - (1990b), "On Use of the EM AIgorithm for Penalized Likelihood
Estimation," Journal of the Royal Statistical Society, Ser. B, 52, 443-
452.
He1fes, H., and Lucantoni, D. M. (1986), "A Markov Modulated Charac-
terization of Packetized Voice and Data Traffic and Related Statistical
Multiplexer Performance," IEEE Journal on Selected Areas in Commu-
nications, 4, 856-868.
Iusem, A. N., and Svaiter, B. F. (1994), "A New Smoothing-Regularization
Approach for a Maximum-Likelihood Estimation Problem," Applied
Mathematics & Optimization: An International Journal, 29, 225-241.
Kruithof, J. (1937), "Telefoonverkeersrekening," De Ingenieur, 52, no. 8,
E15-E25.
Krupp, R. S. (1979), "Properties of Kruithof's Projection Method," The
Bell Systems Technical Journal, 58, 517-538~
Leland, W. E., Taqqu, M. S., Willinger, W., and Wilson, D. V. (1993), "On
the Self-Similar Nature of Ethernet Traffic," presented at SIGCOMM
'93.
Vardi: Network Tomography
Roeder, K., Devlin, B., and Lindsay, B. G. (1989), "Applieation oí"Maxi-
mum Likelihood Methods to Population Genetie Data for the Estimation
of Individual Fertilities," Biometrics, 45, 363-379.
Sherali, H. D., Sivanandan, R., and Hobeika, A. G. (1994), "A Linear Pro-
gramming Approaeh for Synthesizing Origin-Destination Trip Tables
from Link Traffie Volumes," Transportation Research B, 28B, pp. 213-
233.
Shepp, L. A., andVardi, Y. (1982), "Máximum Likelihood Reeonstruetion
for Emission Tomography," IEEE Transactions on Medical Imaging, 1,
113-122.
Silverman, B. W., Jones,M. C., Wilson, J. D., and Nyehka, D. W. (1990),
"A Smoothed EM Approaeh to Indireet Estimation Problems, With Par-
377
tieular Referenee to Stereology and Emission Tomography" (with dis-
eussion), Journal of the Royal Statistical Society, Ser. B, 52, 271-324.
Snyder, D. L., Sehultz, T. 1., and O'Sullivan, 1. A. (1992), "Deblurring Sub-
jeet to Nonnegativity Constraints," IEEE Transactions on Signal Pro-
cesses, 40, 1143-1150.
Vardi, Y., and Lee, D. (1993), "From Image Deblurring to Optimal In-
vestments: Maximum Likelihood Solutions for Positive Linear Inverse
Problems" (with diseussion), Journal of the Royal Statistical Society,
Ser. B, 55, 569-612.
Vardi, Y., Shepp, L. A., and Kaufman, L. (1985), "A Statistieal Model
for Positive Emission Tomography," (with eomments), Journal of the
American Statistical Association, 80, 8-37.