2 INFORMATION-GEOMETRIC NETWORK
INFERENCE
The classical problem of network inference (tomography) is
to determine all the end-to-end flow rates based on the link
rate measurements [1]. For a network with m nodes, the
number of possible flows is mðm 1Þ, while the number of
measured links is usually less than mðm 1Þ, unless the
network is fully connected and all links can be measured.
Therefore, this is a statistical inference problem for underdetermined
systems. Let X denote the vector of end-to-end
information flow rates, where its j-th element xj is the rate
of the j-th source-destination pair. Let Y denote the vector
of link-level rate measurements, where its i-th element yi is
the traffic rate on link i. We can regard both X and Y as random
variables. The randomness in X may be due to the stochastic
packet traffic, whereas the randomness in Y may be
due to its dependency on X and the measurement noise N.
We assume that flow rates take values from a discrete set
X. The size jXjjXj determines the resolution of the network
inference problem, where jXj is the size of X and jXj is the
size of X. Link rate measurements are expressed in terms of
end-to-end flow rates as
Y ¼ AX þ N;
where A is the routing matrix (Ai;j is the fraction of the j-th
flow on the i-th link) and N is the measurement noise.
The dimension of Y is usually smaller than that of X.
Therefore, it is an underdetermined system, where solutions
cannot be uniquely determined. Hence, we pursue a
statistical inference approach, where we infer the probability
distribution pX for X based on the moving average of k
measurements Yk, a (prior) distribution qX on X, and a
expectation ~N over a (prior) distribution on N. Initially, qX
can be chosen as a uniform distribution with the maximum
entropy (i.e., maximum uncertainty). That is, pxj ¼ 1=jXj
for each xj 2 X. Similarly, we can assume small values of
~N
initially.
As we will discuss in Section 5 in detail, the attempt to
better capture and optimization of information flows may
require inference of distributions rather than average values.
With measurements Yk, the distribution pX can be