Many geophysical problems involve inverting data in order to obtain me
aningful descriptions of the Earth's interior. One of the basic charac
teristics of these inverse problems is their non-uniqueness. Since com
putation power has increased enormously in the last few years, it has
become possible to deal with this non-uniqueness by generating and sel
ecting a number of models that all fit the data up to a certain tolera
nce. In this way a solution space with acceptable models is created. T
he remaining task is then to infer the common robust properties of all
the models in the ensemble. In this paper these properties are determ
ined using empirical orthogonal function (EOF) analysis. This analysis
provides a method to search for subspaces in the solution space (ense
mble) that correspond to the patterns of minimum variability. In order
to show the effectiveness of this method, two synthetic tests are pre
sented. To verify the applicability of the analysis in geophysical inv
erse problems, the method is applied to an ensemble generated by a Mon
te Carlo search technique which inverts group-velocity dispersion data
produced by using vertical-component, long-period synthetic seismogra
ms of the fundamental Rayleigh mode. The result shows that EOF analysi
s successfully determines the well-constrained parts of the models and
in effect reduces the variability present in the original ensemble wh
ile still recovering the earth model used to generate the synthetic se
ismograms. Finally, an application of the method to examine the contra
st in upper-mantle S-wave velocity across the Tornquist-Tesseyre Zone
is presented, indicating a significant change in S-wave velocity in th
e upper mantle beneath this zone bordering the East European Platform
and Tectonic Europe, and a significantly thicker crust beneath the Eas
t European Platform.