UQ[PY]LAB user manual
A vine is said to be truncated at level t
∗
∈ {1, . . . , M − 1}, if all pair copulas belonging to
any tree t ≥ t
∗
are the independence pair copula. The rationale for working with truncated
vines is that, often, only conditional probabilities of low order can be reasonably asserted or
reliably inferred from data, whereas imposing complex models for conditional probabilities
of higher order may correspond to over-interpreting or over-fitting data. Furthermore, con-
ditional probabilities of higher order typically have a weaker effect on the joint PDF of the
input random variables than those of lower order, especially if the strongest correlations are
captured in earlier trees, and can thus be neglected.
1.4.5 Copulas inferred from data
In addition to user-defined parametric copulas, UQ[PY]LAB supports the possibility to in-
fer copulas from data. The theory underlying this topic and its usage in UQ[PY]LAB are
extensively covered in the companion UQ[PY]LAB User Manual – Statistical inference.
1.5 Sampling random vectors
Various uncertainty quantification tasks require to generate data according to a specified
joint distribution. For instance, resampling from a known distribution of the inputs to a
system enables statistical estimation of the output by Monte-Carlo or by more sophisticated
strategies.
A number of random sampling strategies (e.g. Monte Carlo sampling, latin hypercube sam-
pling (LHS), pseudorandom sequences, etc.) are available to produce samples in the stan-
dard uniform space, that is, for Z ∼ U([0, 1]
M
). A possible strategy to sample from a different
distribution F
X
is thus to transform a sample z of Z into a sample x of X, if such a trans-
formation exists. Maps of the form (Lebrun and Dutfoy, 2009)
X = T (U ) s.t. X ∼ F
X
, U ∼ F
U
, (1.19)
which transform a random vector U ∼ F
U
(or a sample u thereof) into a random vector
X ∼ F
X
(or a sample x therefore) are called isoprobabilistic transforms. Amongst the many
strategies available to generate samples distributed according to a more general distribution
F
X
, UQ[PY]LAB makes use of isoprobabilistic transforms. Of particular interest for gener-
ating samples of X ∼ F
X
are isoprobabilistic transforms from U ∼ U([0, 1]
M
) to X . In the
following, isoprobabilistic transforms to different target distributions F
X
will be derived. We
will see that different transforms exist depending on the copula of C
X
of F
X
.
1.5.1 Independence copula: probability integral transform
A well known result of probability theory states that, for any random variable X with contin-
uous CDF F
X
, the random variable
U = F
X
(X) (1.20)
UQ[PY]LAB-V1.0-102 - 10 -