Least squares polynomial response surface¶
Instead of replacing the model response with a
local approximation around a given set
of
input parameters as in Taylor expansion, one may seek a global approximation of
over its whole domain of definition. A common
choice to this end is global polynomial approximation. For the sake of
simplicity, a scalar model response
will
be considered from now on. Nonetheless, the following derivations hold
for a vector-valued response.
In the following, one considers global approximations of the model response using:
a linear function, i.e. a polynomial of degree one;
where
is a set of unknown coefficients.
a quadratic function, i.e. a polynomial of degree two.
The two previous equations may be recast as:
where denotes the number of terms, which is equal to
(resp. to
) when
using a linear (resp. a quadratic) approximation, and the family
gathers the constant monomial
, the monomials of degree one
and possibly the
cross-terms
as well as the monomials of degree two
. Using the vector notation
and
,
this can be rewritten:
A global approximation of the model response over its whole
definition domain is sought. To this end, the coefficients
may be computed using a least squares regression approach. In this
context, an experimental design, that is, a set of observations of
input parameters, is required:
as well as the corresponding model evaluations:
The least squares problem is to solve:
where is the cost function, defined as:
Let be the
vector of output observations.
If the design matrix
has full rank,
then the solution is given by the normal equations:
(1)¶
where:
for and
.
A necessary condition for having a solution is that the size
of the experimental design is not less than the number
of
coefficients to estimate.
The Gram matrix
can be
ill-conditionned.
Hence, the best method is not necessarily to invert the Gram matrix,
because the solution may be particularly sensitive to rounding errors.
The least-squares problem is rather solved using more robust numerical methods
such as the singular value decomposition (SVD) or the QR-decomposition.