Interval analysis is a very specific domain: as
a naive user you may want to solve a specific problem but the problem
that you have submitted to `ALIAS` may be already a transformed
version of the problem you want really to solve and this transformation may
be not favorable to the use of interval analysis. Focusing on the
problem you want to solve is important for the use of interval
analysis.

**Rule**: *focus on the problem you want to solve*

Let us give you an example: you have a third order polynomial depending on a set of parameters and you want to find if there are values of these parameters such that the root of the polynomial verifies some properties. A natural way of doing that is to use the available generic analytical form of the root and check if you may find parameters such that satisfy the properties. In terms of interval analysis you have already transformed your problem and this may not be a good idea (the analytical form of the roots of a 3rd order polynomial is badly conditioned).

**Rule**: *it is usually a bad idea to try to adapt to interval
analysis the last step of a solving problem that has already been
transformed to fit another method*

Thinking backward what you want really is to verify if the satisfying satisfy the properties. Hence you should use as unknown not only the but also the : you add an unknown but the calculation will then involve only the evaluation of the polynomial which is much more simple than the evaluation of the analytical form of the roots.

**Rule:** *introducing new variables for simplifying expressions
may be a good strategy*

The rule for the right formulation is to find the right compromise between the complexity of the expressions (very complex expressions will usually lead to a large overestimation of the expressions) and the number of variables.

A good strategy is also to learn what are the main problems that have been addressed in interval analysis so that you can later on transform your problem into another one for which solving algorithms are available.

**Rule:** *investigate which main problems have been addressed in
interval analysis so that you may later on transform your problem into
another one for which solving algorithms are available*

Typical of such approach is to determine if a parametric matrix may include singular matrices. You may be tempted to derive the determinant of the interval matrix and then to determine if there are values of the unknowns that cancel it. But if you have not been able to manually determine these values this probably means that the determinant is quite complicated, and therefore not appropriate for interval analysis. But determining if a parametric matrix may include a singular matrix is a well-known problem in interval analysis and for solving this problem there are efficient methods that do not require the calculation of the determinant (see section 7.4).

Another case of failure occurs when the system is badly
conditioned. This may occur, for example, when you have large
coefficients in an equation while the values of the variable are very
small: as `ALIAS` takes into account the round-off errors you will
get large interval evaluation even for a small width of the intervals
in the input interval (see for example the evaluation of the Wilkinson
polynomial in section 5.9.3.2). There is no standard way to solve the
problem. The first thing you may try is to normalize the unknowns in
order to reduce the size of the coefficients. A second approach may be
to switch a solving problem to an optimization problem.

Another trick for system solving is to combine the initial equations to get new one. For example consider the system:

cons1 := a1 + a2 + a3 + 1 ; cons2 := a1 + 2*a2*x1 + 3*a3*x1^2 + 4*x1^3 ; cons3 := a1 + 2*a2*x2 + 3*a3*x2^2 + 4*x2^3 ; cons4 := a1 + 2*a2*x3 + 3*a3*x3^2 + 4*x3^3 ; cons5 := a1*(x1+x2) + a2*(x1^2 + x2^2) + a3*(x1^3 + x2^3) + x1^4 + x2^4 ; cons6 := a1*(x2+x3) + a2*(x2^2 + x3^2) + a3*(x2^3 + x3^3) + x2^4 + x3^4 ; cons7 := x1 + delta -x2<= 0; cons8 := x2 + delta -x3<= 0;(which is the

**Rule:***Very often introducing additional constraints derived from the
set of initial ones will reduce the solving time*

Interval arithmetics plays evidently a large role in the computation time. The interval evaluation of mathematical operator such as sine, cosine,...is relatively expensive. If similar complex expressions appear more than once in an expression it is usually very efficient to compute it only once and to substitute it in the other places.

**Rule:***try to avoid multiple interval evaluation of the same
complex expressions*