# Linear solvers

Once the integration scheme described how the linear matrix system is built, this system must be solved in order to find the solution at the next time step.

To solve this system, two main categories of algorithms exist: the **direct** solvers and the **iterative** solvers.

## Direct solvers

These solvers aim at finding the exact solution of the system by computing in one single step . To do so, various methods exist to compute the inverse matrix of .

For small-size linear systems, the direct methods will be efficient. Large and sparse systems may imply time-consuming inverse of the matrix . The advantage of direct methods is that they succeed to solve well-conditioned and even some quite ill-conditioned problems. The computation of the inverse of often relies on decomposition of this matrix: Cholesky, LU or LDL and their sparse versions are available.

#### Direct solver implementation

Direct solvers in SOFA are:

- SparseLDLSolver
- LULinearSolver / SparseLUSolver
- CholeskySolver / SparseCholeskySolver
- SVDLinearSolver (Jacobi SVD)
- BTDLinearSolver

#### In the SOFA code

The resolution of the linear system is computed in the `solve()`

function of the LinearSolver. With direct solvers, the integration scheme sucessively calls the two following functions:

implementing the targeted decomposition method:

## Iterative solvers

Contrary to direct solvers, iterative methods converge towards the solution gradually. The solution is approximated at each iteration a little bit more accurately, rather than computed in one single large iteration. With iterative methods, the error esimated in the solution decreases with the number of iterations.

For well-conditioned problems (even large systems), the convergence remains monotonic. However, for ill-conditioned systems, the convergence might be much slower. Since these methods compute the residual at each iteration, the matrix does not have to be built to improve performances (only matrix vector computations). Numerical settings of the solver (maximum number of iterations, tolerance for instance) must be appropriately defined. Two available methods are the conjugate gradient method (using the CGLinearSolver) or the minimal residual method (using the MinResLinearSolver).

#### Iterative solver implementation

Iterative solvers in SOFA are:

- CGLinearSolver
- MinResLinearSolver

#### In the SOFA code

The resolution of the linear system is computed in the `solve()`

function of the LinearSolver. With iterative solvers, the integration scheme only calls the function:

and will handle these vectors as TempVectorContainer and create any new vector using the function *vtmp.createTempVector()* as follows:

Last modified: 5 May 2020