[
https://issues.apache.org/jira/browse/MATH-735?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Sébastien Brisard updated MATH-735:
-----------------------------------
Description:
Before we release version 3.0, I would like to perform some changes to the
general framework for linear iterative solvers.
# Remove interface {{InvertibleRealLinearOperator}}: this was only proposed for
handling preconditioners. Let {{M}} be this preconditioner. In the course of
the iterations, only matrix-vector products of the form {{M^(-1) . y}} are
needed, never direct products {{M . x}}. So it's just as simple to provide
{{M^(-1)}} as a standard {{RealLinearOperator}}, rather than {{M}} as an
{{InvertibleRealLinearOperator}}. This removes a little bit of clutter in the
class hierarchy.
# {{o.a.c.m.util.IterationEvent}} should at the very least have a method to get
the current iteration number. Overwise, I find myself keeping track of the
number of times
{{o.a.c.m.util.IterationManager.fireIterationPerformedEvent(IterationEvent)}}
is called, which is both ugly and painfull.
# {{o.a.c.m.util.linear.IterativeLinearSolverEvent}} should have a method to
access the norm of the residual. I was reluctant to specify such a method in
the interface at the start, because I was not sure any iterative solver
provides an updated estimate of this quantity at each iteration. In fact, I've
realized that most (if not all) solvers do, and it's quite nice to be able to
access this value while monitoring the progress of the inversion (through
{{o.a.c.m.utils.IterationListener}}). Otherwise, one has to compute {{b - A .
x}} inside the listener, which *doubles* the cost of each iteration!
# {{o.a.c.m.util.linear.IterativeLinearSolverEvent}} should have a method to
access the residual (the vector itself, not only its norm). I'm not sure *all*
solvers can do that, so this should be an optional feature, which might be
useful. I see two different possible implementations
## specify {{RealVector
o.a.c.m.util.linear.IterativeLinearSolverEvent.getResidual()}} as an optional
feature (potentially throwing {{UnsupportedOperationException}}), together with
a method {{boolean
o.a.c.m.util.linear.IterativeLinearSolverEvent.providesResidual()}}. I think
some of us do not like {{UnsupportedOperationException}}, but this would be
similar to what was done in {{RealLinearOperator.operateTranspose(RealVector)}}.
## create a new interface {{RealVector
o.a.c.m.util.linear.IterativeLinearSolverWithResidualEvent}}. This is similar
to what is done currently, see {{o.a.c.m.linear.ProvidesResidual}}.
I would love to have some feedback on the last point.
was:
Before we release version 3.0, I would like to perform some changes to the
general framework for linear iterative solvers.
# Remove interface {{InvertibleRealLinearOperator}}: this was only proposed for
handling preconditioners. Let {{M}} be this preconditioner. In the course of
the iterations, only matrix-vector products of the form {{M^(-1) . y}} are
needed, never direct products {{M . x}}. So it's just as simple to provide
{{M^(-1)}} as a standard {{RealLinearOperator}}, rather than {{M}} as an
{{InvertibleRealLinearOperator}}. This removes a little bit of clutter in the
class hierarchy.
# {{o.a.c.m.util.IterationEvent}} should at the very least have a method to get
the current iteration number. Overwise, I find myself keeping track of the
number of times
{{o.a.c.m.util.IterationManager.fireIterationPerformedEvent(IterationEvent)}}
is called, which is both ugly and painfull.
# {{o.a.c.m.util.linear.IterativeLinearSolverEvent}} should have a method to
access the norm of the residual. I was reluctant to specify such a method in
the interface at the start, because I was not sure any iterative solver
provides an updated estimate of this quantity at each iteration. In fact, I've
realized that most (if not all) solvers do, and it's quite nice to be able to
access this value while monitoring the progress of the inversion (through
{{o.a.c.m.utils.IterationListener}}). Otherwise, one has to compute {{b - A .
x}} inside the listener, which *doubles* the cost of each iteration!
# {{o.a.c.m.util.linear.IterativeLinearSolverEvent}} should have a method to
access the residual (the vector itself, not only its norm). I'm not sure *all*
solvers can do that, so this should be an optional feature, which might be
useful. I see two different possible implementations
## specify {{RealVector
o.a.c.m.util.linear.IterativeLinearSolverEvent.getResidual()}} as an optional
feature (potentially throwing {{UnsupportedOperationException}}), together with
a method {{boolean
o.a.c.m.util.linear.IterativeLinearSolverEvent.providesResidual()}}. I think
some of us do not like {{UnsupportedOperationException}}, but this would be
similar to what was done in {{RealLinearOperator.operateTranspose(RealVector)}}.
## create a new interface {{RealVector
o.a.c.m.util.linear.IterativeLinearSolverWithResidualEvent}}.
I would love to have some feedback on the last point.
> Improvements to linear iterative solvers
> ----------------------------------------
>
> Key: MATH-735
> URL: https://issues.apache.org/jira/browse/MATH-735
> Project: Commons Math
> Issue Type: Improvement
> Affects Versions: 3.0
> Reporter: Sébastien Brisard
> Assignee: Sébastien Brisard
> Labels: event, linear, solver
>
> Before we release version 3.0, I would like to perform some changes to the
> general framework for linear iterative solvers.
> # Remove interface {{InvertibleRealLinearOperator}}: this was only proposed
> for handling preconditioners. Let {{M}} be this preconditioner. In the course
> of the iterations, only matrix-vector products of the form {{M^(-1) . y}} are
> needed, never direct products {{M . x}}. So it's just as simple to provide
> {{M^(-1)}} as a standard {{RealLinearOperator}}, rather than {{M}} as an
> {{InvertibleRealLinearOperator}}. This removes a little bit of clutter in the
> class hierarchy.
> # {{o.a.c.m.util.IterationEvent}} should at the very least have a method to
> get the current iteration number. Overwise, I find myself keeping track of
> the number of times
> {{o.a.c.m.util.IterationManager.fireIterationPerformedEvent(IterationEvent)}}
> is called, which is both ugly and painfull.
> # {{o.a.c.m.util.linear.IterativeLinearSolverEvent}} should have a method to
> access the norm of the residual. I was reluctant to specify such a method in
> the interface at the start, because I was not sure any iterative solver
> provides an updated estimate of this quantity at each iteration. In fact,
> I've realized that most (if not all) solvers do, and it's quite nice to be
> able to access this value while monitoring the progress of the inversion
> (through {{o.a.c.m.utils.IterationListener}}). Otherwise, one has to compute
> {{b - A . x}} inside the listener, which *doubles* the cost of each iteration!
> # {{o.a.c.m.util.linear.IterativeLinearSolverEvent}} should have a method to
> access the residual (the vector itself, not only its norm). I'm not sure
> *all* solvers can do that, so this should be an optional feature, which might
> be useful. I see two different possible implementations
> ## specify {{RealVector
> o.a.c.m.util.linear.IterativeLinearSolverEvent.getResidual()}} as an optional
> feature (potentially throwing {{UnsupportedOperationException}}), together
> with a method {{boolean
> o.a.c.m.util.linear.IterativeLinearSolverEvent.providesResidual()}}. I think
> some of us do not like {{UnsupportedOperationException}}, but this would be
> similar to what was done in
> {{RealLinearOperator.operateTranspose(RealVector)}}.
> ## create a new interface {{RealVector
> o.a.c.m.util.linear.IterativeLinearSolverWithResidualEvent}}. This is similar
> to what is done currently, see {{o.a.c.m.linear.ProvidesResidual}}.
> I would love to have some feedback on the last point.
--
This message is automatically generated by JIRA.
If you think it was sent incorrectly, please contact your JIRA administrators:
https://issues.apache.org/jira/secure/ContactAdministrators!default.jspa
For more information on JIRA, see: http://www.atlassian.com/software/jira
