Black-box operators are powerful enough to act as a matrix with a number of standard Matlab functions. Some of this functionality is native to the operators, some have been implemented as adapters, and others work without any particular attention.
This page contains a list of functions that works in a way that is similar to standard Matlab functions. These functions are not included in the normal function list, since their syntax is given by Matlab conventions.
There are a few quirks, however, that one should be aware of. These are explained below. The main place where one needs to be careful is the matrix multiplication, which may be ambigous in certain cases.
Contents | ||
Operators | Support for syntactic elements | |
Basic Information | Information about an operator | |
Operations | Black-box replacements for Matlab operations | |
Analysis | Supported function in matrix analysis | |
Iterative Algorithms | Support for iterative algorithms shipped with Matlab | |
Miscellaneous | Other functions | |
Special considerations | ||
Equation Solving | Caveats with equation solving | |
Unsupported Operations | Functions that are unsupported or behave unexpectedly |
Sometimes it is easiest to illustrate a syntax by example.
We use
A
and B
represent matrices or black-box operators,
c
represent a scalar, and
str
is a string.
Many of the normal operators works with the regular Matlab syntax. This accounts for the fact that a black-box operator can be used as a matrix in many situations. It also makes black-boxes pleasant to work with.
Transposition | |
Complex conjugate transpose | |
Plus | |
, | Minus |
Unary plus | |
Unary minus | |
Horizontal concatenation | |
Vertical concatenation | |
Division by a scalar (see also equation solving) |
BBTools supports matrix multiplication via the normal syntax and . However, to maximize compatibility there are special rules which applies to this product:
A
or B
is a Matlab-matrix, the
product will be evaluated and returned as a normal Matlab matrix.A
or B
is a scalar, which can be
interpreted as a 1-by-1 matrix, then the product will also be
evaluated.A*B
.The convention allows a black-box operator to replace a Matlab matrix in
many circumstances. If A
is a black-box operator and x
is a normal matrix, then y=A*x
evaluates the matrix-vector product,
while bb=A*blackbox(x)
creates a black-box operator.
If you need to create a black-box, please use
bbprod
which always returns a
black-box operator.
Functions that works as in the normal way includes: | |
Size of a black-box operator | |
true for a real operator,
false if it is complex | |
Note: | |
The following functions also works as expected, although the output differ from the normal because a black-box differs from a Matlab matrix: | |
Display evaluation tree, without printing the name | |
As disp , except the name is printed |
These functions creates a black-box operator equivalent to the operation normally performed by Matlab. They generally work as expected, except the result is a black-box.
All operations in BBTools support complex operators.
Kronecker tensor product | |
Replicate and tile an array | |
Operations on complex operators. | |
Complex conjugate | |
Complex real part | |
Complex imaginary part |
Note: kron
and repmat
have counterparts
in BBTools
(bbkron
and
bbrepmat
)
that always produce a black-box operator, even if the input are
Matlab matrices.
Caveat: real
and imag
evaluate two
products with the operator when it evaluates a product (unless they get a
real operator as input). This could potentially lead to an exponential
explosion in the evaluation time, if they are nested.
These functions generally work as expected. However, they are implemented using iterative methods, which may sometimes give a slighly different behavior.
Norm of an operator (warning except for 2-norms) | |
Less accurate, but slighly faster, way to compute the 2-norm | |
Estimate the 1-norm of an operator |
Note: BBTools includes bbnormest
,
which should be used instead of normest
everywhere. It supports both
1-, 2-, and ∞-norms, usually with much better accuracy
for substantially less work than normest
.
Except for the 2-norm, norm
reverts to
bbnormest
with a warning. This is intended
to prevent putting too much trust into an estimate.
Limitation: Currently, BBTools does not include functions to estimate the Frobenius norm of an operator.
Some care have been taken to allow the iterative routines shipped with Matlab, to work with black-box operators. Matlab usually assumes that these functions are intended for sparse matrices, but they work for black-box operators as well.
Find a few eigenvalues and eigenvectors | |
Find a few singular values and vectors | |
Iterative solver (Conjugate Gradients on the Normal Equations) | |
Iterative solver (BiConjugate Gradients) | |
Iterative solver (BiConjugate Gradients Stabilized) | |
Iterative solver (Conjugate Gradients Squared) | |
Iterative solver (Generalized Minimum Residual) | |
Iterative solver (Minimum Residual) | |
Iterative solver (Preconditioned Conjugate Gradients) | |
Iterative solver (Quasi-Minimal Residual) | |
Iterative solver (Symmetric LQ) |
Caveat: In certain modes of operation, eigs
and svds
expects
that the operator can be inverted. This is not supported by BBTools.
This is a catch-all for everything that does not fit comfortably anywhere else. The following functions works as expected.
Sum of elements | |
Average or mean value |
The normal Matlab-syntax for solving a linear system is
or
.
For large-scale problems it is usually necessary to use an iterative
routine, e.g.
bbsolve
, possibly using
a preconditioner to get convergence.
Unfortunately, there is not a single, unattended algorithm that always attains this goal. Among other things, BBTools was written to make it easier to reasearch this kind of algorithms for realistic and relevant large-scale problems.
For this reason BBTools does not support the shorthand notation used
by Matlab. Instead, please call either
bbsolve
or an
iterative algorithms shipped with Matlab.
Basically, nothing is supported unless explicitly stated otherwise. However, there are a few corner cases it is wise to be aware of.
Some common operations goes against the spirit of BBTools.
is not supported directly. Instead you should create an operator via
bbindex
which performs a similar
operation.
Not all functions can be supported from matrix-vector products alone. The following operations will most likely never be supported by BBTools.
Hadamard product, direct product, or array multiply | |
Array power | |
Absolute value |
The following functions are not implemented, and works unexpectedly. They consider a black-box as an array of objects, rather than addressing the black-box they get as input. It is being considered whether it would be appropriate to change this behaviour.
Length of the array of black-boxes | |
true for an empty black-box array |