Convex Iteration

From Wikimization

Revision as of 12:21, 24 November 2011 by Ranjelin (Talk | contribs)
Jump to: navigation, search

Convex iteration is method for constraining rank or cardinality in an otherwise convex optimization problem.
A rank or cardinality constraint is replaced by a linear regularization term added to the objective.
Then two convex problems are iterated until convergence where, ideally, solution to the original problem is found.

constraining cardinality of signed variable

Excert from Chapter 4 Semidefinite Programming:
Consider a feasibility problem equivalent to the classical problem from linear algebra LaTeX: A_{}x_{\!}=_{\!}b ,  but with an upper bound LaTeX: k on cardinality LaTeX: \|x\|_0 :  for vector LaTeX: b\!\in\!\mathcal{R}(A)

LaTeX: \begin{array}{rl}\text{find}&_{}x\in_{}\mathbb{R}^n\\
\mbox{subject to}&A_{}x=b\\

where LaTeX: \|x\|_{0\!}\leq_{_{}\!}k means vector LaTeX: x has at most LaTeX: k nonzero entries; such a vector is presumed existent in the feasible set.
Convex iteration works with a nonnegative variable; absolute value LaTeX: |x| is therefore needed.
We propose that nonconvex problem (1) can be equivalently written as a convex iteration: for LaTeX: \varepsilon a relatively small positive constant,

LaTeX: \begin{array}{cl}\text{minimize}_{x_{}\in_{_{}}\mathbb{R}^{^n},~t_{}\in_{_{}}\mathbb{R}^{^n}}&\langle t_{}\,,\,y+\varepsilon^{}\mathbf{1}\rangle\\
\mbox{subject to}&A_{}x=b\\
&-t\preceq x\preceq_{_{}}t

is iterated with

LaTeX: \begin{array}{cl}\text{minimize}_{y_{}\in_{_{}}\mathbb{R}^{^n}}&\langle t^*,\,y+\varepsilon^{}\mathbf{1}\rangle\\
\mbox{subject to}&0\preceq y\preceq\mathbf{1}\\

to find direction vector LaTeX: y. The cardinality constraint has been moved to the objective as a linear regularization.
The term LaTeX: \langle t\,,_{_{}}\varepsilon^{}\mathbf{1}\rangle in (2) is necessary to determine absolute value LaTeX: |x|_{\!}=_{}t^{*_{}} because vector LaTeX: y can take zero values in its entries.

By initializing direction vector LaTeX: y to LaTeX: (1\!-_{}\!\varepsilon)\mathbf{1}_{}, the first iteration of problem (2) is a 1-norm problem; i.e.,

LaTeX: \begin{array}{ccc}
\begin{array}{cl}\text{minimize}_{x_{}\in_{_{}}\mathbb{R}^{^n},~t_{}\in_{_{}}\mathbb{R}^{^n}}&\langle t\,,_{}\mathbf{1}\rangle\\
\mbox{subject to}&A_{}x=b\\
&-t\preceq x\preceq_{_{}}t
\mbox{subject to}&A_{}x=b

Subsequent iterations of problem (2) engaging cardinality term LaTeX: \langle t_{}\,,\,y\rangle can be interpreted as corrections to this 1-norm problem leading to a LaTeX: 0-norm solution.
Iteration (2) (3) always converges to a locally optimal solution by virtue of a monotonically nonincreasing objective sequence.
There can be no proof of global optimality, defined by an optimal objective equal to 0.
Local solutions are therefore detected by nonzero optimal objective.
Heuristics for reinitializing direction vector LaTeX: y can lead to a globally optimal solution.


Several simple equivalents to linear programs (2) (3) are easily devised, but their geometrical interpretation is not as apparent: e.g., equivalent in the limit LaTeX: \,\varepsilon\!\rightarrow\!0^+\,

LaTeX: \begin{array}{cl}\text{minimize}_{x_{}\in_{_{}}\mathbb{R}^{^n},~t_{}\in_{_{}}\mathbb{R}^{^n}}&\langle t\,,\,y\rangle\\
\text{subject to}&_{}A_{}x=b\\
&-t\preceq x\preceq_{_{}}t

LaTeX: \begin{array}{cl}\text{minimize}_{y_{}\in_{_{}}\mathbb{R}^{^n}}&\langle |x^*|\,,\,y\rangle\\
\text{subject to}&0\preceq y\preceq\mathbf{1}\\
&y^{\rm T}\mathbf{1}=n-_{}k

For a coded numerical example, see Candes.m

Personal tools