What is the exact solution $x_{n \times 1}$ of the following constrained optimization problem \begin{align*} &\min \|A x - b\|^2 \\ s.t.& C x = 0 \end{align*} where $A$ is full column rank $m \times n$ matrix ($m>n$); $b$ is $m \times 1$ matrix; $C$ is full row rank $1 \times n$ matrix?
Asked
Active
Viewed 575 times
3 Answers
1
The problem is given by:
$$ \begin{alignat*}{3} \arg \min_{x} & \quad & \frac{1}{2} \left\| A x - b \right\|_{2}^{2} \\ \text{subject to} & \quad & C x = \boldsymbol{0} \end{alignat*} $$
The Lagrangian is given by:
$$ L \left( x, \nu \right) = \frac{1}{2} \left\| A x - b \right\|_{2}^{2} + {\nu}^{T} C x $$
From KKT Conditions the optimal values of $ \hat{x}, \hat{\nu} $ obeys:
$$ \begin{bmatrix} {A}^{T} A & {C}^{T} \\ C & 0 \end{bmatrix} \begin{bmatrix} \hat{x} \\ \hat{\nu} \end{bmatrix} = \begin{bmatrix} {A}^{T} b \\ \boldsymbol{0} \end{bmatrix} $$
Now all needed is to solve the above with any Linear System Solver.
Royi
- 10,050
-
What happens in the case that the constraints are of the form: Cx <= 0 ?? How it affects the solution proposed here? – noob-mathematician Sep 12 '18 at 15:44
-
@noob-mathematician, Please open a question and link here. I will try to solve it. – Royi Sep 12 '18 at 16:54
-
Related: https://math.stackexchange.com/questions/73712, https://math.stackexchange.com/questions/625850. – Royi Sep 12 '18 at 18:25
-
https://math.stackexchange.com/questions/2915740/linear-least-squares-and-lagrange – noob-mathematician Sep 13 '18 at 15:28
1
$
\def\LR#1{\left(#1\right)}
\def\qiq{\quad\implies\quad}
$Since the constraint is linear it has an explicit solution in terms of the nullspace projector
$$\eqalign{
P &= I-C^+C \qiq P^T=P=P^2 \\
}$$
$$\eqalign{
Cx &= 0 \qiq x = Pu
}$$
where a new unconstrained vector variable $u$ has been introduced.
In terms of $u,\,$ the objective function is likewise unconstrained $$\eqalign{ \phi &= \|APu-b\|^2 \\ }$$ Thus it can be immediately solved using a pseudoinverse $$\eqalign{ u &= \LR{AP}^+b \qiq x = P\LR{AP}^+b = \LR{I-C^+C}\LR{A-AC^+C}^+b \\ }$$ ${\bf NB\!:}\;$ This solution remains valid when $C$ is an arbitrary matrix, not simply a vector.
greg
- 40,033
1
The Lagragian is the following:
$$ L = \sum \limits_{i = 1}^{m}(\sum \limits_{j = 1}^{2}a_{ij}x_{j} - b_{i})^{2} + \lambda \sum \limits_{j = 1}^{n}c_{i}x_{i} $$
$$ \frac{\partial L}{\partial x_{k}} = 2 \sum \limits_{j = 1}^{n}(\sum \limits_{i = 1}^{n}a_{ij}a_{ik})x_{j} - 2 n \sum \limits_{i = 1}^{m}b_{i} a_{ik} + \lambda c_{k} = 0 $$
There are $n + 1$ equations and $n + 1$ unknowns.
Ilya Palachev
- 464
-
1Just to make it clear: $n$ equations come from $\frac{\partial L}{\partial x_k}$, the last is the constraint $Cx=0$. – AndreaCassioli Jul 11 '14 at 11:23