1
$\begingroup$

Let $ \mathcal{O}^{n} $ be the set defined by $ \mathcal{O}^{n} = \left\{ X \in \mathbb{R}^{n \times n} \mid {X}^{T} X = I \right\} $, namely the set of Orthognal Matrices of size $ n \times n $.

I would like to solve the optimization problem, Given $ Y \in \mathbb{R}^{3 \times 3} $:

$$ \begin{align*} \arg \min_{X} \quad & d \left( X, Y \right) \\ \text{subject to} \quad & X \in \mathcal{O}^{3} \end{align*} $$

For my purposes, the type of distance $d$ in the objective does not matter, as long as it indeed a distance-like function (e.g., Bregman distance, norm-squared, etc). Is there a known algorithm for solving such problem for some function $d$?

$\endgroup$
1
2
$\begingroup$

I will solve the problem for the Frobenius Norm:

$$ \begin{align*} \arg \min_{X} \quad & {\left\| X - Y \right\|}_{F}^{2} \\ \text{subject to} \quad & X \in \mathcal{O}^{n} \end{align*} $$

By definition of the Frobenius Norm:

$$\begin{align*} {\left\| X - Y \right\|}_{F}^{2} & = \operatorname{tr} \left( {\left( X - Y \right)}^{T} \left( X - Y \right) \right) & \text{By definition of $ {\left\| \cdot \right\|}_{F}^{2} $} \\ & = \operatorname{tr} \left( {X}^{T} X \right) - 2 \operatorname{tr} \left( {X}^{T} Y\right) + \operatorname{tr} \left( {Y}^{T} Y \right) && \text{} \\ & \Rightarrow \arg \min_{X \in \mathcal{O}^{n}} {\left\| X - Y \right\|}_{F}^{2} = \arg \max_{X \in \mathcal{O}^{n}} \operatorname{tr} \left( {X}^{T} Y \right) & \text{} \\ & = \arg \max_{X \in \mathcal{O}^{n}} \operatorname{tr} \left( {X}^{T} U S {V}^{T} \right) & \text{The SVD of $ Y = U S {V}^{T} $} \\ & = \arg \max_{X \in \mathcal{O}^{n}} \operatorname{tr} \left( {V}^{T} {X}^{T} U S \right) & \text{Cyclic property of the Trace} \end{align*}$$

Defining $ Z = {V}^{T} {X}^{T} U $ we get (Pay attention that $ {Z}^{T} Z = I $, Namely it is an Orthogonal Matrix):

$$ \operatorname{tr} \left( Z S \right) = \sum_{i = 1}^{n} {z}_{ii} {\sigma}_{ii} \leq \sum_{i = 1}^{n} {\sigma}_{i} $$

Where $ {\sigma}_{i} = {s}_{ii} $ the Singular Values of $ Y $ (Diagonal Elements of the Diagonal Matrix $ S $).
The inequality is a result of $ Z $ being an Orthogonal Matrix hence its elements obey $ \left| {z}_{ij} \right| \leq 1 $ and equality is achieved when $ Z = I $.
If $ Z = I \Rightarrow {X}^{T} = V {U}^{T} \Rightarrow X = U {V}^{T} $.

Hence the answer is given by $ X = U {V}^{T} $.

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.