ADMM - proximal vs. augmented lagrangian interpretation: stuck on algebraic step :(

201 Views Asked by At

I'm trying to understand conciliation of the proximal formulation of ADMM with its Augmented Lagrangian (method of multipliers) interpretation. In particular, I'm struggling with page 156 here (https://web.stanford.edu/~boyd/papers/pdf/prox_algs.pdf) where it says "pull the linear terms into the quadratic ones". To be more specific:

$\mathbf y^{T}\mathbf x + (\rho/2) ||\mathbf x- \mathbf z||_2^2$

should become

$(\rho/2) ||\mathbf x- \mathbf z + (1/\rho)\mathbf y||_2^2$

I cannot make sense of that passage. Can someone detail how the equivalence is achieved? Am I missing something obvious?

1

There are 1 best solutions below

0
On BEST ANSWER

The two expressions, i.e. scaled and unscaled forms, are equivalent.

We have

$$\begin{align*} \textbf{y}^T(\textbf{x-z}) + \frac{\rho}{2}\|\textbf{x-z}\|_2^2 &= \frac{\rho}{2}\|\textbf{x-z} + \frac{\textbf{y}}{\rho}\|_2^2 - (\frac{1}{2\rho})\|\textbf{y}\|_2^2 \\ &=\frac{\rho}{2}(\textbf{x-z} + \frac{\textbf{y}}{\rho})^T(\textbf{x-z} + \frac{\textbf{y}}{\rho})- (\frac{1}{2\rho})\|\textbf{y}\|_2^2 \\ &=\frac{\rho}{2}((\textbf{x-z})^T(\textbf{x-z}) + (\textbf{x-z})^T(\frac{1}{\rho})\textbf{y} + (\frac{1}{\rho})\textbf{y}^T(\textbf{x-z}) + (\frac{1}{\rho})\textbf{y}^T(\frac{1}{\rho})\textbf{y}) - (\frac{1}{2\rho})\|\textbf{y}\|_2^2 \\ &=\frac{\rho}{2}(\|\textbf{x-z}\|^2_2 + (\frac{2}{\rho})\textbf{y}^T(\textbf{x-z}) + (\frac{1}{\rho^2})\|\textbf{y}\|^2_2) - (\frac{1}{2\rho})\|\textbf{y}\|_2^2 \\ &=\frac{\rho}{2}\|\textbf{x-z}\|^2_2 + \textbf{y}^T(\textbf{x-z}) + (\frac{1}{2\rho})\|\textbf{y}\|_2^2 - (\frac{1}{2\rho})\|\textbf{y}\|_2^2 \\ &=\frac{\rho}{2}\|\textbf{x-z}\|^2_2 + \textbf{y}^T(\textbf{x-z}) \end{align*}$$

$\textbf{Note}$ that the term $-(\frac{1}{2\rho})\|\textbf{y}\|_2^2$ is generally dropped out from the ADMM iterations because it doesn't depend nither on $\textbf{x}$ nor $\textbf{z}$.