Alternating Direction Method of Multipliers#
Consider the problem
We then augment the objective with a strongly convex term
with \(\rho > 0\). We then have the augemented Lagrangian as
The ADMM algorithm then repeats the following steps
ADMM
\(x^{(k)} = \underset{x}{\mathrm{argmin}}\ L_\rho(x, z^{(k-1)}, u^{(k-1)})\)
\(z^{(k)} = \underset{z}{\mathrm{argmin}}\ L_\rho(x^{(k)}, z, u^{(k-1)})\)
\(u^{(k)} = u^{(k-1)} + \rho(Ax^{(k)} + Bz^{(k)} - c)\)
The ADMM algorithm is also expressed in the scaled form, where we define \(w = u / \rho\), and write the above steps as
ADMM in Scaled Form
\(x^{(k)} = \displaystyle \underset{x}{\mathrm{argmin}}\ f(x) + \frac{\rho}{2}\Big\|Ax + Bz^{(k-1)} - c + w^{(k-1)}\Big\|_2^2\)
\(z^{(k)} = \displaystyle \underset{x}{\mathrm{argmin}}\ g(z) + \frac{\rho}{2}\Big\|Ax^{(k)} + Bz - c + w^{(k-1)}\Big\|_2^2\)
\(w^{(k)} = w^{(k-1)} + Ax^{(k)} + Bz^{(k)} - c\)
Connection to Proximal Operators#
If we consider the problem of
we can write it in ``ADMM” form
We can then write the ADMM updates in scaled form. Note that in this case we have \(A = I\), \(B = -I\), and \(c = \mathbf{0}\). First, we have the \(x\) update as
Then, we have the \(z\) update as
Finally, we have the \(w\) update as