Acceleration
SCS includes Anderson acceleration (AA), which can be used to speed up convergence. AA is a quasiNewton method for the acceleration of fixed point iterations and can dramatically speed up convergence in practice, especially if higher accuracy solutions are desired. However, it can also cause severe instability of the solver and so should be used with caution. It is an open research question how to best implement AA in practice to ensure good performance across all problems and we welcome any contributions in that direction!
Mathematical details
The discussion here is taken from section 2 of our paper. Consider the problem of finding a fixed point of the function \(f: \mathbf{R}^n \rightarrow \mathbf{R}^n\), i.e.,
In our case \(f\) corresponds to one step of DouglasRachford
splitting and the iterate is the \(w^k\) vector, which
converges to a fixed point of the DR operator. At a high level AA, from initial
point \(x_0\) and max memory \(m\) (corresponding to the
acceleration_lookback
setting), works as follows:
In other words, AA produces an iterate that is the linear combination of the last \(m_k + 1\) outputs of the map. Thus, the main challenge is in choosing the weights \(\alpha \in \mathbf{R}^{m_k+1}\). There are two ways to choose them, corresponding to typeI and typeII AA (named for the typeI and typeII Broyden updates). We shall present typeII first.
TypeII AA
Define the residual \(g: \mathbf{R}^n \rightarrow \mathbf{R}^n\) of \(f\) to be \(g(x) = x  f(x)\). Note that any fixed point \(x^\star\) satisfies \(g(x^\star) = 0\). In typeII AA the weights are selected by solving a small least squares problem.
More explicitly, we can reformulate the above as follows:
with variable \(\gamma=(\gamma_0,\dots,\gamma_{m_k1}) \in \mathbf{R}^{m_k}\). Here \(g_i=g(x^i)\), \(Y_k=[y_{km_k}~\dots~y_{k1}]\) with \(y_i=g_{i+1}g_i\) for each \(i\), and \(\alpha\) and \(\gamma\) are related by \(\alpha_0=\gamma_0\), \(\alpha_i=\gamma_i\gamma_{i1}\) for \(1\leq i\leq m_k1\) and \(\alpha_{m_k}=1\gamma_{m_k1}\).
Assuming that \(Y_k\) is full column rank, the solution \(\gamma^k\) to the above is given by \(\gamma^k=(Y_k^\top Y_k)^{1}Y_k^\top g_k\), and hence by the relation between \(\alpha^k\) and \(\gamma^k\), the next iterate of typeII AA can be written as
where \(S_k=[s_{km_k}~\dots~s_{k1}]\), \(s_i=x^{i+1}x^i\) for each \(i\), and
Observe that \(B_k\) minimizes \(\B_kI\_F\) subject to the inverse multisecant condition \(B_kY_k=S_k\), and hence can be regarded as an approximate inverse Jacobian of \(g\). The update of \(x^k\) can then be considered as a quasiNewtontype update, with \(B_k\) being a generalized second (or typeII) Broyden’s update of \(I\) satisfying the inverse multisecant condition.
TypeI AA
In the same spirit, we define typeI AA, in which we find an approximate Jacobian of \(g\) minimizing \(\H_kI\_F\) subject to the multisecant condition \(H_kS_k=Y_k\). Assuming that \(S_k\) is full column rank, we obtain (by symmetry) that
and the update scheme is defined as
assuming \(H_k\) to be invertible. A direct application of the Woodbury matrix identity shows that
where again we have assumed that \(S_k^\top Y_k\) is invertible. Notice that this explicit formula of \(H_k^{1}\) is preferred in that the most costly step, inversion, is implemented only on a small \(m_k\times m_k\) matrix.
In SCS
In SCS both types of acceleration are available, though by default typeI is
used since it tends to have better performance. If you wish to use AA then set
the acceleration_lookback
setting to a nonzero value (10 works well for
many problems and is the default). This setting corresponds to \(m\), the
maximum number of SCS iterates that AA will use to extrapolate to the new point.
To enable typeII acceleration then set acceleration_lookback
to a
negative value, the sign is interpreted as switching the AA type (this is mostly
so that we can test it without fully exposing it the user).
The setting acceleration_interval
controls how frequently AA is applied.
If acceleration_interval
\(=k\) for some integer \(k \geq 1\)
then AA is applied every \(k\) iterations (AA simply ignores the
intermediate iterations). This has the benefit of making AA \(k\) times
faster and approximating a \(k\) times larger memory, as well as improving
numerical stability by ‘decorrelating’ the data. On the other hand, older
iterates might be stale. More work is needed to determine the optimal setting
for this parameter, but 10 appears to work well in practice and is the default.
The details about how the linear systems are solved and updated is abstracted away into the AA package (eg, QR decomposition, SVD decomposition etc). Exactly how best to solve and update the equations is still open.
Regularization
By default we also add a small amount of regularization to the matrices that are being inverted in the above expressions, ie, in the typeII update
for some small \(\epsilon > 0\), and similarly for the typeI update
which is equivalent to adding regularization to the \(S_k^\top S_k\) matrix before using the Woodbury matrix identity. The regularization ensures the matrices are invertible and helps stability. In practice typeI tends to require more regularization than typeII for good performance. The regularization shrinks the AA update towards the update without AA, since if \(\epsilon\rightarrow\infty\) then \(\gamma^\star = 0\) and the AA step reduces to \(x^{k+1} = f(x^k)\). Note that the regularization can be folded into the matrices by appending \(\sqrt{\epsilon} I\) to the bottom of \(S_k\) or \(Y_k\), which is useful when using a QR or SVD decomposition to solve the equations.
Max \(\gamma\) norm
As the algorithm converges to the fixed point the matrices to be inverted
can become illconditioned and AA can become unstable. In this case the
\(\gamma\) vector can become very large. As a simple heuristic we reject
the AA update and reset the AA state whenever \(\\gamma\_2\) is greater
than max_weight_norm
(eg, something very large like \(10^{10}\)).
Safeguarding
We also apply a safeguarding step to the output of the AA step. Explicitly, let \(x^k\) be the current iteration and let \(x_\mathrm{AA} = x^{k+1}\) be the output of AA. We reject the AA step if
where \(\zeta\) is the safeguarding tolerance factor
(safeguard_factor
) and defaults to 1. In other words we reject the step
if the norm of the residual after the AA step is larger than some amount (eg, if
it increases the residual from the previous iterate). After rejecting a step we
revert the iterate to \(x^k\) and reset the AA state.
Relaxation
In some works relaxation has been shown to improve performance. Relaxation replaces the final step of AA by mixing the map inputs and outputs as follows:
where \(\beta\) is the relaxation
parameter, and \(\beta=1\)
recovers vanilla AA. This can be computed using the matrices defined above using
Anderson acceleration API
For completeness, we document the full Anderson acceleration API below.
Functions

AaWork *aa_init(aa_int dim, aa_int mem, aa_int type1, aa_float regularization, aa_float relaxation, aa_float safeguard_factor, aa_float max_weight_norm, aa_int verbosity)
Initialize Anderson Acceleration, allocates memory.
 Parameters:
dim – the dimension of the variable for AA
mem – the memory (number of past iterations used) for AA
type1 – if True use type 1 AA, otherwise use type 2
regularization – typeI and typeII different, for typeI: 1e8 works well, typeII: more stable can use 1e12 often
relaxation – float in [0,2], mixing parameter (1.0 is vanilla)
safeguard_factor – factor that controls safeguarding checks larger is more aggressive but less stable
max_weight_norm – float, maximum norm of AA weights
verbosity – if greater than 0 prints out various info
 Returns:
pointer to AA workspace

aa_float aa_apply(aa_float *f, const aa_float *x, AaWork *a)
Apply Anderson Acceleration. The usage pattern should be as follows:
for i = 0 .. N:
if (i > 0): aa_apply(x, x_prev, a)
x_prev = x.copy()
x = F(x)
aa_safeguard(x, x_prev, a) // optional but helps stability
Here F is the map we are trying to find the fixed point for. We put the AA before the map so that any properties of the map are maintained at the end. Eg if the map contains a projection onto a set then the output is guaranteed to be in the set.
 Parameters:
f – output of map at current iteration, overwritten with AA output
x – input to map at current iteration
a – workspace from aa_init
 Returns:
(+ or ) norm of AA weights vector. If positive then update was accepted and f contains new point, if negative then update was rejected and f is unchanged

aa_int aa_safeguard(aa_float *f_new, aa_float *x_new, AaWork *a)
Apply safeguarding.
This step is optional but can improve stability.
 Parameters:
f_new – output of map after AA step
x_new – AA output that is input to the map
a – workspace from aa_init
 Returns:
0 if AA step is accepted otherwise 1, if AA step is rejected then this overwrites f_new and x_new with previous values