Skip to main content

A Python package for type-II Anderson accelerated Douglas-Rachford splitting (A2DR).

Project description


a2dr is a Python package for solving large-scale non-smooth convex optimization problems with general linear constraints, with separable objective functions accessible through their proximal operators. It exploits the separability of the objective functions and the sparsity in the linear constraints, and utilizes the power of Anderson acceleration to achieve fast and robust convergence and scalability to multiple processors. The current version is 0.2.3.post2.

It is an implementation of type-II Anderson accelerated Douglas-Rachford splitting, based on our paper A. Fu, J. Zhang, and S. Boyd (2019).


To install using pip (recommended), run:

pip install a2dr

To install a2dr from source, first make sure that you have setuptools installed. Then follow the steps below:

  1. Clone the a2dr git repository.
  2. Navigate to the top-level of the cloned directory and run:
python install

To test the installation with nose, first make sure that you have nose installed. Then run:

nosetests a2dr

The requirements are:

Please file an issue on Github if you want Python 2 support.


a2dr solves problems of the following form:

minimize         f_1(x_1) + ... + f_N(x_N)
subject to       A_1x_1 + ... + A_Nx_N = b.

where f_i (i=1,...,N) are convex, closed and proper, and are only accessible through their proximal operators. Notice that f_i can take infinite values, and in particular constraints can be included in the objectives with the help of indicator functions.

Prox-affine forms

The above formulation is also referred to as prox-affine forms in the literature (see e.g., Epsilon). When it is seen as a standard form for generic convex optimization problems, the major advantage of prox-affine forms compared to the more widely used conic forms include:

  • Privacy: suitable for peer-to-peer optimization with privacy requirements.
    • In practice, the data and source code that define the proximal oracle can be securely encrypted (e.g., via compilation) so that privacy is preserved. For example, in Python, we can convert the .py file containing the proximal operator function into an encrypted .so file via the Cython extension.
  • Compactness: straightforward canonicalization/transformation and lower dimensional representations.
    • In general, the prox-affine form often requires less variables than the conic form (see e.g., the portfolio optimization example here). The compactness advantage is also partly exemplified by the comparison between a2dr and SCS in the sparse covariance estimation example in our paper.

For a bit more detailed introduction to prox-affine forms and the comparisons with conic forms, see our companion slides.


After installing a2dr, you can import a2dr using

import a2dr

This module exposes a function a2dr (the solver), which can be used via a2dr.a2dr, or directly imported using

from a2dr import a2dr

The function a2dr is called with the command

a2dr_result = a2dr(p_list,


The arguments p_list, A_list and b correspond to the problem data.

  • p_list is the list of proximal operators of f_i. Each element of p_list is a Python function, which takes as input a vector v and parameter t > 0 and outputs the proximal operator of f_i evaluated at (v,t).
  • A_list is the list of A_i. The lists p_list and A_list must be given in the same order i = 1,...,N.
  • b is the vector b. Notice that A_list and b are optional, and when omitted, the solver recognizes the problem as one without linear constraints. Also notice that in such cases, A_list and b have to be omitted together, and either v_init or n_list has to be provided to declare the dimension of each x_i.

For information on the other optional hyper-parameters, please refer to our companion paper (Algorithm 2) and the source code comments of the function a2dr in


The returned object a2dr_result is a dictionary containing the keys 'x_vals', 'primal', 'dual', 'num_iters' and 'solve_time':

  • The output x_vals is a list of x_1,...,x_N from the iteration with the smallest residuals.
  • primal and dual are arrays containing the primal and dual residual norms for the entire iteration process, respectively.
  • The value num_iters is the total number of iterations, and solve_time is the algorithm runtime.

When the linear equality constraint is infeasible, the solver will print the corresponding flag and return None for all the outputs.

Other tools

The module a2dr also comes with several additional tools that facilitates the transformation of the problems into the required input format described above as well as tests and visualization. In particular, it come with a package for proximal operators, which can be imported via

import a2dr.proximal

It also comes with some tests and visualization tools, which can be imported via

import a2dr.tests


We showcase the usage of the solver function a2dr as well as the the tool packages a2dr.proximal and a2dr.tests with the following example. More examples can be found in the examples/ directory.

# Non-negative least squares (see our companion paper for more details)
import numpy as np
import numpy.linalg
from scipy import sparse
from a2dr import a2dr
from a2dr.proximal import *
from a2dr.tests.base_test import BaseTest

# Problem data.
m, n = 150, 300
density = 0.001
X = sparse.random(m, n, density=density, data_rvs=np.random.randn)
y = np.random.randn(m)

# Convert problem to standard form.
prox_list = [lambda v, t: prox_sum_squares_affine(v, t, F=X, g=y),
A_list = [sparse.eye(n), -sparse.eye(n)]
b = np.zeros(n)

# Solve with DRS.
drs_result = a2dr(prox_list, A_list, b, anderson=False)
# Solve with A2DR.
a2dr_result = a2dr(prox_list, A_list, b, anderson=True)
bt = BaseTest()
bt.compare_total(drs_result, a2dr_result)


If you wish to cite a2dr, please use the following:

    author       = {Fu, A. and Zhang, J. and Boyd, S.},
    title        = {Anderson Accelerated {D}ouglas-{R}achford Splitting},
    journal      = {},
    year         = {2019},

    author       = {Fu, A. and Zhang, J. and Boyd, S.},
    title        = {{a2dr}: Anderson Accelerated {D}ouglas-{R}achford Splitting, version 0.2.3.post2},
    howpublished = {\url{}},
    year         = {2020}

Project details

Download files

Download the file for your platform. If you're not sure which to choose, learn more about installing packages.

Source Distribution

a2dr-0.2.3.post2.tar.gz (34.8 kB view hashes)

Uploaded source

Built Distribution

a2dr-0.2.3.post2-py3-none-any.whl (43.4 kB view hashes)

Uploaded py3

Supported by

AWS AWS Cloud computing Datadog Datadog Monitoring Facebook / Instagram Facebook / Instagram PSF Sponsor Fastly Fastly CDN Google Google Object Storage and Download Analytics Huawei Huawei PSF Sponsor Microsoft Microsoft PSF Sponsor NVIDIA NVIDIA PSF Sponsor Pingdom Pingdom Monitoring Salesforce Salesforce PSF Sponsor Sentry Sentry Error logging StatusPage StatusPage Status page