Douglas-Rachford Splitting (DRS)

• Douglas–Rachford Splitting (DRS) is an iterative technique for solving monotone inclusion problems using resolvent operators and fixed-point iterations.
• It employs a distinct shadow sequence that recovers primal solutions and achieves strong convergence in both consistent and inconsistent settings.
• DRS underpins various optimization algorithms including ADMM and has been extended to robust, decentralized, and nonconvex frameworks with provable rate guarantees.

Douglas–Rachford splitting (DRS) is an iterative splitting method for solving the monotone inclusion

$0\in Ax+Bx,$

where $A$ and $B$ are maximally monotone operators on a real Hilbert space. Originating in the 1956 affine setting and substantially extended by Lions–Mercier in 1979, DRS has become a standard tool for convex feasibility, convex composite optimization, and operator-theoretic formulations related to ADMM. A characteristic feature of the method is the distinction between the governing fixed-point iterates and the associated “shadow” sequence obtained by a resolvent or projection; in many important settings the shadow carries the primal solution information even when the raw iterates converge to a larger fixed-point set or drift in inconsistent problems (Bauschke et al., 2016, Bauschke et al., 2015).

1. Operator-theoretic formulation

For a maximally monotone operator $A$, the resolvent and reflected resolvent are

$$J_A:=(\operatorname{Id}+A)^{-1},\qquad R_A:=2J_A-\operatorname{Id}.$$

The Douglas–Rachford operator for the ordered pair $(A,B)$ is

$$T_{(A,B)}:=\frac12(\operatorname{Id}+R_BR_A)=\operatorname{Id}-J_A+J_BR_A.$$

Starting from $x_0\in X$, the DRS iteration is

$x_{n+1}=Tx_n,\qquad x_n=T^n x_0.$

The associated shadow sequence is

$J_A T^n x_0,$

which reduces to a projection sequence in feasibility problems (Bauschke et al., 2015).

This operator form encodes the solution set through fixed points. In the consistent case,

$A$0

so fixed points of $A$1 are not themselves the primal solutions in general; rather, their resolvent images are. Classical consistent-case theory gives weak convergence of the raw iterates to some fixed point of $A$2, and weak convergence of the shadow sequence to an element of $A$3 (Bauschke et al., 2016, Bauschke et al., 2015).

When $A$4 and $A$5 for proper lower semicontinuous convex functions, DRS becomes a proximal splitting method: $A$6 with $A$7. In this form, DRS sits at the center of a large family of primal, dual, and primal-dual algorithms, including ADMM through standard reformulations (Bauschke et al., 2021, Patrinos et al., 2014).

2. Convex feasibility and the shadow sequence

A particularly important specialization is the convex feasibility problem

$A$8

obtained by choosing

$A$9

with $B$0 nonempty closed convex sets. Then

$B$1

and the Douglas–Rachford operator becomes

$B$2

The shadow sequence is then

$B$3

This is the sequence that typically represents the primal feasibility candidates (Bauschke et al., 2015).

The geometry of DRS in subspace problems is especially explicit. For closed subspaces $B$4, one has

$B$5

and the fixed-point set is exactly

$B$6

Consequently, the raw iterates do not generally converge to $B$7; rather,

$B$8

whereas the shadows satisfy

$B$9

In this precise sense, DRS becomes a best-approximation method through its shadow sequence rather than through its governing iterates (Bauschke et al., 2013).

This distinction resolves a common misconception. DRS is often informally described as “converging to the projection onto the solution set,” but that statement is generally correct only for the shadow sequence. The governing fixed-point iteration and the primal approximation sequence are different objects, and the difference is structurally significant already in the simplest subspace setting (Bauschke et al., 2013).

3. Exact geometry, strong convergence, and rates

For two closed subspaces, the convergence theory is exceptionally sharp. Let $A$0 denote the cosine of the Friedrichs angle between $A$1 and $A$2. Then

$A$3

and the DRS operator satisfies the exact norm identity

$A$4

Likewise, the shadow sequence has the exact rate

$A$5

Thus strong convergence always holds in Hilbert space, and linear convergence occurs exactly under the closed-sum condition, with rate equal to the cosine of the Friedrichs angle (Bauschke et al., 2013).

The subspace theory also admits explicit model examples. For two lines in $A$6,

$A$7

the DRS operator is

$A$8

so

$A$9

This makes the role of the Friedrichs angle completely transparent: the asymptotic contraction factor is precisely $$J_A:=(\operatorname{Id}+A)^{-1},\qquad R_A:=2J_A-\operatorname{Id}.$$0 (Bauschke et al., 2013).

The affine setting is broader than subspaces. For maximally monotone affine relations, the DRS operator is affine nonexpansive, and this rigidity yields strong convergence of the raw iterates themselves: $$J_A:=(\operatorname{Id}+A)^{-1},\qquad R_A:=2J_A-\operatorname{Id}.$$1 In finite dimensions, this convergence is linear. Under additional assumptions such as paramonotonicity and the orthogonality condition $$J_A:=(\operatorname{Id}+A)^{-1},\qquad R_A:=2J_A-\operatorname{Id}.$$2, the shadow sequence converges strongly to the metric projection onto the primal solution set,

$$J_A:=(\operatorname{Id}+A)^{-1},\qquad R_A:=2J_A-\operatorname{Id}.$$3

This sharpens the general Lions–Mercier picture, where one usually has only weak convergence (Bauschke et al., 2016).

Recent rate analysis has added a complementary operator-theoretic perspective. For the relaxed DRS operator

$$J_A:=(\operatorname{Id}+A)^{-1},\qquad R_A:=2J_A-\operatorname{Id}.$$4

an exact worst-case fixed-point residual rate was established for $$J_A:=(\operatorname{Id}+A)^{-1},\qquad R_A:=2J_A-\operatorname{Id}.$$5: $$J_A:=(\operatorname{Id}+A)^{-1},\qquad R_A:=2J_A-\operatorname{Id}.$$6 and this bound is sharp already in convex feasibility. The same work characterizes linear convergence in terms of the error bound

$$J_A:=(\operatorname{Id}+A)^{-1},\qquad R_A:=2J_A-\operatorname{Id}.$$7

which is both sufficient and necessary for linear convergence in distance to the fixed-point set (Abbaszadehpeivasti et al., 8 Sep 2025).

4. Inconsistent problems, infimal displacement, and normal solutions

The inconsistent case, where $$J_A:=(\operatorname{Id}+A)^{-1},\qquad R_A:=2J_A-\operatorname{Id}.$$8, is one of the defining subtleties of DRS. A central object is the infimal displacement vector

$$J_A:=(\operatorname{Id}+A)^{-1},\qquad R_A:=2J_A-\operatorname{Id}.$$9

In convex feasibility with sets $(A,B)$0 and $(A,B)$1, this becomes

$(A,B)$2

The associated normal problem shifts one operator, or one set, by the minimal displacement: $(A,B)$3 In the normal-cone feasibility setting the normal solution set is

$(A,B)$4

Thus the original inconsistent problem is replaced by a canonically shifted consistent one (Bauschke et al., 2015, Bauschke et al., 2021).

For two affine subspaces, this structure becomes exact. One has

$(A,B)$5

so the inconsistent original problem is converted into a consistent DRS problem for $(A,B)$6 and $(A,B)$7. The main convergence theorem then states that for every $(A,B)$8,

$(A,B)$9

The convergence is strong, and if $$T_{(A,B)}:=\frac12(\operatorname{Id}+R_BR_A)=\operatorname{Id}-J_A+J_BR_A.$$0 is closed, it is linear with rate $$T_{(A,B)}:=\frac12(\operatorname{Id}+R_BR_A)=\operatorname{Id}-J_A+J_BR_A.$$1. By contrast, the raw iterates $$T_{(A,B)}:=\frac12(\operatorname{Id}+R_BR_A)=\operatorname{Id}-J_A+J_BR_A.$$2 need not converge and typically drift to infinity in inconsistent convex feasibility (Bauschke et al., 2015).

Beyond affine subspaces, the inconsistent convex-feasibility shadow sequence also admits a general weak-convergence theorem. If

$$T_{(A,B)}:=\frac12(\operatorname{Id}+R_BR_A)=\operatorname{Id}-J_A+J_BR_A.$$3

then

$$T_{(A,B)}:=\frac12(\operatorname{Id}+R_BR_A)=\operatorname{Id}-J_A+J_BR_A.$$4

This completes the weak-convergence picture for convex feasibility shadows in the inconsistent case (Bauschke et al., 2016).

For general convex optimization

$$T_{(A,B)}:=\frac12(\operatorname{Id}+R_BR_A)=\operatorname{Id}-J_A+J_BR_A.$$5

the corresponding normal problem is

$$T_{(A,B)}:=\frac12(\operatorname{Id}+R_BR_A)=\operatorname{Id}-J_A+J_BR_A.$$6

with $$T_{(A,B)}:=\frac12(\operatorname{Id}+R_BR_A)=\operatorname{Id}-J_A+J_BR_A.$$7. Under assumptions including $$T_{(A,B)}:=\frac12(\operatorname{Id}+R_BR_A)=\operatorname{Id}-J_A+J_BR_A.$$8 and solvability of the normal problem, the primal shadow satisfies

$$T_{(A,B)}:=\frac12(\operatorname{Id}+R_BR_A)=\operatorname{Id}-J_A+J_BR_A.$$9

where $x_0\in X$0 minimizes $x_0\in X$1, while the companion shadow satisfies

$x_0\in X$2

and the objective values converge to the optimal value of the shifted problem. The same analysis decomposes the minimal displacement as

$x_0\in X$3

linking domain mismatch and range mismatch to asymptotic behavior of primal and dual shadows (Bauschke et al., 2021).

5. Envelope viewpoints and nonconvex DRS

A major development in DRS theory is the introduction of the Douglas–Rachford envelope (DRE). For convex composite minimization

$x_0\in X$4

with smooth $x_0\in X$5, the DRE is

$x_0\in X$6

This function is continuously differentiable even when $x_0\in X$7 is nonsmooth, and for $x_0\in X$8 it satisfies

$x_0\in X$9

When $x_{n+1}=Tx_n,\qquad x_n=T^n x_0.$0 is convex quadratic, DRS is exactly a scaled gradient method on the DRE, which permits transfer of smooth optimization techniques to DRS, including explicit parameter tuning and accelerated variants. In that setting the recommended tuning is

$x_{n+1}=Tx_n,\qquad x_n=T^n x_0.$1

and an accelerated DRS variant achieves an $x_{n+1}=Tx_n,\qquad x_n=T^n x_0.$2 objective bound (Patrinos et al., 2014).

The DRE also plays a central role in nonconvex analysis. For

$x_{n+1}=Tx_n,\qquad x_n=T^n x_0.$3

with $x_{n+1}=Tx_n,\qquad x_n=T^n x_0.$4 and $x_{n+1}=Tx_n,\qquad x_n=T^n x_0.$5 proper lsc, the DRE yields a sufficient decrease inequality for DRS and PRS, leading to global residual convergence and subsequential convergence to stationary points under explicit parameter ranges. The same analysis shows that the parameter bounds are tight whenever the relaxation parameter lies in $x_{n+1}=Tx_n,\qquad x_n=T^n x_0.$6 (Themelis et al., 2017).

In weakly convex optimization, the envelope viewpoint becomes even closer to the Moreau-envelope interpretation of the proximal point method. If $x_{n+1}=Tx_n,\qquad x_n=T^n x_0.$7 is $x_{n+1}=Tx_n,\qquad x_n=T^n x_0.$8-smooth and $x_{n+1}=Tx_n,\qquad x_n=T^n x_0.$9 is proper lsc and $J_A T^n x_0,$0-weakly convex, then the Douglas–Rachford envelope epi-approximates the original objective as $J_A T^n x_0,$1. Under boundedness of the generated sequence, DRS yields convergence of DRE values and convergence of cluster points to critical points of $J_A T^n x_0,$2. Under a local error bound and the proper separation of isocost surfaces, the DRE gap converges $J_A T^n x_0,$3-linearly and the iterates converge $J_A T^n x_0,$4-linearly to a critical point (Atenas, 2023).

This suggests a unifying interpretation: in convex settings the DRE exposes best-approximation and rate structure, while in weakly convex settings it supplies a descent function strong enough to replace the convex fixed-point geometry that classical DRS analysis relies on.

6. Generalizations, variants, and graph-structured extensions

Modern work has pushed DRS far beyond its classical form without abandoning its basic resolvent architecture. One direction is parameter generalization. In the convex optimization setting, all frugal, no-lifting, fixed-point-encoding resolvent splittings have the three-line form

$J_A T^n x_0,$5

and unconditional convergence holds exactly when

$J_A T^n x_0,$6

Classical relaxed DRS is the special case $J_A T^n x_0,$7, $J_A T^n x_0,$8. This identifies a larger family of unconditionally convergent DRS-type methods specific to convex optimization (Nilsson et al., 24 Nov 2025).

A second direction is robustness beyond monotonicity. For the inclusion

$J_A T^n x_0,$9

with $A$00 maximally $A$01-monotone and $A$02 maximally $A$03-monotone, including the regime “strongly monotone + weakly monotone,” an adaptive DRS scheme uses

$A$04

with the coupling

$A$05

If $A$06 and the admissibility conditions of the paper hold, the iterates converge weakly to a fixed point and the shadow recovers the unique primal solution when $A$07; if one operator is Lipschitz, global linear convergence follows (Dao et al., 2018).

Inexactness is another major extension. A fully inexact DRS method with relative error tolerance allows both proximal subproblems to be solved approximately within $A$08-enlargements and residual tests, while still guaranteeing weak convergence of the generated sequences to a point in the extended solution set, provided a solution exists. A semi-inexact variant solves the first subproblem exactly and the second inexactly under a related criterion (Svaiter, 2018).

Finally, DRS has acquired graph-based and decentralized forms. For sums of $A$09 maximal monotone operators, bilevel-graph constructions generate unconditionally stable frugal resolvent splittings with the minimal $A$10-fold lifting required by general lower bounds. These schemes are instances of degenerate preconditioned proximal point methods, and they recover Ryu-type and Malitsky–Tam-type multi-operator splittings as special cases (Bredies et al., 2022). In decentralized smooth optimization over compact submanifolds, a decentralized DRS algorithm and its inexact variant integrate nonconvex DRS, gradient tracking, and projection onto a manifold; under proximal smoothness of the manifold constraint, both methods achieve an $A$11 rate for consensus and stationarity measures (Deng et al., 2023).

Across these variants, the persistent structural ideas are resolvent-based decomposition, shadow recovery of primal information, and fixed-point iterations whose convergence is often best understood through auxiliary geometry: Friedrichs angles, infimal displacement vectors, envelopes, or graph-induced preconditioners.