Distributed Asynchronous Primal-Dual Optimization for Supply-Chain Networks

Laksh Patel; Neel Shanbhag

REVIEW 4 major objections 5 minor 21 references

Distributed Asynchronous Primal-Dual Optimization for Supply-Chain Networks

T0 review · 4 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read A fully asynchronous primal–dual algorithm for supply-chain flow allocation is claimed to converge almost surely to a saddle point with an optimal O(K^{-1/2}) ergodic duality gap despite unbounded sublinear messaging delays.

desk verdict The paper's central summability claim is arithmetically false, so the main theorem collapses, but the asynchronous primal-dual idea for sublinear delays is worth a second look if the proof can be fixed. read the letter →

arxiv 2506.08024 v1 pith:F3WQWVTZ submitted 2025-05-29 math.OC

Laksh Patel , Neel Shanbhag This is my paper

classification math.OC MSC 90C2590C3590B0668W15

keywords asynchronousdistributedoptimizationprimal-dualsplittingnetworkflowallocationsupply-chainnetworkssublinearmessagedelaysLyapunovanalysisdualitygapconvex

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper introduces DAPD-SCO, a fully asynchronous primal–dual algorithm for allocating flows over a directed acyclic supply-chain network, in which edge agents update flows by projected gradient descent and retailer agents update dual prices by projected gradient ascent using only local, potentially stale information. The central claim is that, whenever message delays grow more slowly than the square root of the iteration count, the iterates converge almost surely to a saddle point of the Lagrangian and the ergodic duality gap decays as O($K^{{-1/2}}$), matching centralized lower bounds. The authors position this as the first asynchronous primal–dual method for network flow problems that tolerates unbounded sublinear delays while giving explicit convergence and rate guarantees. If the claim holds, supply chains can coordinate flows without global synchronization or bounded-latency assumptions, which matters for warehouses, carriers, and retailers communicating over packet-switched links with packet loss and high staleness.

What carries the argument

The carrying object is the composite Lyapunov function $V^k = \|x^k - x^*\|^2 + \|\lambda^k - \lambda^*\|^2$. Through the primal descent lemma and dual descent lemma, the one-step change of $V^k$ is bounded by the negative of step-size-weighted saddle-point progress terms plus an error term $E_k$ that collects the squared step sizes and the cross-terms $\alpha_k S_k$ and $\beta_k T_k$, where $S_k$ and $T_k$ are sums of step sizes over the delay windows. The argument then applies the Robbins–Siegmund almost-supermartingale theorem: if $E_k$ is summable, the progress terms are summable too, yielding convergence and the $O(K^{-1/2})$ ergodic rate. This decomposition isolates everything caused by stale information into a single series, which is what makes the unbounded-delay analysis possible.

What would settle it

Compute the cumulative delay-error sum under the paper's own step sizes with delays set to $k^{0.4}$; if the series diverges, the summability lemma underlying the Lyapunov argument fails for a delay regime the paper claims to cover.

Watch

Extended reading notes

Core claim

The paper's central claim is Theorem 2 and Theorem 3: under Slater's condition and convexity/Lipschitz assumptions (F1)–(F3), with step sizes $\alpha_k = \beta_k = 1/\sqrt{k+1}$ and delays $\delta_{\max}(k), \Delta_{\max}(k) = o(k^{1/2})$, the iterates $(x^k, \lambda^k)$ converge almost surely to a saddle point of the Lagrangian, and the ergodic duality gap is $O(K^{-1/2})$ with explicit constants, matching centralized lower bounds. The proof constructs a composite Lyapunov function $V^k = \|x^k - x^*\|^2 + \|\lambda^k - \lambda^*\|^2$ and derives per-step primal and dual descent inequalities in which all delay-induced error is collected into terms $S_k = \sum_{s=k-\delta_{\max}(k)}^{k-1} \beta_s$ and $T_k = \sum_{s=k-\Delta_{\max}(k)}^{k-1} \alpha_s$ (Lemmas 1 and 2). Lemma 3 asserts that the resulting error series is summable under the chosen step sizes and delay growth, and the Robbins–Siegmund lemma (Lemma 4) then converts the almost-supermartingale inequality into almost-sure convergence and controls the ergodic gap. The same Lyapunov framework is extended to bounded observation noise, time-varying jointly connected graphs, and slowly drifting cost or capacity parameters.

Load-bearing premise

The load-bearing premise is that the cumulative error caused by stale information, accumulated over the whole run, is finite; the paper derives this from the assumption that delays grow more slowly than the square root of the iteration count.

Editorial extensions

If this is right

Supply-chain flow allocation can proceed without any global synchronization or bounded-delay assumption; each agent simply works with the latest stale message it has received.
The ergodic duality-gap rate $O(K^{-1/2})$ matches centralized lower bounds, so the asynchrony does not degrade the asymptotic worst-case rate on this metric.
Reaching an ergodic duality gap of $\varepsilon$ requires $O(\varepsilon^{-2})$ iterations, with $O(|E|+|R|)$ scalar messages per iteration.
The Lyapunov analysis extends the same guarantees to bounded observation noise, time-varying jointly connected communication graphs, and slowly drifting capacities or costs.
The intended operating regime is a directed acyclic supply-chain graph with single-commodity flows, which is exactly the setting covered by the assumptions.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

A direct check on the proof is whether the cumulative delay-error series stays finite for delays like $k^{0.4}$; the paper's assumptions allow such delays, so this is the step that determines where the theoretical guarantee actually holds.
The analysis covers single-commodity flows on directed acyclic graphs; extending to multi-commodity flows would likely need additional dual variables and stronger coupling assumptions, so practical scope may be narrower than the general supply-chain framing suggests.
Because the rate is ergodic, real-time use would rely on averaged iterates; last-iterate convergence at the same rate is not addressed.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

The paper has a good high-level idea — extending asynchronous primal-dual methods to supply-chain flow with sublinearly growing delays — and the writing is clear. The Lyapunov-based decomposition is a sensible organizing principle, and the literature review is competent. Unfortunately, the central proof is broken.

The load-bearing error is Lemma 3. For δmax(k) = O(k^γ), γ < 1/2, the paper claims S_k = O(k^{γ−1/2}) = o(k^{−1/2}). The equality is false for every γ > 0, and even for γ = 0 you get S_k = Θ(k^{−1/2}). With α_k = 1/√(k+1), the product α_k S_k = Θ(k^{γ−1}), so ∑ α_k S_k diverges for all γ ≥ 0. This kills the Robbins–Siegmund step in Theorem 2 and the O(K^{−1/2}) rate in Theorem 3. The proof as written does not establish convergence or any rate.

There are also two reproducibility issues. The assumptions (F1)–(F3) are cited throughout but never stated, so Lemmas 1–2 cannot be independently checked. The simulations use a different problem (quadratic costs, Ax = b, fixed step sizes, bounded delay τ = 5) than the theory (linear costs, box constraints, diminishing steps, sublinear delays), and there is no code or data. The empirical claims are consistent with the theory only in a loose, heuristic sense.

What the paper does well: it identifies a genuine gap in the literature and proposes a plausible algorithmic structure. The error in Lemma 3 is not a minor typo; it is a real summability mistake. But the underlying idea — using stale prices and flows with sublinearly growing delays — is coherent, and a corrected proof might well succeed with a more careful bound on the delayed error or a different step-size normalization.

Who should read this: researchers working on asynchronous distributed optimization who want to see a concrete attempt at unbounded delays in primal-dual methods. It is a good cautionary example of why asymptotic claims need checking, but it is not a reliable source for a theorem or a rate.

Recommendation: I would not accept this paper as is. The main result is unproven, and the empirical section does not offset that. But I would not desk-reject it either: a careful referee could help the authors identify the summability flaw and either fix it or state the conditions under which the claim holds. If the proof can be repaired, the paper would be a modest but real contribution to a subfield that needs more work.

Referee Report

4 major / 5 minor

Summary. The paper proposes DAPD-SCO, a distributed asynchronous primal-dual algorithm for supply-chain flow allocation on directed acyclic networks. Each edge-agent performs projected gradient descent on a flow variable using delayed prices, while each retailer-agent performs projected gradient ascent on a dual multiplier using delayed flows. The main theoretical claims are almost-sure convergence to a saddle point and an ergodic duality-gap rate of O(K^{-1/2}) under sublinearly growing delays, established through a Lyapunov decomposition and a Robbins-Siegmund argument. The paper also reports simulations on a three-tier network comparing DAPD-SCO with synchronous primal-dual, ADMM, and gradient-push baselines.

Significance. If the theoretical results were valid, the paper would make a genuine contribution: it would be the first asynchronous primal-dual method for network flow allocation that tolerates unbounded but sublinear delays while retaining a central-lower-bound-matching rate. The algorithmic idea is clearly described, and the use of a single Lyapunov framework for both primal and dual descent is appealing. However, the central summability lemma contains an arithmetic error that invalidates the proofs of both main theorems, and the assumptions on which the descent estimates rest are never stated. As it stands, the theoretical contribution is not established, so the significance can only be assessed conditionally.

major comments (4)

[Section 6.2, Lemma 3] The claim that S_k = sum_{s=k-delta_max(k)}^{k-1} (s+1)^{-1/2} satisfies S_k = O(k^{gamma-1/2}) = o(k^{-1/2}) is false for every gamma >= 0. Since the sum has delta_max(k)+1 terms each of size Theta(k^{-1/2}), one obtains S_k = Theta(k^{gamma-1/2}); for gamma = 0 this is Theta(k^{-1/2}) when the delay is bounded away from zero, not o(k^{-1/2}), and for gamma > 0 it is omega(k^{-1/2}). Consequently alpha_k S_k = Theta(k^{gamma-1}), whose series diverges for every gamma < 1/2. Therefore the error series sum E_k in Eq. (6) is not summable, the hypothesis sum v_k < infinity in Lemma 4 fails at the exact point used in Theorem 2, and the O(sqrt(K)) bound on sum_{k=0}^{K-1} E_k used in Theorem 3 is unsupported. A successful repair would require a sharper decay estimate on the delayed-price error than the crude sum of beta_s, or a different step-size schedule; neither appears in the manuscript.
[Sections 4, 5, and 6] Assumptions (F1)-(F3) are cited throughout the proofs of Lemmas 1 and 2 and in Theorems 2 and 3, but they are never stated anywhere in the manuscript. Lemma 1's proof begins 'By (F1)' and then invokes '(F2)' and '(F3)', while Lemma 2 similarly relies on '(F3)' for dual residuals; without knowing what these assumptions are, the reader cannot verify that the algorithm's projections, gradients, and delay model satisfy them. These assumptions are load-bearing for every descent estimate, so they must be stated explicitly and checked against the algorithm.
[Section 6.3, proof of Theorem 2] The inference 'since sum alpha_k = sum beta_k = infinity, the only way sum u_k can converge is if Delta^x_k -> 0 and Delta^lambda_k -> 0' is not valid. A nonnegative sequence a_k can satisfy sum alpha_k a_k < infinity with sum alpha_k = infinity without converging to zero; for example, a_k = 1 on a sparse subsequence whose weighted sum is finite. Additional structure on Delta^x_k and Delta^lambda_k, such as monotonicity or a supplementary argument, is needed. The subsequent statement that 'primal-dual saddle-point properties then imply (x^k, lambda^k) -> (x*, lambda*)' also requires justification when the saddle point is not unique.
[Section 3.3 versus Section 6.2] The algorithm's delay assumption is delta_max(k), Delta_max(k) = o(k^{1/2}), but Lemma 3 only treats the strictly stronger condition O(k^gamma) with gamma < 1/2. A delay such as delta_max(k) = k^{1/2} / log k is o(k^{1/2}) yet is not O(k^gamma) for any gamma < 1/2. Thus the proof as written does not cover the full delay model claimed in the abstract and introduction. Either the delay model must be narrowed or the summability argument must be extended to the stated o(k^{1/2}) class.

minor comments (5)

[Section 3.1 and Section 3.2] The pseudo-code uses a parameter tau for the circular buffer and 'max delay tau', but Section 3.3 later uses delta_j(k) and Delta_ji(k) without relating them to tau; the buffer depth tau is never connected to the unbounded-delay model.
[Abstract and Section 10.6] The abstract and introductory discussion refer to a baseline called 'Asynchronous Distributed Decision-Making (ADDM)', while the experimental section lists 'Alternating Direction Method of Multipliers (ADMM)' as a baseline; this inconsistency makes it unclear which method was actually evaluated.
[Section 10] The simulations use fixed step sizes alpha_i = 0.01, beta = 0.05, bounded staleness tau = 5, and an agent-wise quadratic objective with equality constraint Ax = b, which is a different algorithmic and problem setting from the theoretical model with diminishing step sizes, sublinear delays, and box constraints. The numerical experiments therefore do not directly validate the theorems in the paper.
[Section 7.1] The iteration-complexity constant is written as C = V^0 + G^2 + D^2 + eC, but eC is never defined, despite the paper's claim of explicit constants.
[Section 4, Proof of Lemma 1] The proof absorbs a factor |E| into G^2 and later absorbs U into G, so the constant G changes meaning within the proof; this should be flagged explicitly if the constants are meant to be explicit and tracked.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the main defects are an arithmetic error in Lemma 3 and missing assumptions, not a self-referential derivation.

full rationale

No circular step is present. The paper's convergence and rate claims are derived from a Lyapunov recursion (Lemmas 1 and 2), the Robbins-Siegmund lemma (Lemma 4), and a claimed summability of the error series (Lemma 3). These ingredients are standard external results, and the paper contains no load-bearing self-citations: the reference list includes no prior work by the authors. The ergodic O(K^{-1/2}) rate is not obtained by fitting constants to data or by defining the error term in terms of the conclusion. The serious problems are non-circular: Lemma 3 asserts S_k = O(k^{gamma-1/2}) = o(k^{-1/2}), but for gamma >= 0 the sum S_k has Theta(k^{gamma}) terms of size Theta(k^{-1/2}), so S_k = Theta(k^{gamma-1/2}) and, with alpha_k = 1/sqrt(k+1), sum alpha_k S_k diverges for every gamma >= 0; thus the summability of E_k is not established. Separately, assumptions (F1)-(F3) are invoked but never stated, and the simulation section uses fixed step sizes and bounded delays despite the theory assuming diminishing steps and unbounded sublinear delays. These are correctness and completeness gaps, not cases where a prediction reduces by construction to its inputs. Accordingly, the circularity score is 0.

Assumptions & free parameters 3 free parameters · 5 assumptions · 0 invented entities

The central theory rests on a Lyapunov-Robbins-Siegmund framework with unstated regularity assumptions (F1-F3) and a stepsize/delay interaction whose error terms are not summable under the stated sublinear-delay model. The proof's incorrect asymptotic in Lemma 3 means the reader cannot reconstruct a valid derivation from the manuscript. No new physical or mathematical entities are introduced.

free parameters (3)

step-size schedule alpha_k = beta_k = 1/sqrt(k+1) = 1/sqrt(k+1)
Chosen by hand to balance primal-dual progress against delay-error accumulation. The summability of the delay-error series depends on this exact schedule, and the claimed convergence fails for growing delays under this choice.
simulation fixed step sizes alpha_i = 0.01, beta = 0.05 = 0.01 / 0.05
Selected for the experiments, not derived from the theory. The numerical results in Table 1 may depend on these choices.
simulation max delay tau = 5 and loss rate 10% = 5 iterations / 10% loss
Simulation parameters. The experiment uses bounded delay, which is a different regime from the theoretical unbounded-sublinear-delay model.

assumptions (5)

standard math Sion's minimax theorem and Slater's condition imply strong duality and existence of a saddle point
Invoked in Theorem 1, Section 2.5. This is a standard convex-analysis result and is not a source of circularity.
ad hoc to paper Assumptions (F1)-(F3), referenced in Lemmas 1-2 but never stated
The proof relies on projection nonexpansiveness, convexity, and Lipschitz bounds but the assumptions are never enumerated, so the bounds G and D are not tied to explicit hypotheses.
domain assumption Delays grow sublinearly: delta_max(k), Delta_max(k) = o(k^(1/2))
Section 3.3. This is the core asynchrony model, but as used in Lemma 3 it is too weak to make the delay-error series summable.
ad hoc to paper Crude delay-error bound |lambda*_j - tilde_lambda^k_j| <= G sum_{s=k-delta_j(k)}^{k-1} beta_s (and the analogous primal bound), without residual decay
Step 4 of Lemma 1 and Step 2 of Lemma 2. This bound ignores the decay of the residuals and leads to a divergent error series for growing delays.
domain assumption Bounded gradients and bounded iterates via constants G, D, U, Lambda_max
Used throughout the proofs but not stated as explicit assumptions before appearing in the bounds.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Distributed Asynchronous Primal-Dual Optimization for Supply-Chain Networks." pith.science (2026). https://pith.science/paper/F3WQWVTZ

@misc{pith2026250608024,
  author       = {Pith},
  title        = {Pith review of: Distributed Asynchronous Primal-Dual Optimization for Supply-Chain Networks},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/F3WQWVTZ}},
  note         = {Machine review of arXiv:2506.08024}
}

abstract

Distributed supply-chain optimization demands algorithms that can cope with unreliable communication, unbounded messaging delays, and geographically dispersed agents while still guaranteeing convergence with provable rates. In this work, we introduce DAPD-SCO (Distributed Asynchronous Primal-Dual Optimization for Supply-Chain Networks), a fully asynchronous primal-dual scheme for network flow allocation over directed acyclic supply-chain graphs. Each edge agent independently updates its local flow by projected gradient descent, and each retailer agent independently updates its dual multiplier by projected gradient ascent, using only potentially stale information whose delays can grow sublinearly. Under standard convexity and Slater's conditions and without any global synchronization or bounded-delay assumptions, we prove almost-sure convergence to a saddle point and establish an ergodic duality gap rate of $O(K^{-1/2})$, matching centralized lower bounds. Our analysis uses a Lyapunov-based decomposition that isolates delay-induced errors and handles time-varying communication topologies, bounded noise, and slowly drifting cost or capacity parameters. Extensive simulations on realistic three-tier networks show that DAPD-SCO outperforms synchronous primal-dual methods, Asynchronous Distributed Decision-Making (ADDM), and gradient-push, achieving faster convergence, lower communication overhead, and robust performance under packet loss and high staleness.

Figures

Figures reproduced from arXiv: 2506.08024 by the authors.

**Figure 1.** depicts the overall agent-based architecture of DAPD–SCO on a simple 2-warehouse, 4-retailer supply chain. Each link is handled by an edge-agent that controls flow xij at cost cij , and each retailer Rk maintains a dual price λk [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗

**Figure 2.** Convergence behavior of the algorithm: (a) primal variables and (b) dual [PITH_FULL_IMAGE:figures/full_fig_p013_2.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

21 extracted references · 21 canonical work pages

[1]

Prentice-Hall (1989)

Bertsekas, D., Tsitsiklis, J.: Parallel and Distributed Computation: Numerical Methods. Prentice-Hall (1989)

work page 1989
[2]

IEEE Trans

Tsitsiklis, J., Bertsekas, D.: Distributed asynchronous deterministic and stochastic gradient optimization algorithms. IEEE Trans. Autom. Control 31(9), 803–812 (1986)

work page 1986
[3]

PhD thesis, MIT (1984)

Tsitsiklis, J.: Problems in decentralised decision making and computation . PhD thesis, MIT (1984)

work page 1984
[4]

Tsitsiklis, J.: Asynchronous stochastic approximation and Q-learning. Mach. Learn. 16(3), 185–202 (1994)

work page 1994
[5]

In: ICML (2018)

Lian, X., et al.: Asynchronous decentralized parallel stochastic gradient descent. In: ICML (2018)

work page 2018
[6]

IEEE Trans

Nedi´ c, A., et al.: Network topology and communication complexity of decentralized optimization. IEEE Trans. Autom. Control 65(9), 3785–3800 (2020)

work page 2020
[7]

Cambridge Univ

Boyd, S., Vandenberghe, L.: Convex Optimization . Cambridge Univ. Press (2004)

work page 2004
[8]

Chambolle, A., Pock, T.: A first-order primal-dual algorithm for convex problems with applications to imaging. J. Math. Imaging Vis. 40(1), 120–145 (2011)

work page 2011

Show all 21 references

[9]

Nemirovski, A., et al.: Robust stochastic approximation approach to stochastic programming. SIAM J. Optim. 19(4), 1574–1609 (2009)

2009
[10]

In: NIPS (2014)

Wang, X., Banerjee, A.: Online alternating direction method. In: NIPS (2014)

2014
[11]

Automatica 107, 195–203 (2019)

Xu, J., Zhu, M.: Distributed asynchronous primal–dual algorithm for convex opti- mization over networks. Automatica 107, 195–203 (2019)

2019
[12]

Shi, W., et al.: EXTRA: An exact first-order algorithm for decentralized consensus optimization. SIAM J. Optim. 25(2), 944–966 (2015)

2015
[13]

Automatica 75, 295–304 (2017)

Chang, T., Hong, M.: Multi-agent distributed optimization via double-layer com- munication. Automatica 75, 295–304 (2017)

2017
[14]

IEEE Trans

Nedi´ c, A., ¨Ozdaglar, A.: Distributed subgradient methods for multi-agent opti- mization. IEEE Trans. Autom. Control 54(1), 48–61 (2009)

2009
[15]

Johansson, B., et al.: A randomized incremental subgradient algorithm for dis- tributed optimization in networked systems. SIAM J. Optim. 20(3), 1157–1170 (2009)

2009
[16]

IEEE Trans

Kar, S., Moura, J.: Distributed consensus with imperfect communication. IEEE Trans. Signal Process. 57(1), 355–369 (2009)

2009
[17]

Automatica 46(2), 427–433 (2010)

Zhu, M., Mart ´ ınez, S.: Dynamic average consensus. Automatica 46(2), 427–433 (2010)

2010
[18]

SIAM (2014)

Beck, A.: Introduction to Nonlinear Optimization . SIAM (2014)

2014
[19]

Unpublished manuscript (2018)

Nedi´ c, A., Olshevsky, A., Ozdaglar, A.: Joint-connectivity in time-varying graphs. Unpublished manuscript (2018)

2018
[20]

Pacific J

Sion, M.: On general minimax theorems. Pacific J. Math. 8(1), 171–176 (1958)

1958
[21]

In: Serfling, R

Robbins, H., Siegmund, D.: A convergence theorem for nonnegative almost super- martingales and some applications. In: Serfling, R. J. (ed.) Optimizing Methods in Statistics, pp.233–257. Academic Press (1971)

1971

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.

[1] [1]

Prentice-Hall (1989)

Bertsekas, D., Tsitsiklis, J.: Parallel and Distributed Computation: Numerical Methods. Prentice-Hall (1989)

work page 1989

[2] [2]

IEEE Trans

Tsitsiklis, J., Bertsekas, D.: Distributed asynchronous deterministic and stochastic gradient optimization algorithms. IEEE Trans. Autom. Control 31(9), 803–812 (1986)

work page 1986

[3] [3]

PhD thesis, MIT (1984)

Tsitsiklis, J.: Problems in decentralised decision making and computation . PhD thesis, MIT (1984)

work page 1984

[4] [4]

Tsitsiklis, J.: Asynchronous stochastic approximation and Q-learning. Mach. Learn. 16(3), 185–202 (1994)

work page 1994

[5] [5]

In: ICML (2018)

Lian, X., et al.: Asynchronous decentralized parallel stochastic gradient descent. In: ICML (2018)

work page 2018

[6] [6]

IEEE Trans

Nedi´ c, A., et al.: Network topology and communication complexity of decentralized optimization. IEEE Trans. Autom. Control 65(9), 3785–3800 (2020)

work page 2020

[7] [7]

Cambridge Univ

Boyd, S., Vandenberghe, L.: Convex Optimization . Cambridge Univ. Press (2004)

work page 2004

[8] [8]

Chambolle, A., Pock, T.: A first-order primal-dual algorithm for convex problems with applications to imaging. J. Math. Imaging Vis. 40(1), 120–145 (2011)

work page 2011