Many distributed systems coordinate self-interested parties by running a splitting algorithm. Examples include data owners jointly fitting a model, and firms sharing a capacity-constrained network. A splitting algorithm breaks one big optimization into small local problems tied together by a shared "consensus" variable. The fastest versions, adaptive ADMM (the alternating direction method of multipliers), let each participant's coupling strength, called the penalty parameter, tune itself from that participant's own recent updates. This note makes one point about that design. In the common consensus form, a node's penalty is also its voting weight in the shared variable, so a self-tuned penalty amounts to a self-reported voting weight. A participant who distorts the updates it reports can inflate its own weight. When the algorithm is stopped early, as it always is in practice, that weight pulls the result toward the participant's preferred outcome.
We propose to measure this lever, separate it from the ordinary market power an agent already has, and close it with a simple rule. The rule borrows the oldest idea in truthful mechanism design, that an agent's report should not set its own price. Here it becomes: an agent's report should not set its own penalty. The question is whether this "report-insulated" adaptation keeps most of the speed of self-tuned adaptation while removing the extra gain from manipulation.
Core Claim
Fix the algorithm to be consensus ADMM. Each node $i$ holds a local copy $x_i$ and a scaled dual variable $u_i$. The shared variable is updated, in the standard consensus form with no regularizer on it, as a penalty-weighted average:
$$z^{k+1} = \frac{\sum_i \tau_i^k \,(x_i^{k+1} + u_i^k)}{\sum_i \tau_i^k},$$
where $\tau_i^k$ is node $i$'s penalty at round $k$. In Adaptive Consensus ADMM, $\tau_i^k = 1/\sqrt{\alpha_i\beta_i}$. Here $\alpha_i$ and $\beta_i$ are Barzilai–Borwein curvature estimates, that is, ratios of how much a node's gradient changed to how much its variable changed between rounds. Both are computed entirely from node $i$'s own reported $\Delta x_i$ and $\Delta u_i$.
The fixed point of consensus ADMM does not depend on the penalties. So if the algorithm is run to exact convergence, this channel is harmless. At a finite stopping round $K$, however, the weights shape $z^K$. A node that reports updates suggesting flat local curvature receives a large $\tau_i$ and drags the stopped solution toward its own optimum.
Our claim has three parts. First, the extra gain $\epsilon_{\text{adapt}}(K)$ that a best-responding agent gets from this channel is measurable. It can be separated from its static market-power gain $\epsilon_{\text{static}}(K)$, which an agent has even under fixed penalties. We expect $\epsilon_{\text{adapt}}(K)$ to be a non-trivial share of total manipulation gain at the communication budgets where adaptive ADMM is actually useful. Second, this gain disappears if $\tau_i$ is computed only from other agents' reports. We call this the leave-one-out rule. Its convergence still follows from the bounded-adaptivity argument of Adaptive Consensus ADMM, because that argument constrains how fast penalties may change, not whether they equal the node's own curvature. Third, the price of insulation is an empirical speed question, and that is what the experiments settle.
Relation to Prior Work
Two lines of work motivate the gap without filling it. Faithful Implementations of Distributed Algorithms and Control Laws shows that indirect Groves mechanisms make strategic agents execute distributed dual-subgradient updates faithfully. Its guarantee is asymptotic: the gain from cheating vanishes as iterations go to infinity, with static step sizes. Mechanism Design for Large Scale Network Utility Maximization layers price-based payments on top of existing solvers and shows the incentive to misreport shrinks as $O(1/I^2)$ in the number of agents $I$. Neither work has an augmented-Lagrangian algorithm, and neither has parameters that move during the run. The parameter channel we study therefore does not exist in their models.
Three papers supply the methods we reuse. Adaptive ADMM with Spectral Penalty Parameter Selection gives the spectral rule and its correlation safeguard. Adaptive Consensus ADMM gives the per-node version and its $O(1/k)$ guarantee under bounded adaptivity. Decentralized Resource Allocation via Dual Consensus ADMM gives a resource-allocation testbed. It proves convergence for any fixed penalty and explicitly leaves penalty tuning open. All three assume obedient nodes, and the incentive analysis is what we add.
The design principle comes from the decoupling result behind A New Strategy-Proof Greedy-Allocation Combinatorial Auction Protocol. That result keeps strategy-proofness, even under approximate allocation, by setting each agent's price independently of its own report. We apply the same move to an algorithmic parameter rather than a price.
Procurement Auctions with Predictions is the closest example of a mechanism with a tunable parameter. Its parameters are static predictions fixed before the auction, whereas ours evolve with the reports themselves.
Research Program
The work has four pieces.
The first piece is a formal model. Strategic agents interact with consensus ADMM in two settings, each with its own utility. In federated consensus estimation, agents have no payments and each agent's utility is the negative of its own-data loss at the stopped model $z^K$. In capacity-constrained resource allocation, run through dual consensus ADMM, each agent's utility is its private valuation minus a payment. That payment follows either a price-based rule (shadow price times allocation, as in Zhang et al.) or a Groves/Clarke rule (as in Huang et al.). In both settings agents act by reporting their per-round updates.
The second piece is an attack procedure that estimates best-response gains. An attacker controls one node's reports. We optimize its distortion either by gradient ascent through the unrolled $K$-round algorithm or by black-box search over low-dimensional distortion families, such as rescaling $\Delta x_i$ or adding drift. To separate the adaptive channel from the static one, the attacker is run twice. In the first run its reports feed into everything, which gives the total gain. In the second run the attacker's distortion is applied to the shared variable while its penalty is frozen at its truthful value, which gives $\epsilon_{\text{static}}$. The difference between the two is $\epsilon_{\text{adapt}}$.
The third piece is the insulated rule and two cheaper variants. In the leave-one-out rule, $\tau_i$ is the geometric mean of the other nodes' spectral estimates. The global-aggregate rule is a single AADMM-style penalty computed from pooled quantities, so each agent's influence on it scales as $1/I$. The clipped rule is self-tuned ACADMM with a narrow band on $\tau_i$. All variants keep bounded-adaptivity safeguards, so the existing convergence proof applies.
The fourth piece is analysis. We will try to bound $\epsilon_{\text{adapt}}(K)$ by the sensitivity of $z^K$ to the weight vector, multiplied by the attainable weight distortion. The aim is to show that this quantity is zero under leave-one-out and $O(1/I)$ under the global rule. A related question is whether, as asked in Gap 3 of the file, the gain is monotone in $K$ or shows transient windows of large arbitrage.
Baselines
The naive reference point is fixed-penalty consensus ADMM with a well-chosen constant, combined with each payment rule. It has no adaptive channel, so its manipulation gain is pure $\epsilon_{\text{static}}$. It is also the honest extension of the static-step approach of Huang et al. to ADMM. Its weakness is speed. Adaptive ADMM reports that vanilla ADMM failed to converge on 13 of 23 realistic datasets where adaptive ADMM converged on all. To be fair to this baseline, its constant gets a grid search with the same tuning budget the adaptive methods get for their safeguards.
The strongest existing method is self-tuned Adaptive Consensus ADMM, because it is the speed frontier. On News20 Elastic Net it reports convergence in 78 iterations, while standard ADMM did not converge within 1,000. We run it with the published safeguards on the same data partitions and with the same stopping budgets $K$. A proposal "beats" it if the insulated rule reduces total manipulation gain to within noise of the fixed-penalty baseline while needing at most roughly 1.5 times ACADMM's iterations to tolerance. If the slowdown is larger, insulation is a bad trade, and we will report it as such.
The ablations isolate the key ingredient. The clipped self-tuned rule tests whether merely limiting how far the weight can move is enough, without removing the report from the weight's computation. The global-aggregate rule tests whether diluting influence is as good as removing it. Residual balancing is included as a commonly used heuristic. Adaptive ADMM reports it to be sensitive to rescaling, and the research file argues that rescaling is exactly the cheapest manipulation available to an agent.
In resource allocation we also include the Clarke-pivot Groves mechanism. It serves as the incentive gold standard against which price-based payments are judged.
Why not just…?
Why not just use a Groves tax and keep ACADMM unchanged? Under a Groves tax, each agent's utility equals reported social welfare. A strategic node then wants the stopped solution to be good for everyone, and the penalty channel becomes harmless or even benign. This is a real alternative, and we include it as an arm rather than argue it away. We predict that self-tuned and insulated rules show nearly equal gains under Groves. The cost is elsewhere. The Clarke pivot needs one counterfactual run per agent, the mechanism can run large deficits, and the federated setting has no payments at all. Insulation therefore matters for price-based and payment-free deployments. If experiments showed Groves cheap enough at our scales, the contribution would narrow to those deployments, and we would state that.
Why not just run until convergence? At the exact fixed point the weights are irrelevant, so the channel closes. The reason adaptive penalties exist is to stop early under communication limits. We therefore measure $\epsilon_{\text{adapt}}(K)$ across a sweep of $K$ from small budgets up to tight tolerance. If the gain is negligible already at the budgets where ACADMM's speedup is realized, the problem is not practically important, and we will say so.
Why not just clip the penalties to a narrow band? Clipping bounds the attainable weight distortion. The same bound also limits the curvature adaptation that produces ACADMM's speedup, so clipping trades speed for safety along a single dial. The clipped arm is swept over band widths, which traces its speed–manipulation curve. Leave-one-out is only worth proposing if it lies strictly above that curve.
Why not just use fixed-penalty dual decomposition as in Huang et al.? It has no adaptive channel. However, the evidence above from Adaptive ADMM and Decentralized Resource Allocation via Dual Consensus ADMM shows that fixed-parameter first-order methods are slow or fail. Slow convergence also weakens its own asymptotic incentive guarantee at any realistic budget. The fixed-penalty baseline measures this directly.
Proposed Experiments
For federated consensus estimation, we use the distributed Elastic Net and logistic regression tasks on News20, RCV1, and Realsim from Adaptive Consensus ADMM. Data is partitioned across 8 to 64 owners in two ways: a uniform split, and a non-uniform split by label or topic so that owners genuinely prefer different models. One owner, and then a coalition, runs the attack procedure. We report three quantities against $K$: total gain, $\epsilon_{\text{static}}$, and $\epsilon_{\text{adapt}}$. Gain is measured as the attacker's own-loss reduction relative to truthful play. We also report the welfare loss imposed on others and the iterations to a fixed tolerance under truthful play.
For resource allocation, we use the synthetic network setup of Decentralized Resource Allocation via Dual Consensus ADMM with randomly drawn concave valuations, varying the number of agents $I$. Payments follow the price-based rule and the Clarke rule. Here we also check the $O(1/I^2)$ decay of Zhang et al.. The file has no reported numeric result on Lacuna for this decay, nor for the simulations in Huang et al., so we will reproduce both rather than compare against reported values. The test is whether $\epsilon_{\text{adapt}}$ decays at the same rate under self-tuned penalties or more slowly.
The decisive check has two parts. The first is whether, on the non-uniform News20 split at budgets where ACADMM reaches tolerance and fixed ADMM does not, self-tuned ACADMM shows a statistically clear $\epsilon_{\text{adapt}} > 0$. The second is whether leave-one-out drives that component to zero within about 1.5 times ACADMM's iteration count. If the first fails, the lever is theoretical and the project should stop. If the second fails because insulation is too slow, the result is still a useful negative finding: speed from self-tuned coupling cannot be separated from manipulability. In that case we would turn the analysis toward characterizing that trade-off precisely.
References
- Adaptive Consensus ADMM for Distributed Optimization
- Faithful Implementations of Distributed Algorithms and Control Laws
- Mechanism Design for Large Scale Network Utility Maximization
- Adaptive ADMM with Spectral Penalty Parameter Selection
- Decentralized Resource Allocation via Dual Consensus ADMM
- A New Strategy-Proof Greedy-Allocation Combinatorial Auction Protocol and Its Extension to Open Ascending Auction Protocol
- Procurement Auctions with Predictions: Improved Frugality for Facility Location