AdaSDBO: Adaptive Single-Loop Decentralized BO

• AdaSDBO is a decentralized bilevel optimization algorithm that requires no tuning of problem parameters, enabling adaptive single-loop updates using only neighbor communication.
• It employs hierarchical adaptive stepsizes and scalar stepsize tracking to simultaneously update the upper-level, lower-level, and auxiliary hypergradient variables.
• The method achieves a near-optimal convergence rate of O(1/T) up to polylogarithmic factors, outperforming prior decentralized and double-loop bilevel approaches.

AdaSDBO, short for Adaptive Single-loop Decentralized Bilevel Optimization, is an algorithm for decentralized bilevel optimization in which a network of agents cooperatively solves a bilevel problem using only neighbor-to-neighbor communication. It is introduced in "Problem-Parameter-Free Decentralized Bilevel Optimization" and is characterized as fully problem-parameter-free in the sense that its stepsizes do not require prior knowledge of lower-level strong convexity, smoothness, or graph spectral parameters. The method operates in a single-loop manner, updating the upper-level variable, the lower-level variable, and an auxiliary variable for hypergradient computation simultaneously, rather than nesting inner solves for the lower-level subproblems (Zhai et al., 28 Oct 2025).

1. Problem setting and mathematical formulation

AdaSDBO is designed for decentralized bilevel problems of the form

$$\min_{x\in\mathbb{R}^p} \Phi(x) = f(x,y^*(x)) := \frac1n\sum_{i=1}^n f_i(x,y^*(x)), \qquad \text{s.t. } y^*(x)=\arg\min_{y\in\mathbb{R}^q} l(x,y):=\frac1n\sum_{i=1}^n l_i(x,y).$$

Each agent $i$ knows only its local upper-level loss $f_i$ and lower-level loss $l_i$. The agents communicate over a graph $\mathcal G=(\mathcal N,\mathcal E)$, where $\mathcal N=\{1,\dots,n\}$, using a mixing matrix $W=(w_{ij})\in\mathbb R^{n\times n}$ satisfying

$$W\mathbf 1 = \mathbf 1,\qquad \mathbf 1^\top W=\mathbf 1^\top,\qquad \rho_W:=\|W-\mathbf J\|_2^2<1,$$

with

$\mathbf J = \frac1n \mathbf 1\mathbf 1^\top .$

The paper studies the deterministic nonconvex–strongly-convex regime: the lower-level functions $l_i(x,y)$ are strongly convex in $i$0, whereas the upper-level functions $i$1 may be nonconvex. Under this structure, the exact hypergradient is

$i$2

To avoid explicit inversion of the lower Hessian, the paper introduces the auxiliary linear-system objective

$i$3

whose minimizer approximates

$i$4

Given approximations $i$5 and $i$6, the resulting surrogate hypergradient is

$i$7

This yields three coupled subproblems: the upper-level optimization in $i$8, the lower-level problem defining $i$9, and the auxiliary linear-system problem defining $f_i$0. The coupling among these three objects is central to the design of AdaSDBO (Zhai et al., 28 Oct 2025).

2. Algorithmic structure

AdaSDBO is a single-loop decentralized algorithm. At every iteration, each agent updates its local copies of $f_i$1, $f_i$2, and $f_i$3 simultaneously. The paper emphasizes that a naive adaptive rule applied independently to each variable is ineffective because $f_i$4 depends on sufficiently accurate $f_i$5, $f_i$6 depends on both $f_i$7 and $f_i$8, and decentralized local adaptive stepsizes can damage consensus.

For agent $f_i$9 at iteration $l_i$0, the local quantities are

$l_i$1

$l_i$2

$l_i$3

Each agent maintains scalar accumulators $l_i$4 through cumulative squared gradient norms: $l_i$5

$l_i$6

The defining mechanism is a system of hierarchical adaptive stepsizes. The hierarchy is described as follows: $l_i$7 receives the basic AdaGrad-style rate, $l_i$8 is slowed to respect $l_i$9, and $\mathcal G=(\mathcal N,\mathcal E)$0 is slowed further to respect both $\mathcal G=(\mathcal N,\mathcal E)$1 and $\mathcal G=(\mathcal N,\mathcal E)$2. The control coefficients $\mathcal G=(\mathcal N,\mathcal E)$3 are arbitrary positive constants and are stated not to depend on problem parameters. The paper introduces the shorthand

$\mathcal G=(\mathcal N,\mathcal E)$4

After the local updates, each agent mixes both variables and accumulators: $\mathcal G=(\mathcal N,\mathcal E)$5 The variable $\mathcal G=(\mathcal N,\mathcal E)$6 is then projected onto $\mathcal G=(\mathcal N,\mathcal E)$7: $\mathcal G=(\mathcal N,\mathcal E)$8

A notable implementation point is that the extra communication required for stepsize tracking involves only scalar accumulators, whereas gradient-tracking methods communicate full-dimensional tracker states. The paper also states explicitly that AdaSDBO uses no classical gradient tracking, no momentum, and no Neumann-series Hessian inversion; its hypergradient approximation is based instead on the auxiliary variable $\mathcal G=(\mathcal N,\mathcal E)$9 solving the linear system in single-loop fashion (Zhai et al., 28 Oct 2025).

3. Distinguishing features relative to prior methods

The paper positions AdaSDBO against decentralized bilevel methods including DBO, MDBO, MA-DSBO, and SLDBO, and against centralized single-loop or adaptive references including FSLA and AID. Its main comparative claim is that prior decentralized methods with $\mathcal N=\{1,\dots,n\}$0-type rates require tuned fixed or decaying stepsizes based on quantities such as $\mathcal N=\{1,\dots,n\}$1, $\mathcal N=\{1,\dots,n\}$2, and $\mathcal N=\{1,\dots,n\}$3, whereas AdaSDBO removes that dependence.

Method	Required quantities
DBO	$\mathcal N=\{1,\dots,n\}$4
MDBO	$\mathcal N=\{1,\dots,n\}$5
FSLA	$\mathcal N=\{1,\dots,n\}$6
AID	$\mathcal N=\{1,\dots,n\}$7
SLDBO	$\mathcal N=\{1,\dots,n\}$8
AdaSDBO	none

The expression problem-parameter-free is used in a specific technical sense. It means that the algorithm’s stepsizes do not require tuning based on unknown problem constants such as lower-level strong convexity $\mathcal N=\{1,\dots,n\}$9, smoothness and Lipschitz constants $W=(w_{ij})\in\mathbb R^{n\times n}$0, $W=(w_{ij})\in\mathbb R^{n\times n}$1, $W=(w_{ij})\in\mathbb R^{n\times n}$2, $W=(w_{ij})\in\mathbb R^{n\times n}$3, or the network connectivity quantity $W=(w_{ij})\in\mathbb R^{n\times n}$4. The paper motivates this by noting that such constants are often unavailable, expensive to estimate, privacy-sensitive, or topology-dependent in realistic decentralized systems.

The algorithmic differences emphasized by the paper are fourfold. First, AdaSDBO is simultaneously single-loop and parameter-free. Second, it uses hierarchical stepsizes tailored to the bilevel dependency structure rather than a uniform AdaGrad prescription. Third, it introduces stepsize tracking because local adaptive denominators differ across agents and create a new source of perturbation in the decentralized setting. The tracking recursion is written as

$W=(w_{ij})\in\mathbb R^{n\times n}$5

where $W=(w_{ij})\in\mathbb R^{n\times n}$6 stacks squared accumulators and $W=(w_{ij})\in\mathbb R^{n\times n}$7 stacks squared gradient norms. Fourth, its proof adapts a two-stage AdaGrad-style analytical framework to the coupled decentralized bilevel setting with consensus and stepsize inconsistency errors.

The paper also describes AdaSDBO as the first parameter-free decentralized bilevel method. That claim is the paper’s own positioning statement and should be read as a claim about the literature comparison conducted there (Zhai et al., 28 Oct 2025).

4. Convergence guarantees and analytical mechanism

The main theoretical guarantee is Theorem 1. Under Assumptions 1–3, for any positive constants

$W=(w_{ij})\in\mathbb R^{n\times n}$8

the iterates satisfy a bound on the time-averaged stationarity measure

$W=(w_{ij})\in\mathbb R^{n\times n}$9

of order

$$W\mathbf 1 = \mathbf 1,\qquad \mathbf 1^\top W=\mathbf 1^\top,\qquad \rho_W:=\|W-\mathbf J\|_2^2<1,$$0

This is presented as matching well-tuned state-of-the-art methods up to polylogarithmic factors.

Corollary 1 states that obtaining an $$W\mathbf 1 = \mathbf 1,\qquad \mathbf 1^\top W=\mathbf 1^\top,\qquad \rho_W:=\|W-\mathbf J\|_2^2<1,$$1-stationary point requires

$$W\mathbf 1 = \mathbf 1,\qquad \mathbf 1^\top W=\mathbf 1^\top,\qquad \rho_W:=\|W-\mathbf J\|_2^2<1,$$2

and therefore the gradient complexity per agent is

$$W\mathbf 1 = \mathbf 1,\qquad \mathbf 1^\top W=\mathbf 1^\top,\qquad \rho_W:=\|W-\mathbf J\|_2^2<1,$$3

The theorem is supported by several intermediate controls. The lower-level and auxiliary residuals admit logarithmic cumulative bounds, which regulate how well $$W\mathbf 1 = \mathbf 1,\qquad \mathbf 1^\top W=\mathbf 1^\top,\qquad \rho_W:=\|W-\mathbf J\|_2^2<1,$$4 approximate the exact lower-level and linear-system solutions. The averaged accumulators satisfy logarithmic growth bounds such as

$$W\mathbf 1 = \mathbf 1,\qquad \mathbf 1^\top W=\mathbf 1^\top,\qquad \rho_W:=\|W-\mathbf J\|_2^2<1,$$5

which prevent adaptive denominators from growing too quickly. Consensus error, defined through

$$W\mathbf 1 = \mathbf 1,\qquad \mathbf 1^\top W=\mathbf 1^\top,\qquad \rho_W:=\|W-\mathbf J\|_2^2<1,$$6

satisfies

$$W\mathbf 1 = \mathbf 1,\qquad \mathbf 1^\top W=\mathbf 1^\top,\qquad \rho_W:=\|W-\mathbf J\|_2^2<1,$$7

The stepsize inconsistency terms induced by local adaptive denominators are also shown to be of order $$W\mathbf 1 = \mathbf 1,\qquad \mathbf 1^\top W=\mathbf 1^\top,\qquad \rho_W:=\|W-\mathbf J\|_2^2<1,$$8 on average.

The proof strategy begins with descent on the averaged upper objective $$W\mathbf 1 = \mathbf 1,\qquad \mathbf 1^\top W=\mathbf 1^\top,\qquad \rho_W:=\|W-\mathbf J\|_2^2<1,$$9, then decomposes the update error into upper-level stationarity, lower-level and auxiliary approximation errors, consensus errors, and stepsize discrepancy perturbations. Strong convexity of the lower level is used to convert tracking errors in $\mathbf J = \frac1n \mathbf 1\mathbf 1^\top .$0 and $\mathbf J = \frac1n \mathbf 1\mathbf 1^\top .$1 into residual bounds involving $\mathbf J = \frac1n \mathbf 1\mathbf 1^\top .$2 and $\mathbf J = \frac1n \mathbf 1\mathbf 1^\top .$3. A two-stage adaptive analysis, analogous to AdaGrad-Norm arguments, introduces threshold constants and stopping indices so that after sufficiently large accumulators are reached, the adaptive denominators dominate curvature-dependent terms. The logarithmic overhead ultimately originates from telescoping and log-type bounds for sums of squared gradients divided by adaptive denominators.

A central point in the theory is that quantities such as $\mathbf J = \frac1n \mathbf 1\mathbf 1^\top .$4, $\mathbf J = \frac1n \mathbf 1\mathbf 1^\top .$5, smoothness constants, $\mathbf J = \frac1n \mathbf 1\mathbf 1^\top .$6, and bounded derivative constants enter the analysis, but not the algorithmic tuning. This distinction is essential to the paper’s use of the term problem-parameter-free (Zhai et al., 28 Oct 2025).

5. Assumptions, computational profile, and empirical behavior

The assumptions are standard for decentralized consensus analysis and implicit-differentiation bilevel theory, but they are also restrictive. The communication matrix must satisfy

$\mathbf J = \frac1n \mathbf 1\mathbf 1^\top .$7

For each agent $\mathbf J = \frac1n \mathbf 1\mathbf 1^\top .$8, $\mathbf J = \frac1n \mathbf 1\mathbf 1^\top .$9 and $l_i(x,y)$0 are assumed twice continuously differentiable; $l_i(x,y)$1 is $l_i(x,y)$2-Lipschitz continuous; $l_i(x,y)$3 and $l_i(x,y)$4 are Lipschitz with constants $l_i(x,y)$5 and $l_i(x,y)$6; and $l_i(x,y)$7 and $l_i(x,y)$8 are $l_i(x,y)$9-Lipschitz. The lower-level function $i$00 is $i$01-strongly convex with respect to $i$02. The paper also remarks that these assumptions imply bounded derivative quantities such as $i$03, $i$04, $i$05, and $i$06. Initialization requires positive accumulator values, and the theorem allows any positive initialization.

Per iteration, each agent computes $i$07, $i$08, $i$09, the Hessian-vector product $i$10, and the mixed Jacobian-vector term $i$11. Thus, the method avoids explicit matrix inversion but still requires second-order derivative operations or Hessian-vector products. Communication per iteration consists of $i$12, $i$13, $i$14, and three scalar accumulators. Memory consists of $i$15, $i$16, $i$17, and the three scalar accumulators.

The experiments cover decentralized bilevel hyperparameter optimization and decentralized meta-learning. Baselines include SLDBO, MA-DSBO, MDBO, and DBO. In synthetic hyperparameter optimization, the local objectives are

$i$18

with $i$19, and heterogeneity is controlled by sampling $i$20. The reported outcomes are that AdaSDBO converges faster than baselines for $i$21 and $i$22, consistently outperforms the double-loop methods DBO and MA-DSBO, and remains strong under higher heterogeneity ($i$23).

For decentralized hyperparameter optimization on MNIST and Fashion-MNIST, the paper uses cross-entropy-based objectives,

$i$24

$i$25

and reports that AdaSDBO maintains competitive convergence and robust performance as network size changes. The stepsize-robustness experiment varies stepsizes from

$i$26

for 1000 rounds; AdaSDBO is reported to maintain stable test accuracy over a much wider range than baseline methods.

Additional experiments vary $i$27 and compare ring, ladder, and random topologies. Stronger connectivity improves all methods, but AdaSDBO remains best or among the best across topologies. For $i$28, the reported MNIST accuracies are $i$29 for ring, $i$30 for ladder, and $i$31 for random topology; corresponding Fashion-MNIST accuracies are $i$32, $i$33, and $i$34. In decentralized meta-learning on CIFAR-10 under a decentralized MAML-style setup, the paper reports train/test accuracy pairs of $i$35 for $i$36, $i$37 for $i$38, and $i$39 for $i$40, all compared favorably against SLDBO (Zhai et al., 28 Oct 2025).

6. Interpretation, limitations, and open directions

The principal significance attributed to AdaSDBO is that it combines three features that are usually difficult to obtain simultaneously in decentralized bilevel optimization: single-loop updates, no tuning based on $i$41, $i$42, or $i$43, and a convergence rate of

$i$44

up to a $i$45 factor. A plausible implication is that the method is particularly attractive when the communication topology is known only operationally, when estimating graph or curvature constants is impractical, or when hyperparameter tuning itself is a major engineering burden.

At the same time, the paper explicitly restricts its formal theory to the deterministic full-gradient setting and requires the lower level to be strongly convex in $i$46. These conditions are standard within a large part of implicit-differentiation bilevel analysis, but they limit direct applicability to broader stochastic or merely convex lower-level settings. The method also uses second-order objects in the form of Hessian-vector and Jacobian-vector products, so it is not a purely first-order decentralized optimizer.

A common misunderstanding would be to equate problem-parameter-free with the absence of all algorithmic choices. The paper does not make that stronger claim. It still uses positive coefficients $i$47 and positive initial accumulators $i$48. The theorem states that any positive values work for convergence, but the paper also notes that practical speed may still vary. Another possible misconception is that AdaSDBO is a gradient-tracking method in the standard decentralized-optimization sense; the paper states explicitly that it is not. Its distinctive mechanism is instead the combination of hierarchical adaptive rates with scalar stepsize-tracking communication.

The paper identifies likely use cases including decentralized hyperparameter optimization, collaborative learning, and decentralized meta-learning. It also points to two natural future directions: extension to stochastic settings and to lower-level problems that are convex rather than strongly convex. These directions are consistent with the main structural limitations of the current theory and provide the most immediate agenda for follow-up work (Zhai et al., 28 Oct 2025).