Utility-Optimized Block Design (uBD)

• Utility-optimized block design (uBD) is a strategy where block creation is driven by explicit, domain-specific utility criteria, such as privacy trade-off, latency, or estimation risk.
• uBD methodologies span areas like local differential privacy, distributed quantum compilation, and statistical experimental design, using techniques such as beam search and dual decomposition.
• The framework addresses trade-offs between optimality and computational cost, offering both exact and near-optimal solutions to improve efficiency, robustness, and performance.

Searching arXiv for the cited papers to ground the article and confirm metadata. Utility-Optimized Block Design (uBD) denotes a block-centered design philosophy in which the atomic design object is a block and the choice, weighting, scheduling, or decomposition of blocks is driven by an explicit utility criterion. In the narrowest and most explicit sense, the term appears in discrete distribution estimation under utility-optimized local differential privacy (ULDP), where a “utility-optimized block design (uBD)” scheme is proved to achieve the exact asymptotic minimax privacy-utility trade-off (Yoon et al., 29 Sep 2025). In a broader and explicitly stated interpretive sense, the same label or closely related language is used to organize work on distributed-quantum scheduling, local differential privacy mechanisms, blocked experimental design, network experiments, and workload-aware block-size tuning, where “utility” may mean asymptotic estimation risk, information volume, teleportation count, latency, aberration, Price of Anarchy, or runtime (Yun et al., 20 May 2026).

1. Scope and conceptual structure

The most stable feature across the literature is that a uBD problem selects or constructs blocks under a utility objective rather than treating blocking as a purely organizational convenience. In some papers the block is a combinatorial subset, in others a group of operations, a treatment composition, a bipartite subgraph, a resource-block count, or a contiguous algorithmic segment. The associated utility is likewise domain-specific: privacy-utility trade-off in ULDP and LDP (Yoon et al., 29 Sep 2025, Park et al., 2023); future teleportation savings and latency reduction in distributed quantum computers (Yun et al., 20 May 2026); pseudo-Bayesian $D_A$- or $D_s$-information in blocked experiments with dependent non-normal responses (Mueller et al., 2018); and runtime minimization in square-root decomposition (Zhao, 4 Jun 2026).

Setting	Block object	Utility notion
ULDP distribution estimation	block designs on the sensitive subset	exact asymptotic minimax privacy-utility trade-off
Distributed quantum computing	groups of CNOTs compiled together	teleportation count and latency
Copula-based blocked experiments	block treatment compositions $\zeta_i$	robust $D_A$- or $D_s$-optimality
Resource allocation in cellular systems	integer-valued resource blocks	product of user utilities
Square-root decomposition	block size $B$	runtime, regret, speedup

Several papers explicitly state that they do not use the phrase “utility-optimized block design,” while also describing their contributions as fitting naturally into a broader uBD perspective (Mueller et al., 2018, Chisaki et al., 2023, Singh et al., 29 Jan 2025). This suggests that uBD is not a single formalism shared across fields, but a recurring design pattern in which block structure is chosen or exploited according to a stated performance functional.

2. Privacy-preserving estimation and the explicit uBD formalism

The paper that explicitly defines uBD studies discrete distribution estimation under ULDP, where the domain is partitioned into a sensitive set $\mathcal{X}_{\mathrm S}=[v]$ and a non-sensitive set $$\mathcal{X}_{\mathrm N}=\mathcal{X}\setminus\mathcal{X}_{\mathrm S}$$ (Yoon et al., 29 Sep 2025). A mechanism is $(\mathcal{X}_{\mathrm S},\epsilon)$-ULDP if its output alphabet is partitioned into protected outputs $\mathcal{Y}_{\mathrm P}$, which satisfy the LDP-type inequality

$D_s$0

and invertible outputs $D_s$1, each of which reveals a unique non-sensitive input. The uBD construction is an extremal ULDP mechanism built from block designs on the sensitive subset, together with invertible outputs for non-sensitive symbols. Its design parameters are a mixture $D_s$2 over block sizes $D_s$3, and the exact asymptotic privacy-utility trade-off is characterized by

$D_s$4

The associated estimator is a score-based linear estimator, and optimized uBD attains the converse exactly. A simple uBD scheme corresponds to $D_s$5; the $D_s$6 case is exactly uRR. The paper also isolates closed-form regimes in which uRR is asymptotically optimal and regimes in which a simple uBD based on the LDP-optimal block size on the sensitive subset is optimal (Yoon et al., 29 Sep 2025).

This explicit ULDP uBD sits on top of an earlier LDP block-design program. In “Exactly Optimal and Communication-Efficient Private Estimation via Block Designs” the mechanism is induced by a $D_s$7-block design with channel

$D_s$8

and the paper proves that, for block-design schemes, the asymptotic risk depends only on $D_s$9, not directly on $\zeta_i$0 (Park et al., 2023). Exact optimality holds when $\zeta_i$1, while symmetric block designs with $\zeta_i$2 achieve the minimum communication cost among unbiased or consistent schemes. Because exact block-design existence is sparse, the same paper introduces regular and pairwise-balanced designs (RPBDs), which relax uniformity and yield near-optimal privacy-utility trade-offs over a much larger set of $\zeta_i$3 pairs (Park et al., 2023).

A further refinement appears in the shared-randomness setting. “Achieving the Exactly Optimal Privacy-Utility Trade-Off with Low Communication Cost via Shared Randomness” shows that a resolution of a PUT-optimal block-design mechanism preserves the exact same risk while reducing communication cost (Nam et al., 2023). The Baranyai resolution achieves communication cost

$\zeta_i$4

which the paper proves is the minimum among PUT-optimal resolutions of block design schemes, while the cyclic shift resolution offers an explicit alternative with communication cost at most $\zeta_i$5 bits (Nam et al., 2023). In uBD terms, the privacy utility is fixed and communication is optimized by decomposing the block design rather than changing its estimation geometry.

3. Utility-optimized block design in distributed quantum compilation

In distributed quantum computing (DQC), ATHENA uses the phrase “utility-optimized block design” for a compiler architecture built around blocks of CNOTs that are optimized collectively because one relocation can make several non-local gates local (Yun et al., 20 May 2026). Prior compilers such as AutoComm and QuComm already use blocks, but ATHENA argues that naïve block-level scheduling is too myopic because teleportations change future qubit locations. The paper quantifies this using a $\zeta_i$6 DQC study in which consecutive RELOCATEs involving a qubit are separated on average by $\zeta_i$7–$\zeta_i$8 blocks, and in QAOA-FC the average separation is $\zeta_i$9 blocks, of which $D_A$0 are local-only (Yun et al., 20 May 2026).

ATHENA’s first core mechanism is Utility-driven Lookahead with Multi-Candidate Block Scheduling (UMS). A future block has utility for the current block $D_A$1 iff it shares at least one qubit with $D_A$2; these are “overlapping qubits.” Rather than taking the next $D_A$3 blocks, ATHENA scans forward and includes only useful future blocks until it reaches a fixed lookahead size $D_A$4, default $D_A$5, or the program ends. It then scores candidate scheduling decisions with a future-cost estimate

$D_A$6

where $D_A$7 is block distance and $D_A$8 by default, so a block $D_A$9 blocks ahead contributes weight $D_s$0. UMS keeps multiple candidate schedules alive in a solution tree and prunes to Top-$D_s$1 candidates with beam width $D_s$2. Commitment is deferred until the full program is scheduled, and the compiler returns the schedule with the fewest teleportations (Yun et al., 20 May 2026).

The second mechanism is EPR-Capacity-Aware Early Scheduling (EES), which moves otherwise-ready RELOCATEs and future CNOTs earlier when dependencies allow and EPR capacity is available. The paper reports that on a $D_s$3 DQC with 240-qubit programs, $D_s$4–$D_s$5 of teleportations are delayed in prior scheduling styles, with average waiting times of $D_s$6–$D_s$7 ms depending on benchmark. EES does not change teleportation count, but it changes when those operations happen, increasing concurrency (Yun et al., 20 May 2026).

The utility variables are explicit. ATHENA measures effective teleportation count

$D_s$8

with $D_s$9, and schedule latency $B$0. Under the neutral-atom timing model, RELOCATE is $B$1 ms and Re-CNOT is $B$2 ms; non-local CNOTs are $B$3–$B$4 slower and roughly $B$5 more error-prone than local CNOTs (Yun et al., 20 May 2026). Across benchmarks and DQC sizes, ATHENA reduces teleportations by $B$6 on average and up to $B$7, and reduces latency by $B$8 on average and up to $B$9 compared to the state of the art. The paper is also explicit that this search is more expensive than QuComm’s, with compilation $\mathcal{X}_{\mathrm S}=[v]$0 to $\mathcal{X}_{\mathrm S}=[v]$1 slower in some scaling experiments (Yun et al., 20 May 2026).

4. Statistical experimental design and causal inference

In statistical design, uBD appears mainly as information-based or precision-based block optimization. “Copula-based robust optimal block designs” considers blocked experiments with dependent non-normal responses and treats utility as expected inferential value under uncertainty about marginal and dependence parameters (Mueller et al., 2018). The design variable is a block treatment composition

$\mathcal{X}_{\mathrm S}=[v]$2

and the objective is the robust $\mathcal{X}_{\mathrm S}=[v]$3-criterion

$\mathcal{X}_{\mathrm S}=[v]$4

or the corresponding $\mathcal{X}_{\mathrm S}=[v]$5-criterion for a parameter subset. The method combines a marginal GLM-type model with a copula for within-block dependence, uses approximate designs over weighted support blocks, and computes designs numerically with a version of the Fedorov–Wynn algorithm implemented in docopulae (Mueller et al., 2018). The paper emphasizes that copula-based dependence changes the geometry of the optimal block design and that designs obtained under independence or working-correlation assumptions can lose efficiency.

“E- and R-optimality of block designs for treatment-control comparisons” gives an exact structural characterization of one important precision-based utility (Rosa, 2018). For approximate designs, $\mathcal{X}_{\mathrm S}=[v]$6-optimality holds iff

$\mathcal{X}_{\mathrm S}=[v]$7

Thus every block assigns half of its mass to control and the test treatments are equireplicated. The paper also shows that $\mathcal{X}_{\mathrm S}=[v]$8-optimality is equivalent to optimality for the contrast $\mathcal{X}_{\mathrm S}=[v]$9 and to minimizing the sum of variances plus absolute covariances of the treatment-control estimators. In the same model, all approximate $$\mathcal{X}_{\mathrm N}=\mathcal{X}\setminus\mathcal{X}_{\mathrm S}$$0-optimal designs and a large class of $$\mathcal{X}_{\mathrm N}=\mathcal{X}\setminus\mathcal{X}_{\mathrm S}$$1-optimal exact designs are also $$\mathcal{X}_{\mathrm N}=\mathcal{X}\setminus\mathcal{X}_{\mathrm S}$$2-optimal (Rosa, 2018). Here utility is not a learned score but a precisely defined variance criterion.

“Design-based Causal Inference for Incomplete Block Designs” supplies another set of objective ingredients rather than a direct optimizer (Koo et al., 2024). In a general incomplete block design (IBD), each block receives exactly $$\mathcal{X}_{\mathrm N}=\mathcal{X}\setminus\mathcal{X}_{\mathrm S}$$3 treatments with $$\mathcal{X}_{\mathrm N}=\mathcal{X}\setminus\mathcal{X}_{\mathrm S}$$4, and the paper derives unbiased estimators and conservative variance estimators in a finite-population design-based framework. For balanced incomplete block designs (BIBDs), it proves a clean variance interpolation: $$\mathcal{X}_{\mathrm N}=\mathcal{X}\setminus\mathcal{X}_{\mathrm S}$$5 This shows that a BIBD sits between cluster randomization and complete block designs, and the paper further notes that a deliberately unbalanced IBD may outperform a symmetric BIBD for a focal contrast because it allocates more blocks to the relevant pair (Koo et al., 2024).

Blocked order-of-addition experiments extend the same logic to robustness against aliasing and confounding. “Efficient and Robust Block Designs for Order-of-Addition Experiments” proposes the composite word length pattern

$$\mathcal{X}_{\mathrm N}=\mathcal{X}\setminus\mathcal{X}_{\mathrm S}$$6

and ranks designs lexicographically by minimum aberration (Lin, 8 Jun 2025). Here $$\mathcal{X}_{\mathrm N}=\mathcal{X}\setminus\mathcal{X}_{\mathrm S}$$7 measures aliasing between the mean and $$\mathcal{X}_{\mathrm N}=\mathcal{X}\setminus\mathcal{X}_{\mathrm S}$$8th-order position effects, while $$\mathcal{X}_{\mathrm N}=\mathcal{X}\setminus\mathcal{X}_{\mathrm S}$$9 measures confounding between block effects and $(\mathcal{X}_{\mathrm S},\epsilon)$0th-order position effects. The paper then develops constructions based on orthogonal Latin squares and component orthogonal arrays (COAs), together with exchange steps over Latin squares or rows of Latin squares. Its simulations report power and type I error rates comparable to those of full block OofA designs, despite much smaller blocked designs (Lin, 8 Jun 2025).

5. Combinatorial, geometric, and bipartite block constructions

Some uBD-relevant work is purely combinatorial but still utility-driven. “Optimality and Constructions of Spanning Bipartite Block Designs” studies edge-effect estimation on a complete bipartite graph $(\mathcal{X}_{\mathrm S},\epsilon)$1 and defines a Spanning Bipartite Block Design (SBBD) by replication and concurrence conditions on edges (Chisaki et al., 2023). The key statistical result is that SBBD$(\mathcal{X}_{\mathrm S},\epsilon)$2 is variance balanced whenever all basic contrasts are estimable, and that a semi-regular SBBD is $(\mathcal{X}_{\mathrm S},\epsilon)$3-optimum relative to the class $(\mathcal{X}_{\mathrm S},\epsilon)$4. The construction uses $(\mathcal{X}_{\mathrm S},\epsilon)$5-designs and ordered designs, and a BIBD with prime power number of blocks yields an $(\mathcal{X}_{\mathrm S},\epsilon)$6-optimum semi-regular or regular SBBD (Chisaki et al., 2023). In this setting, utility is average precision of estimable edge contrasts.

“Optimal BIBD-extended designs” studies what happens when one starts from a binary design $(\mathcal{X}_{\mathrm S},\epsilon)$7 and repeatedly adds the blocks of a BIBD $(\mathcal{X}_{\mathrm S},\epsilon)$8 (Cakiroglu et al., 2019). If the base design has non-trivial Laplacian eigenvalues $(\mathcal{X}_{\mathrm S},\epsilon)$9, then after adding $\mathcal{Y}_{\mathrm P}$0 copies of the BIBD the non-trivial eigenvalues become

$\mathcal{Y}_{\mathrm P}$1

The paper derives exact formulas for the $\mathcal{Y}_{\mathrm P}$2- and $\mathcal{Y}_{\mathrm P}$3-values of the extended design and shows that, after sufficient extension, the stable order of base designs determines both $\mathcal{Y}_{\mathrm P}$4- and $\mathcal{Y}_{\mathrm P}$5-optimality. In the graph case $\mathcal{Y}_{\mathrm P}$6, it proves that the path becomes both $\mathcal{Y}_{\mathrm P}$7- and $\mathcal{Y}_{\mathrm P}$8-optimal after extension when $\mathcal{Y}_{\mathrm P}$9, and the cycle becomes both $D_s$00- and $D_s$01-optimal after extension when $D_s$02 (Cakiroglu et al., 2019). Balanced augmentation therefore acts as a spectral regularizer that can collapse criterion disagreement.

The geometric version of the same idea appears in “Maximal Orthoplectic Fusion Frames from Mutually Unbiased Bases and Block Designs” (Bodmann et al., 2016). There, a block $D_s$03 of size $D_s$04 defines a coordinate projection

$D_s$05

and inside one basis the overlap is exactly

$D_s$06

Between different mutually unbiased bases the overlap becomes $D_s$07. The paper uses cohesive block designs to control within-basis intersections, shows that the orthoplex bound for fusion frames is

$D_s$08

when the number of projections is large, and proves that maximal orthoplectic fusion frames require $D_s$09 (Bodmann et al., 2016). In this setting, the block design is the exact combinatorial device that enforces optimal projection overlap.

6. Networks, resource allocation, and systems tuning

A different but closely related line of work treats utility design itself as the optimization target. “Optimal Utility Design with Arbitrary Information Networks” studies resource-allocation games in which an agent observes only a subset of other agents and receives a local utility

$D_s$10

The information network induces classes of similar agents, and the paper develops one linear program to compute the exact Price of Anarchy for a fixed design and another linear program to optimize the utility design itself (Singh et al., 29 Jan 2025). Because utilities are classwise, local, and anonymous in observed occupancy, the framework is naturally block-like: blocks are information-equivalent classes, and utility is worst-case equilibrium efficiency.

“Optimal block designs for experiments on networks” addresses the same theme from the design-of-experiments side (Koutra et al., 2019). Its Network Block Model (NBM) combines direct treatment effects, block effects, and spillovers: $D_s$11 It then defines two $D_s$12-optimality criteria, one for direct treatment contrasts and one for network-effect contrasts, and searches over treatment allocations with a Point Exchange on Networks (PEN) algorithm. Blocks may be externally specified or inferred by spectral clustering on the random-walk normalized Laplacian $D_s$13, with the number of communities selected by modularity maximization (Koutra et al., 2019). The paper reports large efficiency losses and bias risk when network interference is ignored, especially for estimating spillovers.

In communications, “A Utility Proportional Fairness Radio Resource Block Allocation in Cellular Networks” gives perhaps the clearest non-statistical utility formulation (Ghorbanzadeh et al., 2014). The problem is to allocate integer-valued resource blocks $D_s$14 to $D_s$15 users so as to maximize

$D_s$16

subject to $D_s$17, with sigmoidal utilities for real-time applications and logarithmic utilities for delay-tolerant traffic. The paper solves a continuous relaxation by dual decomposition and then applies floor/ceiling boundary mapping, reducing discrete search from $D_s$18 to at most $D_s$19 (Ghorbanzadeh et al., 2014). The block is the indivisible resource unit, and utility is application-level QoS encoded directly into $D_s$20.

At the systems level, “Workload-Aware Autotuning of Block Size in Square-Root Decomposition” turns block size into a runtime utility variable (Zhao, 4 Jun 2026). Starting from

$D_s$21

the paper argues that the textbook $D_s$22 rule is asymptotically natural but not always fastest in implementation. It defines regret relative to the best candidate block size in a finite set, trains predictors on $D_s$23, and finds that a full-feature KNN-9 policy reduces mean regret from $D_s$24 to $D_s$25 and yields a paired geometric-mean speedup of $D_s$26 over fixed $D_s$27 blocking (Zhao, 4 Jun 2026). A confidence gate using $D_s$28 with threshold $D_s$29 keeps most of the gain while reducing slowdowns. Here the block is a contiguous algorithmic segment, and utility is runtime rather than inference or privacy.

7. Recurring principles, trade-offs, and limitations

Across these literatures, the recurring design move is to turn a block into the atomic action and then attach a utility metric to that action. The metric may be teleportation cost and future relocation utility defined by overlapping qubits (Yun et al., 20 May 2026), asymptotic minimax MSE under ULDP (Yoon et al., 29 Sep 2025), exact or near-optimal privacy-utility trade-off under LDP (Park et al., 2023), pseudo-Bayesian information volume (Mueller et al., 2018), variance-based optimality criteria (Rosa, 2018), lexicographic aberration (Lin, 8 Jun 2025), Price of Anarchy (Singh et al., 29 Jan 2025), or runtime regret (Zhao, 4 Jun 2026). The computational patterns are likewise recurrent: beam-search-like candidate retention in DQC (Yun et al., 20 May 2026), dual decomposition and boundary mapping in cellular RB allocation (Ghorbanzadeh et al., 2014), linear programs over classwise utility parameters (Singh et al., 29 Jan 2025), exchange algorithms on networks (Koutra et al., 2019), and combinatorial constructions from BIBDs, ordered designs, Latin squares, and MUBs (Cakiroglu et al., 2019, Bodmann et al., 2016).

The trade-offs are consistently explicit. ATHENA’s search improves teleportations and latency but is much slower to compile (Yun et al., 20 May 2026). ULDP uBD achieves the exact asymptotic optimum, but outside closed-form regimes the saddle point $D_s$30 is obtained numerically (Yoon et al., 29 Sep 2025). Exact LDP block designs can be sparse enough to motivate RPBD relaxations (Park et al., 2023). Resolution-based communication reduction preserves privacy utility, but the exact optimal privacy-utility-communication trade-off with shared randomness beyond resolution-based schemes remains open (Nam et al., 2023). Copula-based robust blocked design is practically focused on blocks of size two (Mueller et al., 2018). In workload-aware tuning, short-prefix variants do not produce a successful low-overhead online tuner in the current prototype (Zhao, 4 Jun 2026). In incomplete-block causal inference, the theory provides estimators and variance formulas rather than a direct combinatorial optimizer (Koo et al., 2024).

This suggests that uBD is best understood as a family of block-structured optimization principles rather than a single universal method. What remains invariant is the pairing of three elements: an admissible block class, a domain-specific utility functional, and an optimization or certification mechanism showing why one block structure is preferable to another.