Adjacent-optimal Design Overview

• Adjacent-optimal design is an optimization strategy defined by leveraging local adjacency constraints to improve experimental, combinatorial, and physical system performance.
• Its methodologies include iterative local update rules, equivalence optimality conditions, and bi-level optimization techniques to achieve near-global efficiency with localized adjustments.
• Applications span ordered treatment designs, multinomial adjacent-category models, scheduling problems, communication networks, and quantum compilation scenarios.

Adjacent-optimal design denotes a family of optimization problems in which optimality is defined relative to an adjacency structure. In the literature, that structure may be the ordering of treatments, adjacency of response categories, the use of consecutive group sizes, adjacency of graph vertices, physical proximity of engineered subsystems, or nearest-neighbour hardware constraints. The expression therefore does not designate a single universally standardized formalism; rather, it names a recurring locality principle that is instantiated differently across statistics, combinatorics, control, electromagnetics, communications, and quantum compilation (Huda et al., 2014, Bu et al., 2017, Miller et al., 31 Jul 2025, Bigdeli et al., 2015, Chalaki et al., 2019, Meijer et al., 2024).

1. Domain scope and core meanings

The main uses of the term cluster around a small number of technical patterns. In statistical design, adjacency usually refers either to consecutive treatment contrasts or to adjacent response categories. In combinatorial design, it refers to schedules or colorings constrained by local neighborhood relations. In engineering, it refers to systems whose performance is dominated by interaction between nearby components, such as adjacent buildings, intersections, base stations, sources, or qubits (Huda et al., 2014, Bu et al., 2017, Gravier et al., 2020, Miller et al., 31 Jul 2025, Bigdeli et al., 2015, Jiang et al., 2023).

Context	Adjacent object	Optimization target
Small-block experimental design	Consecutive treatments	Minimize $\operatorname{tr}\{M(p)^{-1}\}$
Multinomial logistic design	Adjacent categories	Maximize $|F|$
Social golfer scheduling	Adjacent group sizes	Maximize rounds
Graph edge-coloring	Adjacent vertices	Minimize number of colors
Structural, traffic, ISAC, EM, quantum systems	Nearby physical entities	Application-specific performance under adjacency constraints

This heterogeneity matters. A direct statistical meaning is given by ordered-treatment block design for comparing $\tau_i-\tau_{i-1}$ (Huda et al., 2014). A different but still statistical meaning appears in D-optimal design for adjacent-categories logit models (Bu et al., 2017). Elsewhere, “adjacent-optimal” is best understood as an analogy: designs are optimal because they respect and exploit a local adjacency relation rather than a global all-to-all one (Bigdeli et al., 2015, Meijer et al., 2024).

2. Ordered treatments and consecutive-contrast block design

The most direct experimental-design usage appears in small-block design with naturally ordered treatments $1,\dots,v$, where the inferential target is the set of consecutive contrasts

$\tau_i-\tau_{i-1}, \qquad 2\le i\le v.$

Under the usual additive fixed-effects block model,

$y = \mu 1 + X\tau + Z\beta + \varepsilon,$

the treatment information matrix is

$C = R-\frac{1}{k}NN',$

and the contrast system is encoded by the $(v-1)\times v$ matrix $L$ with $L_{i,i}=-1$, $|F|$0. With

$|F|$1

the approximate design information matrix is

$|F|$2

and A-optimality is formulated as

$|F|$3

The paper proves the equivalence condition that a design measure $|F|$4 is A-optimal iff $|F|$5 is nonsingular and

$|F|$6

and derives the multiplicative update

$|F|$7

with stopping tolerance $|F|$8 (Huda et al., 2014).

A distinctive feature of this adjacent-treatment problem is that the candidate block class includes nonbinary blocks, and the paper emphasizes that these cannot in general be ignored. Most optimal measures are supported on binary blocks, but the cases $|F|$9, $\tau_i-\tau_{i-1}$0 require nonbinary support. Exact designs are then obtained by scaling and rounding the optimal design measure, with efficiency benchmark

$\tau_i-\tau_{i-1}$1

This leads to highly efficient exact designs even when the number of blocks is small; reported examples include efficiencies $\tau_i-\tau_{i-1}$2, $\tau_i-\tau_{i-1}$3, and $\tau_i-\tau_{i-1}$4, and several nested families for staged experimentation (Huda et al., 2014).

The resulting adjacent-optimal designs are structurally local. Contiguous or near-contiguous blocks such as $\tau_i-\tau_{i-1}$5, $\tau_i-\tau_{i-1}$6, and overlapping chains of such blocks receive high mass, because they carry direct information about neighboring differences. This sharply distinguishes the problem from all-pairs comparison design, where more global balance criteria dominate (Huda et al., 2014).

3. Adjacent categories and continuous nearby refinement

A second statistical meaning concerns multinomial logistic models with adjacent-categories logits,

$\tau_i-\tau_{i-1}$7

The general framework covers proportional odds, non-proportional odds, and partial proportional odds assumptions, and develops three equivalent Fisher-information representations: $\tau_i-\tau_{i-1}$8 For adjacent-categories models, the paper gives an especially simple structure in which $\tau_i-\tau_{i-1}$9 for $1,\dots,v$0, which simplifies $1,\dots,v$1 and the D-optimality calculations. Two conclusions are central: a feasible design may contain fewer experimental settings than parameters, and even for a minimally supported design a uniform allocation is not optimal in general. The paper develops a lift-one algorithm for approximate designs and an exchange algorithm for exact designs (Bu et al., 2017).

The adjacent-categories setting is conceptually different from consecutive-treatment design. Here adjacency is not among design points but among response categories. Nevertheless, the local structure is similar: the information and optimal allocation depend on how neighboring categories are contrasted, not on a fully symmetric treatment of all outcome pairs. A plausible implication is that adjacency can enter optimal design either on the treatment axis or on the response axis, with different but formally comparable consequences for Fisher-information structure (Bu et al., 2017).

A further extension appears in locally optimal continuous design for nonlinear models on a continuous design space. The paper “An Adaptive Algorithm based on High-Dimensional Function Approximation to obtain Optimal Designs” is explicit that it is not about adjacent-optimal design in the classical exact-design sense; it instead develops ADA-GPR for locally optimal continuous designs, using the Kiefer–Wolfowitz Equivalence Theorem and a Gaussian-process surrogate for the sensitivity function (Seufert et al., 2021). With

$1,\dots,v$2

the support-search step seeks

$1,\dots,v$3

but approximates $1,\dots,v$4 over the continuous space by GPR, thereby reducing expensive Jacobian evaluations. The paper describes this as a continuous-space support-point augmentation and refinement procedure rather than an adjacent-exchange algorithm. This suggests an indirect adjacent-optimal interpretation: each iteration makes a nearby structural modification to the current design—one new support point plus weight reallocation—while retaining global search over the continuous domain (Seufert et al., 2021).

4. Combinatorial and graph-theoretic adjacent optimality

In combinatorial design theory, adjacency often refers to allowable group sizes or local distinguishability constraints. The most direct example is the Social Golfer Problem with Adjacent Group Sizes (SGA), where the allowed block sizes are

$1,\dots,v$5

An SGA instance

$1,\dots,v$6

arranges $1,\dots,v$7 points into $1,\dots,v$8 rounds with exactly $1,\dots,v$9 blocks of size $\tau_i-\tau_{i-1}, \qquad 2\le i\le v.$0 and $\tau_i-\tau_{i-1}, \qquad 2\le i\le v.$1 blocks of size $\tau_i-\tau_{i-1}, \qquad 2\le i\le v.$2 per round, with no repeated pair in a block. The paper’s central constructional idea is to derive such schedules from a superior allocation on $\tau_i-\tau_{i-1}, \qquad 2\le i\le v.$3 points with block size $\tau_i-\tau_{i-1}, \qquad 2\le i\le v.$4 by deleting points, or from an inferior allocation on $\tau_i-\tau_{i-1}, \qquad 2\le i\le v.$5 points with block size $\tau_i-\tau_{i-1}, \qquad 2\le i\le v.$6 by adding points. Resolvable combinatorial designs—RBIBDs, RGDDs, URDs, RTDs/ITDs, and MOLS/MOLRs—are used to obtain optimal or best-known schedules, and a complete catalogue is given for up to $\tau_i-\tau_{i-1}, \qquad 2\le i\le v.$7 players (Miller et al., 31 Jul 2025).

A different graph-theoretic meaning appears in adjacent vertex-distinguishing edge-colorings of circulant graphs. A proper edge-coloring is adjacent vertex-distinguishing if adjacent vertices have different incident color sets, and the minimum number of colors is denoted $\tau_i-\tau_{i-1}, \qquad 2\le i\le v.$8. For circulant graphs

$\tau_i-\tau_{i-1}, \qquad 2\le i\le v.$9

which are $y = \mu 1 + X\tau + Z\beta + \varepsilon,$0-regular, the trivial lower bound is

$y = \mu 1 + X\tau + Z\beta + \varepsilon,$1

The main theorem proves that for most sufficiently large $y = \mu 1 + X\tau + Z\beta + \varepsilon,$2, specifically under the stated threshold and when $y = \mu 1 + X\tau + Z\beta + \varepsilon,$3 or $y = \mu 1 + X\tau + Z\beta + \varepsilon,$4,

$y = \mu 1 + X\tau + Z\beta + \varepsilon,$5

The remaining uncovered regime is

$y = \mu 1 + X\tau + Z\beta + \varepsilon,$6

which the paper leaves open (Gravier et al., 2020).

These works show that “adjacent-optimal” in combinatorics can mean either minimal deviation from equal group sizes or minimal resource overhead for local distinguishability. In both cases, optimality is driven by what can be guaranteed for neighboring or nearly neighboring objects rather than by a global unconstrained packing or coloring objective (Miller et al., 31 Jul 2025, Gravier et al., 2020).

5. Physical adjacency as a design constraint

In structural optimization of adjacent buildings, the problem is to place dampers between corresponding floors and choose their coefficients to minimize the maximum inter-story drift

$y = \mu 1 + X\tau + Z\beta + \varepsilon,$7

The paper formulates a bi-level optimization: an outer “inserting dampers” method chooses floors, while an inner derivative-free optimizer tunes coefficients using GA, MADS, or RAGS. It reports that MADS with warm-start gives the best balance of accuracy and speed, and that there is a threshold effect: in most cases behavior within $y = \mu 1 + X\tau + Z\beta + \varepsilon,$8 of optimum can be achieved using $y = \mu 1 + X\tau + Z\beta + \varepsilon,$9 or fewer dampers, and only one problem required more than $C = R-\frac{1}{k}NN',$0 (Bigdeli et al., 2015).

For connected and automated vehicles at adjacent intersections, adjacency appears as spatial coupling across nearby conflict zones. One paper formulates a zone-based upper-level MILP that minimizes each vehicle’s exit time through two adjacent intersections, with lower-level energy-optimal control

$C = R-\frac{1}{k}NN',$1

between scheduled zone times; simulations report average travel-time reductions of $C = R-\frac{1}{k}NN',$2–$C = R-\frac{1}{k}NN',$3 relative to signalized intersections (Chalaki et al., 2019). A related multi-intersection framework computes recursive arrival times and lane choices in an upper level, then solves a lower-level optimal control problem with interior-point constraints; simulations show average travel-time reduction $C = R-\frac{1}{k}NN',$4 to $C = R-\frac{1}{k}NN',$5, delay reduction roughly $C = R-\frac{1}{k}NN',$6 to $C = R-\frac{1}{k}NN',$7, and fuel-consumption improvement $C = R-\frac{1}{k}NN',$8 to $C = R-\frac{1}{k}NN',$9 in a symmetric corridor (Chalaki et al., 2020).

In adjacent ISAC base stations, the issue is mutual downlink interference affecting both communications and radar sensing. The collaborative precoding problem is posed under SINR, transmit-power, and constant-modulus or similarity constraints, relaxed to an SDP in the joint optimization algorithm and decomposed into four interference-suppression subproblems in the sequential optimization algorithm. The main practical conclusion is that collaborative design among adjacent BSs improves both average communication rate and detection probability relative to non-collaborative operation (Jiang et al., 2023).

Electromagnetic absorber design provides another interpretation. When an antenna is placed adjacent to a Dallenbach absorber, the illumination is modeled as a weighted angular spectrum of propagating oblique plane waves. With normalized $(v-1)\times v$0, the paper derives the weighted broad-angle bound

$(v-1)\times v$1

and optimizes layered absorbers to approach that bound. In the 6G dipole example, the flatness metric improves from $(v-1)\times v$2 to $(v-1)\times v$3 for $(v-1)\times v$4 and to $(v-1)\times v$5 for $(v-1)\times v$6 at $(v-1)\times v$7, showing robust mitigation of scattering-induced pattern distortion (Firestein et al., 2023).

Quantum compilation imposes perhaps the most literal adjacency constraint: two-qubit gates can act only on physically adjacent qubits. The nearest neighbour compliance problem asks for a sequence of qubit orders $(v-1)\times v$8 that complies with every gate while minimizing

$(v-1)\times v$9

the total number of inserted SWAPs. By exploiting automorphisms of the relevant Cayley graphs, the paper reduces the original LP to an equivalent generalized network flow problem and proves polynomial-time solvability for several highly symmetric architectures, including the star $L$0. Numerical tests indicate reductions in variables and constraints of at least $L$1 on average, and star-architecture instances with up to $L$2 qubits and more than $L$3 gates can be solved in very short computation time (Meijer et al., 2024).

6. Methodological patterns, misconceptions, and open directions

Across these literatures, adjacency is not a single criterion but a structural restriction on information flow, resource sharing, or feasibility. In ordered-treatment design it restricts the contrast system to neighboring treatments; in adjacent-categories models it restricts logits to neighboring response categories; in combinatorics it restricts allowable block sizes or local distinguishability; in physical systems it restricts coupling to nearby components or legal operations (Huda et al., 2014, Bu et al., 2017, Miller et al., 31 Jul 2025, Meijer et al., 2024).

Several recurrent methodological patterns follow. One is the use of equivalence-type optimality conditions and iterative local updates, as in the multiplicative algorithm for consecutive-treatment block design and the sensitivity-function search for locally optimal continuous design (Huda et al., 2014, Seufert et al., 2021). Another is bi-level decomposition, visible in adjacent buildings, adjacent intersections, and ISAC precoding, where combinatorial or scheduling decisions are separated from continuous parameter optimization (Bigdeli et al., 2015, Chalaki et al., 2019, Jiang et al., 2023). A third is the exploitation of symmetry or resolvability, as in SGA constructions and symmetry-reduced quantum routing (Miller et al., 31 Jul 2025, Meijer et al., 2024).

A common misconception is that “adjacent-optimal design” names one classical method. The sources do not support that reading. The term is direct and canonical in some settings, especially ordered-treatment block design (Huda et al., 2014), but indirect or analogical in others. The ADA-GPR paper, for example, explicitly states that it is not about adjacent-optimal design in the classical discrete exact-design sense, even though its iterative support-point refinement is closely related in spirit (Seufert et al., 2021).

Another misconception is that local problems force trivial or uniform solutions. The opposite occurs repeatedly. Uniform allocation is not generally D-optimal in multinomial adjacent-categories models (Bu et al., 2017). Nonbinary blocks can be genuinely optimal in consecutive-treatment block design (Huda et al., 2014). Full-height damper placement is usually unnecessary once coefficient optimization is allowed (Bigdeli et al., 2015). Exact nearest-neighbour routing in quantum circuits remains highly nontrivial unless symmetry is exploited (Meijer et al., 2024).

Taken together, these works suggest that adjacent-optimal design is best understood as a broad design paradigm organized around locality. The mathematical object that is “adjacent” changes from paper to paper, but the central question remains stable: how should one allocate experiments, groups, colors, control actions, materials, signals, or hardware moves when only neighboring interactions are scientifically relevant, physically allowed, or computationally affordable?