Accordion: Multi-Domain Perspectives

• Accordion is a multi-disciplinary concept characterized by alternating expansion and compression in systems from graph theory to quantum computing and machine learning.
• In graph theory, accordion graphs feature a 4-regular structure with circulant properties that inform studies on Hamiltonicity, matchings, and isomorphism.
• Across applied domains, frameworks named Accordion optimize resource use, enhance communication fidelity, and support adaptive model reasoning in distributed and quantum systems.

Searching arXiv for relevant papers on “Accordion” across the domains represented in the source material. “Accordion” denotes several distinct technical constructs across mathematics, computer science, quantum computing, networking, cloud data systems, biological model assembly, and atomic physics. In graph theory, an accordion graph $A[n,k]$ is a 4-regular graph on two $n$-cycles connected by vertical and diagonal spokes, introduced and analyzed for Hamiltonicity, matchings, circulant structure, and graph isomorphism (Gauci et al., 2020, Gauci et al., 2021). In machine learning and systems, “Accordion” names frameworks for adaptive gradient communication (Agarwal et al., 2020), communication-aware model delivery in next-generation networks (Ayed et al., 2023), self-compressing large-language-model reasoning (Yang et al., 3 Feb 2026), and intra-query runtime elasticity for OLAP engines (Zhang et al., 25 Feb 2025). In quantum compilation, Accordion is an end-to-end framework for compiling fermionic Hamiltonians under hardware-connectivity constraints (Gao et al., 31 May 2026). In systems biology, ACCORDION is a workflow for clustering literature-extracted interactions and constructing executable regulatory models (Ahmed et al., 2020). In mathematical logic, the Accordion is a counter-example for the non-conservativity of the linear approximation of the infinitary $\lambda$-calculus (Cerda et al., 2023). In cold-atom physics, an optical accordion is a standing-wave lattice with dynamically tunable spacing used to realize two-dimensional Bose gases with tunable confinement and interaction strength (Ville et al., 2016). The term also appears in operator-algebra classification, where “accordion spaces” are finite $T_0$-spaces on which filtered $K$-theory admits a reduced complete invariant (Arklint et al., 2013).

1. Accordion graphs in discrete mathematics

Accordion graphs $A[n,k]$ are defined for integers $n\ge 3$ and $1\le k\le \lfloor n/2\rfloor$ on the vertex set

$V(A[n,k])=\{u_1,\dots,u_n\}\cup\{v_1,\dots,v_n\},$

with edge set consisting of an outer $n$-cycle, an inner $n$0-cycle, vertical spokes $n$1, and diagonal spokes $n$2, with indices modulo $n$3 (Gauci et al., 2021, Gauci et al., 2020). The construction yields a 4-regular graph and can be viewed as a slight modification of the Cartesian product of two cycles (Gauci et al., 2021).

The family was introduced in connection with the Pairing-Hamiltonian property and perfect matchings. A pairing of a graph $n$4 is a perfect matching of the complete graph $n$5, and $n$6 is Pairing-Hamiltonian if every such pairing can be completed by a perfect matching of $n$7 to form a Hamiltonian cycle of $n$8 (Gauci et al., 2020). Within this family, $n$9 is the usual $\lambda$0-antiprism; it is PMH for all $\lambda$1, and PH exactly for $\lambda$2. Also, $\lambda$3 is isomorphic to the 4-regular graph obtained by doubling each vertex of an $\lambda$4-cycle and has the full PH property for every $\lambda$5 (Gauci et al., 2020).

The relation to Cartesian products is explicit. If $\lambda$6, then deleting the break-edges

$\lambda$7

produces a graph isomorphic to

$\lambda$8

which links accordion graphs to structured grid-like topologies and to prior circulant-product classification results (Gauci et al., 2020).

2. Circulant structure and isomorphism classification

A central theorem gives the precise conditions under which $\lambda$9 is circulant. For $T_0$0 and $T_0$1, the accordion graph $T_0$2 is circulant if and only if exactly one of the following holds: $T_0$3 is odd; $T_0$4 is even and $T_0$5 is odd; or $T_0$6 and $T_0$7 is even (Gauci et al., 2021). Equivalently, $T_0$8 is not circulant precisely when both $T_0$9 and $K$0 are even with $K$1 (Gauci et al., 2020).

The same classification can be expressed via explicit quartic circulants. If $K$2 is odd, then

$K$3

If $K$4 is odd and $K$5 is even, then

$K$6

If $K$7 and $K$8 is even, then

$K$9

(Gauci et al., 2020).

The graph is bipartite if and only if both $A[n,k]$0 and $A[n,k]$1 are even, and therefore the only bipartite circulant accordions occur in the case $A[n,k]$2 with $A[n,k]$3 even (Gauci et al., 2021). For isomorphism among accordion graphs themselves, isomorphic graphs must share the same $A[n,k]$4, since $A[n,k]$5 (Gauci et al., 2021). Fixing $A[n,k]$6 and $A[n,k]$7, one has

$A[n,k]$8

if and only if

$A[n,k]$9

and

$n\ge 3$0

(Gauci et al., 2021).

A key structural lemma states that if $n\ge 3$1, then the $n\ge 3$2 spokes form two vertex-disjoint $n\ge 3$3-cycles; if $n\ge 3$4, no such disjoint cycle decomposition of the spoke set exists (Gauci et al., 2021). This spoke-cycle interchange is fundamental in the proof strategy. The authors show that an isomorphism between two parameter choices must map some cycle-edge to a spoke, otherwise it would preserve the two canonical $n\ge 3$5-cycles and force equality of the parameters. The resulting analysis yields the modular congruence above (Gauci et al., 2021).

3. Adaptive compression, model delivery, and LLM reasoning

In distributed optimization, Accordion is a wrapper around gradient-compression or batch-size scheduling schemes that switches between “high-fidelity” and “low-fidelity” regimes by detecting critical learning regimes via a gradient-norm criterion (Agarwal et al., 2020). Let

$n\ge 3$6

The current epoch is declared critical if

$n\ge 3$7

with $n\ge 3$8 used throughout (Agarwal et al., 2020). During critical regimes, the method chooses $n\ge 3$9; otherwise it uses $1\le k\le \lfloor n/2\rfloor$0, where $1\le k\le \lfloor n/2\rfloor$1 is the compression parameter or, in the extension, the batch size (Agarwal et al., 2020). Across tasks on 4-node GPU clusters, it maintains similar model accuracy to uncompressed training while reducing total communicated floats by up to $1\le k\le \lfloor n/2\rfloor$2 and wall-clock training time by up to $1\le k\le \lfloor n/2\rfloor$3 over static baselines (Agarwal et al., 2020).

In communication-aware ML for future networks, Accordion redesigns training so that only a subset of the hidden-layer blocks is updated at each SGD iteration, while always updating the final classification layer (Ayed et al., 2023). The trained model contains nested sub-models of varying sizes, and a server-side size-to-accuracy lookup table maps a transmitted fraction $1\le k\le \lfloor n/2\rfloor$4 to expected error $1\le k\le \lfloor n/2\rfloor$5. A user request specifies desired error $1\le k\le \lfloor n/2\rfloor$6, latency $1\le k\le \lfloor n/2\rfloor$7, and estimated rate $1\le k\le \lfloor n/2\rfloor$8; the server selects the largest $1\le k\le \lfloor n/2\rfloor$9 satisfying

$V(A[n,k])=\{u_1,\dots,u_n\}\cup\{v_1,\dots,v_n\},$0

where $V(A[n,k])=\{u_1,\dots,u_n\}\cup\{v_1,\dots,v_n\},$1 is the full-model size in bits (Ayed et al., 2023). The framework also supports incremental enhancement requests that transmit only additional layers (Ayed et al., 2023).

In large-language-model reasoning, “Accordion-Thinking” introduces a structured reasoning format

$V(A[n,k])=\{u_1,\dots,u_n\}\cup\{v_1,\dots,v_n\},$2

where $V(A[n,k])=\{u_1,\dots,u_n\}\cup\{v_1,\dots,v_n\},$3 is a detailed derivation and $V(A[n,k])=\{u_1,\dots,u_n\}\cup\{v_1,\dots,v_n\},$4 is a concise summary (Yang et al., 3 Feb 2026). In Unfold mode, step $V(A[n,k])=\{u_1,\dots,u_n\}\cup\{v_1,\dots,v_n\},$5 is conditioned on all prior details and summaries: $V(A[n,k])=\{u_1,\dots,u_n\}\cup\{v_1,\dots,v_n\},$6 whereas in Fold mode it is conditioned only on compressed history: $V(A[n,k])=\{u_1,\dots,u_n\}\cup\{v_1,\dots,v_n\},$7 (Yang et al., 3 Feb 2026). Since $V(A[n,k])=\{u_1,\dots,u_n\}\cup\{v_1,\dots,v_n\},$8, Fold mode reduces attention cost from

$V(A[n,k])=\{u_1,\dots,u_n\}\cup\{v_1,\dots,v_n\},$9

to

$n$0

(Yang et al., 3 Feb 2026). Reinforcement learning is applied using GRPO without KL, with trajectory-level reward $n$1 based on answer correctness and summary-format validity (Yang et al., 3 Feb 2026). On five math benchmarks, Fold-RL and Mix-RL recover Fold-mode performance to match Unfold-RL, described as “lossless compression”; under a 48 GB memory cap, Fold mode achieves $n$2 tok/s versus $n$3 tok/s for Unfold-RL, and the paper reports $n$4 throughput while maintaining accuracy on a 48 GB GPU memory configuration (Yang et al., 3 Feb 2026). Human annotation on 20 samples found that 19/20 summaries fully capture the required information for subsequent reasoning (Yang et al., 3 Feb 2026).

A plausible implication is that, across these machine-learning usages, “Accordion” consistently names a mechanism that alternates between expanded and compressed operating regimes rather than a single fixed operating point.

4. Cloud-native data systems and quantum circuit compilation

In cloud-native OLAP, Accordion is presented as the first Intra-Query Runtime Elasticity query engine, built atop the Presto execution model (Zhang et al., 25 Feb 2025). The system can adjust a query’s Degree of Parallelism during execution without pausing data processing (Zhang et al., 25 Feb 2025). Its coordinator gathers streaming runtime statistics such as scan rates, exchange buffer fill levels, and CPU and network utilization, and a what-if predictor estimates how remaining execution time changes under a different DOP (Zhang et al., 25 Feb 2025). For a stage with current DOP $n$5, remaining time

$n$6

and proposed DOP $n$7, the predicted remaining time is

$n$8

with $n$9 for stateless stages and $n$00 for hash-join stages (Zhang et al., 25 Feb 2025). The auto-tuner solves

$n$01

where $n$02 (Zhang et al., 25 Feb 2025). On TPC-H SF100, dynamic intra-stage tuning reduced Q3 from 740 s to 194 s, and dynamic task or driver changes were implemented without stopping the pipeline (Zhang et al., 25 Feb 2025).

In fermionic quantum simulation, Accordion is an end-to-end compilation framework that co-designs the fermion-to-qubit mapping with circuit synthesis and hardware routing (Gao et al., 31 May 2026). It fixes the Jordan–Wigner mapping and exploits the structural regularity of the resulting Pauli strings. For full-rank all-to-all electronic-structure Hamiltonians, the method proves $n$03 gate count and circuit depth, matching the $n$04 lower bound imposed by the second-excitation terms in the UCCSD ansatz (Gao et al., 31 May 2026). The mapped double-excitation terms share the same four $n$05-operator positions and contiguous $n$06 prefixes, enabling a grouping hierarchy into Large, Medium, Mini, and Atomic groups (Gao et al., 31 May 2026). The scheduling uses remapping by SWAPs, cancellation of CNOT ladders, and odd-even transposition sort, with amortized $n$07 depth per string across transitions (Gao et al., 31 May 2026). On linear, IBM heavy-hex (“Boston”), and square-grid (“Miami”) architectures, Accordion reduces CNOT count by up to 79.6% on linear, 59.6% on Boston, and 36.7% on Miami, and critical-path depth by up to 77.2% on linear, 54.6% on Boston, and 34.8% on Miami (Gao et al., 31 May 2026).

These two systems usages share an end-to-end co-design principle. In the OLAP engine, execution control, buffering, and DOP prediction are jointly engineered (Zhang et al., 25 Feb 2025). In quantum compilation, mapping choice, Pauli grouping, routing, and scheduling are jointly engineered (Gao et al., 31 May 2026). This suggests that “Accordion” is also used as a label for architectures that exploit runtime or structural regularity through hierarchical adjustment.

5. Biological model assembly and formal analysis

ACCORDION, expanded as “Automated Clustering Conditional On Relating Data of Interactions tO a Network,” is a tool and methodology for automatically assembling new models or expanding existing ones from published literature (Ahmed et al., 2020). Its workflow has six stages: information extraction, clustering of Candidate Extension Interactions, model assembly, simulation, formal analysis via statistical model checking, and model selection (Ahmed et al., 2020).

Starting from a query, literature APIs return papers whose abstracts or full texts are processed by a machine-reading engine such as REACH to extract regulator-target events (Ahmed et al., 2020). These are filtered and partitioned into corroborations, contradictions, and extensions, with only extensions retained as CEIs (Ahmed et al., 2020). Given a baseline influence graph $n$08, the union graph

$n$09

is formed with

$n$10

(Ahmed et al., 2020). Markov Clustering is then applied using a column-stochastic matrix

$n$11

with iterative expansion $n$12 and inflation

$n$13

until convergence (Ahmed et al., 2020).

Clusters are filtered by the existence of a return path that starts and ends in $n$14 while otherwise remaining in the cluster; viable clusters and some minimal merges define Candidate Executable Models (Ahmed et al., 2020). These are simulated as Boolean or multi-valued dynamical systems $n$15, and desired properties are encoded in BLTL, for example

$n$16

or

$n$17

(Ahmed et al., 2020). For each model and property, statistical model checking estimates

$n$18

and the global score is

$n$19

under an independence assumption (Ahmed et al., 2020).

In a T-cell differentiation case study, three CEI scenarios were examined: fully automated with 171 CEIs, semi-automated with 81 CEIs, and semi-manual with 54 CEIs (Ahmed et al., 2020). MCL with $n$20 and $n$21 yielded 22, 11, and 9 clusters respectively, and return-path filtering plus pairwise merges produced 27, 22, and 16 Candidate Executable Models (Ahmed et al., 2020). Using 27 BLTL properties and $n$22 DiSH runs, the top semi-manual ACCORDION model achieved $n$23 and satisfied 24/27 properties, compared with $n$24 and 25/27 for the manually extended golden model (Ahmed et al., 2020).

6. Optical accordions, infinitary $n$25-calculus, and accordion spaces

In ultracold-atom experiments, an optical accordion is a standing-wave lattice whose fringe spacing is tuned by changing the crossing angle $n$26 of two interfering 532 nm beams (Ville et al., 2016). For wavelength $n$27, the fringe spacing is

$n$28

so varying $n$29 from $n$30 to $n$31 tunes $n$32 from roughly $n$33 down to $n$34 (Ville et al., 2016). The vertical confinement obeys

$n$35

and measured oscillation frequencies ranged from $n$36 to $n$37 as $n$38 increased (Ville et al., 2016). The full $n$39 compression took about 130 ms, with no measurable atom loss and extra heating $n$40 per compression (Ville et al., 2016). Combined with a flat-bottom in-plane trap, the apparatus realized uniform 2D Bose gases with tunable interaction parameter

$n$41

varied from about 0.08 to 0.26 (Ville et al., 2016).

In the infinitary $n$42-calculus, the Accordion is a specific closed infinitary term $n$43 constructed from $n$44, $n$45, $n$46, Church numerals, successor, and a family of $n$47 operators (Cerda et al., 2023). It serves as a counter-example to conservativity of the ordinary Taylor approximation. Cerda and Vaux show that although there is no infinitary $n$48-reduction

$n$49

there is a reduction of sets of resource approximants

$n$50

in the finitary resource calculus (Cerda et al., 2023). The failure arises because every finite approximant tracks only a finite unfolding, yet any genuine infinitary $n$51-sequence would have to contract redexes whose depth tends to infinity; the Accordion instead forces contraction at bounded depth, in fact depth 0, throughout the head-reduction pattern (Cerda et al., 2023). Restricting to uniform reductions restores conservativity (Cerda et al., 2023).

In $n$52-algebra classification, “accordion spaces” are finite $n$53-spaces whose Hasse diagrams form a zig-zag of order relations without short cycles of length 3 (Arklint et al., 2013). On such spaces, filtered $n$54-theory collapses to a smaller invariant. For real-rank-zero algebras over an accordion space $n$55, the full filtered $n$56-theory $n$57 is equivalent to reduced filtered $n$58-theory $n$59, denoted $n$60 (Arklint et al., 2013). For stable Kirchberg $n$61-algebras of real rank zero whose simple subquotients lie in the UCT bootstrap class,

$n$62

as modules over the reduced category $n$63 (Arklint et al., 2013). A standard example is the three-point $n$64-shaped space $n$65 with $n$66 and $n$67 (Arklint et al., 2013).

7. Cross-domain patterns and terminological structure

The term “Accordion” is therefore not a single scientific object but a recurrent metaphor and naming convention. In several domains, it denotes controlled alternation between expansion and compression: optical-lattice spacing in ultracold gases (Ville et al., 2016), stepwise detail and summary in LLM reasoning (Yang et al., 3 Feb 2026), communication fidelity in distributed learning (Agarwal et al., 2020), model fraction delivery in communication-aware inference (Ayed et al., 2023), and DOP scaling during query execution (Zhang et al., 25 Feb 2025). In graph theory and finite-space topology, the name is structural rather than operational, referring to a repeated zig-zag or spoke-cycle arrangement (Gauci et al., 2021, Arklint et al., 2013). In infinitary rewriting, it names a construction whose repeated stretch-and-compress behavior is visible in its head-reduction sequence (Cerda et al., 2023).

A plausible implication is that the persistence of the term across fields reflects a common abstraction: a system whose effective state is repeatedly reconfigured between larger and smaller representations while preserving a controlled invariant. In graph theory that invariant is unlabeled graph isomorphism under constrained relabelings (Gauci et al., 2021). In reduced filtered $n$68-theory, it is classification data under compression of the invariant (Arklint et al., 2013). In LLM reasoning, it is answer quality under compressed context (Yang et al., 3 Feb 2026). In OLAP and distributed training, it is task accuracy or SLA satisfaction under reduced communication or resource expenditure (Agarwal et al., 2020, Zhang et al., 25 Feb 2025). In quantum compilation, it is asymptotically optimal $n$69 synthesis under structured regularity (Gao et al., 31 May 2026).