Iterative Compaction Regime: Mechanisms & Applications

Updated 12 June 2026

Iterative compaction regime is a process that progressively densifies and reconfigures materials, data structures, or memory systems through repeatable, controlled operations.
In granular media, the regime follows stretched-exponential or logarithmic densification laws, highlighting dynamic crossovers and glassy aging phenomena in response to cyclic perturbations.
In databases and machine learning, iterative compaction optimizes performance by incrementally merging files or synthesizing compact cache representations to maintain efficiency.

An iterative compaction regime denotes any system in which material (often disordered or composite) is progressively densified, reorganized, or memory-reduced through a sequence of controlled, repeatable operations, rather than via a single monolithic transformation. This concept arises in diverse scientific contexts, including granular media under repeated taps or cycles, memory architectures in LLMs under recurrent compaction, and storage engines that merge data via incremental partial compactions. Across fields, iterative compaction exhibits characteristic signatures: slow, sometimes glassy relaxation toward a steady state; emergent boundaries between distinct dynamical regimes; and universal laws—such as stretched-exponential or logarithmic densification—reflecting the complex landscape of accessible microstates or configurations.

1. Iterative Compaction in Granular Media

A. Cyclic Compression, Tapping, and Thermal Cycling

In granular physics, iterative compaction is exemplified by systems wherein beds of particles are subjected to repeated driving, such as mechanical taps, oscillatory compression, or thermal cycling. Compaction proceeds through a competition between local particle rearrangements, frictional relaxation, and force-chain reorganizations, with dynamical laws depending on both the nature and amplitude of the perturbation.

For virtually all protocols (mechanical tapping, cyclic compression/release, thermal cycling), the evolution of the global packing fraction $\phi(N)$ with the number of cycles $N$ is well described by the Kohlrausch–Williams–Watts (KWW) stretched-exponential law:

$\phi(N) = \phi_\infty - [\phi_\infty - \phi_0] \exp\left( - (N/\alpha)^\beta \right)$

with $\phi_0$ the initial loose-packing density, $\phi_\infty$ the steady-state density, $\alpha$ a characteristic relaxation time in cycles, and $0 < \beta < 1$ a stretching exponent reflecting the width of the relaxation time distribution. In some systems—especially under weak forcing—the logarithmic law

$\phi(N) \simeq \phi_0 + C \ln N$

or a double-exponential fit offers improved description of ultra-slow tail behavior (Luo et al., 23 Jan 2025, Santos et al., 2023, Luo et al., 2 Nov 2025).

A crossover in dynamical regime is ubiquitous. For example, in stress-controlled cyclical compression of frictional spheres, there exists a critical pressure $P_c$ below which the packing very rapidly saturates to a fixed $\phi$ within a few cycles ("constant volume-fraction" regime), while above $N$ 0 the packing fraction continues to increase slowly over many cycles ("continued densification" regime), with stretched-exponential relaxation (Santos et al., 2023). A similar two-stage pathway is seen in packings of geometrically complex grains such as hexapods: initial densification is dominated by local interlocking, after which further compaction occurs via medium-range glassy ordering of polytetrahedra (Luo et al., 2 Nov 2025). Under thermal cycling, compaction dynamics similarly crossover from rapid initial relaxation (dominated by easy local rearrangements) to asymptotically slow glassy aging, with $N$ 1 increasing and characteristic timescales decreasing as cycle amplitude grows (Luo et al., 23 Jan 2025).

Table: Compaction Fitting Laws and Regimes

Model	Functional Form	Regime/Signature
KWW	$N$ 2	Intermediate cycles, broad timescale distribution
Logarithmic	$N$ 3	Asymptotic slow aging, glass-like
Double-Exp	$N$ 4	Two-stage (fast/slow) mechanisms

Parameter values and crossovers are explicitly measured in (Santos et al., 2023, Luo et al., 23 Jan 2025, Luo et al., 2 Nov 2025).

B. Grain-Scale Mechanics and Transition Points

The iterative compaction regime admits sharp microscopic signatures. In bonded-particle models of granular matter (Clemmer et al., 2023), the system transitions from a jamming-dominated regime—characterized by power-law scalings of pressure and coordination above critical density—to a deformation-dominated regime at $N$ 5, where elastic strain and grain asphericity grow rapidly and frictional differences vanish. The Poisson's ratio of the structure exhibits non-monotonic evolution, reflecting a complex interplay between bulk and shear moduli as the system leaves the hard-particle regime.

For frictional grains, the fraction of fully mobilized sliding contacts $N$ 6 governs the attainable density more directly than confining pressure; reducing $N$ 7 via sufficient cycling underlies the approach to dense steady states (Santos et al., 2023).

C. Stochastic Event Statistics and Glassy Phenomenology

Disordered packings under iterative compaction exhibit glassy aging phenomena. When a granular pile is rapidly quenched, densification progresses via rare, record-breaking irreversible events ("quakes," i.e., increases in a grain's contact number), whose occurrence rate decelerates hyperbolically, $N$ 8, leading to a log-Poisson process—the universal aging law in structural glasses. This entails logarithmic densification, $N$ 9, and aging correlators that scale with $\phi(N) = \phi_\infty - [\phi_\infty - \phi_0] \exp\left( - (N/\alpha)^\beta \right)$ 0 (waiting time ratio) (Gago et al., 2020).

2. Iterative Compaction in Data Structures

A. LSM-Trees and Incremental Merging

In log-structured merge (LSM) tree databases, an iterative compaction regime refers to reorganization policies that incrementally combine ("compact") subsets of files between adjacent levels rather than performing global, full-level merges (Sarkar et al., 2022, Mathieu et al., 2014). Each compaction event selects a subset of sorted runs or files to merge into the next level, guided by system-configurable primitives:

Compaction trigger: event when compaction is initiated (e.g., level overflow, sorted-run count, staleness).
Data layout: structuring of runs per level (leveling, tiering, hybrid).
Compaction granularity: amount of data moved per job (file-level, multi-file, level-wide).
Data movement policy: selection strategy for files to be compacted (round-robin, least-overlap, etc.)

Iterative compaction is contrasted with monolithic approaches, where compactions operate on an entire level in one step. The iterative regime allows finer-grained, distributed overhead management and can be systematically designed by tuning the above primitives (Sarkar et al., 2022).

Theoretical and empirical analysis show trade-offs: finer-grained, frequent iterative compactions lower tail latency and smooth write stalls, while larger, less frequent ones can deliver better average write amplification or read amplification, depending on layout and workload. Formal analysis of online policies achieves worst-case competitive guarantees (e.g., $\phi(N) = \phi_\infty - [\phi_\infty - \phi_0] \exp\left( - (N/\alpha)^\beta \right)$ 1-competitive algorithms for $\phi(N) = \phi_\infty - [\phi_\infty - \phi_0] \exp\left( - (N/\alpha)^\beta \right)$ 2-level LSMs), and average-case optimality on i.i.d. streaming workloads (Mathieu et al., 2014).

Regime	Layout	Granularity	Movement Policy	Ideal Use Case
Full	Leveling	Level	N/A	Latency-agnostic, write-heavy
LO+1	Leveling	File	Least-Overlap	Latency-sensitive, look-up focused
Tier	Tiering	Sorted-runs	N/A	High-update, batch-pruning

B. Optimization and Design Guidelines

Iterative compaction policies are not universally optimal; selection of regime must be matched to workload profile and service objectives. Key design takeaways include the need for open tuning of compaction primitives, dynamic regime switching in response to workload changes, and leveraging hybrid or delete-aware regimes to optimize both average and tail performance for ingestion, lookup, and compliance (Sarkar et al., 2022).

3. Iterative Compaction in Machine Learning Systems

A. KV Cache Compaction in LLMs

In long-horizon LLMs, iterative compaction addresses the unsustainable memory growth of the key-value (KV) cache by repeatedly synthesizing a compact representation of past context as new tokens arrive (O'Neill et al., 5 Jun 2026). In this regime, after each chunk of input, a lightweight yet expressive compactor module (e.g., a per-layer Perceiver) maps recent cache entries into a fixed number of compressed slots, causing the effective cache size to grow sublinearly with sequence length.

Key properties distinguishing iterative compaction in this setting:

On-the-fly reuse: Compaction can be triggered repeatedly as context grows, keeping KV cache size tractable.
Amortized inference: The compactor is trained once and applied in each iteration with fixed overhead, unlike per-context optimization (non-amortized) synthesis.
Synthesis vs. selection: Iterative amortized synthesis methods recover more model fidelity under aggressive cache budgets than subset-selection or per-document synthesis strategies at long horizons.

Experimental benchmarks demonstrate that with appropriately trained iterative compactors, long-context tasks (QA, summarization) preserve significant fractions of full-context performance (e.g., 74–95% of summarization gain at $\phi(N) = \phi_\infty - [\phi_\infty - \phi_0] \exp\left( - (N/\alpha)^\beta \right)$ 3– $\phi(N) = \phi_\infty - [\phi_\infty - \phi_0] \exp\left( - (N/\alpha)^\beta \right)$ 4 compression), outperforming prior approaches, especially at extreme sequence lengths (O'Neill et al., 5 Jun 2026).

4. Structural and Algorithmic Iterative Compaction

A. Graph Algorithms: Connectivity-Preserving Compaction

In combinatorial optimization, iterative compaction provides a framework for recursive reduction of graph instances while preserving key structural properties (e.g., $\phi(N) = \phi_\infty - [\phi_\infty - \phi_0] \exp\left( - (N/\alpha)^\beta \right)$ 5-connectivity). Given a highly connected input graph, local operations (edge deletion/matching, contraction, stable sets) are applied to shrink the graph by a constant fraction at each step, ensuring that critical properties are intact. The process iterates until the instance is trivially solvable, at which point solutions/reductions are reconstructed on the original graph in reverse ("uncompaction") (Kawarabayashi et al., 2015).

Such iterative compaction yields efficient algorithms for problems such as 2-disjoint-paths: repeatedly reducing the problem (while preserving triple-connectivity or planarity obstructions), with each iteration effecting only local, connectivity-safe modifications.

B. Null-Space Guided Compaction

In frictional granular assemblies, iterative compaction following vibration or mechanical perturbation can be mapped to deterministic paths in configuration space associated with null-spaces of the instantaneous rigidity matrix. Each compaction burst corresponds to a system transition through the null-space (floppy modes) until an additional contact is formed, reducing the dimensionality. This observable branch structure leads to a discrete—a quasi-tree—progression toward isostatic, jammed configurations (Hubard et al., 2015).

5. Universalities, Regime Boundaries, and Open Problems

The iterative compaction regime, despite context-specific realizations, displays strong universalities:

Stretched-exponential, logarithmic, and double-exponential densification laws are observed in granular compaction under both mechanical and thermal driving, and in algorithmic systems with distributed timescales (Luo et al., 23 Jan 2025, Santos et al., 2023, Luo et al., 2 Nov 2025).
Phase transitions and crossovers—such as from friction-dominated to deformation-dominated mechanics in granular packings, or from local to glassy order in composite particles—are sharp, reproducible boundaries separating distinct regimes of compaction (Clemmer et al., 2023, Luo et al., 2 Nov 2025).
Governing role of microscopic topology: Whether via sliding-contact network topology in grains, record-driven quakes in jammed piles, or matching subgraph structure in graph algorithms, local configuration entropies control macroscopic compaction progress (Gago et al., 2020, Santos et al., 2023, Kawarabayashi et al., 2015).
Aging and memory effects: Non-ergodicity, protocol dependence, and memory retention of process history (e.g., annealing ramp rate, number of cycles, or compaction history) appear across granular and glassy systems (Gago et al., 2020, Luo et al., 23 Jan 2025).

Open questions remain in quantifying these universalities for more complex particle shapes, strongly coupled thermal-mechanical cycling, and systems with multi-field compaction processes (e.g., coupled grain friction and thermal expansion), as well as in the application of iterative compaction strategies to scaling challenges in machine learning and data storage.

References

(Clemmer et al., 2023) Clemmer et al., "A soft departure from jamming: the compaction of deformable granular matter under high pressures"
(Gago et al., 2020) Abate, Makse, "Universal Features of Annealing and Aging in Compaction of Granular Piles"
(Santos et al., 2023) Santos et al., "Granular packing simulation protocols: tap, press and relax"
(Luo et al., 2 Nov 2025) Schuerpf et al., "Competition between Glassy Five-Fold Structures and Locally Dense Packing Structures Governs Two-Stage Compaction of Granular Hexapods"
(O'Neill et al., 5 Jun 2026) Sun, Hoffmann, "Still: Amortized KV Cache Compaction in a Single Forward Pass"
(Luo et al., 23 Jan 2025) Luo et al., "Compaction of Granular Columns under Thermal Cycling"
(Sarkar et al., 2022) Jelmini et al., "Constructing and Analyzing the LSM Compaction Design Space"
(Mathieu et al., 2014) Iacono, Kamath, "Bigtable Merge Compaction"
(Kawarabayashi et al., 2015) Kawarabayashi, Li, Reed, "Connectivity Preserving Iterative Compaction and Finding 2 Disjoint Rooted Paths in Linear Time"
(Hubard et al., 2015) Lois, Bandi et al., "Experiments demonstrate that the null space of the rigidity matrix determines grain motion during vibration-induced compaction"
(Gago et al., 2010) Gago et al., "High intensity tapping regime in a frustrated lattice gas model of granular compaction"