Exlump Aggregations: Controlled Procedures

• Exlump aggregations are controlled procedures that condense fine-grained objects (e.g., spatial points, quadratic constraints) into coarser representations while preserving key properties such as local proximity or convexity.
• They are applied across domains—from generating spatial heatmaps via the Numericized Histogram Score and convex-hull recovery in quadratic programs to dynamic grouping in split-apply-combine systems and ε-approximate fluid lumpability in process algebras.
• These methods explicitly model invariants, error bounds, and trade-offs (e.g., bias versus variance), ensuring that the aggregation process delivers computational tractability with controlled approximation.

Searching arXiv for the cited papers to ground the article in current metadata. Exlump aggregations denote a family of aggregation and lumping procedures that replace heterogeneous, fine-grained objects by coarser constructs while preserving a task-specific semantics such as local spatial proximity, valid convex-hull relaxations, statistically adequate grouped estimates, reduced ODE dynamics, or compact summaries of aggregate-query outputs. Across the cited work, the phrase appears in several distinct but related senses: hidden preferential spatial aggregation quantified by the Numericized Histogram Score (NHS), nonnegative aggregation of quadratic inequalities, split-apply-combine with dynamic grouping, and $\epsilon$-approximate fluid lumpability; related database systems extend the same operational idea to summarizing and explaining aggregate answers (Nguyen, 2017, Dey et al., 2021, Loo, 2024, Tschaikowski et al., 2014, Wen et al., 2018, Savva et al., 2018).

1. Scope and conceptual structure

A common misconception would be to treat exlump aggregation as a single standardized algorithm. Across the cited work, it is not. The term is used alongside “extended-lumping,” “split-apply-combine with dynamic grouping,” “$\epsilon$-approximate fluid lumpability,” and practical advice on “exlumping” large sets of aggregate answers. This suggests that the expression functions as an umbrella label for controlled aggregation procedures rather than as a unique formalism (Loo, 2024, Tschaikowski et al., 2014, Wen et al., 2018).

A plausible common structure is the following. One begins with a fine-grained object family: spatial points, quadratic constraints, group labels, process-algebra derivatives, or aggregate-query answers. One then defines an aggregation operator: histogram numericization, nonnegative multiplier combination, hierarchical collapse maps, lumpable partitions, or cluster generalization. The result is retained only if a target property is preserved or certified, such as a normalized heatmap score, coverage of $\operatorname{conv} S$, satisfaction of a quality test $\beta$, an $O(\epsilon)$ trajectory bound, or guaranteed coverage of the top-$L$ answers. In all cases, aggregation is not merely compression; it is a mechanism for making hidden structure computationally tractable.

2. Spatial preferential aggregation via the Numericized Histogram Score

In spatial point data, exlump aggregation is used to detect “hidden preferential aggregations amid heterogeneity.” The NHS algorithm converts the histogram distribution of shortest distances between objects into a continuous variable that can be rendered as a spatial heatmap. For every ordered class pair $A \to B$, one selects a focal object $p \in A$, computes Euclidean distances from $p$ to every object in class $B$, sorts them, takes the $\epsilon$0 smallest distances, and bins those distances into a fixed set of $\epsilon$1 contiguous bins with shared edges across the analysis. If $\epsilon$2 is the count in bin $\epsilon$3 and $\epsilon$4 is a monotonic decreasing weight, then the adjusted proximity score is $\epsilon$5, the maximum possible adjusted proximity is $\epsilon$6, and the NHS saturation score is

$\epsilon$7

A standard linear weighting is

$\epsilon$8

with $\epsilon$9; the described implementation used $\operatorname{conv} S$0 bins, $\operatorname{conv} S$1, $\operatorname{conv} S$2, and $\operatorname{conv} S$3 (Nguyen, 2017).

The method then “repaints” each object by its NHS score. Two visualization modes are described. In the single-color transparency option, $\operatorname{conv} S$4 is mapped to opacity $\operatorname{conv} S$5 of a false color, so that $\operatorname{conv} S$6 appears fully opaque and $\operatorname{conv} S$7 nearly transparent. In the diverging two-color option, one first computes the uniform-distribution threshold $\operatorname{conv} S$8 from a hypothetical histogram with $\operatorname{conv} S$9 for all $\beta$0, optionally applies a monotonic transform such as $\beta$1, centers the transformed score as $\beta$2, and then maps negative and positive deviations around that pivot to opposite sides of a diverging color scale. The uniform threshold must be computed with the same $\beta$3 and $\beta$4 as the real data, and if $\beta$5 is monotonic then relative ordering is preserved (Nguyen, 2017).

The significance of NHS lies in its explicitly spatial character. PCA and MCA provide information about which features in multidimensional data aggregate, but they do not encode spatial coordinates and do not provide in situ spatial information about those aggregations. NHS instead condenses each object’s local distance-distribution shape into a scalar and reprojects that scalar into the original geometry. The resulting heatmaps reveal loco-regional “pockets” of strong co-aggregation or relative dispersion that would be averaged out in global eigenspace methods. The paper frames digital pathology as a primary use case and also points to microbial biofilms, neuron–glia interactions, multi-lineage tumor cell tracking, immune cell–pathogen encounters, subcellular granule or protein cluster mapping, ecology, and urban analytics. The stated caveats are equally central: the choice of $\beta$6 trades off sensitivity against noise; bin edges and $\beta$7 determine the relevant length scales; $\beta$8 may be linear, exponential, or custom-tuned; Euclidean geometry is assumed unless distance is redefined; dense or sparse regions may saturate raw NHS; and dataset comparisons require identical $\beta$9, bin edges, and weight schedules (Nguyen, 2017).

3. Aggregation of quadratic inequalities and convex-hull recovery

In nonconvex quadratic programming, exlump aggregation refers to the construction of valid quadratic inequalities by nonnegative aggregation of the inequalities defining a feasible region. If

$O(\epsilon)$0

then an aggregation is a single quadratic inequality

$O(\epsilon)$1

with associated relaxation

$O(\epsilon)$2

The intended use is to choose multipliers $O(\epsilon)$3 so that $O(\epsilon)$4 is valid for $O(\epsilon)$5, or so that an intersection of aggregated sets exactly recovers $O(\epsilon)$6 (Dey et al., 2021).

The sharpest positive result is for two quadratics. Under mild regularity, including nonemptiness and the exclusion of the case in which all $O(\epsilon)$7 are simultaneously PSD, the convex hull is obtained by at most two aggregated inequalities:

$O(\epsilon)$8

In homogeneous form, the underpinning is the classical S-lemma for two quadratics:

$O(\epsilon)$9

For three strict quadratics, the cited result requires the positive definite linear combination condition

$L$0

and $L$1. Defining

$L$2

where $L$3 encodes “at most one negative eigenvalue” so that $L$4 is a “semi-convex cone” after homogenization, one obtains

$L$5

An analogous closed-inequality statement holds for

$L$6

under PDLC, $L$7, and the condition $L$8 (Dey et al., 2021).

The limitations are structurally important. The three-quadratic theorem may require a potentially infinite family of aggregated inequalities. The paper also gives counterexamples showing failure beyond three quadratics and failure when PDLC is violated: there are sets described by four strict quadratics in $L$9 for which

$A \to B$0

so no finite or infinite family of such aggregated inequalities recovers the convex hull. Algorithmically, separation of a point $A \to B$1 reduces to a small SDP of size $A \to B$2, and aggregated cuts $A \to B$3 can be generated dynamically inside cutting-plane methods, global nonlinear solvers, or branch-and-cut for mixed-integer quadratics (Dey et al., 2021).

4. Dynamic grouping in split-apply-combine systems

In data analysis, exlump aggregation appears as split-apply-combine with dynamic grouping. The formal problem begins with a finite record set $A \to B$4, an aggregation function $A \to B$5, a finite set of target labels $A \to B$6, and a surjection $A \to B$7 assigning each record to an initial group. It also requires a Boolean test $A \to B$8 that decides whether a subset is “large enough” or otherwise of “sufficient quality,” together with a collapsing scheme

$A \to B$9

With composites

$p \in A$0

and pullbacks

$p \in A$1

the objective is: for each target label $p \in A$2, find the smallest $p \in A$3 such that

$p \in A$4

and output the tuple $p \in A$5. If no such $p \in A$6 exists, the result is missing or NA (Loo, 2024).

The algorithmic interpretation is a “fall through” hierarchy of coarsenings. Each initial group is tested at level $p \in A$7; if the test fails, the group label is collapsed to the next coarser level, the corresponding pullback subset is recomputed, and the test is applied again. The paper states best-case time $p \in A$8 when every label passes at $p \in A$9, worst-case time $p$0 when every label fails until depth $p$1, and overall time $p$2 when $p$3 is constant. Space is linear in $p$4, and intermediate subsets can be streamed (Loo, 2024).

The implementation is provided by the R package accumulate. The described examples include a toy collapse sequence $p$5 with min_records(3), and a small-area-estimation example in which mean turnover per (sbi,size) is required to have at least 10 non-missing industrial values, with fallback to sbi, then sbi2, then sbi1. The guarantees and trade-offs are explicit: every returned aggregate satisfies $p$6; if no collapse succeeds, no unsafe guess is returned; collapsing to coarser groups reduces variance but can introduce bias; and any Boolean test on subset statistics, as well as any multi-step or unbalanced hierarchy, can be encoded via the formula or table interface (Loo, 2024).

5. Extended differential aggregations and $p$7-approximate fluid lumpability

In process algebra and reaction-network semantics, exlump aggregation is formalized as $p$8-approximate fluid lumpability. The framework is FEPA (Fluid Extended Process Algebra), in which a model is built from sequential fluid atoms with syntax

$p$9

and parallel composition $B$0 synchronizes on $B$1 using either $B$2 for PEPA semantics or $B$3 for the law of mass action. A population function over the derivative set evolves according to the induced ODE, and FEPA is well-posed if every synchronization in $B$4 can fire in isolation somewhere, ensuring that the ODE is globally Lipschitz and has a unique solution (Tschaikowski et al., 2014).

Two exact notions precede the approximate one. Exact fluid lumpability (EFL) requires a partition of fluid atoms into blocks such that, under identical initial conditions and suitable bijections between derivative sets, corresponding trajectories remain identical for all $B$5. Ordinary fluid lumpability (OFL) is weaker: only the sums of trajectories within each block must be recovered exactly by a smaller autonomous ODE. Both notions can be characterized by semi-isomorphism of derivation graphs, both are congruences for $B$6, and both admit efficient, often bottom-up, algorithms to find maximal lumpable partitions (Tschaikowski et al., 2014).

$B$7-approximate lumpability relaxes the perfect symmetry requirement. If one can perturb the model parameters by at most $B$8 so as to restore exact or ordinary lumpability, then the original model stays close to the perturbed lumpable one. For two ODE systems

$B$9

with $\epsilon$00 and $\epsilon$01, the paper states

$\epsilon$02

For FEPA, the resulting perturbation theorem gives constants $\epsilon$03 such that

$\epsilon$04

Thus, over a fixed finite interval, the deviation is $\epsilon$05 when initial conditions coincide (Tschaikowski et al., 2014).

The practical procedure is to identify candidate blocks of near-isomorphic fluid atoms, define a nominal parameter vector $\epsilon$06 by averaging or selecting representative rates within each block, and solve the reduced ODE for the perturbed exactly lumpable model $\epsilon$07. The original model $\epsilon$08 is then $\epsilon$09-approximately lumpable. The paper’s worked example uses mass-action semantics with rates

$\epsilon$10

nominal rates $\epsilon$11 and $\epsilon$12, and initial populations

$\epsilon$13

After replacing all $\epsilon$14 by their average, numerical integration over $\epsilon$15 with time-step $\epsilon$16 shows that for $\epsilon$17 the $\epsilon$18-EFL error grows linearly in $\epsilon$19 and remains below $\epsilon$20 up to $\epsilon$21, while the $\epsilon$22-OFL error is negligible, $\epsilon$23, across the entire $\epsilon$24 range. Similar behavior is reported under PEPA min-semantics, with slightly larger errors (Tschaikowski et al., 2014).

6. Summarizing and explaining aggregate answers

Database-oriented work extends the same lumping intuition to the outputs of aggregate queries. One line of work summarizes high-valued GROUP BY answers by clusters over attribute values and “don’t care” symbols. If $\epsilon$25 is the output of an aggregate SQL query ordered by descending value, and $\epsilon$26 denotes the top-$\epsilon$27 tuples, then each cluster is an $\epsilon$28-tuple

$\epsilon$29

covering tuple $\epsilon$30 when each component either matches $\epsilon$31 or is *. The optimization problem seeks a set $\epsilon$32 maximizing

$\epsilon$33

subject to four constraints: $\epsilon$34, coverage of all top-$\epsilon$35 tuples, pairwise diversity

$\epsilon$36

and incomparability so that no cluster in $\epsilon$37 covers another. The decision and optimization versions are NP-hard once $\epsilon$38 or $\epsilon$39 is part of the input. The paper exploits the semi-lattice induced by the coverage order, uses least common ancestors for merges, and proposes Bottom-Up Greedy, Fixed-Order Greedy, and Hybrid algorithms, together with delta-judgment, lazy cluster generation, and hashing. The reported experiments give initialization in $\epsilon$40 ms–$\epsilon$41 s as $\epsilon$42 rises to $\epsilon$43, run times $\epsilon$44 ms on MovieLens and $\epsilon$45 s on TPC-DS, and speedups of $\epsilon$46 to $\epsilon$47 from the optimizations. The user study with $\epsilon$48 reports that about $\epsilon$49 of subjects preferred the cluster-based summarization over decision trees (Wen et al., 2018).

A second line of work explains aggregate queries for exploratory analytics by replacing scalar answers with local parametric functions. XAXA models a center-radius selection

$\epsilon$50

and an aggregate response

$\epsilon$51

It jointly optimizes location representatives $\epsilon$52, radius representatives $\epsilon$53, and piecewise-linear parameters $\epsilon$54 through the sum of a center-quantization objective $\epsilon$55, a radius-quantization objective $\epsilon$56, and local squared-error losses $\epsilon$57. Within each cell $\epsilon$58, the explanation takes the max-hinge form

$\epsilon$59

Training is performed online by monitoring aggregate queries and their answers, without database access during explanation serving. The paper reports $\epsilon$60 and NRMSE $\epsilon$61 for COUNT queries, $\epsilon$62 and NRMSE $\epsilon$63 for AVG queries, KL-divergence below $\epsilon$64 bits per query, cosine similarity of local slopes above $\epsilon$65, per-explanation latency under $\epsilon$66 ms on a single core, model size around $\epsilon$67 MB, and online-phase training throughput above $\epsilon$68 queries/sec (Savva et al., 2018).

These database formulations do not present exlump aggregation as a canonical term. A plausible implication is that they generalize the same underlying operation: compress a large answer space into structured surrogates that preserve coverage, diversity, or predictive fidelity, thereby enabling interactive exploration without forcing direct inspection of all original aggregates.

7. Limits, guarantees, and recurring trade-offs

Across these domains, exlump aggregation is governed by explicit admissibility conditions rather than by unrestricted merging. In NHS, the parameters $\epsilon$69, $\epsilon$70, bin edges, and weight schedule determine the length scales and sensitivity of the spatial score. In quadratic convexification, the positive theorems depend on the number of inequalities and, for three quadratics, on PDLC and auxiliary spectral conditions. In dynamic grouping, every returned estimate must satisfy the user-defined test $\epsilon$71, and failure to pass the test yields NA rather than extrapolation. In $\epsilon$72-lumpability, the reduced model is justified only to the extent that perturbations are small and the Lipschitz growth factor remains controlled. In aggregate summarization and explanation, feasibility and interpretability are traded against optimization hardness, model complexity, or approximation error (Nguyen, 2017, Dey et al., 2021, Loo, 2024, Tschaikowski et al., 2014, Wen et al., 2018, Savva et al., 2018).

The resulting controversies are methodological rather than rhetorical. One concerns locality versus globality: NHS emphasizes in situ geometry that PCA and MCA omit, whereas query-answer summarization emphasizes coverage and diversity over exact enumeration. Another concerns exactness versus approximation: two-quadratic aggregation and OFL/EFL provide exact structural reduction under specific conditions, whereas three-quadratic aggregation may require infinitely many cuts and $\epsilon$73-lumpability provides only controlled error bounds. A third concerns bias versus stability: dynamic grouping guarantees minimum support but can introduce the classic bias–variance trade-off of small area estimation. Taken together, these patterns indicate that exlump aggregation is best understood as a technical design principle: aggregate only through a map whose invariants, error bounds, or failure modes are themselves explicitly modeled.