Composable Sets: Theory and Applications

• Composable sets are set-valued summaries that allow independent outputs to be merged while maintaining key objectives such as determinant maximization or submodular utility.
• They leverage diverse constructs—core-sets, witness sets, and compressed representations—to provide efficient and approximate solutions in distributed and streaming settings.
• Applications span optimization, uncertainty modeling, and query processing, enabling robust performance through preservation of crucial invariants under composition.

Composable sets are set-valued summaries, representations, or operators designed so that independently produced outputs can be merged, or sequentially composed, without leaving the underlying formalism. In optimization, the dominant meaning is the composable core-set: each block of a partitioned dataset is compressed to a small subset whose union still preserves the objective up to a prescribed approximation factor. In other research lines, composability means closure of set-valued queries under function composition, exact or approximate combination of overlapping credal sets, compact binary encodings derived from final witness sets, or compressed integer sets that support bulk set algebra directly on compressed words (Gollapudi et al., 2023, Vejnarová et al., 2017, Phadke, 25 Nov 2025, Colantonio et al., 2010, Johnson et al., 2019).

1. Formal scope of composability

The literature does not use a single universal definition of composable sets. Instead, several mathematically distinct notions recur. In composable core-sets, a map returns a subset of each local dataset, and the union of those subsets preserves an objective approximately. In credal-set theory, a binary operator combines two convex sets of probability distributions on overlapping variable domains and returns another credal set on the union domain. In CRAQL, a query is a function $Q:\mathcal P(\mathcal T)\to\mathcal P(\mathcal T)$, so outputs remain sets of AST subtrees and can therefore be fed into subsequent queries. In REWA, arbitrary transformation pipelines terminate in finite witness sets $W(v)$, and the theory applies only to the final witness overlap. In CONCISE, the emphasis is operational rather than approximation-theoretic: compressed integer sets remain closed under union, intersection, XOR, complement, and difference without decompression (Gollapudi et al., 2023, Vejnarová et al., 2017, Phadke, 25 Nov 2025, Colantonio et al., 2010, Johnson et al., 2019).

Setting	Formal object	Composition mechanism
Composable core-sets	subset-valued map on data blocks	union of local outputs preserves objective approximately
Credal sets	$M_1\triangleright M_2$	operator on overlapping marginal models
CRAQL	$Q:\mathcal P(\mathcal T)\to\mathcal P(\mathcal T)$	function composition $Q_2\circ Q_1$
REWA	final finite witness sets $W(v)$	transformation pipelines ending in witness sets
CONCISE	compressed bitmap of an integer set	bitwise set algebra on compressed words

A plausible implication is that composability is best understood as a closure principle coupled to a preservation guarantee. The preserved quantity depends on the domain: determinant value, submodular utility, top-$k$ ranking, marginal consistency, or exact set-algebra semantics.

2. Composable core-sets as distributed set summaries

For determinant maximization, an $\alpha$-composable core-set is defined by a mapping

$$c:\{\text{finite subsets of }\mathbb R^d\}\to \{\text{subsets of same size}\}$$

such that for every collection $P_1,\dots,P_m$,

$W(v)$0

Here each $W(v)$1 is compressed to a small subset, often of size $W(v)$2, and the union of the local outputs approximates the global optimum on the full union (Gollapudi et al., 2023).

The same pattern appears in randomized composable core-sets for submodular maximization. Under random clustering, an algorithm $W(v)$3 that outputs at most $W(v)$4 elements from each local set is an $W(v)$5-approximate randomized composable core-set if

$W(v)$6

The expectation is over the random partition, and $W(v)$7 denotes the best value attainable by a subset of $W(v)$8 of size at most $W(v)$9 (Mirrokni et al., 2015).

A closely related definition is used for diversity on multi-dataset streams. For a nonnegative diversity function $M_1\triangleright M_2$0, a randomized algorithm $M_1\triangleright M_2$1 is an $M_1\triangleright M_2$2-approximate composable core-set construction if for every partition $M_1\triangleright M_2$3, with $M_1\triangleright M_2$4,

$M_1\triangleright M_2$5

In that work, $M_1\triangleright M_2$6 is the remote-edge diversity $M_1\triangleright M_2$7 (Wang et al., 2023).

These definitions induce the same systems interpretation. Each machine or stream shard emits a local set summary, communication is proportional to the size of the summary rather than the original data, and a central coordinator solves the reduced instance on the union of the summaries. In determinant maximization this yields one-round distributed or streaming algorithms; in submodular maximization it gives two-round MapReduce and random-order streaming algorithms; in diversity streaming it motivates online summarization of multiple streams (Indyk et al., 2019, Mirrokni et al., 2015, Wang et al., 2023).

3. Determinant maximization as the canonical case

Given $M_1\triangleright M_2$8 and $M_1\triangleright M_2$9, determinant maximization asks for a size-$Q:\mathcal P(\mathcal T)\to\mathcal P(\mathcal T)$0 subset whose squared volume

$Q:\mathcal P(\mathcal T)\to\mathcal P(\mathcal T)$1

is maximized. Equivalently,

$Q:\mathcal P(\mathcal T)\to\mathcal P(\mathcal T)$2

The problem is also the MAP-inference task for determinantal point processes, since a $Q:\mathcal P(\mathcal T)\to\mathcal P(\mathcal T)$3-DPP assigns each $Q:\mathcal P(\mathcal T)\to\mathcal P(\mathcal T)$4-subset probability proportional to the determinant of its Gram matrix (Gollapudi et al., 2023, Indyk et al., 2019).

Three local constructions dominate the composable-coreset literature for this problem. Spectral spanners provide the first near-optimal approximation of order $Q:\mathcal P(\mathcal T)\to\mathcal P(\mathcal T)$5; the earlier greedy analysis yielded an $Q:\mathcal P(\mathcal T)\to\mathcal P(\mathcal T)$6 composable guarantee; local search improved this to $Q:\mathcal P(\mathcal T)\to\mathcal P(\mathcal T)$7; and the more recent greedy analysis shows that greedy itself is a $Q:\mathcal P(\mathcal T)\to\mathcal P(\mathcal T)$8-composable coreset with an explicit factor $Q:\mathcal P(\mathcal T)\to\mathcal P(\mathcal T)$9, improving over the previous $Q_2\circ Q_1$0 bound (Indyk et al., 2018, Indyk et al., 2019, Gollapudi et al., 2023).

Method	Local output size	Composable guarantee
Spectral spanner	$Q_2\circ Q_1$1	$Q_2\circ Q_1$2
Greedy, earlier analysis	$Q_2\circ Q_1$3	$Q_2\circ Q_1$4
Local Search	$Q_2\circ Q_1$5	$Q_2\circ Q_1$6
Greedy, improved analysis	$Q_2\circ Q_1$7	$Q_2\circ Q_1$8

The improved greedy analysis is driven by a local-optimality theorem. If $Q_2\circ Q_1$9 and $W(v)$0, then for every $W(v)$1,

$W(v)$2

Thus no single swap can increase the volume by more than $W(v)$3. Earlier work had already shown that any $W(v)$4-locally-optimal mapping yields an $W(v)$5-composable coreset with

$W(v)$6

and substituting $W(v)$7 gives the greedy guarantee above (Gollapudi et al., 2023).

Empirical results sharpen the contrast between worst-case and observed behavior. On random $W(v)$8-point samples from MNIST or GENES, for $W(v)$9 up to $k$0 the largest swap improvement $k$1 remains below $k$2. Even when $k$3 grows to $k$4 on smaller point sets, the observed $k$5 stays under $k$6, and when the base-set size varies from $k$7 to $k$8 for fixed $k$9, the measured $\alpha$0 is always under $\alpha$1. Earlier experiments had also found that Local Search improves over Greedy by $\alpha$2–$\alpha$3 offline, and that as composable core-sets Local Search improves over Greedy by $\alpha$4–$\alpha$5 on average, while outperforming the LP-based method in $\alpha$6 of runs (Gollapudi et al., 2023, Indyk et al., 2019).

4. Extensions: constraints, randomization, and streaming diversity

The determinant-maximization framework extends to partition constraints and more general matroidal structure. For $\alpha$7 partitioned into $\alpha$8, with quotas $\alpha$9 summing to $$c:\{\text{finite subsets of }\mathbb R^d\}\to \{\text{subsets of same size}\}$$0, constrained determinant maximization asks for

$$c:\{\text{finite subsets of }\mathbb R^d\}\to \{\text{subsets of same size}\}$$1

When $$c:\{\text{finite subsets of }\mathbb R^d\}\to \{\text{subsets of same size}\}$$2, a local maximum of size $$c:\{\text{finite subsets of }\mathbb R^d\}\to \{\text{subsets of same size}\}$$3 per part yields a coreset of size $$c:\{\text{finite subsets of }\mathbb R^d\}\to \{\text{subsets of same size}\}$$4 with approximation $$c:\{\text{finite subsets of }\mathbb R^d\}\to \{\text{subsets of same size}\}$$5, and a matching $$c:\{\text{finite subsets of }\mathbb R^d\}\to \{\text{subsets of same size}\}$$6 lower bound shows that no smaller composable coreset with finite approximation exists in this regime. When $$c:\{\text{finite subsets of }\mathbb R^d\}\to \{\text{subsets of same size}\}$$7, a peeling algorithm gives a coreset of size $$c:\{\text{finite subsets of }\mathbb R^d\}\to \{\text{subsets of same size}\}$$8 with approximation $$c:\{\text{finite subsets of }\mathbb R^d\}\to \{\text{subsets of same size}\}$$9, and the paper states that this approximation factor is tight. The same framework extends to strongly Rayleigh distributions, several experimental-design problems, and laminar matroid constraints, where a coreset of size at most $P_1,\dots,P_m$0 with approximation $P_1,\dots,P_m$1 is obtained for cover number $P_1,\dots,P_m$2 (Mahabadi et al., 2022).

Randomization changes the feasible approximation landscape for submodular maximization. With random clustering, a simple greedy algorithm yields a $P_1,\dots,P_m$3-approximate randomized composable core-set of size $P_1,\dots,P_m$4 under a cardinality constraint. By allowing each machine to select $P_1,\dots,P_m$5 items, with $P_1,\dots,P_m$6, and using the $P_1,\dots,P_m$7 post-processing, the distributed factor improves to $P_1,\dots,P_m$8 for monotone submodular maximization. This contrasts with a known $P_1,\dots,P_m$9 impossibility result for non-randomized composable core-sets of size $W(v)$00 (Mirrokni et al., 2015).

Streaming diversity gives a different operational profile. For remote-edge diversity, the paper presents two brute-force composable constructions on multiple streams. The first, Brute-Force $W(v)$01-Replacement, uses $W(v)$02 time and $W(v)$03 space per stream. The second, $W(v)$04-Adjacency Replacement, reduces each insertion to $W(v)$05 time by storing a lower-triangular adjacency matrix, at the cost of $W(v)$06 space. The work recalls the standard two-set composable bound

$W(v)$07

for remote-edge diversity and reports runtime regressions on learned $W(v)$08-dimensional CIFAR-10 vectors split into $W(v)$09 streams of $W(v)$10 each (Wang et al., 2023).

5. Witness sets, compressed integer sets, and composable query sets

REWA formulates composability around finite witness sets. Each $W(v)$11 is assigned a finite set $W(v)$12, with $W(v)$13, and similarity is driven by witness overlap

$W(v)$14

A REWA encoder maps each witness to $W(v)$15 bit positions in an $W(v)$16-bit vector $W(v)$17, under semi-random bit assignments and conditional independence, so that for $W(v)$18,

$W(v)$19

Under an overlap-gap condition between true top-$W(v)$20 neighbors and non-neighbors, the main reducibility theorem gives

$W(v)$21

or in big-$W(v)$22 notation $W(v)$23, sufficient for preserving the top-$W(v)$24 ranking with probability at least $W(v)$25. The central compositional claim is that the theory does not inspect how $W(v)$26 was built: any sequence

$W(v)$27

is valid once the final witness sets are finite and satisfy the overlap gap (Phadke, 25 Nov 2025).

CONCISE addresses composability for integer sets stored as compressed bitmaps. Using 32-bit words with literal and fill representations, including a mixed-fill case with one exception bit, it supports UNION, INTERSECTION, XOR, AND-NOT, and COMPLEMENT directly on compressed words. Building is $W(v)$28, binary operations are $W(v)$29 in the compressed lengths of the inputs, and COMPLEMENT is $W(v)$30. Compared with WAH, CONCISE reduces required memory up to $W(v)$31, has worst-case $W(v)$32 word per set bit rather than $W(v)$33, and in the experiments matches or improves on WAH in computation time while remaining suitable for batch set algebra in data warehousing, information retrieval, and data mining (Colantonio et al., 2010).

CRAQL shows a third notion of composability in which the carrier itself is a set domain. If $W(v)$34 is the universe of AST subtrees, a query is

$W(v)$35

and composition is ordinary function composition: $W(v)$36 Closure, associativity, and identity follow directly from that typing discipline. Because inputs and outputs are both sets of ASTs, complex repository analyses can be chained while retaining tree-specific operations such as outmost, inmost, and directly in; the implementation further uses node-type indexes, binding caches, and position/span indexes to make such composition efficient (Johnson et al., 2019).

6. Composition of credal sets

In imprecise-probability theory, a credal set $W(v)$37 is a closed convex set of probability distributions on $W(v)$38, represented as the convex hull of its extreme points. Given credal sets $W(v)$39 and $W(v)$40, their composition $W(v)$41 is a credal set on $W(v)$42. The construction compares the common marginals on $W(v)$43 using Euclidean distance; when absolute continuity holds, the combined distribution takes the Bayes-type form

$W(v)$44

and otherwise one uses extreme points of the vacuous extension of $W(v)$45 (Vejnarová et al., 2017).

Several structural properties are explicit. The operator is always defined. If the two credal sets are projective, meaning that they agree on the common marginal,

$W(v)$46

then composition reduces to the common extension and is symmetric. If $W(v)$47, one recovers the strong-independence product

$W(v)$48

The operator satisfies marginalization, idempotence, and associativity on perfect sequences, but is non-commutative in general when the common marginals differ (Vejnarová et al., 2017).

This version of composability differs sharply from union-based core-sets. The output is not a smaller summary; it is a higher-dimensional model obtained by gluing together lower-dimensional assessments. A plausible implication is that “composable sets” spans both lossy summarization and exact model synthesis, depending on whether the preservation target is objective value or probabilistic coherence.

7. Limits, lower bounds, and recurring misconceptions

A recurring misconception is that composability implies exact preservation. The dominant results are approximate. For determinant maximization, spectral spanners achieve $W(v)$49, which is almost optimal in the worst case, and the corresponding lower bound shows that no method of size $W(v)$50 can beat $W(v)$51 up to lower-order terms. For constrained determinant maximization, no smaller coreset than size $W(v)$52 can achieve finite approximation when $W(v)$53, while for $W(v)$54 the stated $W(v)$55 behavior is tight (Indyk et al., 2018, Mahabadi et al., 2022).

A second misconception is that local optimization automatically composes. The submodular literature gives the clearest counterpoint: non-randomized core-sets of size $W(v)$56 face an $W(v)$57 impossibility barrier, whereas random clustering restores constant-factor guarantees such as $W(v)$58 and $W(v)$59. Likewise, the determinant-maximization literature required either spectral spanners, local search, or a new swap analysis of greedy before near-optimal composability guarantees could be proved (Mirrokni et al., 2015, Gollapudi et al., 2023).

A third misconception is that composition is always commutative or semantics-preserving without hypotheses. Credal-set composition is non-commutative in general, REWA requires an overlap-gap condition together with min-entropy and conditional-independence assumptions, and CONCISE sacrifices random-access efficiency even as it preserves bulk set algebra. REWA also guarantees only top-$W(v)$60 ranking preservation, not exact distance reconstruction, and explicitly notes that $W(v)$61, $W(v)$62, and $W(v)$63 may be hard to estimate in advance and that the framework is static rather than adaptive to concept drift (Vejnarová et al., 2017, Phadke, 25 Nov 2025, Colantonio et al., 2010).

Taken together, these results suggest a stable taxonomy. Union-based composable core-sets are approximation devices for large-scale optimization; witness and bitmap constructions are composable representations for similarity and set algebra; CRAQL provides closure of query semantics under composition; and credal-set composition is an operator for assembling multidimensional uncertainty models. The unifying thread is not a single algorithmic recipe, but the preservation of a useful invariant under set-level combination.