Weak-Strong Distance Oracle Model

• The weak-strong distance oracle model is a metric-optimization framework that uses two types of queries—cost-effective weak oracles and expensive strong oracles—to approximate distances under adversarial corruption.
• It achieves provable approximations for classic problems such as k-center, k-median, MST, and fair clustering by strategically limiting costly strong queries while robustly aggregating noisy weak data.
• Key algorithmic techniques include heavy-light decompositions, median-based estimators, and assignment-preserving coresets that together stabilize inference from imprecise weak distances.

The weak-strong distance oracle model is a metric-optimization framework in which the underlying metric $d$ is not directly available to the algorithm. Instead, distance information is accessed through two heterogeneous interfaces: a weak oracle, which is cheap but unreliable, and a strong oracle, which is exact but expensive. The central algorithmic objective is therefore not merely to optimize running time, but to obtain provable approximations while minimizing the number of strong-oracle queries. The model was introduced for clustering and minimum spanning tree problems (Bateni et al., 2023), extended to fair clustering and coreset construction (Braverman et al., 14 Jun 2025), and subsequently refined with improved guarantees for $k$-means and $k$-center (Pradhan et al., 20 Feb 2026). It is distinct from other oracle formalisms that use the word “oracle,” such as exact alignment oracles for edit distance (Charalampopoulos et al., 2021) or abstention-based weak comparison oracles for content search (Kazemi et al., 2018).

1. Formal definition

In the standard formulation, one is given a metric space $(X,d)$ or $(\mathcal X,d)$ on $n$ points, but the algorithm does not know $d$ directly. It may query a weak oracle and a strong oracle (Bateni et al., 2023).

The weak oracle returns an approximate distance value $\tilde d(x,y)$. In the clustering/MST formulation, for any pair $(x,y)\in X^2$, it returns the true distance with probability $1-\delta$, and with probability $k$0 it may be arbitrarily corrupted; corruptions are independent across pairs and are fixed once drawn, so repeated queries to the same pair return the same weak value (Bateni et al., 2023). In the fair-clustering formulation, the weak oracle returns $k$1 with probability $k$2 and an arbitrary adversarial value with probability $k$3, again with randomness drawn once, so repeated queries and majority voting are not assumed to help (Braverman et al., 14 Jun 2025).

The strong oracle has two standard variants. A strong point query reveals a point’s exact strong information; once two points have both been queried, their true distance can be computed exactly. A strong edge query or distance query returns $k$4 directly for a specified pair $k$5 (Bateni et al., 2023). If an algorithm uses $k$6 strong point queries, then it can be simulated with $k$7 strong distance queries (Braverman et al., 14 Jun 2025).

Oracle type	Input	Guarantee
Weak oracle	$k$8	True distance with probability $k$9 or $k$0; otherwise arbitrary value
Strong point query	$k$1	Exact strong information for $k$2; exact distances among queried points become available
Strong edge/distance query	$k$3	Exact value $k$4

A defining feature of the model is that the weak oracle is not assumed to have benign additive or multiplicative noise. Failure events may be adversarial, and persistence means that re-querying does not average out errors (Bateni et al., 2023).

2. Optimization objectives and problem formulations

The model has been studied for several canonical metric problems. The original results cover $k$5-center, $k$6-median, $k$7-means, and MST (Bateni et al., 2023). The fair-clustering extension studies $k$8-fair clustering with overlapping groups $k$9, where $(X,d)$0 denotes the maximum number of groups any point belongs to (Braverman et al., 14 Jun 2025).

For clustering, the standard $(X,d)$1-objective is

$(X,d)$2

with $(X,d)$3 for $(X,d)$4-median and $(X,d)$5 for $(X,d)$6-means (Braverman et al., 14 Jun 2025). In the fair setting, for a center set $(X,d)$7 and assignment function $(X,d)$8, every group $(X,d)$9 and center $(\mathcal X,d)$0 must satisfy

$(\mathcal X,d)$1

This imposes disparate-impact style proportionality constraints within each cluster (Braverman et al., 14 Jun 2025).

The objective shift induced by the model is fundamental. In settings where exact similarity is expensive—large-scale embeddings, cross-attention-style scoring, or costly verification—the algorithmic question becomes: how many exact queries are really necessary to achieve a desired approximation guarantee (Bateni et al., 2023). This is the organizing principle behind both the original clustering/MST results and the later fair-clustering coreset constructions (Braverman et al., 14 Jun 2025).

3. Core algorithmic ideas

The earliest clustering algorithms in the model use robustification mechanisms that turn weak distances into stable surrogates. For $(\mathcal X,d)$2-center, the algorithm follows a sample, carve, and cover strategy based on heavy and light optimal clusters: a cluster is heavy if it contains at least $(\mathcal X,d)$3 points, so heavy clusters cover at least $(\mathcal X,d)$4 of the data. The method samples $(\mathcal X,d)$5 points, queries them strongly, performs greedy ball carving, and then uses medians of weak distances to estimate distances to heavy balls. Because each round covers a constant fraction of the remaining points, only $(\mathcal X,d)$6 rounds are needed (Bateni et al., 2023).

For $(\mathcal X,d)$7-median and $(\mathcal X,d)$8-means, the 2023 work builds a small weighted coreset using a heavy-ball nearest distance proxy. Instead of relying on a raw nearest sampled distance, it looks for a ball around a sampled point containing at least $(\mathcal X,d)$9 sampled points, takes a median of weak distances from the new point to those witnesses, and adds a safety margin proportional to the ball radius. This proxy upper-bounds the true distance to the current coreset and is robust to corruption (Bateni et al., 2023).

The fair-clustering paper develops a different but related framework based on assignment-preserving coresets. Its high-level strategy is ring-based sampling plus peeling: obtain a rough clustering using weak information, partition points around each center into distance rings, identify heavy rings using a small number of strong queries, sample from heavy rings to form the coreset, and peel off processed points recursively (Braverman et al., 14 Jun 2025). The key point is that fair clustering requires preserving not only ordinary costs but also feasible assignments under prescribed mass constraints.

The 2026 refinement introduces a reusable local estimator. For $n$0-means, if $n$1 contains at least $n$2 points, it defines

$n$3

and proves that, with high probability, $n$4. For a center set $n$5, it uses

$n$6

which is an upper bound on $n$7 with high probability (Pradhan et al., 20 Feb 2026). This estimator enables a weak-strong adaptation of oversampling $n$8-means++, where sampling probabilities are proportional to $n$9 rather than the exact $d$0-distribution.

A recurrent theme across these methods is that weak queries provide inexpensive global coverage, while strong queries are concentrated on small witness sets, sampled centers, or coreset support. This suggests a common design pattern: exact geometry is used sparingly to stabilize structures that are otherwise inferred from noisy distances.

4. Performance guarantees and lower bounds

The original clustering/MST paper gives constant-factor approximations with nearly optimal strong-query complexity (Bateni et al., 2023). For $d$1-center it obtains a $d$2-approximation using

$d$3

strong point queries. For $d$4-median and $d$5-means it gives $d$6-approximations using

$d$7

strong point queries, with an extra $d$8-type overhead appearing in the presentation after parameter guessing. The corresponding edge-query complexity is $d$9 (Bateni et al., 2023).

The same paper proves that these query bounds are essentially tight for bounded-approximation clustering: any bounded-approximation algorithm for $\tilde d(x,y)$0-center, $\tilde d(x,y)$1-median, or $\tilde d(x,y)$2-means needs at least

$\tilde d(x,y)$3

strong point queries, or equivalently

$\tilde d(x,y)$4

strong edge queries (Bateni et al., 2023).

For MST, the landscape is qualitatively different. If the weak-oracle distances themselves form a metric, then one can compute an $\tilde d(x,y)$5-approximation using no strong oracle queries at all. A matching lower bound shows that any better-than-$\tilde d(x,y)$6 approximation requires

$\tilde d(x,y)$7

strong point queries. In the non-metric weak-oracle setting, any $\tilde d(x,y)$8-approximation requires $\tilde d(x,y)$9 strong point queries (Bateni et al., 2023).

The 2026 work improves the clustering guarantees in the weak-strong setting using the median estimator described above (Pradhan et al., 20 Feb 2026).

Problem	Guarantee	Strong-query complexity
$(x,y)\in X^2$0-center	$(x,y)\in X^2$1-approximation	$(x,y)\in X^2$2 strong point queries
$(x,y)\in X^2$3-median / $(x,y)\in X^2$4-means	$(x,y)\in X^2$5-approximation	$(x,y)\in X^2$6 strong point queries
MST, metric weak oracle	$(x,y)\in X^2$7-approximation	$(x,y)\in X^2$8 strong queries
$(x,y)\in X^2$9-means, improved	Constant-factor approximation	$1-\delta$0 strong-oracle queries
$1-\delta$1-center, improved	$1-\delta$2-approximation	$1-\delta$3 strong-oracle queries

These improved results are reported in edge-query style. The $1-\delta$4-means algorithm adapts oversampling $1-\delta$5-means++ and proves a bi-criteria guarantee of

$1-\delta$6

with the same asymptotic query complexity; the $1-\delta$7-center algorithm is a weak-strong form of greedy ball carving and succeeds with probability at least $1-\delta$8 (Pradhan et al., 20 Feb 2026). The same paper also proves that in the weak-oracle-only model, any randomized algorithm giving an $1-\delta$9-approximation for $k$00-means with probability at least $k$01 requires

$k$02

queries in expectation (Pradhan et al., 20 Feb 2026).

5. Fair clustering and assignment-preserving coresets

The fair-clustering extension studies $k$03-fair clustering in the weak-strong oracle model and gives the first $k$04-coresets for fair $k$05-median using only

$k$06

queries to the strong oracle (Braverman et al., 14 Jun 2025).

The main structural notion is an assignment-preserving coreset. For a weighted set $k$07, such a coreset satisfies

$k$08

for every $k$09 and every valid assignment constraint $k$10 (Braverman et al., 14 Jun 2025). This is the correct abstraction because fairness constraints can be reduced to assignment constraints.

The reduction from assignment-preserving to fair coresets yields the scaling

$k$11

where $k$12 is the maximum number of groups any point belongs to (Braverman et al., 14 Jun 2025). For fair $k$13-median, the paper proves that with high probability it computes a $k$14 fair coreset of size

$k$15

using

$k$16

strong oracle point queries,

$k$17

weak queries, and time

$k$18

(Braverman et al., 14 Jun 2025).

For the disjoint-group or assignment-preserving case, it gives size

$k$19

using

$k$20

strong point queries and

$k$21

weak queries (Braverman et al., 14 Jun 2025). The framework also extends to general $k$22-clustering for constant $k$23, yielding a $k$24-coreset of size roughly

$k$25

with the appendix theorem stating

$k$26

and similar strong-oracle complexity (Braverman et al., 14 Jun 2025).

These results show that fairness constraints do not force a prohibitive exact-query cost. In this model, fairness can be handled through coreset machinery that remains strongly query-efficient.

6. Relation to other oracle models and common distinctions

The weak-strong distance oracle model should not be conflated with every use of the term “distance oracle.” In the exact alignment setting of edit distance and LCS, the oracle in question is a preprocessed exact shortest-path oracle on a planar alignment graph. That work explicitly does not introduce a formal weak-vs-strong distance oracle distinction; rather, it preprocesses two strings so that exact substring-alignment queries can be answered in $k$27 preprocessing time and $k$28 query time, or with $k$29 space and $k$30 query time (Charalampopoulos et al., 2021). A plausible implication is that this alignment oracle is best viewed as an exact specialized planar-graph distance oracle, not as an instance of the weak-strong model.

A second nearby but distinct line of work studies weak comparison oracles rather than weak distance oracles. In that model, a query asks which of two objects is closer to a hidden target, and the oracle may abstain when the comparison is not sufficiently separated. Formally, for parameter $k$31, the oracle returns $k$32 if $k$33, returns $k$34 symmetrically, and may return $k$35 in the ambiguous zone. The oracle is non-malicious: it either answers correctly when confident or abstains (Kazemi et al., 2018). This is qualitatively different from the weak-strong distance model, where the weak oracle may output an arbitrary corrupted numeric distance.

The most persistent misconception concerns the meaning of “weak.” In the weak-strong distance oracle model, weakness does not mean small unbiased noise or an approximation guarantee of the form $k$36. It means that some weak answers may be adversarially corrupted, and that those corruptions may be persistent across repeated queries (Bateni et al., 2023). This is precisely why the literature relies on medians over carefully chosen witness sets, heavy/light decompositions, assignment-preserving coresets, and localized exact querying, rather than on simple repetition or majority voting (Braverman et al., 14 Jun 2025).

Taken together, the literature positions the weak-strong distance oracle model as a query-complexity framework for metric optimization under heterogeneous access to geometry: cheap but unreliable weak distances, expensive but exact strong distances, and algorithms designed to exploit that asymmetry with provable guarantees.