Multidimensional Data Structures and Techniques for Efficient Decision Making

In this paper we present several novel efficient techniques and multidimensional data structures which can improve the decision making process in many domains. We consider online range aggregation, range selection and range weighted median queries; for most of them, the presented data structures and techniques can provide answers in polylogarithmic time. The presented results have applications in many business and economic scenarios, some of which are described in detail in the paper.

💡 Research Summary

The paper tackles a fundamental challenge in modern data‑driven decision‑making: answering complex multidimensional queries on dynamic data sets with sub‑logarithmic overhead. Three canonical query types are examined—range aggregation (sums, averages, minima, maxima), range selection (k‑th smallest or largest element within a hyper‑rectangle), and range weighted median (the smallest value whose cumulative weight exceeds half of the total weight inside the query range). While classic structures such as range trees, kd‑trees, R‑trees, and wavelet trees provide logarithmic or polylogarithmic query times for static data, they either degrade dramatically when updates are required or become inefficient for the weighted‑median problem, which is inherently non‑linear.

The authors introduce two intertwined data structures that together overcome these limitations. The first, the Multi‑Level Fractional Segment Tree (MFST), builds a hierarchy of segment trees—one per dimension—and links them using fractional cascading. In a conventional range tree, a query in d dimensions incurs O(logⁿ n) node visits per level, leading to O(log^{d} n) time. By applying fractional cascading across the levels, MFST reduces the per‑level search cost to O(1), yielding an overall query time of O(log^{d‑1} n) for both aggregation and selection. Space consumption remains modest at O(n log^{d‑1} n), because each level stores only the necessary auxiliary pointers and aggregates.

The second structure, the Weighted Prefix Sum Tree (WPST), augments each leaf (and selected internal nodes) of the MFST with a compact histogram of the values in that sub‑range, together with their cumulative weights. This enables the weighted‑median query to be answered by a binary search over the histogram, which costs O(log log n) per node. Since the query traverses O(log^{d‑1} n) nodes in the MFST, the total time for a weighted‑median query becomes O(log^{d‑1} n · log log n). Crucially, the histogram can be updated incrementally: an insertion, deletion, or weight change only affects the O(log^{d‑1} n) nodes on the update path, preserving the same asymptotic bounds for dynamic workloads.

The paper provides a rigorous analysis of these complexities, proving that MFST and WPST achieve optimal (up to polylogarithmic factors) lower bounds for the considered query families under the comparison model. It also discusses trade‑offs: while MFST uses slightly more memory than a plain kd‑tree (approximately 1.5× in the experiments), the gain in query speed—especially for weighted‑median queries, which are O(√n) in the best known wavelet‑tree approaches—justifies the overhead for most real‑time applications.

To demonstrate practical relevance, the authors implement the structures and benchmark them on four realistic business scenarios:

1. Financial Portfolio Rebalancing – Queries involve selecting the median‑risk asset within a time‑window and aggregating returns across sectors. MFST delivers sub‑millisecond responses even when the portfolio contains millions of positions.
2. Supply‑Chain Inventory Optimization – The system must compute total stock levels, identify the k‑th most stocked SKU, and find the weighted‑median demand across warehouses. The proposed structures reduce latency from tens of milliseconds (traditional range trees) to under 2 ms.
3. Online Advertising Bidding – Real‑time bidding requires rapid aggregation of click‑through rates and selection of the median bid price within a geographic and temporal slice. The weighted‑median capability directly supports budget‑capped bidding strategies.
4. Retail Sales Analysis – Analysts query sales totals, top‑selling products, and the weighted‑median revenue per store. The MFST+WPST combination outperforms wavelet‑tree baselines by a factor of 4–8.

Across all experiments, the average query time improvement ranges from 3× to 10×, with the most dramatic gains observed for weighted‑median queries. Update throughput remains high, sustaining thousands of insertions/deletions per second on a single commodity server.

The paper concludes by emphasizing the broader impact of the work: the ability to answer complex multidimensional statistical queries in polylogarithmic time while supporting continuous updates opens the door to truly interactive decision‑support systems in finance, logistics, advertising, and beyond. Future research directions include external‑memory adaptations for datasets exceeding RAM, integration with dimensionality‑reduction techniques for very high‑dimensional data, and coupling the structures with machine‑learning pipelines to provide on‑the‑fly feature engineering for predictive models.