📈 economics

Efficient tensor bases for pairwise comparisons

This paper introduces the first orthogonal basis for additively consistent subspaces in pairwise comparisons theory, utilizing a minimal-support tensor basis to derive new composite formulae for logarithmic, Saaty, and SVD projections.

Original authors: Konrad Kułakowski, Ryszard Smarzewski

Published 2026-08-27

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Konrad Kułakowski, Ryszard Smarzewski

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Decision-making often requires us to weigh options that cannot be measured with a ruler or a scale. When a committee must choose between three different locations for a new park, or a manager must rank five potential projects, they rely on comparing items in pairs. They might decide that location A is twice as important as location B, or that project X is significantly better than project Y. These judgments are collected into a grid, a square table where every cell holds a value representing the relative importance of one item over another. The goal is to turn this collection of subjective opinions into a single, clear list of priorities. However, human judgment is rarely perfect. A person might say A is better than B, and B is better than C, but then mistakenly claim C is better than A. This internal contradiction, known as inconsistency, creates a fog that makes it difficult to extract a reliable ranking from the data. For decades, mathematicians and decision scientists have struggled to find the best way to clear that fog and find the true underlying order hidden within these imperfect comparisons.

The core challenge lies in the nature of the data itself. Because these comparisons are multiplicative—meaning if A is twice as important as B, then B is half as important as A—the mathematics involved is complex and non-linear. To simplify the problem, researchers often transform these multiplicative values into additive ones, much like turning a complex curve into a straight line to make it easier to draw. In this additive world, the goal becomes finding a "perfectly consistent" version of the messy data, a version where all the comparisons fit together logically without any contradictions. This process is essentially a search for the closest possible fit between the real, flawed data and an ideal, logical structure. The difficulty has always been finding a way to do this calculation efficiently and accurately, especially when the data is large or the inconsistencies are deep.

In a recent study, researchers Konrad Kułakowski and Ryszard Smarzewski have constructed a new mathematical tool that solves a long-standing problem in this field. They developed the first explicit set of building blocks, or a basis, for the space of perfectly consistent data. Imagine trying to describe every possible shape in a room using only a few specific, standard shapes. For years, mathematicians had a set of shapes that worked, but they were awkward and difficult to use together because they overlapped in complicated ways. Kułakowski and Smarzewski have now created a new set of shapes that are perfectly independent of one another, meaning they do not overlap or interfere. This new set allows them to break down any messy set of comparisons into its consistent parts with extreme precision and speed. Their method provides a direct, step-by-step formula to calculate the best possible ranking from the data, eliminating the need for slow, repetitive guessing games that previous methods required.

The significance of this discovery extends beyond just finding a faster way to calculate. The researchers used their new tool to re-examine three different methods that have been used for years to rank items: the traditional eigenvector method, a logarithmic approach, and a technique based on singular value decomposition. By applying their new orthogonal basis, they were able to show exactly how these three methods relate to one another. They found that the singular value decomposition method, which is based on breaking a matrix down into its fundamental components, is actually a combination of two simpler processes. It turns out that this method is unique because it satisfies two different mathematical criteria at once, acting as both a distance-based measure and an eigenvector-based measure. This finding challenges the long-held belief that one specific method is always superior to the others. The authors demonstrate that no single approach works best in every situation; the choice of method depends on the specific nature of the data and the type of error present.

The paper also addresses a common criticism of the most popular method, the eigenvector approach, which is widely used in strategic planning and finance. The researchers show that this method can fail to distinguish between certain types of data, particularly when the data resembles a random distribution of probabilities. In such cases, the method might produce a result that ignores the actual input entirely. By contrast, their new orthogonal projection offers a more robust alternative that does not suffer from this specific blind spot. The study provides closed-form formulas, meaning the answers can be calculated directly without iteration, which makes the process stable and reliable for computer implementation. This is a crucial improvement for real-world applications where decision-makers need quick, trustworthy results from complex data sets.

Ultimately, this work reshapes how we understand the mathematics of preference. It moves the field away from relying on a single, dominant technique toward a more nuanced understanding that different tools are needed for different problems. The researchers have provided a clear map of the mathematical landscape, showing exactly where the different methods overlap and where they diverge. Their construction of the orthogonal basis is not just a theoretical curiosity; it is a practical engine that can drive more accurate decision-making in fields ranging from psychology to market research. By clarifying the relationships between these methods, the study empowers practitioners to choose the right tool for their specific problem, ensuring that the final rankings reflect the true intent of the decision-makers rather than the limitations of the calculation method. The work confirms that while no single method is perfect for every scenario, having a clear understanding of their strengths and weaknesses allows for much better outcomes in the complex world of human judgment.

Technical Summary: Efficient Tensor Bases for Pairwise Comparisons

Problem Statement
The paper addresses the fundamental problem of approximating inconsistent pairwise comparison (PC) matrices with consistent ones within the Analytic Hierarchy Process (AHP). While multiplicative PC matrices are standard, the authors focus on the additive representation obtained via logarithmic transformation. The core mathematical challenge identified is the construction of an efficient, explicit orthogonal basis for the subspace $A_n$ of all additively consistent matrices in $\mathbb{R}^{n \times n}$ . Previous attempts, such as those by Koczkodaj and Orłowski (1997), provided bases that lacked minimal support or relied on recurrence relations (Gram-Schmidt) that were computationally inefficient and difficult to generalize for $n > 4$ . The authors aim to solve the problem of finding an orthogonal basis $\{B_1, \dots, B_{n-1}\}$ for $A_n$ that allows for closed-form expressions of orthogonal projections, thereby facilitating efficient "windowing" (projection) of PC matrices.

Methodology
The authors employ a tensor-based approach to linear algebra within the context of pairwise comparisons.

Tensor Representation: The study utilizes tensor products ( $\otimes$ ) and tensor sums/differences ( $\oplus, \ominus$ ) to represent matrices. An additively consistent matrix $A$ is represented as $y \ominus y = [y_i - y_j]$ , where $y$ is an additive priority vector.
Basis Construction: The authors define a specific basis $\{A_1, \dots, A_{n-1}\}$ for $A_n$ with minimal support, where $A_k = e_k \ominus e_k$ (using standard basis vectors $e_k$ ).
Orthogonalization: Instead of relying on the numerically unstable or complex Gram-Schmidt recurrence relations, the authors derive a closed-form solution for the orthogonal basis $\{B_1, \dots, B_{n-1}\}$ . They define vectors $x_k \in \mathbb{R}^n$ with specific component structures ( $x_k^i = 1$ for $i=k$ , $1/(n-k+1)$ for $i < k$ , and $0$ otherwise) and construct the orthogonal basis matrices as $B_k = x_k \ominus x_k$ .
Projection Derivation: Using this explicit orthogonal basis, the authors derive closed-form formulas for the orthogonal projection $P_A$ $P_{A}$ of any skew-symmetric matrix onto $A_n$ $A_{n}$ . This is extended to the multiplicative domain to define three types of "windowing" (projections) of PC matrices onto the consistent subgroup $C_n$ $C_{n}$ :
- Orthogonal Windowing ( $P_C$ ): Derived from the orthogonal projection in the additive space.
- Saaty Windowing ( $R_C$ ): Based on the Perron eigenvector method.
- SVD Windowing ( $T_C$ ): Based on Singular Value Decomposition.

Key Contributions and Results

First Explicit Orthogonal Basis: The paper presents the first explicit, closed-form orthogonal basis for the additively consistent subspace $A_n$ that possesses minimal support. This resolves a long-standing problem regarding the efficient computation of such bases.
Closed-Form Projection Formulas: The authors provide explicit formulas for the orthogonal projection $P_A$ and the corresponding multiplicative projection $P_C$ . These formulas are computable via a finite sum of scalar products, offering a direct algorithm of polynomial complexity that avoids iterative methods.
Composite Nature of Projections:
- The SVD projection ( $T_C$ ) is proven to be a composite projection: $T_C M = P_C (T_G M) = R_C (T_G M)$ , where $T_G$ is a tensor projection onto the group of rank-1 tensor products.
- The paper establishes that for tensor products of positive vectors ( $a \otimes b$ ), the Saaty projection and SVD projection coincide: $R_C(a \otimes b) = T_C(a \otimes b) = a \oslash a$ .
Comparison of Windowing Methods: The study theoretically compares the Orthogonal, Saaty, and SVD windowing methods. It demonstrates that no single method is universally superior. For instance, the Saaty projection fails to distinguish between stochastic matrices, while the SVD projection is unique in being both norm-based and eigenvector-based.
Weighted Frobenius Norm: The results are generalized to include weighted Frobenius inner products, allowing for the incorporation of specific weight vectors (e.g., derived from orthogonal polynomial theory or the Perron eigenvector) to handle randomly disturbed data.

Significance and Claims
The authors claim that their construction of the orthogonal basis provides a new theoretical framework for understanding and computing priority vectors in AHP. By establishing a direct link between the orthogonal projection in the additive space and the multiplicative projection, the paper offers a unified perspective on logarithmic, Saaty, and SVD methods.

The paper modestly asserts that these theoretical findings provide "additional arguments" against the claim of the absolute superiority of the eigenvector method (Saaty's approach), supporting the conclusion by Golany and Kress that no windowing method is optimal in all cases. The authors emphasize that while the results are formal, they yield direct, non-iterative algorithms for computing consistent approximations. They conclude that further studies are needed, particularly regarding the choice of matrix norms (beyond the standard Frobenius norm) and numerical comparisons of these windowing methods in real-world applications. The work does not propose new experimental applications but rather lays the mathematical groundwork for more efficient and theoretically grounded approximations in pairwise comparison theory.

More like this