The bottom line: a popular decision-making tool has a hidden reliability problem, and a new optimization technique appears to fix it. The Analytic Hierarchy Process (AHP) is used across procurement, engineering design, healthcare, and public policy to turn a decision-maker's pairwise comparisons — "is A more important than B, and by how much?" — into a ranked list of priorities. One of the methods for extracting that ranking, called Direct Least Squares (DLS), can quietly return different "correct" answers depending on where the underlying algorithm happens to start searching.
Here's the mechanics of why that happens. AHP starts from a pairwise comparison matrix, where an entry $a_{ij}$ represents how much more important criterion $i$ is judged to be than criterion $j$. In a perfectly consistent world, that matrix would satisfy $a_{ij} \cdot a_{jk} = a_{ik}$ for every triple, and recovering a clean priority vector would be trivial. Real human judgment is rarely that tidy — people give inconsistent, sometimes cyclical, comparisons — and DLS tries to find a priority vector $w$ that minimizes the squared discrepancy between the matrix and the ratios it implies, roughly $\min_w \sum_{i,j} (a_{ij} w_j - w_i)^2$. The problem is that this objective is not convex, so under enough inconsistency it has multiple distinct global minima rather than one clear winner.
That's not a cosmetic quirk — it's a stability problem with real consequences. If two analysts run the same pairwise comparison data through the same DLS method but their optimization routines happen to initialize differently, they can land on two different priority rankings, both mathematically "optimal" by the same criterion. In a context like budget allocation, supplier selection, or medical treatment prioritization, that means the tool can produce contradictory recommendations from identical inputs — a serious problem for any process that needs to be defensible or repeatable.
The paper's proposed fix, Anchored Regularized Direct Least Squares (ARDLS), doesn't throw out DLS. Instead it tethers the optimization to a second, independently well-behaved solution — an "anchor" — and penalizes the DLS objective for straying too far from it. The anchors considered include several classical AHP priority operators that already have a unique, closed-form answer: simple normalization of the matrix, the Eigenvector method (Saaty's original approach), Singular Value Decomposition, Cosine Maximization, and the Pseudo-Inverse Gram Matrix solution, which is the closed-form answer to Weighted Least Squares.
In objective terms, ARDLS looks like the original DLS loss plus a quadratic penalty term tied to an anchor vector $w^{*}$ produced by one of those established methods: $$\min_w \; \sum_{i,j} (a_{ij} w_j - w_i)^2 \;+\; \lambda \lVert w - w^{*} \rVert^2$$ where $\lambda$ controls how strongly the solution is pulled toward the anchor. Because the anchor term is strictly convex, adding it breaks the symmetry that let multiple global minima coexist in plain DLS — it tilts the optimization landscape so there is exactly one lowest point, and that point is provably unique. In plain terms: instead of letting the algorithm wander toward whichever minimum it happens to find first, ARDLS gives it a reliable reference point to stay close to while still optimizing.
According to the paper, numerical experiments and simulations show that ARDLS not only restores uniqueness but also reduces root mean square error relative to the established anchor operators used on their own — meaning the anchored version isn't just more stable, it can fit the original comparison data more faithfully than the classical methods it borrows from. That combination — guaranteed uniqueness plus improved accuracy — is what the authors point to as the practical case for treating ARDLS as a drop-in replacement for plain DLS in AHP implementations.
The bigger takeaway for practitioners building or auditing AHP-based tools is procedural, not just mathematical: any pipeline relying on DLS for priority derivation should be checked for solution instability under high-inconsistency inputs, since a single dataset could otherwise yield materially different decisions depending on solver initialization. Regularizing around a well-established, closed-form anchor is a comparatively low-cost way to close that gap, and it's a pattern — take a nonconvex objective, tether it to a convex, uniquely-solvable reference — that shows up elsewhere in optimization whenever a preferred model has better properties on paper than it does in practice.
Sources: Anchored Regularized Direct Least Squares (ARDLS): Integrating Established Prioritization Operators for Priority Elicitation in the Analytic Hierarchy Process, arXiv:2608.21187
