Noisy Nonreciprocal Pairwise Comparisons: Scale Variation, Noise Calibration, and Admissible Ranking Regions

Abstract: Pairwise comparisons are widely used in decision analysis, preference modeling, and evaluation problems. In many practical situations, the observed comparison matrix is not reciprocal. This lack of reciprocity is often treated as a defect to be corrected immediately. In this article, we adopt a different point of view: part of the nonreciprocity may reflect a genuine variation in the evaluation scale, while another part is due to random perturbations. We introduce an additive model in which the unknown underlying comparison matrix is consistent but not necessarily reciprocal. The reciprocal component carries the global ranking information, whereas the symmetric component describes possible scale variation. Around this structured matrix, we add a random perturbation and show how to estimate the noise level, assess whether the scale variation remains moderate, and assign probabilities to admissible ranking regions in the sense of strict ranking by pairwise comparisons. We also compare this approach with the brutal projection onto reciprocal matrices, which suppresses all symmetric information at once. The Gaussian perturbation model is used here not because human decisions are exactly Gaussian, but because observed judgment errors often result from the accumulation of many small effects. In such a context, the central limit principle provides a natural heuristic justification for Gaussian noise. This makes it possible to derive explicit estimators and probability assessments while keeping the model interpretable for decision problems.

• The paper presents a structured model that decomposes observed pairwise comparison matrices into latent ranking, scale deformation, and stochastic noise.
• It provides explicit least-squares estimators and diagnostics, such as the global scale deformation ratio and ranking impact index, to quantify systematic deviations.
• The approach yields calibrated probabilistic ranking regions, contrasting with traditional reciprocal projection methods that overestimate uncertainty.

Noisy Nonreciprocal Pairwise Comparisons: Structured Modeling, Noise Calibration, and Implications for Ranking Uncertainty

Introduction and Motivation

Pairwise comparison (PC) methods constitute foundational tools in preference measurement, decision analysis, and multicriteria evaluation. Traditional PC analysis presumes reciprocity, i.e., $a_{ji} = -a_{ij}$, where $a_{ij}$ is the (additive or log-ratio) comparison of alternative $i$ over $j$. Additive consistency ($a_{ik} = a_{ij} + a_{jk}$) further enables a latent score vector representation $a_{ij} = u_i - u_j$, which supports unique strict global rankings.

Empirical PC matrices, however, universally depart from perfect reciprocity due to inconsistency—random errors, hesitations, contextual dependencies, or systematic effects in the elicitation scale. Standard practice employs brute-force projection of observed matrices onto the space of reciprocal matrices, effectively discarding the symmetric content (i.e., all deviations from antisymmetric structure), treating any nonreciprocity as noise.

This work presents a paradigm shift: not all nonreciprocity is random. The author introduces a scale-deformed additive model admitting a structured symmetric component, interpretable as moderate scale variation rather than error. Residuals, after accounting for latent ranking and scale deformation, are explicitly modeled as Gaussian perturbations, which justifies estimable and interpretable probabilistic calibrations of ranking uncertainty.

Model Formulation and Decomposition

Let $X = (x_{ij})_{i, j = 1}^n$ be the observed PC matrix. The core model is:

$$x_{ij} = u_i - u_j + s_i + s_j + \varepsilon_{ij},\quad i \neq j$$

where:

• $u \in \mathbb{R}^n$ is the centered ($\sum_i u_i = 0$) latent ranking vector;
• $a_{ij}$0 is the centered scale-deformation vector ($a_{ij}$1), encoding persistent, structured deviation from reciprocity;
• $a_{ij}$2 are independent residual perturbations (with possible pairwise correlation between $a_{ij}$3 and $a_{ij}$4).

This structure yields a “scale-deformed” matrix that is not, in general, reciprocal or additively consistent, thereby generalizing the classical PC framework.

The matrix $a_{ij}$5 admits a unique decomposition into antisymmetric $a_{ij}$6 and symmetric $a_{ij}$7 parts:

$a_{ij}$8

with

$a_{ij}$9

The latent ranking is estimated from $i$0, and the scale deformation from $i$1 via constrained least squares. Explicit solutions:

$i$2

Uncertainty Quantification: Residual Analysis and Noise Calibration

After fitting $i$3 and $i$4, the residual matrix is:

$i$5

Residual noise is characterized by two quadratic indicators:

• Residual reciprocity: $i$6
• Residual triangle inconsistency: $i$7

Under the Gaussian assumption (justified by the central limit principle), the variance $i$8 and residual interdependence parameter $i$9.

Residual analysis is only meaningful after structured components are removed. Direct application to $j$0 confounds signal and noise, yielding overestimated uncertainties.

Diagnostics: Scale Deformation Influence

Two diagnostics are introduced:

• Global scale deformation ratio: $j$1 (with $j$2, $j$3), measuring the overall structural deviation from reciprocity.
• Ranking impact index: $j$4, quantifying the maximal possible effect of scale deformation on any pairwise ranking decision.

Interpretation is decisive: large $j$5 or $j$6 indicates non-negligible, decision-relevant scale variation that cannot be safely treated as noise.

Ranking Regions: Probabilistic Output Beyond Point Estimates

The ranking vector $j$7 supplies a central estimate. However, due to residual uncertainty, nearby vectors may yield different rankings.

Ranking regions are convex polytopes in the centered score space (combinatorially equivalent to the $j$8 strict orderings). The calibrated Gaussian distribution

$j$9

enables calculation (via Monte Carlo) of the probability that the consensus ranking is unique, versus possible neighbor rankings, and supports computation of top-$a_{ik} = a_{ij} + a_{jk}$0 probabilities, pairwise precedence probabilities, and ranking region entropy.

Comparison with Brutal Reciprocal Projection

A salient result is that the central ranking estimate from both the structured method and naive reciprocal projection coincide in least-squares settings. The distinction emerges in uncertainty quantification:

• The structured approach separates and interprets the scale-deformation vector before noise calibration, yielding focused, calibrated ranking region probabilities.
• Brutal projection absorbs all symmetric structure as noise, systematically overestimating residual variance and inflating ranking uncertainty—especially when $a_{ik} = a_{ij} + a_{jk}$1 or $a_{ik} = a_{ij} + a_{jk}$2 are not small.

This effect is quantitatively pronounced in regimes where structured scale deformation is moderate or large; in this regime, ignoring symmetric information leads to unjustifiable diffusions of uncertainty and can mislead decision support.

Numerical Illustration and Statistical Consequences

A detailed numerical and Monte Carlo analysis demonstrate:

1. Even purely random perturbations, after reciprocal projection, can yield sign patterns incompatible with any strict ranking, or change the true ranking with non-trivial probability (cf. Table, $a_{ik} = a_{ij} + a_{jk}$3).
2. Systematically increasing scale deformation at fixed noise leaves the central estimate unchanged but induces an artificial increase of apparent residual noise and spurious positive pairwise dependence when using brutal projection.
3. The structured model accurately calibrates uncertainty, isolating genuine noise from systematic effects.

Scenario	Structured Output	Brutal Projection Output
Small deformation	Accurate $a_{ik} = a_{ij} + a_{jk}$4, low $a_{ik} = a_{ij} + a_{jk}$5	Similar to structured
Moderate deformation	Accurate $a_{ik} = a_{ij} + a_{jk}$6, moderate $a_{ik} = a_{ij} + a_{jk}$7, uncertainty localized	Overestimated $a_{ik} = a_{ij} + a_{jk}$8, spurious uncertainty
Strong deformation	Diagnostics warn against projection	Unreliable probabilities

Failure to differentiate structured and stochastic nonreciprocity can result in poor advice for decision-makers in multicriteria settings.

Theoretical and Practical Implications

The theoretical implications are as follows:

• Scale-deformation provides a mechanism for structured asymmetry in PC matrices. Its impact must be separated from random error to correctly infer both central ranking and uncertainty.
• The usual mathematical guarantee that most "consistency corrections" recover the underlying true ranking (as sometimes implicitly assumed in classical AHP literature) does not hold under moderate/large structured nonreciprocity.

Practically, the results necessitate a diagnostic approach:

• If both $a_{ik} = a_{ij} + a_{jk}$9 and $a_{ij} = u_i - u_j$0 are small, reciprocal projection is acceptable and numerically efficient.
• If either is moderate or large, the probabilistic consequences of forced reciprocity are misleading and the structured methodology provides a more robust and interpretable solution with reliable uncertainty estimates.

Implications for Future Developments

Future research directions include:

• Adapting these methods to incomplete, sparse, or multiplicative PC matrices.
• Extending the noise model beyond Gaussianity, e.g., for heavy tails or correlated/residual heteroscedastic effects.
• Application in dynamic, repeated, or group-elicitation contexts to disentangle persistent scaling biases from transient inconsistency.
• Theoretical integration with robust multicriteria decision analysis and sensitivity analysis frameworks.

Conclusion

This work rigorously develops a structured decomposition of noisy, nonreciprocal PC matrices, enabling explicit statistical distinction between latent ranking, scale deformation, and random noise. The approach provides interpretable diagnostics for deciding when classical reciprocal correction suffices and when deeper modeling is essential. It hence addresses a critical gap in the analysis and interpretation of PC data, with direct implications for the reliability of rankings and their associated uncertainty in decision analysis contexts.

The presented analysis indicates that the symmetric (scale-deformation) content of a PC matrix, if non-negligible, carries interpretable information and must not be automatically suppressed as noise. By providing explicit least-squares estimators, residual diagnostics, and a calibrated probabilistic framework over ranking regions, this work sets a new standard for robust, interpretable, and actionable ranking under judgmental uncertainty.