Scoring Rules as Least-Squares Estimators

Technical Summary: Scoring Rules as Least-Squares Estimators

Problem Statement
The paper addresses the theoretical relationship between scoring rules and the cosine similarity rule in social choice theory. Scoring rules determine a social ranking by aggregating individual scores assigned to alternatives based on their position in voters' preference orderings (e.g., the Borda rule). Recently, Kawada (2018) established a remarkable equivalence theorem stating that for any given score vector, the resulting scoring rule is identical to the corresponding cosine similarity rule. While Kawada's original proof relied on a direct analysis of the cosine similarity optimization problem, the authors of this note seek to provide a more transparent and elementary proof of this equivalence.

Methodology
The authors employ a least-squares characterization of the arithmetic mean to re-derive Kawada's result. The methodology proceeds as follows:

1. Least-Squares Foundation: The paper utilizes the well-known property that the arithmetic mean of a set of vectors is the unique minimizer of the sum of squared Euclidean distances to those vectors. Formally, for a set of vectors $\{p_i\}_{i \in N}$, the function $\Phi(x) = \sum_{i \in N} \|x - p_i\|^2$ is uniquely minimized at $x = \bar{p} = \frac{1}{n}\sum p_i$.
2. Vector Representation: Individual voter rankings are mapped to score vectors $s^\theta_i \in \mathbb{R}^V$ based on a fixed score vector $\theta$.
3. Optimization Equivalence: The authors demonstrate that minimizing the least-squares objective (sum of squared distances) is mathematically equivalent to maximizing the sum of dot products $\sum (s^\theta_i)^\top x$ when $x$ is constrained to a sphere of fixed radius.
4. Norm Invariance: A critical step in the proof is observing that for any voter $i$, the Euclidean norm of the score vector $\|s^\theta_i\|$ is constant (denoted as $R$), regardless of the specific ranking. This allows the dot product sum to be rewritten in terms of cosine similarity: $\sum (s^\theta_i)^\top x = R\rho \sum C(s^\theta_i, x)$, where $\rho = \|x\|$.
5. Synthesis: By combining the least-squares minimization property with the constant-norm property of score vectors, the authors show that the optimizer of the cosine similarity maximization problem is precisely the arithmetic mean of the score vectors.

Key Contributions and Results

• Alternative Proof of Kawada (2018): The paper provides a simpler, geometric proof of the equivalence between scoring rules and cosine similarity rules.
• Geometric Interpretation: The proof reveals that the arithmetic mean of the score vectors, $\bar{s}^\theta = \frac{1}{n} \sum s^\theta_i$, is the unique solution to both the least-squares estimation problem and the cosine similarity optimization problem.
• Equivalence Confirmation: Since the social ranking in a scoring rule is determined by the coordinates of the aggregate score vector (which is proportional to the arithmetic mean), and the cosine similarity rule is determined by the coordinates of the optimizer (which is also the arithmetic mean), the two rules necessarily induce the same social ranking.
• Formal Result: Theorem 1 confirms that for every score vector $\theta$, the scoring rule (2) coincides with the cosine similarity rule (4).

Significance and Claims
The authors claim that their approach offers a transparent geometric interpretation of scoring rules. By framing scoring rules as least-squares estimators, the paper clarifies why the cosine similarity rule must coincide with the corresponding scoring rule: both are optimization representations sharing the arithmetic mean as their common optimizer.

The paper positions this result within the broader context of optimization-based formulations in social choice, citing classic examples like the Kemeny–Young rule and recent work on the Schulze rule. The authors modestly suggest that this framework may provide a useful perspective for understanding other social choice procedures, though they do not propose specific new applications or experimental validations in this note. The primary contribution is the theoretical unification of these two concepts through elementary optimization principles.