<p>The standard <i>n</i>-simplex <InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(\Delta ^n\)</EquationSource> <EquationSource Format="MATHML"><math> <msup> <mi mathvariant="normal">Δ</mi> <mi>n</mi> </msup> </math></EquationSource> </InlineEquation> serves as a natural geometric space for representing discrete probability distributions and compositional data. Quantifying dissimilarity between points <InlineEquation ID="IEq2"> <EquationSource Format="TEX">\(P, Q \in \Delta ^n\)</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mi>P</mi> <mo>,</mo> <mi>Q</mi> <mo>∈</mo> <msup> <mi mathvariant="normal">Δ</mi> <mi>n</mi> </msup> </mrow> </math></EquationSource> </InlineEquation> is essential, with metrics respecting ordinal structures being particularly valuable. The 1-Wasserstein distance <InlineEquation ID="IEq3"> <EquationSource Format="TEX">\(W_1\)</EquationSource> <EquationSource Format="MATHML"><math> <msub> <mi>W</mi> <mn>1</mn> </msub> </math></EquationSource> </InlineEquation> excels in this regard but lacks a direct geometric interpretation as a path length on the simplex. This paper addresses this gap by defining a canonical path from <i>P</i> to <i>Q</i> in an “outside-in” recursive manner using dimension-reducing projections. We prove that the length of this path, under a weighted Euclidean metric with weights <InlineEquation ID="IEq4"> <EquationSource Format="TEX">\(\textbf{a} = (a_1, \dots , a_n)\)</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mi mathvariant="bold">a</mi> <mo>=</mo> <mo stretchy="false">(</mo> <msub> <mi>a</mi> <mn>1</mn> </msub> <mo>,</mo> <mo>⋯</mo> <mo>,</mo> <msub> <mi>a</mi> <mi>n</mi> </msub> <mo stretchy="false">)</mo> </mrow> </math></EquationSource> </InlineEquation> encoding ordinal spacings, equals the weighted 1-Wasserstein distance <InlineEquation ID="IEq5"> <EquationSource Format="TEX">\(W_{1,\textbf{a}}(P, Q)\)</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <msub> <mi>W</mi> <mrow> <mn>1</mn> <mo>,</mo> <mi mathvariant="bold">a</mi> </mrow> </msub> <mrow> <mo stretchy="false">(</mo> <mi>P</mi> <mo>,</mo> <mi>Q</mi> <mo stretchy="false">)</mo> </mrow> </mrow> </math></EquationSource> </InlineEquation>. Furthermore, we situate this result within Finsler geometry, demonstrating that <InlineEquation ID="IEq6"> <EquationSource Format="TEX">\(W_{1,\textbf{a}}\)</EquationSource> <EquationSource Format="MATHML"><math> <msub> <mi>W</mi> <mrow> <mn>1</mn> <mo>,</mo> <mi mathvariant="bold">a</mi> </mrow> </msub> </math></EquationSource> </InlineEquation> is the geodesic distance induced by a continuous Finsler metric on <InlineEquation ID="IEq7"> <EquationSource Format="TEX">\(\Delta ^n\)</EquationSource> <EquationSource Format="MATHML"><math> <msup> <mi mathvariant="normal">Δ</mi> <mi>n</mi> </msup> </math></EquationSource> </InlineEquation>. The canonical path thus emerges as a non-smooth length-minimizing geodesic.</p>

错误:搜索内容不能为空,请输入英文关键词
错误:关键词超出字数限制,请精简
高级检索

A dyadic path construction as a geometric realization of the 1-Wasserstein distance on the simplex

  • Beniamino Cappelletti-Montano,
  • Monica Musio

摘要

The standard n-simplex \(\Delta ^n\) Δ n serves as a natural geometric space for representing discrete probability distributions and compositional data. Quantifying dissimilarity between points \(P, Q \in \Delta ^n\) P , Q Δ n is essential, with metrics respecting ordinal structures being particularly valuable. The 1-Wasserstein distance \(W_1\) W 1 excels in this regard but lacks a direct geometric interpretation as a path length on the simplex. This paper addresses this gap by defining a canonical path from P to Q in an “outside-in” recursive manner using dimension-reducing projections. We prove that the length of this path, under a weighted Euclidean metric with weights \(\textbf{a} = (a_1, \dots , a_n)\) a = ( a 1 , , a n ) encoding ordinal spacings, equals the weighted 1-Wasserstein distance \(W_{1,\textbf{a}}(P, Q)\) W 1 , a ( P , Q ) . Furthermore, we situate this result within Finsler geometry, demonstrating that \(W_{1,\textbf{a}}\) W 1 , a is the geodesic distance induced by a continuous Finsler metric on \(\Delta ^n\) Δ n . The canonical path thus emerges as a non-smooth length-minimizing geodesic.