# The Fisher-Rao geometry of beta distributions applied to the study of canonical moments

This paper studies the Fisher-Rao geometry on the parameter space of beta distributions. We derive the geodesic equations and the sectional curvature, and prove that it is negative. This leads to uniqueness for the Riemannian centroid in that space. We use this Riemannian structure to study canonical moments, an intrinsic representation of the moments of a probability distribution. Drawing on the fact that a uniform distribution in the regular moment space corresponds to a product of beta distributions in the canonical moment space, we propose a mapping from the space of canonical moments to the product beta manifold, allowing us to use the Fisher-Rao geometry of beta distributions to compare and analyze canonical moments.

## Authors

• 4 publications
• 4 publications
• ### Fisher-Rao geometry of Dirichlet distributions

In this paper, we study the geometry induced by the Fisher-Rao metric on...
05/12/2020 ∙ by Alice Le Brigant, et al. ∙ 0

• ### Classifying histograms of medical data using information geometry of beta distributions

In this paper, we use tools of information geometry to compare, average ...
06/03/2020 ∙ by Alice Le Brigant, et al. ∙ 0

• ### Measures of goodness of fit obtained by canonical transformations on Riemannian manifolds

The standard method of transforming a continuous distribution on the lin...
11/12/2018 ∙ by P. E. Jupp, et al. ∙ 0

• ### Method of Moments Histograms

Uniform bin width histograms are widely used so this data graphic should...
09/10/2019 ∙ by James S. Weber, et al. ∙ 0

• ### Optimal Uncertainty Quantification on moment class using canonical moments

We gain robustness on the quantification of a risk measurement by accoun...
11/30/2018 ∙ by Jerome Stenger, et al. ∙ 0

• ### Fisher-Rao distance on the covariance cone

The Fisher-Rao geodesic distance on the statistical manifold consisting ...
10/29/2020 ∙ by Joseph Wells, et al. ∙ 0

• ### An information-geometric approach to feature extraction and moment reconstruction in dynamical systems

We propose a dimension reduction framework for feature extraction and mo...
04/05/2020 ∙ by Suddhasattwa Das, et al. ∙ 20

##### This week in AI

Get the week's most popular data science and artificial intelligence research sent straight to your inbox every Saturday.

## 1 Introduction

The differential geometric approach to probability theory and statistics has met increasing interest in the past years, from the theoretical point of view as well as in applications. In this approach, probability distributions are seen as elements of a differentiable manifold, on which a metric structure is defined through the choice of a Riemannian metric. Two very important ones are the Wasserstein metric, central in optimal transport, and the Fisher-Rao metric (also called Fisher information metric), essential in information geometry. Unlike optimal transport, information geometry is foremost concerned with parametric families of probability distributions, and defines a Riemannian structure on the parameter space using the Fisher information matrix

[5]

. In parameter estimation, the Fisher information can be interpreted as the quantity of information on the unknown parameter contained in the model. As the Hessian of the well-known Kullback-Leibler divergence, it measures through the notion of curvature the capacity to distinguish between two different values of the parameter. Rao

[9] showed that it could be used to locally define a scalar product on the space of parameters, interpretable as a Riemannian metric. An important feature of this metric is that it is invariant under any diffeomorphic change of parameterization. In fact, considering the infinite-dimensional space of probability densities on a given manifold , there is a unique metric, which also goes by the name Fisher-Rao, that is invariant with respect to the action of the diffeomorphism group of [3, 2]. This metric induces the regular Fisher information metric on the finite dimensional submanifolds corresponding to the parameterized statistical models of interest in information geometry. Arguably the most famous example of Fisher-Rao geometry of a statistical model is that of the Gaussian model, which is hyperbolic. The multivariate Gaussian case, among other models, has also received a lot of attention [1, 11].

In this work, we are interested in beta distributions, a family of probability measures on

used to model random variables defined on a compact interval in a wide variety of applications. Up to our knowledge, the information geometry of beta distributions has not yet received much attention. In this paper, we give new results and properties for this geometry, and its curvature in particular. Interestingly, this geometric framework yields new by-product tools to study the set of all moments of compactly supported probability measures on the real line. This is achieved through the so-called canonical moments representation

[4], an alternative to the usual moment representation of a probability distribution that satisfies interesting symmetries and invariance properties.

The paper is organized as follows. Section 2

deals with the study of the Fisher-Rao geometry of beta distributions. We derive the geodesic equations, prove that sectional curvature is negative, give some bounds and observe a geometrical manifestation of the central limit theorem. Section

3 deals with the application to canonical moments. After a brief presentation of these objets, we propose a representation in the product beta manifold, allowing us to use the Fisher-Rao geometry of beta distributions to compare and analyze canonical moments.

## 2 Geometry of the beta manifold

### 2.1 The beta manifold

Information geometry is concerned with parametric families of probability distributions, i.e. sets of distributions with densities with respect to a common dominant measure parameterized by a parameter member of a given set . That is, a collection of measures of the kind

 PΘ={pθμ,θ∈Θ}.

We assume that is a non empty open subset of . Associated to any such family is the Fisher information matrix, defined for all as

 I(θ)=[E(∂2∂θi∂θjlnp(X;θ))]1≤i,j≤d.

As an open subset of , is a differentiable manifold and can be equipped with a Riemannian metric using this quantity. This gives the Fisher information metric on the parameter space

 GFθ(u,v)=utI(θ)v,θ∈Θ,u,v∈TθΘ≃Rd,

where

denotes the transpose of the vector

. By extension, we talk of the Fisher geometry of the parameterized family , and of the Riemannian manifold .

In this paper, we are interested in beta distributions, a family of probability distributions on with density with respect to the Lebesgue measure parameterized by two positive scalars

 pα,β(x)=Γ(α+β)Γ(α)Γ(β)xα−1(1−x)β−1,x∈[0,1].

We consider the Riemannian manifold composed of the parameter space and the Fisher metric , and by extension denote by beta manifold the pair , where is the family of beta distributions

 B={B(α,β)=pα,β(⋅)dx,α>0,β>0}.

Here denotes the Lebesgue measure on . The distance between two beta distributions is then defined as the geodesic distance associated to the Fisher metric in the parameter space

 dF(B(α,β),B(α′,β′))=infγ∫10√gF(˙γ(t),˙γ(t))dt,

where the infimum is taken over all paths such that and .

### 2.2 The Fisher-Rao metric

The beta distributions are part of an exponential family and so the general term of the Fisher-Rao metric depends on second order derivatives of the underlying potential function. Denoting by the matrix form of ,

 gF(α,β)=−Hessφ(α,β), (1)

where is the potential function

 φ(α,β)=lnΓ(α)+lnΓ(β)−lnΓ(α+β).

Proposition 1

describes the metric tensor and Proposition

2 the geodesic equations.

###### Proposition 1.

The matrix representation of the Fisher-Rao metric on the space of beta distributions is given by

 gF(α,β)=[ψ′(α)−ψ′(α+β)−ψ′(α+β)−ψ′(α+β)ψ′(β)−ψ′(α+β)]

where denotes the digamma function, i.e. .

###### Proof.

This follows from straightforward computations. ∎

###### Proposition 2.

The geodesic equations are given by

 ¨α+a(α,β)˙α2+b(α,β)˙α˙β+c(α,β)˙β2=0, ¨β+a(β,α)˙β2+b(β,α)˙α˙β+c(β,α)˙α2=0,

where

 a(x,y) =12d(x,y)(ψ′′(x)ψ′(y)−ψ′′(x)ψ′(x+y)−ψ′(y)ψ′′(x+y)), b(x,y) =−1d(x,y)ψ′(y)ψ′′(x+y), c(x,y) =12d(x,y)(ψ′′(y)ψ′(x+y)−ψ′(y)ψ′′(x+y)), d(x,y) =ψ′(x)ψ′(y)−ψ′(x+y)(ψ′(x)+ψ′(y)).
###### Proof.

The geodesic equations are given by

 ¨α+Γααα˙α2+2Γααβ˙α˙β+Γαββ˙β2=0 (2) ¨β+Γβαα˙α2+2Γβαβ˙α˙β+Γβββ˙β2=0

where the ’s denote the Christoffel symbols of the second kind. These can be obtained from the Christoffel symbols of the first kind and the coefficients of the inverse of the metric matrix

 Γkij=Γijlgkl.

Here we have used the Einstein summation convention. Since the Fisher metric is a Hessian metric, the Christoffel symbols of the first kind can be obtained as

 Γijk=12φijk,

where is the potential function (1). Straightforward computation yields the desired equations. ∎

Notice that when , both geodesic equations (2

) yield a unique ordinary differential equation

 ¨γ+(a(γ,γ)+b(γ,γ)+c(γ,γ))˙γ2=0.

The line of equation is therefore a geodesic for the Fisher metric. More precisely, we have the following corollary obtained directly from Proposition 2.

###### Corollary 1.

The line of equation , where

 ¨γ+ψ′(γ)ψ′′(γ)−4ψ′(γ)ψ′′(2γ)2(ψ′(γ)2−2ψ′(γ)ψ′(2γ))˙γ2=0,

is a geodesic for the Fisher metric.

### 2.3 Some properties of the polygamma functions

In order to further study the geometry of the beta manifold, we will need a few technical results on the polygamma functions. The polygamma functions are the successive derivatives of the logarithm of the Euler Gamma function , i.e.

 ψ(m−1)(x):=dmdxmlnΓ(x),m≥1.

Their series representation is given by:

 ψ(m)=(−1)m+1m!∑k≥01(k+x)m+1.

In the sequel, we are mostly interested by the first three, i.e.

 ψ′(x)=∑k≥01(k+x)2,ψ′′(x)=−2∑k≥01(k+x)3,ψ′′′(x)=6∑k≥01(k+x)4,

and we will use the following equivalents in the neighborhood of zero, given by the first term of their series

 ψ′(x)∼x→01x2,ψ′′(x)∼x→0−2x3,ψ′′′(x)∼x→06x4. (3)

In the neighborhood of infinity, we will need the following expansions

 ψ(x)=x→+∞ln(x)−12x+o(1x2), (4) ψ′(x)=x→+∞1x+12x2+o(1x2), ψ′′(x)=x→+∞−1x2−1x3+o(1x3).

### 2.4 Curvature of the Fisher-Rao metric

In this section, we prove our main result, that is that the sectional curvature of the beta manifold is negative.

###### Proposition 3.

The sectional curvature of the Fisher metric is given by:

 K(α,β)=ψ′′(α)ψ′′(β)ψ′′(α+β)4d(α,β)2(ψ′(α)ψ′′(α)+ψ′(β)ψ′′(β)−ψ′(α+β)ψ′′(α+β)),
###### Proof.

The sectional curvature of a Hessian metric is given by

 K=14(detg)2R1212

where

 R1212=−φββ(φαααφαββ−φ2ααβ)+φαβ(φαααφβββ−φααβφαββ)+φαα(φααβφβββ−φ2αββ).

Computing the partial derivatives of the potential function gives

 φααα =ψ′′(α)−ψ′′(α+β), φβββ =ψ′′(β)−ψ′′(α+β), φααβ =φαββ=−ψ′′(α+β),

and the determinant of the metric is given by

 detg(α,β)=ψ′(α)ψ′(β)−ψ′(α+β)(ψ′(α)+ψ′(β)).

This gives

 K=ψ′′(α+β)(ψ′(α)ψ′′(β)+ψ′′(α)ψ′(β))−ψ′(α+β)ψ′′(α)ψ′′(β)4(d(α,β))2.

Factorizing the numerator by yields the desired result. ∎

###### Proposition 4.

The asymptotic behavior of the sectional curvature is given by

 limβ→0K(α,β) =limβ→0K(β,α)=34−ψ′(α)ψ′′′(α)2ψ′′(α)2, limβ→∞K(α,β) =limβ→∞K(β,α)=αψ′′(α)+ψ′(α)4(αψ′(α)−1)2.

Moreover, we have the following limits

 limα,β→0K(α,β)=0,limα,β→∞K(α,β)=−12, limα→0,β→∞K(α,β)=limα→∞,β→0K(α,β)=−14.
###### Proof.

Let us fix , and denote the varying parameter of the beta distribution. The asymptotic behavior of the sectional curvature can be obtained by separately examining its numerator and the metric determinant appearing at the denominator

 K(α,x)=N(α,x)4d(α,x)2.

Using a first order Taylor development of in and the equivalent (3), we deduce the following expansion for the determinant around zero

 d(α,x)=ψ′x(ψ′α−ψ′α+x)−ψ′a+xψ′a=x→0−ψ′′αx+o(1x).

Similarly, writing the numerator of the sectional curvature as

 N(α,x) :=ψ′′α+x(ψ′′αψ′x+ψ′αψ′′x)−ψ′α+xψ′′αψ′′x =(ψ′′α+x−ψ′′α)(ψ′′αψ′x+ψ′αψ′′x)+(ψ′a−ψ′a+x)ψ′′αψ′′x+ψ′′α(ψ′′αψ′x+ψ′αψ′′x)−ψ′aψ′′αψ′′x,

we get the following behavior around zero

 N(α,x) =x→0xψ′′′α(ψ′′αx2−2ψ′(α)x3)+x(ψ′′α)22x3+ψ′′α(ψ′′αx2−2ψ′αx3)+2ψ′αψ′′αx3+o(1x2) =x→03(ψ′′α)2−2ψ′αψ′′′αx2+o(1x2).

This yields the desired expression for the limit of as . Now, in the neighborhood of infinity, the expansions (4) yield the following behavior for the determinant

 d(α,x) =ψ′αψ′β−ψ′α+β(ψ′α+ψ′β) =x→+∞ψ′α(1x+12x2)−(1α+x+12(α+x)2)(ψ′α+1x)+o(1x2) =x→+∞αψ′α−1x2+o(1x2),

while an expansion of the numerator gives

 N(α,x) =x→+∞−(1(a+x)2+1(a+x)3)(ψ′′αx+ψ′′α2x2−ψ′αx2) +(1a+x+12(a+x)2)ψ′′α(1x2+1x4)+o(1x3) =x→+∞αψ′′α+ψ′αx4+o(1x4),

yielding again the desired limit for . Finally, approximating by when , we get

 limα→0,β→+∞K(α,β) =limα→0αψ′′α+ψ′α4(αψ′α−1)2=limα→0−1/α24/α2=−14, limα,β→+∞K(α,β)= limα,β→+∞−1/(2α2)1/α2=−12, limα,β→0K(α,β)= 34−12limα,β→06/α6(−2/α3)2=0,

which completes the proof. ∎

We can now show the following property.

###### Proposition 5.

The sectional curvature is negative and bounded from below.

###### Proof.

Recall that in its most factorized form, the sectional curvature is given by

 K(α,β)=ψ′′(α)ψ′′(β)ψ′′(α+β)4d(α,β)2(ψ′(α)ψ′′(α)+ψ′(β)ψ′′(β)−ψ′(α+β)ψ′′(α+β)),

Since is negative, the first factor is negative and so there remains to prove that the function is sub-additive, i.e.

 ψ′(α)ψ′′(α)+ψ′(β)ψ′′(β)−ψ′(α+β)ψ′′(α+β)≥0.

This has been shown recently in [12] (Corollary 4). Now, to show that it is bounded from below, set

 k1(α):=limβ→0K(α,β)=34−ψ′(α)ψ′′′(α)2ψ′′(α)2, k2(α):=limβ→+∞K(α,β)=αψ′′(α)+ψ′(α)4(αψ′(α)−1)2.

and are continuous functions on , and according to Proposition 4 they have finite limits at the boundaries

 limα→0k1(α)=0,limα→+∞k1(α)=−14, limα→0k2(α)=−14,limα→+∞k2(α)=−12.

Therefore, they are bounded, i.e., there exist negative finite constants and such that for all ,

 limβ→0K(α,β)>M1,limβ→+∞K(α,β)>M2.

Setting , notice that is a continuous function on due to the continuity of in both its variables and the invertibility of the limit and infimum. For this last reason, we also obtain

 limβ→0infα∈R∗+K(α,β)=infα∈R∗+limβ→0K(α,β)>M1, limβ→+∞infα∈R∗+K(α,β)=infα∈R∗+limβ→+∞K(α,β)>M2,

i.e., has finite limits at the boundaries and is therefore bounded, in particular from below

 infβ∈R∗+infα∈R∗+K(α,β)>−∞.

The fact that the beta manifold has negative curvature is particularly interesting for the computation of Riemannian centroids such as Fréchet or Karcher means [6, 7]. In general, the Fréchet mean on a Riemannian manifold is not unique. However, when the curvature is negative, there is no cut locus and uniqueness holds. In this context, it is defined for any given sequence of probability measure as

 ¯B=argminB∈Bn∑i=1dF(B,Bi)2.

This quantity can be computed using a gradient descent algorithm the Karcher flow algorithm.

### 2.5 A lower bound on the determinant of the metric

The determinant of the metric is the key ingredient to volume computations. In this section, a lower bound of this determinant is computed, which is also its asymptotic value.

###### Proposition 6.

The determinant of the information metric matrix admits the following integral representation:

 |g(α,β)|=∫R+∫10x(1−x)(1−e−tx)(1−e−t(1−x))((eβtx−1)(e−αt(1−x))−1)e−(α+β)tdxdt (5)
###### Proof.

The polygamma function of order can be expressed as an integral [8] :

 ψ(n)(x)=(−1)n+1∫R+tn1−e−te−xtdt,x>0 (6)

The determinant expands as:

 |g(α,β)|=(ψ′(α)−ψ′(α+β))(ψ′(β)−ψ′(α+β))−ψ′(α+β)2 (7)

Using the integral 6, it comes:

 ψ′(α)−ψ′(α+β)=∫R+t1−e−te−αt−e−(α+β)tdt=∫R+t1−e−t(eαt−1)e−(α+β)tdt (8)

The difference is thus equal to the laplace transform at of the function:

 t1−e−t(eαt−1) (9)

Using the convolution theorem [10], it comes:

 (ψ′(α)−ψ′(α+β))(ψ′(β)−ψ′(α+β))=∫R+(∫t0x(t−x)(1−e−x)(1−e−(1−x))(eβx−1)(eα(t−x)−1)dx)e−(α+β)tdt=∫R+t3(∫10x(1−x)(1−e−tx)(1−e−t(1−x))(eβtx−1)(eαt(1−x)−1)dx)e−(α+β)tdt (10)

The same procedure can be applied to the integral expression of to obtain:

 ψ′2(α+β)=∫R+t3(∫10x(1−x)(1−e−tx)(1−e−t(1−x))dx)e−(α+β)tdt (11)

Combining 10 and 11 gives:

 |g(α,β)|=∫R+t3(∫10x(1−x)(1−e−tx)(1−e−t(1−x))[(eβtx−1)(eαt(1−x)−1)−1]dx)e−(α+β)tdt (12)

Building on the integral representation of Proposition 6, it is possible to derive a lower bound for the determinant, which is also its asymptotic value.

###### Proposition 7.

The following lower bound holds:

 |g(α,β)|>1+α+β2αβ(α+β)2 (13)
###### Proof.

The hyperbolic cotangent satisfies:

 cothx2=1+e−x1−e−x (14)

and so:

 11−e−x=12+12cothx2 (15)

Letting:

 h(x,t)=x2(1+cothx2) (16)

The integral expression 12 is rewritten as:

 |g(α,β)|=∫R+t3(∫10h(x,t)h((1−x)t)[(eβtx−1)(eαt(1−x)−1)−1]dx)e−(α+β)tdt (17)

Since:

 cothx>1x,x>0

it comes:

 h(x,t)h((1−x)t)>x(1−x)4(1+2tx)(1+2t(1−x))>14(x+2t)(1−x+2t) (18)

a lower bound for 17 is thus given by:

 |g(α,β)|>∫R+t3(∫1014(x+2t)(1−x+2t)[(eβtx−1)(eαt(1−x)−1)−1]dx)e−(α+β)tdt (19)

The inner term:

 I(t)=t3∫1014(x+2t)(1−x+2t)[(eβtx−1)(eαt(1−x)−1)−1]dx

 I(t)=14a3b3(a−b)3I1(t)+I2(t)−I3(t)−I4(t)+I5(t)−I6(t) (20)

with:

 I1(t)=a6(−(2b2(t+2)(ebt−1)+bt(ebt+1)−2ebt+2)) (21) I2(t)=a5b(4b2(t+2)(ebt−1)+3bt(ebt+1)−6ebt+6) (22) I3(t)=2a4b2(b2(t+2)(−(eat−ebt))+bt(ebt+2)−3ebt+3) (23) I4(t)=2a3b4(2b(t+2)(eat−1)−t(eat+2)) (24) I5(t)=a2b4(2b2(t+2)(eat−1)−3bt(eat+1)−6eat+6) (25) I6(t)=2b6(eat−1)+ab5(eat(bt+6)+bt−6) (26)

Performing the outer integration yields finally:

 |g(α,β)|>1+α+β2αβ(α+β)2 (27)

thus completing the proof. ∎

### 2.6 A geometric view point of the central limit theorem

The central limit theorem tells us that once re-centered, a beta distribution converges at rate

to a centered normal distribution

 √n(B(nα,nβ)−αα+β)→n→∞N(0,ab(a+b)3).

For a fixed , the line corresponds to all the beta distributions of mean . Asymptotically, we retrieve a hyperbolic distance between two distributions on this line.

###### Proposition 8.

When for a fixed , the metric is asymptotically

 ds2=dα22α2+o(1α2).

This means

 dF(B(nα,nλα),B(nα′,nλα′))→n→∞dF(N(0,α),N(0,α′))
###### Proof.

The infinitesimal element of length is given by

 ds2=(ψ′(α)−ψ′(α+β))dα2+(ψ′(β)−ψ′(α+β))dβ2−2ψ′(α+β)dαdβ,

and so when for a fixed ,

 ds2=G(α)dα2

where

 G(α)=ψ′(α)+λ2ψ′(λα)−(1+λ)2ψ′((1+λ)α).

When , we have asymptotically using (4)

 G(α)=1α+12α2+