Minimum Wasserstein Distance Estimator under Finite Location-scale Mixtures

1 Introduction

Let $F={f(⋅|\boldmathθ):\boldmathθ∈Θ}$ be a parametric distribution family with density function $f(⋅|\boldmathθ)$ with respect to some $σ$ -finite measure. Denote by $G=∑Kk=1wk{\boldmathθk}$

a distribution assigning probability

w_{k}

\boldmathθk∈Θ

. A distribution with the following density function

f(x|G)=∫f(x|\boldmathθ)dG(\boldmathθ)=K∑k=1wkf(x|\boldmathθk)

is called a finite $F$ mixture. We call $f(x|\boldmathθ)$ the subpopulation density function, $θ$ the subpopulation parameter, and $w_{k}$ the mixing weight of the $k$ th subpopulation. We use $F(x|\boldmathθ)$ and $F (x | G)$

for the cumulative distribution functions (CDF) of

f(x|\boldmathθ)

and

f (x | G)

respectively. Let

GK={G:G=K∑k=1wk{\boldmathθk},0≤wk≤1,K∑k=1wk=1,\boldmathθk∈Θ}

be a space of mixing distributions with at most $K$ support points. A mixture distribution of (exactly) order $K$ has its mixing distribution $G$ being a member of $G_{K} - G_{K - 1}$ .

We study the problem of learning the mixing distribution $G$ given a set of independent and identically distributed (IID) observations $X = {x_{1}, x_{2}, \dots, x_{N}}$ from a mixture $f (x | G)$ . Throughout the paper, we assume the order of $G$ is known and $F$ is a known location-scale family. That is,

f(x|\boldmathθ)=1σf0(x−μσ)

for some probability density function

f_{0} (x)

with

x \in R

with respect to Lebesgue measure where

\boldmathθ=(μ,σ)

with

Θ = {R \times R^{+}}

Finite mixture models provide a natural representation of heterogeneous population that is believed to be composed of several homogeneous subpopulations (Pearson, 1894; Schork et al., 1996). They are also useful for approximating distributions with unknown shapes which are particularly relevant in image generation (Kolouri et al., 2018), image segmentation (Farnoosh and Zarpak, 2008), object tracking (Santosh et al., 2013), and signal processing (Plataniotis and Hatzinak, 2000).

In statistics, the most fundamental task is to learn the unknown parameters. In early days, the method of moments was the choice for its ease of computation

(Pearson, 1894) under finite mixture models. Nowadays, the maximum likelihood estimate (MLE) is the first choice due to its statistical efficiency and the availability of an easy-to-use EM-algorithm. Under a finite location-scale mixture model, the log-likelihood function of

G

is given by

(1)

At an arbitrary mixing distribution $G_{ϵ} = 0.5 {(x_{1}, ϵ)} + 0.5 {(0, 1)},$ we have $ℓ_{N} (G_{ϵ} | X) \to \infty$ as $ϵ \to 0$ . Hence, the MLE of $G$ is not well defined or is ill defined. Various remedies, such as penalized maximum likelihood estimate (pMLE), has been proposed to overcome this obstacle (Chen et al., 2008; Chen and Tan, 2009)

. At the same time, MLE can be thought of a special minimum distance estimator. It minimizes a specific Kullback-Leibler divergence between the empirical distribution and the assumed model

F

. Other divergences and distances have been investigated in the literature as in Choi (1969); Yakowitz (1969); Woodward et al. (1984); Clarke and Heathcote (1994); Cutler and Cordero-Brana (1996); Deely and Kruse (1968). Recently, the Wasserstein distance has drawn increased attention in machine learning community due to its intuitive interpretation and good geometric properties (Evans and Matsen, 2012; Arjovsky et al., 2017). The Wasserstein distance based estimator for learning finite mixture models is absent in the literature.

Are there any benefits to learn finite location-scale mixtures by the minimum Wasserstein distance estimator (MWDE)? This paper answers this question from several angles. We find that the MWDE is consistent and derive a numerical solution under finite location-scale mixtures. We compare the robustness of the MWDE with pMLE in the presence of outliers and mild model mis-specifications. We conclude that the MWDE suffers some efficiency loss against pMLE in general without obvious gain in robustness. Through this paper, we better understand the pros and cons of the MWDE under finite location-scale mixtures. We reaffirm the general superiority of the likelihood based learning strategies even for the non-regular finite location-scale mixtures.

In the next section, we first introduce the Wasserstein distance and some of its properties. This is followed by a formal definition of the MWDE, a discussion of its existence and consistency under finite location-scale mixtures. In Section 2.4, we give some algebraic results that are essential for computing $2$ -Wasserstein distance between the empirical distribution and the finite location-scale mixtures. We then develop a BFGS algorithm scheme for computing the MWDE of the mixing distribution. In addition, we briefly review the penalized likelihood approach and its numerical issues. In Section 3, we characterize the efficiency properties of the MWDE relative to pMLE in various circumstances via simulation. We also study their robustness when the data contains outliers, is contaminated or when the model is mis-specified. We then apply both methods in an image segmentation example. We conclude the paper with a summary in Section 4.

2 Wasserstein Distance and the Minimum Distance Estimator

2.1 Wasserstein Distance

Wasserstein distance is a distance between probability measures. Let $Ω$ be a Polish space endowed with a ground distance $D (\cdot, \cdot)$ and $P (Ω)$ the space of Borel probability measures on $Ω$ . Let $η \in P (Ω)$ be a probability measure. If for some $p > 0$ ,

\int_{Ω} D^{p} (x, x_{0}) η (d x) < \infty,

for some (and thus any) $x_{0} \in Ω$ , we say $η$ has finite $p$ th moment. Denote by $P_{p} (Ω) \subset P (Ω)$ the space of probability measures with finite $p$ th moment. For any $η, ν \in P (Ω)$ , we use $Π (η, ν)$ to denote the space of the bivariate probability measures on $Ω \times Ω$ whose marginals are $η$ and $ν$ . Namely,

Π (η, ν) = {π \in P (Ω^{2}) : \int_{Ω} π (x, d y) = η (x), \int_{Ω} π (d x, y) = ν (y)} .

The $p$ -Wasserstein distance is defined as follows.

Definition 2.1 ( $p$ -Wasserstein distance).

For any $η, ν \in P_{p} (Ω)$ with $p \geq 1$ , the $p$ th Wasserstein distance between $η$ and $ν$ is

W_{p} (η, ν) = {inf π \in Π (η, ν) \int_{Ω^{2}} D^{p} (x, y) π (d x, d y)}^{1 / p} .

Suppose $X$ and $Y$

are two random variables whose distributions are

F

and

G

and induced probability measures are

η

and

ν

. We regard the

p

-Wasserstein distance between

η

and

ν

also the distance between random variables or distributions:

W_{p} (X, Y) = W_{p} (F, G) = W_{p} (η, ν)

The $p$ -Wasserstein distance is a distance on $P_{p} (Ω)$ as shown by Villani (2003, Theorem 7.3). For any $η, ν, ξ \in P_{p} (Ω)$ , it has the following properties:

Non-negativity: $W_{p} (η, ν) \geq 0$ and $W_{p} (η, ν) = 0$ if and only if $η = ν$ ;
Symmetry: $W_{p} (η, ν) = W_{p} (ν, η)$ ;
Triangular inequality: $W_{p} (η, ν) \leq W_{p} (η, ξ) + W_{p} (ξ, ν)$ .

The Wasserstein distance has many nice properties. Let us denote $η_{n} d ⟶ η$ for convergence in distribution or measure. Villani (2003, Theorem 7.1.2) shows that it has the following properties:

Property 1. For any $q \geq p \geq 1$ , $W_{q} (η, ν) \geq W_{p} (η, ν)$ .
Property 2. $W_{p} (η_{n}, η) \to 0$ as $n \to \infty$ if and only if both
- $η_{n} d ⟶ η$ , and
- $\int D^{p} (x, x_{0}) η_{n} (d x) \to \int D^{p} (x, x_{0}) η (d x)$ for some (and thus any) $x_{0} \in Ω$ .

Computing the Wasserstein distance involves a challenging optimization problem in general but has a simple solution under a special case. Suppose $Ω$ is the space of real numbers, $D (x, y) = | x - y |$ , and $F$ and $G$ are univariate distributions. Let $F^{- 1} (t) := inf {x : F (x) \geq t}$ and $G^{- 1} (t) := inf {x : G (x) \geq t}$ for $t \in [0, 1]$

be their quantile functions. We can easily compute the Wasserstein distance based on the following property.

Property 3. $W_{p} (F, G) = {\int_{0}^{1} | F^{- 1} (t) - G^{- 1} (t) |^{p} d t}^{1 / p}$ .

2.2 Minimum Wasserstein Distance Estimator

Let $W_{p} (\cdot, \cdot)$ be the $p$ -Wasserstein distance with ground distance $D (x, y) = | x - y |$ for univariate random variables. Let $X = {x_{1}, x_{2}, \dots, x_{N}}$ be a set of IID observations from finite location-scale mixture $f (x | G)$ of order $K$ and $F_{N} (x) = N^{- 1} \sum_{n = 1}^{N} 1 (x_{n} \leq x)$ be the empirical distribution. We introduce the MWDE of the mixing distribution $G$ that is

{^G}_{N}^{MWDE} = a r g i n f G \in G_{K} W_{p} (F_{N} (\cdot), F (\cdot | G)) = a r g i n f G \in G_{K} W_{p}^{p} (F_{N} (\cdot), F (\cdot | G)) .

(2)

As we pointed out earlier, the MLE is not well defined under finite location-scale mixtures. Is the MWDE well defined? We examine the existence or sensibility of the MWDE. We show that the MWDE exists when $f_{0} (\cdot)$ satisfies certain conditions.

Assume that $f_{0} (0) > 0$ , $f_{0} (x)$ is bounded, continuous, and has finite $p$ th moment. Under these conditions, we can see

0 \leq W_{p} (F_{N} (\cdot), F (\cdot | G)) < \infty

for any $G \in G_{K}$ . When $N \leq K$ , the solution to (2) merits special attention. Let $G_{ϵ} = \sum_{n = 1}^{N} (1 / N) {(x_{n}, ϵ)}$ be a mixing distribution assigning probability $1 / N$ on $\boldmathθn=(xn,ϵ)$ . When $ϵ \to 0$ , each subpopulation in the mixture $f (x | G_{ϵ})$ degenerates to a point mass at $x_{n}$ . Hence, as $ϵ \to 0$ ,

W_{p} (F_{N} (\cdot), F (\cdot | G_{ϵ})) \to 0.

Since none of $G \in G_{K}$ has zero-distance from $F_{N} (\cdot)$ , the MWDE does not exist unless we expand $G_{K}$ to include $G_{0} = \sum_{n = 1}^{N} (1 / N) {(x_{n}, 0)} = lim G_{ϵ}$ . To remove this technical artifact, in the MWDE definition we expand the space of $σ$ to $[0, \infty)$ . We denote by $F (\cdot | (θ_{0}, 0))$ a distribution with point mass at $x = θ_{0}$ . With this expansion, $G_{0}$ is the MWDE when $N \leq K$ .

Let $δ = inf {W_{p} (F_{N} (\cdot), F (\cdot | G)) : G \in G_{K}}$ . Clearly, $0 \leq δ < \infty$ . By definition, there exists a sequence of mixing distributions $G_{m} \in G_{K}$ such that $W_{p} (F_{N} (\cdot), F (\cdot | G_{m})) \to δ$ as $m \to \infty$ . Suppose one mixing weight of $G_{m}$ has limit 0. Removing this support point and rescaling, we get a new mixing distribution sequence and it still satisfies $W_{p} (F_{N} (\cdot), F (\cdot | G_{m})) \to δ$ . For this reason, we assume that its mixing weights have non-zero limits by selecting converging subsequence if necessary to ensure the limits exist. Further, when the mixing weights of $G_{m}$ assume their limiting values while keeping subpopulation parameters the same, we still have $W_{p} (F_{N} (\cdot), F (\cdot | G_{m})) \to δ$ as $m \to \infty$ . In the following discussion, we therefore discuss the sequence of mixing distributions whose mixing weights are fixed.

Suppose the first subpopulation of $G_{m}$ has its scale parameter $σ_{1} \to \infty$ as $m \to \infty$ . With the boundedness assumption on $f_{0} (x)$ , the mass of this subpopulation will spread thinly over entire $R$ because $σ_{1}^{- 1} f_{0} ((x - μ_{1}) / σ_{1}) \to 0$ uniformly. For any fixed finite interval, [ $a, b$ ], this thinning makes

F(b|\boldmathθ1)−F(a|\boldmathθ1)→0

as $m \to \infty$ . It implies that for any given $t \in (0, 0.5)$ , we have

|F−1(t|\boldmathθ1)|+|F−1(1−t|\boldmathθ1)|→∞.

This further implies for any $t \in (0, w_{1} / 2)$ , we have

| F^{- 1} (t | G_{m}) | + | F^{- 1} (1 - t | G_{m}) | \to \infty

as $m \to \infty$ . In comparison, the empirical quantile satisfies $x_{(1)} \leq F_{N}^{- 1} (t) \leq x_{(N)}$ for any $t$ . By Property 3 of $W_{p} (\cdot, \cdot)$ , these lead to $W_{p} (F_{N} (\cdot), F (\cdot | G_{m})) \to \infty$ as $m \to \infty$ . This contradicts the assumption $W_{p} (F_{N} (\cdot), F (\cdot | G_{m})) \to δ$ . Hence, $σ_{1} \to \infty$ is not a possible scenario of $G_{m}$ nor $σ_{k} \to \infty$ for any $k$ .

Can a subpopulation of $G_{m}$ instead have its location parameter $μ \to \infty$ ? For definitiveness, let this subpopulation correspond to $\boldmathθ1$ . Note that at least $w_{1} {1 - F_{0} (0)}$ -sized probability mass of $F (x | G_{m})$ is contained in the range $[μ_{1}, \infty)$ . Because of this, when $μ_{1} \to \infty$ , we have $F^{- 1} (1 - t | G_{m}) \to \infty$ for $t = w_{1} {1 - F_{0} (0)} / 2$ . Therefore, $W_{p} (F_{N} (\cdot), F (\cdot | G_{m})) \to \infty$ by Property 3. This contradicts $W_{p} (F_{N} (\cdot), F (\cdot | G_{m})) \to δ < \infty$ . Hence, $μ_{1} \to \infty$ is not a possible scenario of $G_{m}$ either. For the same reason, we cannot have $μ_{k} \to \pm \infty$ for any $k$ .

After ruling out $μ_{k} \pm \infty$ and $σ_{k} \to \infty$ , we find $G_{m}$ has a converging subsequence whose limit is a proper mixing distribution in $G_{K}$ . This limit is then an MWDE and the existence is verified.

The MWDE may not be unique and the mixing distribution may lead to a mixture with degenerate subpopulations. We will show that the MWDE is consistent as the sample size goes to infinity. Thus, having degenerated subpopulations in the learned mixture is a mathematical artifact and also a sensible solution. In contrast, no matter how large the sample size becomes, there are always degenerated mixing distributions with unbounded likelihood values.

2.3 Consistency of MWDE

We consider the problem when $X = {x_{1}, \dots, x_{N}}$ are IID observations from a finite location-scale mixture of order $K$ . The true mixing distribution is denoted as $G^{*}$ . Assume that $f_{0} (x)$ is bounded, continuous, and has finite $p$ th moment. We say the location-scale mixture is identifiable if

F (x | G_{1}) = F (x | G_{2})

for all $x$ given $G_{1}, G_{2} \in G_{K}$ implies $G_{1} = G_{2}$ . We allow subpopulation scale $σ = 0$ . The most commonly used finite locate-scale mixtures, such as the normal mixture, are well known to be identifiable (Teicher, 1961). Holzmann et al. (2004) give a sufficient condition for the identifiability of general finite location-scale mixtures. Let $φ (\cdot)$

be the characteristic function of

f_{0} (t)

. The finite location-scale mixture is identifiable if for any

σ_{1} > σ_{2} > 0

{lim}_{t \to \infty} φ (σ_{1} t) / φ (σ_{2} t) = 0

We consider the MWDE based on $p$ -Wasserstein distance with ground distance $D (x, y) = | x - y |$ for some $p \geq 1$ . The MWDE under finite location-scale mixture model as defined in (2) is asymptotically consistent.

Theorem 2.1.

With the same conditions on the finite location-scale mixture and same notations above, we have the following conclusions.

For any sequence $G_{m} \in G_{K}$ and $G^{*} \in G_{K}$ , $W_{p} (F (\cdot | G_{m}), F (\cdot | G^{*})) \to 0$ implies $G_{m} d ⟶ G^{*}$ as $m \to \infty$ .
The MWDE satisfies $W_{p} (F (\cdot | G^{*}), F (\cdot | {^G}_{N}^{MWDE})) \to 0$ as $N \to \infty$ almost surely.
The MWDE is consistent: $W_{p} ({^G}_{N}^{MWDE}, G^{*}) \to 0$ as $N \to \infty$ almost surely.

Proof.

We present these three conclusions in the current order which is easy to understand. For the sake of proof, a different order is better. For ease presentation, we write $F^{*} = F (\cdot | G^{*})$ and $^G = {^G}_{N}^{MWDE}$ in this proof.

We first prove the second conclusion. By the triangular inequality and the definition of the minimum distance estimator, we have

W_{p} (F^{*}, F (\cdot | {^G}_{N})) \leq W_{p} (F_{N}, F^{*}) + W_{p} (F_{N}, F (\cdot | {^G}_{N})) \leq 2 W_{p} (F_{N}, F^{*}) .

Note that $F_{N}$ is the empirical distribution and $F^{*}$ is the true distribution, we have $F_{N} (x) \to F^{*} (x)$ uniformly in $x$ almost surely. At the same time, under the assumption that $F_{0} (x)$ has finite $p$ th moment, $F^{*} (x)$ also has finite $p$ th moment. The $p$ th moment of $F_{N} (x)$ converges to that of $F^{*} (x)$ almost surely. Given the ground distance $D (x, y) = | x - y |$ , the $p$

th moment in Wasserstein distance sense is the usual moments in probability theory. By Property 2, we conclude

W_{p} (F_{N}, F (\cdot | G^{*})) \to 0

as both conditions there are satisfied.

Conclusion 3 is implied by Conclusions 1 and 2. With Conclusion 2 already established, we need only prove Conclusion 1 to complete the whole proof. By Helly’s lemma (Van der Vaart, 2000, Lemma 2.5) again, $G_{m}$ has a converging subsequence though the limit can be a sub-probability measure. Without loss of generality, we assume that $G_{m}$ itself converges with limit $~ G$ . If $~ G$ is a sub-probability measure, so would be $F (\cdot | ~ G)$ . This will lead to

W_{p} (F (\cdot | G_{m}), F (\cdot | G^{*})) \to W_{p} (F (\cdot | ~ G), F (\cdot | G^{*})) \neq 0

which violates the theorem condition. If $~ G$ is a proper distribution in $G_{K}$ and

W_{p} (F (\cdot | ~ G), F (\cdot | G^{*})) = 0,

then by identifiability condition, we have $~ G = G^{*}$ . This implies $G_{m} \to G^{*}$ and completes the proof. ∎

The multivariate normal mixture is another type of location-scale mixture. The above consistency result of MWDE can be easily extended to finite multivariate normal mixtures.

Theorem 2.2.

Consider the problem when $X = {x_{1}, \dots, x_{N}}$ are IID observations from a finite multivariate normal mixture distribution of order $K$ and ${^G}_{N}^{MWDE}$ is the minimum Wasserstein distance estimator defined by (2). Let the true mixing distribution be $G^{*}$ . The MWDE is consistent: $W_{p} ({^G}_{N}^{MWDE}, G^{*}) \to 0$ as $N \to \infty$ almost surely.

The rigorous proof is long though the conclusion is obvious. We offer a less formal proof based on several well known probability theory results:

A multivariate random variable sequence $Y_{n}$ converges in distribution to $Y$ if and only if $a^{τ} Y_{n}$ converges to $a^{τ} Y$
for any unit vector
$a$ ;
If $Y$ is multivariate normal if and only if $a^{τ} Y$ is normal for all $a$ ;
The normal distribution has finite moment of any order.

Let $X_{m}$ be a random vector with distribution $F (\cdot | G_{m})$ for some $G_{m} \in G_{K}$ , $m = 0, 1, 2, \dots$ , in a general mixture model setting. Suppose as $m \to \infty$ , with the notation we introduced previously,

W_{p} (X_{m}, X_{0}) \to 0.

Then for any unit vector $a$ , based on property 2 of the Wasserstein distance and the result (I), we can see that

W_{p} (a^{τ} X_{m}, a^{τ} X_{0}) \to 0.

Next, we apply this result to normal mixture so that $F (\cdot | G_{m})$ becomes $Φ (\cdot | G_{m})$ which stands for a finite multivariate normal mixture with mixing distribution $G_{m}$ . In this case, $X_{m}$ is a random vector with distribution $Φ (\cdot | G_{m})$ . Let $(\boldmathμk,Σk)$ be generic subpopulation parameters. We can see that the distribution of $a^{τ} X_{m}$ , $Φ_{a} (\cdot | G_{m})$ is a finite normal mixture with subpopulation parameters $(aτ\boldmathμk,aτΣka)$ , and mixing weights the same as those of $G_{m}$ . Let the mixing distributions after projection be $G_{m, a}$ and $G_{0, a}$ .

By the same argument in the proof of Theorem 2.1,

W_{p} (Φ (\cdot | {^G}_{N}), Φ (\cdot | G^{*})) \to 0

almost surely as $N \to \infty$ . This implies

W_{p} (Φ_{a} (\cdot | {^G}_{N}), Φ_{a} (\cdot | G^{*})) \to 0

almost surely as $N \to \infty$ for any $a$ . Hence, by Conclusion 1 of Theorem 2.1, ${^G}_{N, a} d ⟶ {^G}_{a}^{*}$ almost surely for any unit vector $a$ . We therefore conclude the consistency result: ${^G}_{N} d ⟶ {^G}^{*}$ almost surely.

2.4 Numerical Solution to MWDE

Both in applications and in simulation experiments, we need an effective way to compute the MWDE. We develop an algorithm that leverages the explicit form of the Wasserstein distance between two measures on $R$ for the numerical solution to the MWDE. The strategy works for any $p$ -Wasserstein distance but we only provide specifics for $p = 2$ as it is the most widely used.

Let $Y$ be a random variable with distribution $f_{0} (\cdot)$

. Denote the mean and variance of

Y

μ_{0} = E (Y)

and

σ_{0}^{2} = Var (Y)

. Recall that

G = \sum_{k = 1}^{K} w_{k} {(μ_{k}, σ_{k})}

. Let

x_{(1)} \leq x_{(2)} \leq \dots \leq x_{(N)}

be the order statistics,

^{2} = N^{- 1} \sum_{n = 1}^{N} x_{n}^{2}

, and

ξ_{n} = F^{- 1} (n / N | G)

be the

(n / N)

th quantile of the mixture for

n = 0, 1, \dots, N

. Let

T (x) = \int_{- \infty}^{x} t f_{0} (t) d t

and define

	$Δ F_{n k}$
	$Δ T_{n k}$	$= T (\frac{ξ_{n} - μ_{k}}{σ_{k}}) - T (\frac{ξ_{n - 1} - μ_{k}}{σ_{k}}) .$

When $p = 2$ , we expand the squared $W_{2}$ distance, $W_{N}$ , between the empirical distribution and $F (\cdot | G)$ as follows:

$W_{N} (G)$	$=$	$W_{2}^{2} (F_{N} (\cdot), F (\cdot \| G))$
	$=$	$\int_{0}^{1} {F_{N}^{- 1} (t) - F^{- 1} (t \| G)}^{2} d t$
	$=$	$^{2} + K \sum k = 1 w_{k} {μ_{k}^{2} + σ_{k}^{2} (μ_{0}^{2} + σ_{0}^{2}) + 2 μ_{k} σ_{k} μ_{0}}$
		$- 2 \sum k w_{k} {μ_{k} N \sum n = 1 x_{(n)} Δ F_{n k} + σ_{k} N \sum n = 1 x_{(n)} Δ T_{n k}} .$

The MWDE minimizes $W_{N} (G)$ with respect to $G$ . The mixing weights and subpopulation scale parameters in this optimization problem have natural constraints. We may replace the optimization problem with an unconstrained one by the following parameter transformation:

	$σ_{k} = exp (τ_{k}),$
	$w_{k} = exp (t_{k}) / {K \sum j = 1 exp (t_{j})}$

for $k = 1, 2, \dots, K$ . We may then minimize $W_{N}$ with respect to ${(μ_{k}, τ_{k}, t_{k}) : k = 1, 2, \dots, K}$ over the unconstrained space $R^{3 K}$ . Furthermore, we adopt the quasi-Newton BFGS algorithm (Nocedal and Wright, 2006, Section 6.1). To use this algorithm, it is best to provide the gradients of $W_{N} (G)$ , which are given as follows:


	$\frac{\partial}{\partial μ_{j}} W_{N} = 2 w_{j} {μ_{j} + σ_{j} μ_{0} - N \sum n = 1 x_{(n)} Δ F_{n j}},$
	$\frac{\partial}{\partial τ_{j}} W_{N} = 2 w_{j} {σ_{j} (μ_{0}^{2} + σ_{0}^{2}) + μ_{j} μ_{0} - N \sum n = 1 x_{(n)} Δ T_{n j}} \frac{\partial σ_{j}}{\partial τ_{j}}$

for $j = 1, 2, \dots, K$ where

	$\frac{\partial}{\partial w_{k}} W_{N} =$	${μ_{k}^{2} + σ_{k}^{2} (μ_{0}^{2} + σ_{0}^{2}) + 2 μ_{k} σ_{k} μ_{0}} - 2 N - 1 \sum n = 1 {x_{(n + 1)} - x_{(n)}} ξ_{n} F (ξ_{n} \| μ_{k}, σ_{k})$
		$- 2 {μ_{k} N \sum n = 1 x_{(n)} Δ F_{n k} + σ_{k} N \sum n = 1 x_{(n)} Δ T_{n k}} .$

Since $W_{N} (G)$ is non-convex, the algorithm may find a local minimum of $W_{N} (G)$ instead of a global minimum as required for MWDE. We use multiple initial values for the BFGS algorithm, and regard the one with the lowest $W_{N} (G)$ value as the solution. We leave the algebraic details in the Appendix.

This algorithm involves computing the quantiles $ξ_{n}$ and $Δ T_{n j}$ repeatedly which may lead to high computational cost. Since $ξn∈[minkF−1(n/N|\boldmathθk),maxkF−1(n/N|\boldmathθk)]$ , it can be found efficiently via a bisection method. Fortunately, $T (x)$ has simple analytical forms under two widely used location-scale mixtures which make the computation of $Δ T_{n j}$ efficient:

When $f_{0} (t) = (2 π)^{- 1 / 2} exp (- x^{2} / 2)$ which is the density function of the standard normal, we have $t f_{0} (t) = - f_{0}^{'} (t)$ . In this case, we find

$T (x) = - f_{0} (x) .$
For finite mixture of location-scale logistic distributions, we have

$f_{0} (t) = \frac{exp (- x)}{(1 + exp (- x))^{2}}$

and

$T (x) = \int_{- \infty}^{x} t f_{0} (t) d t = \frac{x}{1 + exp (- x)} - log (1 + exp (x)) .$ (3)

2.5 Penalized Maximum Likelihood Estimator

A well investigated inference method under finite mixture of location-scale families is the pMLE (Tanaka, 2009; Chen et al., 2008). Chen et al. (2008) consider this approach for finite normal mixture models. They recommend the following penalized log-likelihood function

p ℓ_{N} (G | X) = ℓ_{N} (G | X) - a_{N} \sum k {s_{x}^{2} / σ_{k}^{2} + log σ_{k}^{2}}

for some positive $a_{N}$ and sample variance $s_{x}^{2}$ . The log-likelihood function is given in (1). They suggest to learn the mixing distribution $G$ via pMLE defined as

{^G}_{N}^{pMLE} = a r g s u p p ℓ_{N} (G | X) .

The size of $a_{N}$ controls the strength of the penalty and a recommended value is $N^{- 1 / 2}$ . Regularizing the likelihood function via a penalty function fixes the problem caused by degenerated subpopulations (i.e. some $σ_{k} = 0$ ). The pMLE is shown to be strongly consistent when the number of components has a known upper bound under the finite normal mixture model.

The penalized likelihood approach can be easily extended to finite mixture of location-scale families. Let $f_{0} (\cdot)$ be the density function in the location-scale family as before. We may replace the sample variance $s_{x}^{2}$

in the penalty function by any scale-invariance statistic such as the sample inter-quartile range. This is applicable even if the variance of

f_{0} (\cdot)

is not finite.

We can use the EM algorithm for numerical computation. Let $z_{n} = (z_{n 1}, \dots, z_{n K})$ be the membership vector of the $n$ th observation. That is, the $k$ th entry of $z_{n}$ is 1 when the response value $x_{n}$ is an observation from the $k$ th subpopulation and 0 otherwise. When the complete data ${(z_{n}, x_{n}), n = 1, 2, \dots, N}$ are available, the penalized complete data likelihood function of $G$ is given by

p ℓ_{N}^{c} (| X) = N \sum n = 1 K \sum k = 1 z_{n k} log {\frac{}{w_{k}} σ_{k} f_{0} (\frac{x_{i} - μ_{k}}{σ_{k}})} - a_{N} \sum k {s_{x}^{2} / σ_{k}^{2} + log (σ_{k}^{2})} .

Given the observed data $X$ and proposed mixing distribution $G^{(t)}$ , we have the conditional expectation

w_{n k}^{(t)} = E (z_{n k} | X, G^{(t)}) = \frac{w_{k}^{(t)} f (x_{n} | μ_{k}^{(t)}, σ_{k}^{(t)})}{\sum_{j = 1}^{K} w_{j}^{(t)} f (x_{n} | μ_{j}^{(t)}, σ_{j}^{(t)})} .

After this E-step, we define

Q (G | G^{(t)}) =

N \sum n = 1 K \sum k = 1 w_{n k}^{(t)} log {\frac{w_{k}}{σ_{k}} f_{0} (\frac{x_{n} - μ_{k}}{σ_{k}})} - a_{N} \sum k {s_{x}^{2} / σ_{k}^{2} + log (σ_{k}^{2})} .

Note that the subpopulation parameters are well separated in $Q (\cdot | \cdot)$ . The M-step is to maximize $Q (G | G^{(t)})$ with respect to $G$ . The solution is given by the mixing distribution $G^{(t + 1)}$ with mixing weights

w_{k}^{(t + 1)} = N^{- 1} N \sum n = 1 w_{n k}^{(t)}

and the subpopulation parameters

\boldmathθ(t+1)k=argminθ{∑nw(t)nk{logσ−f(xn|\boldmathθ)}+aN{s2x/σ2+logσ2}}

(4)

with the notational convention $\boldmathθ=(μ,σ)$ .

For general location-scale mixture, the M-step (4) may not have a closed form solution but it is merely a simple two-variable function. There are many effective algorithms in the literature to solve this optimization problem. The EM-algorithm for pMLE increases the value of the penalized likelihood after each iteration. Hence, it should converge as long as the penalized likelihood function has an upper bound. We do not give a proof as it is a standard problem.

3 Experiments

We now study the performance of MWDE and pMLE under finite location-scale mixtures. We explore the potential advantages of the MWDE and quantify its efficiency loss, if any, by simulation experiments. Consider the following three location-scale families (Chen et al., 2020):

Normal distribution: $f_{0} (x) = (2 π)^{- 1 / 2} exp (- x^{2} / 2)$ . Its mean and variance are given by $μ_{0} = 0$ and $σ_{0}^{2} = 1$ .
Logistic distribution: $f_{0} (x) = exp (- x) / (1 + exp (- x))^{2}$ . Its mean and variance are given by $μ_{0} = 0$ and $σ_{0}^{2} = π^{2} / 3$ .
Gumbel distribution (type I extreme-value distribution): $f_{0} (x) = exp (- x - exp (- x))$ . Its mean and variance are given by $μ_{0} = γ$ and $σ_{0}^{2} = π^{2} / 6$ where $γ$ is the Euler constant.

We will also include a real data example to compare the image segmentation result of using the MWDE and pMLE.

3.1 Performance Measure

For vector valued parameters, the commonly used performance metric of their estimators is the mean squared error (MSE). A mixing distribution with finite and fixed support points can be regarded as a real-valued vector in theory. Yet the mean squared errors of the mixing weights, the subpopulation means, and the subpopulation scales are not comparable in terms of the learned finite mixture. In this study, we use two performance metrics specific for finite mixture models. Let $^G$ and $G^{*}$ be the learned mixing distribution and the true mixing distribution. We use $L_{2}$ distance between the learned mixture and the true mixture as the first performance metric. The $L_{2}$ distance between two mixtures $f (\cdot | G)$ and $f (\cdot | ~ G)$ is defined to be

L_{2} (f (\cdot | G), f (\cdot | ~ G)) = {w^{τ} S_{G G} w - 2 w^{τ} S_{G ~ G} ~ w + {~ w}^{τ} S_{~ G ~ G} ~ w}^{1 / 2}

where $S_{G G}, S_{G ~ G}$ and $S_{~ G ~ G}$ are three square matrices of size $K \times K$ with their $(n, m)$ th elements given by

∫f(x|\boldmathθn)f(x|\boldmathθm)dx,∫f(x|\boldmathθn)f(x|~\boldmathθm)dx,∫f(x|~\boldmathθn)f(x|~% \boldmathθm)dx.

Given an observed value $x$

of a unit from the true mixture population, by Bayes’ theorem, the most probably membership of this unit is given by

k∗(x)=argmaxk{w∗kf∗(x|\boldmathθ∗k)}.

Following the same rule, if $^G$ is the learned mixing distribution, then the most likely membership of the unit with observed value $x$ is

^k(x)=argmaxk{^wkf(x|^\boldmathθk)}.

We cannot directly compare $k^{*} (x)$ and $^k (x)$ because the subpopulation themselves are not labeled. Instead, the adjusted rand index (ARI) is a good performance metric for clustering accuracy. Suppose the observations in a dataset are divided into $K$ clusters $A_{1}, A_{2}, \dots, A_{K}$ by one approach, and $K^{'}$ clusters $B_{1}, B_{2}, \dots, B_{K^{'}}$ by another. Let $N_{i} = # (A_{i}), M_{j} = # (B_{j}), N_{i j} = # (A_{i} B_{j})$ for $i, j = 1, 2, \dots, K$ , where $# (A)$ is the number of units in set $A$ . The ARI between these two clustering outcomes is defined to be

ARI = \frac{\sum_{i, j} (\frac{N_{i j}}{2}) - {(\frac{N}{2})}^{- 1} \sum_{i, j} (\frac{N_{i}}{2}) (\frac{M_{j}}{2})}{\frac{1}{2} \sum_{i} (\frac{N_{i}}{2}) + \frac{1}{} 2 \sum_{j} (\frac{M_{j}}{2}) - {(\frac{N}{2})}^{- 1} \sum_{i, j} (\frac{N_{i}}{2}) (\frac{M_{j}}{2})} .

When the two clustering approaches completely agree with each other, the ARI value is $1$ . When data are assigned to clusters randomly, the expected ARI value is $0$ . ARI values close to 1 indicate a high degree of agreement. We compute ARI based on clusters formed by $k^{*} (x)$ and $^k (x)$ .

For each simulation, we choose or generate a mixing distribution $G^{* (r)}$ , then generate a random sample from mixture $f (x | G^{* (r)})$ . This is repeated $R$ times. Let ${^G}^{(r)}$ be the learned $G$ based on the $r$ th data set. We obtain the two performance metrics as follows:

Mean $L_{2}$ distance:

$ML2 = R^{- 1} R \sum r = 1 L_{2} (f (\cdot | {^G}^{(r)}), f (\cdot | G^{* (r)})) .$
Mean adjusted rand index:

$MARI = R^{- 1} R \sum r = 1 ARI ({^G}^{(r)}, G^{* (r)}) .$

The lower the ML2 and the higher the MARI, the better the estimator performs.

3.2 Performance under Homogeneous Model

The homogeneous location-scale model is a special mixture model with a single subpopulation $K = 1$ . Both MWDE and MLE are applicable for parameter estimation. There have been no studies of MWDE in this special case in the literature. It is therefore of interest to see how MWDE performs under this model.

Under three location-scale models given earlier, the MWDE has closed analytical forms. Using the same notation introduced, their analytical forms are as follows.

Normal distribution:

$^μMWDE=¯x, ^σMWDE=N∑n=1x(n){f0(ξn−1)−f0(ξn)}.$
Logistic distribution:

$^μMWDE=¯x, ^σMWDE=3π2N∑n=1x(n){T(ξn)−T(ξn−1)}$

where $T (x)$ is given in (3).
Gumbel distribution:

$^μMWDE={1−γr}−1{¯x−γT}, ^σMWDE=T−r^μMWDE$

where

$T = {γ^{2} + π^{2} / 6}^{- 1} N \sum n = 1 x_{(n)} \int_{ξ_{n -}}^{ξ_{n}}$

Minimum Wasserstein Distance Estimator under Finite Location-scale Mixtures

Homogeneity testing under finite location-scale mixtures

The Sketched Wasserstein Distance for mixture distributions

Approximation of probability density functions via location-scale finite mixtures in Lebesgue spaces

Identifiability and optimal rates of convergence for parameters of multiple types in finite mixtures

A unified study for estimation of order restricted location/scale parameters under the generalized Pitman nearness criterion

Distributed Learning of Finite Gaussian Mixtures

Estimating finite mixtures of semi-Markov chains: an application to the segmentation of temporal sensory data

1 Introduction

2 Wasserstein Distance and the Minimum Distance Estimator

2.1 Wasserstein Distance

Definition 2.1 ( $p$ -Wasserstein distance).

2.2 Minimum Wasserstein Distance Estimator

2.3 Consistency of MWDE

Theorem 2.1.

Proof.

Theorem 2.2.

2.4 Numerical Solution to MWDE

2.5 Penalized Maximum Likelihood Estimator

3 Experiments

3.1 Performance Measure

3.2 Performance under Homogeneous Model

Minimum Wasserstein Distance Estimator under Finite Location-scale Mixtures

Related Research

Homogeneity testing under finite location-scale mixtures

The Sketched Wasserstein Distance for mixture distributions

Approximation of probability density functions via location-scale finite mixtures in Lebesgue spaces

Identifiability and optimal rates of convergence for parameters of multiple types in finite mixtures

A unified study for estimation of order restricted location/scale parameters under the generalized Pitman nearness criterion

Distributed Learning of Finite Gaussian Mixtures

Estimating finite mixtures of semi-Markov chains: an application to the segmentation of temporal sensory data

1 Introduction

2 Wasserstein Distance and the Minimum Distance Estimator

2.1 Wasserstein Distance

Definition 2.1 (p-Wasserstein distance).

2.2 Minimum Wasserstein Distance Estimator

2.3 Consistency of MWDE

Theorem 2.1.

Proof.

Theorem 2.2.

2.4 Numerical Solution to MWDE

2.5 Penalized Maximum Likelihood Estimator

3 Experiments

3.1 Performance Measure

3.2 Performance under Homogeneous Model

Definition 2.1 ( $p$ -Wasserstein distance).