I Introduction
Finite mixture models are a primary workhorse of machine learning and pattern recognition and are frequently used in a wide range of applications (e.g.
[14, 17, 7]). By far the most common approach to estimating the parameters of a mixture model is the ExpectationMaximization (EM) algorithm that iteratively improves the likelihood of the data at every iteration
[5, 3]. The widespread use of EM for solving such problems is despite its wellknown dependence on initial conditions [2]. In fact the common practice is to rerun the algorithm multiple times and choose the local maximum with highest likelihood [14]. Figure 1 shows a simple synthetic dataset similar to a problem set at Brown University (http://cs.brown.edu/courses/cs1955). Depending on the initial conditions, EM may converge to the correct parameters (figure 1c) or very different parameters (figure 1b).In order to better understand the behavior of EM, a number of recent papers have analyzed the case of EM for Gaussian mixture models when the Gaussians are
well separated and in high dimensions. For such cases it can be shown that EM will converge rapidly to the correct parameters [4, 1]. An alternative to EM are momentbased methods which go back to the days of Pearson
[16] and seek parameters whose predicted moments exactly match the empirical moments. Often the number of moments needed to be matched grow rapidly with the size of the mixture and lead to nonlinear equations. Recent spectral methods [9] bypass many of the problems of traditional moment methods but they require specific forms for the covariance. The data in figure 1 contains three elliptical Gaussians in two dimensions and the densities of the Gaussians overlap significantly. Can we globally maximize the likelihood for such data?A trivial solution to maximizing the likelihood of a Gaussian mixture model is to set the mean of one of the Gaussians on a single datapoint and shrink the covariance to zero. This solution can increase the likelihood of the data arbitrarily [4, 14]. Thus the problem of finding the global maximum only makes sense if we restrict the set of possible models. In this paper, we adopt the restriction that each of the models must belong to a large, discrete set of possible parameters. Can we find the maximum likelihood solution within this constrained set of parameters? This strategy of restricting the search to a large, discrete set of possible parameters has been successful in the exemplar approach to clustering. Two prominent examples are kmedoids [10] and affinity propagation [6]
which show that by using discrete optimization one can often obtain better clustering solutions compared to kmeans. Lashkari and Golland
[12] presented a convex formulation of exemplar based clustering using a variant of EM. Nowozin and Gokhan extended this approach by allowing the algorithm to suggest additional exemplars beyond the training data [15].Even with the restriction that the model parameters belong to a discrete set the problem is seemingly combinatorial: we still need to find the “active” parameters out of the possibilities and the mixing weights on each of the models. In this paper, we show that a simple upper bound on the likelihood can be computed using convex optimization and that this bound is guaranteed to be tight in the generative setting as the number of data points goes to infinity. This bound can then be used to assess the quality of a solution achieved by any mixture estimation algorithm, provided it satisfies the constraints. Thus for any dataset our method allows us to find a finite mixture model together with a datasetspecific bound on how far the likelihood of this mixture is from the global optimum of the likelihood.
Ii Problem Formulation
In a finite mixture model we consider models of the form:
(1) 
Given a dataset
we seek to find a vector of mixing weights of size
and parameter vectors to maximize the likelihood of the data. In the constrained maximum likelihood problem we are also given a set of possible parameter vectors and the problem becomes:(2) 
Iii An Upper Bound
We can rewrite the constrained ML problem (equation 2) as follows:
(3) 
Where we now represent the solution as a vector of length with only nonzero elements. This is obviously equivalent to the original MLE problem. By relaxing the cardinality constraint we get an immediate upper bound:
(4) 
Observation 1: The calculation of the upper bound (equation 4) is equivalent to solving a convex optimization problem.
Proof: This result is well known(e.g. [8, 12].) and we sketch the proof for completeness. We can rewrite the problem as:
(5) 
where is a fixed matrix. Direct calculation shows that this problem is concave.
Observation 2: The calculation of the upper bound (equation 4) can be performed using a variant of the EM algorithm. For this variant, the solution does not depend on initialization.
Proof: This result is also well known(e.g. [8, 12].) and we sketch the proof for completeness. It is known that at every iteration of EM the likelihood increases or stays the same, and that if EM converges it is a local maximum of the likelihood. Since for this problem the log likelihood is concave any local maximum is also globally optimal.
The form of EM for this restricted setting is:
(6)  
(7) 
Helmbold et al. [8] show how to speed up the convergence of these iterations by using overrelaxation: with . We refer to this algorithm (iterating equations 67) as the “convex EM algorithm”.
To summarize, by removing the requirement that the mixture be a finite mixture, or equivalently that the vector has only nonzero components, we obtain a concave problem that can be solved using EM regardless of initial condition. But why should this bound be tight? After all, the space of
sparse probability vectors of length
is a vanishing fraction of the space of all probability vectors of length . It would therefore seem that by allowing a much larger search space we will obtain higher log likelihood and thus the bound will never be tight. In the next section, we show that this intuition is wrong and in fact the bound will be tight under reasonable conditions.Iiia Tightness of Upper Bound
We start with the generative setting. We assume we are given a dataset which are IID samples from a finite mixture density . We also assume that we are given a class of possible components and that all components of the generating distribution are members of the discrete set . We now show that under these assumptions the bound is guaranteed to be tight as the number of points goes to infinity and give a characterization of the looseness of the bound for finite .
We denote by
Definition: The normalized log likelihood that a weight vector gives to a dataset is denoted by
Definition: For a dataset and a fixed set of models we denote the upper bound computed by maximizing equation 4 by:
(8) 
The main idea of the proof is to view the normalized log likelihood as an estimate of the negative crossentropy between and the distribution . Clearly as the size of the dataset goes to infinity converges to the negative cross entropy . (Recall that for any function , converges to ). To state the tightness of the upper bound we will need to characterize the worstcase difference between the log likelihood and the negative cross entropy between and .
Definition: The cross entropy estimation error is defined as:
(9) 
Standard results (e.g. [11]) show that for any will approach zero as goes to infinity (since is a sample average and the negative cross entropy is its expected value, the expected squared distance between them decreases as ).
We can now state our main result.
Theorem 1: In the generative setting, let be a sparse vector that represents the generative distribution then
(10) 
In particular as goes to infinity the bound is tight and .
Proof:
(13)  
(14) 
where we have used the fact that the cross entropy between any two distributions is minimal when . Now, by the definition of the entropy estimation error we also know that . Substituting into the right hand side of equation 14 shows that .
Figure 2 illustrates the result in theorem 1 for the data shown in Figure 1 where the class of models contains more than a million different models. Even though the upper bound is optimized over all possible probability vectors without any explicit sparsity penalty, the bound becomes tight as the amount of data grows and already at points per cluster the true model achieves about 98% of the upper bound.
What happens when the data is not generated by a finite mixture whose components are in the discrete set ? It can be shown that the tightness of the bound degrades gracefully: as long as there exists a finite mixture which is close to the generating distribution (i.e. has KL divergence at most from the generating distribution) then as the number of points goes to infinity the bound will be at most away from the log likelihood of the close finite mixture. We formalize this in the following theorem.
Theorem 2: let be a sparse vector such that is closest to the generating distributing and then:
(15) 
In particular as goes to infinity the gap between the bound and what can be achieved by a finite mixture in the discrete set approaches .
Proof: This follows by substituting the fact that the KL between and is at most into equation 14 and again using the definition of .
Figure 2df illustrates this theorem. Here the data is generated by a mixture of three rectangles while the class of models is the same class of a million Gaussians. Even though there is no way to represent the generating distribution as a mixture of three Gaussians, the upper bound that is computed using convex optimization becomes tighter as the number of points increases and already at there exist a mixture of three Gaussians that attains more than 98% of the upper bound.
We note some properties of theorems 1 and 2.

Theorem 1 shows that for any fixed set of models the true generating distribution optimizes the constrained likelihood problem as the number of points goes to infinity. This should be contrasted with unconstrained maximum likelihood estimation for Gaussian mixture models where by shrinking the variances arbitrarily we can achieve higher likelihood than the true parameters.

The asymptotic tightness of the bound calculated using convex optimization does not depend on how “well separated” the mixture components are.
Iv Obtaining a Lower Bound
Theorems 1 and 2 show that we can recover an asymptotically tight upper bound on the likelihood of the best sparse weight vector using convex optimization, i.e. by running the convex EM algorithm. This bound can then be used to assess the quality of any algorithm that estimates a mixture model from data, provided the algorithm’s output satisfies the constraint that all models come from the large, discrete set of models.
The simplest algorithm for estimating a finite mixture that satisfies the constraints is projected EM. We simply run the standard EM algorithm and then “project” the result into the constraint set: replace each model from the mixture found by EM with the closest model in the discrete set. As shown by [2], if EM is initialized close to the correct solution it will converge to that solution and so we use the standard method of using multiple random restarts and choose the EM solution that gives highest likelihood after projection.
If the log likelihood found by projected EM (or any other algorithm) equals the upper bound then we have provably found the MLE. In practice, we need to define a numerical threshold that measures how far the lower bound is from the upper bound. We do this by defining the optimality ratio. this is simply the ratio between the log likelihood of a discrete solution and the upper bound. Since the ratio can be affected by adding a constant to the log likelihood, we “calibrate” the log likelihood by subtracting from it the average log likelihood of a random mixture model for this data.
where is the average value of a random vector with cardinality . Thus an optimality ratio of zero means that the proposed solution is no better than a random vector and an optimality ratio of one means that the proposed solution is optimal.
V Experiments
We first performed experiments with synthetic two dimensional data. In low dimensions, one can perform exhaustive enumeration
over mixture parameters. A two dimensional Gaussian is defined by 5 numbers (two coordinates of the mean, two eigenvalues of the covariance, and one rotation for the covariance). We used 3,000 possible means,
possible eigenvalues and possible angles to create a set of one million possible Gaussians. We generated data from finite mixtures whose parameters where chosen randomly (and hence not part of the discrete set) and ran convex EM to obtain an upper bound as well as projected EM to obtain lower bounds. Figure 3a shows an example of a generated data and figure 3d shows a subset of the discrete set of models.Figure 3bc show two solutions found by EM for this dataset. Both solutions give high likelihood to the data but our convex upper bound allows us to say that one is subpotimal (achieves only 90% optimality ratio) while the other is closer to optimal (96% optimality ratio). Without our bound, there would be no easy way to tell how suboptimal the solution in figure 3b is. Figure 3e shows the optimality ratio between the log likelihood found by EM and our computed upper bound on the optimal solution. We plot these ratios as a function of the separation of the mixture Gaussians [4]. A 2separated mixture corresponds roughly to almost completely separated Gaussians, whereas a mixture that is or separated contains Gaussians which overlap significantly [4]. As can be seen, in many cases the optimality ratio is close to 100%. Thus our convex upper bound calculation allows us to prove that we have found the global optimum of the log likelihood for these problems. Figure 3f shows the improvement in optimality ratio obtained when we run EM with 1000 random restarts as compared to a single run. This improvement is plotted as a function of the optimality ratio calculated after a single run of EM. As can be expected, when the optimality ratio is low, there is room for significant improvement by running EM with more restarts. Again, in the absence of a bound on the log likelihood it would be impossible to tell whether additional restarts will improve the likelohood or not.
A popular use of mixture models in computer vision is in
image segmentation where each segment is assumed to be modeled by a five dimensional Gaussian — x,y and RGB (e.g. [7]). Since this is a five dimensional problem exhaustive search for creating the set of models is infeasible. Instead we generate a discrete set of a quarter of a million possible models by taking rectangular image patches of different sizes and robustly fitting a Gaussian to the pixels within that patch. Image segmentation is then performed by fitting a Gaussian mixture model with 5 components to the pixel data.We ran projected EM on the 100 images in the Berkeley Segmentaton Dataset [13]. Figure 4a shows a similar plot to what we saw in the synthetic data. When the optimality ratio is far from one, there is potential for significant improvement in the log likelihood values by rerunning EM. Figures 4bd show a particular example. After 10 restarts the optimality ratio is 85% and so rerunning EM with multiple restarts improves the log likelihood significantly (and also leads to more natural looking segmentation). Figures 4ch show two examples where the bound allows us to assess how far we are from optimality. Even though this data was not generated by a mixture of Gaussians, our convex upper bound allows us to give an optimality certificate for some of the images and also gives an indication regarding whether it is worthwhile to run EM with more restarts.
Vi Discussion
Finite mixture models are widely used in machine learning and in additional application areas. The standard method for estimating the parameters of mixture models from data is the EM algorithm which is known to sometimes find local optima. In this paper we have discussed a method that can globally optimize the constrained maximum likelihood problem: when each model’s parameters come from a discrete set of possible parameters (the models can have different forms, e.g. some could be Gaussians and others of uniform distribution). We have shown that an upper bound can be found using a simple convex variant of EM and gave conditions under which this bound is guaranteed to be tight. This bound can be used to assess the performance of any estimation algorithm including projected EM.
The strongest assumption made by our algorithm is that we can generate a finite discrete set of models so that a mixture that only used models inside the discrete set can approximate the true distribution. In low dimensions this discrete set can be generated by exhaustively covering a range of possible parameter values but in high dimensions this is obviously intractable. Nevertheless our experiments in image segmentation have shown that it is often possible to generate this class of models by estimating models based on subsets of the data.
Whenever optimization methods are applied to realworld problems and the methods are not perfect, the question arises: are the failures due to the optimization method used or to the cost function being optimized? For discrete optimization methods (e.g. MAP in graphical models), the answer to this question can often be obtained using the linear programming relaxation (e.g.
[18, 19]) which gives a problemdependent bound on the optimal value of the optimization problem. To the best of our knowledge, our paper is the first to give an analogous bound for the problem of maximizing the likelihood of a finite mixture model.Acknowledgment
This work has been supported by the the ISF. The authors wish to thank the anonymous reviewers for their helpful comments.
References

[1]
Sanjeev Arora and Ravi Kannan.
Learning mixtures of arbitrary gaussians.
In
Proceedings of the thirtythird annual ACM symposium on Theory of computing
, pages 247–257. ACM, 2001.  [2] Sivaraman Balakrishnan, Martin J Wainwright, and Bin Yu. Statistical guarantees that algorithm: From population to samplebased analysis. arXiv preprint arXiv:1408.2156, 2014.
 [3] Christopher M Bishop et al. Pattern recognition and machine learning. 2006.
 [4] Sanjoy Dasgupta. Learning mixtures of gaussians. In Foundations of Computer Science, 1999. 40th Annual Symposium on, pages 634–644. IEEE, 1999.
 [5] Arthur P Dempster, Nan M Laird, et al. Maximum likelihood from incomplete data via the em algorithm. Journal of the Royal statistical Society, 39(1):1–38, 1977.
 [6] Brendan J Frey and Delbert Dueck. Clustering by passing messages between data points. science, 315(5814):972–976, 2007.
 [7] Hayit Greenspan, Jacob Goldberger, and Lenny Ridel. A continuous probabilistic framework for image matching. Computer Vision and Image Understanding, 84(3):384–406, 2001.
 [8] David P. Helmbold, Yoram Singer, Robert E. Schapire, and Manfred K. Warmuth. A comparison of new and old algorithms for a mixture estimation problem. In COLT, 1995.
 [9] Daniel Hsu and Sham M Kakade. Learning mixtures of spherical gaussians: moment methods and spectral decompositions. In ITCS ’13.
 [10] Leonard Kaufman and Peter Rousseeuw. Clustering by means of medoids. NorthHolland, 1987.
 [11] S.M. Kay. Fundamentals of Statistical Signal Processing: Estimation theory. Number v. 1 in Fundamentals of Statistical Signal Processing. PrenticeHall PTR, 1998.
 [12] Danial Lashkari and Polina Golland. Convex clustering with exemplarbased models. In NIPS ’07.
 [13] D. Martin, C. Fowlkes, D. Tal, and J. Malik. A database of human segmented natural images and its application to evaluating segmentation algorithms and measuring ecological statistics. In ICCV, 2001.
 [14] Geoffrey McLachlan and David Peel. Finite mixture models. John Wiley & Sons, 2004.

[15]
Sebastian Nowozin and Gökhan Bakir.
A decoupled approach to exemplarbased unsupervised learning.
In ICML ’08.  [16] Karl Pearson. Contributions to the mathematical theory of evolution. Philosophical Transactions of the Royal Society of London. A, pages 71–110, 1894.
 [17] Carsten Rother, Vladimir Kolmogorov, and Andrew Blake. Grabcut: Interactive foreground extraction using iterated graph cuts. In ACM TOG ’04.

[18]
Alexander M. Rush, David Sontag, Michael Collins, and Tommi S. Jaakkola.
On dual decomposition and linear programming relaxations for natural language processing.
In Proceedings of the 2010 Conference on Empirical Methods in Natural Language Processing, EMNLP 2010, pages 1–11, 2010.  [19] Martin J. Wainwright and Michael I. Jordan. Graphical models, exponential families, and variational inference. Foundations and Trends in Machine Learning, 1(12):1–305, 2008.
Comments
There are no comments yet.