Partial Knowledge of Probability Distributions 1.4 2003/06/11 2003/09/12 12:01:38.115 GMT-5 Don Johnson dhj@rice.edu Don Johnson dhj@rice.edu Eileen Krause erkrause@rice.edu Kyle Clarkson kclarks@rice.edu Elizabeth Gregory lizzardg@rice.edu Kevin Duh kevinduh@rice.edu Mariyah Poonawala mariyah@rice.edu Matthew Jeanes mjeanes@rice.edu Jeffrey Silverman jsilv@rice.edu In previous chapters, we assumed we knew the mathematical form of the probability distribution for the observations under each model; some of these distribution's parameters were not known and we developed decision rules to deal with this uncertainty. A more difficult problem occurs when the mathematical form is not known precisely. For example, the data may be approximately Gaussian, containing slight departures from the ideal. More radically, so little may be known about an accurate model for the data that we are only willing to assume that they are distributed symmetrically about some value. We develop model evaluation algorithms in this section that tackle both kinds of problems. However, be forewarned that solutions to such general models come at a price: the more specific a model can be that accurately describes a given problem, the better the performance. In other words, the more specific the model, the more the signal processing algorithms can be tailored to fit it with the obvious result that we enhance the performance. However, if our specific model is in error, our neatly tailored algorithms can lead us drastically astray. Thus, the best approach is to relax those aspects of the model which seem doubtful and to develop algorithms that will cope well with worst-case situations should they arise ("And they usually do," echoes every person experienced in the vagaries of data). These considerations lead us to consider nonparametric variations in the probability densities compatible with out assessment of model accuracy and to derive decision rules that minimize the impact of the worse-case situation.

Worst-Case Probability Distributions In model evaluation problems, there are "optimally" hard problems, those where the models are the most difficult to distinguish. The impossible problem is to distinguish models that are identical. In this situation, the conditional densities of the observed data are equal and the likelihood ratio is constant for all possible values of the observations. It is obvious that identical models are indistinguishable; this elaboration suggest that in terms of the likelihood ratio, hard problems are those in which the likelihood ratio is constant. Thus, "hard problems" are those in which the class of conditional probability densities has a constant ratio for wide ranges of observed data values. The most relevant model evaluation problem for us is the discrimination between two models that differ only in the means of statistically independent observations: the conditional densities of each observation are related as p r l ℳ 1 r l p r l ℳ 0 r l m . Densities that would make this model evaluation problem hard would satisfy the functional equation x m x m p x m C m p x where C m is quantity depending on the mean m, but not the variable x.The uniform density does not satisfy this equation as the domain of the function p · is assumed to be infinite. For the probability densities satisfying this equation, any value of the observed datum which has a value greater than m cannot be used to distinguish the two models. If one considers only those zero-mean densities p · which are symmetric about the origin, then by symmetry the likelihood ratio would also be constant for x 0 . Hypotheses having these densities could only be distinguished when the oberservations lay in the interval 0 m ; such model evaluation problems are hard! From the functional equation, we see that the quantity C m must be inversely proportional to p m (substitute x m into the equation). Incorporating this fact into our functional equation, we find that the only solution is the exponential function. z z 0 p z m C m p z ∝ p z z If we insist that the density satisfying the functional equation by symmetric, the solution is the so-called Laplacian (or double-exponential) density. p z z 1 2 σ 2 z σ 2 2 When this density serves as the underlying density for our hard model-testing problem, the likelihood ratio has the form (Huber; 1965, Huber; 1981, Poor pp.175-187) Λ r l m σ 2 2 r l 0 2 r l m σ 2 2 0 r l m m σ 2 2 m r l Indeed, the likelihood ratio is constant over much of the range of values of r l , implying that the two models are very similar over those ranges. This worst-case result will appear repeatedly as we embark on searching for the model evaluation rules that minimize the effect of modeling errors on performance.

P.J. Huber A robust version of the probability ratio test Ann. Math. Stat. 1965 36 1753-1758 P.J. Huber Robust Statistics John Wiley and Sons 1981 New York H.V. Poor An Introduction to Signal Detection and Estimation Springer-Verlag 1988 New York