Independence with Lower and Upper Probabilities

Lonnie ChrismanUAI 1996

TL;DR

This paper investigates the challenging problem of representing "independence" within the framework of lower and upper probabilities. The authors demonstrate that standard definitions of independence often fail to hold in this representation, leading to counterintuitive results, and propose an "ontological interpretation" as a way to reconcile these limitations.

What Problem Does This Paper Address?

In probability theory, two events A and B are considered independent if knowing the outcome of one event doesn't change our beliefs about the probability of the other. This is typically formalized by the factorization property: P(A ∩ B) = P(A)P(B). However, when we move beyond precise probabilities to "lower and upper probabilities" (which represent intervals of belief, often due to incomplete knowledge), capturing this notion of independence becomes surprisingly difficult. The paper highlights that standard Bayesian updating of these probability intervals can lead to situations where events that are supposed to be independent appear to influence each other, creating a paradox. The core problem is understanding when and how independence can be meaningfully represented in these less precise probability frameworks.

Methods

The paper explores two main approaches to defining independence for lower and upper probabilities:

  • Irrelevance: This is the idea that knowing event B doesn't change the probability bounds of event A, i.e., P(A|B) = P(A) and P(A|B) = P(A).
  • Factorization: This is the direct extension of the probabilistic definition, P(A ∩ B) = P(A)P(B) and P(A ∩ B) = P(A)P(B).

The authors analyze the compatibility of these definitions, particularly for 2-monotone lower probabilities (a common and useful class). They use mathematical theorems and examples, like a coin-tossing scenario, to illustrate when these independence properties fail. A key methodological contribution is the proposal of an "ontological interpretation" of lower probabilities, contrasting it with the more common "epistemological interpretation."

Datasets

Not discussed in the paper. The paper uses theoretical examples and mathematical derivations rather than empirical datasets.

Key Insights and Ideas

  • Independence is Tricky with Intervals: The core insight is that standard notions of independence break down when using lower and upper probabilities. Learning about one "independent" event can drastically alter the probability bounds of another, which seems paradoxical.
  • The Paradox of Dilation: When conditioning on an event that is supposed to be independent, the probability bounds can "dilate" (widen), further suggesting a loss of independence.
  • Interpretation Matters: The paper argues that the apparent paradoxes arise from interpreting probability bounds as representing epistemological indeterminacy (lack of knowledge). If, instead, bounds represent ontological indeterminacy (uncertainty due to structural approximations in a model), the results become natural and expected.
  • Ontological vs. Epistemological Indeterminacy: This is a crucial distinction. Epistemological indeterminacy is well-represented by precise probabilities. Ontological indeterminacy, arising from simplifying assumptions in a model (like assuming independence when it's not perfectly true), is what lower/upper probabilities are proposed to capture.
  • Abstraction as a Key Concept: The paper introduces the idea of a lower probability being an "abstraction" of a more detailed probability distribution, relative to independence assertions. This abstraction process introduces ontological indeterminacy.

Key Results

  • Theorems on Independence Failure: The paper presents theorems showing that for 2-monotone lower probabilities, independence properties (both irrelevance and factorization) can only hold in degenerate cases: when the probability is a point probability (no uncertainty) or completely vacuous (bounds are 0 and 1).
  • Coin Tossing Example: A detailed example of two independent coins demonstrates how conditioning on the outcome of one coin leads to counterintuitive changes in the probability bounds for the other, illustrating the failure of standard independence.
  • Ontological Interpretation Resolves Paradox: The proposed ontological interpretation explains why these counterintuitive results occur: the independence assumption itself is an approximation (ontological indeterminacy), and the resulting bounds reflect this.

Limitations and Future Work

  • Focus on 2-Monotone Probabilities: While the paper discusses limitations extending beyond 2-monotone functions, much of the formal analysis is centered on this class. Further work could explore independence for more general lower probabilities.
  • Generalizing Independence Concepts: The paper acknowledges that other concepts of independence exist for lower probabilities and convex sets of probabilities, suggesting avenues for future research.
  • Abstraction Beyond Single Assertions: The paper mentions the possibility of generalizing the concept of abstraction to sets of conditional independence assertions, which is noted as being beyond the scope of the current work.
  • Practical Applications: While the paper provides a theoretical framework, exploring how this ontological interpretation can be practically applied in areas like AI planning or decision-making would be valuable future work.

Summary updated

BibTeX
@inproceedings{chrisman1996independence,
  author = {Lonnie Chrisman},
  title = {{Independence with Lower and Upper Probabilities}},
  booktitle = {Conference on Uncertainty in Artificial Intelligence},
  year = {1996},
}
Figure