# «Abstract In this paper, we give a frequency interpretation of negative probability, as well as for extended probability, demonstrating that to a ...»

Interpretations of Negative Probabilities

M. Burgin

Department of Mathematics

University of California, Los Angeles

405 Hilgard Ave.

Los Angeles, CA 90095

Abstract

In this paper, we give a frequency interpretation of negative probability, as

well as for extended probability, demonstrating that to a great extent these

new types of probabilities, behave as conventional probabilities. Extended

probability comprises both conventional probability and negative probability.

The frequency interpretation of negative probabilities gives supportive evidence to the axiomatic system built in (Burgin, 2009) for extended probability as it is demonstrated in this paper that frequency probabilities satisfy all axioms of extended probability.

Keywords: probability; negative probability; extended probability; axiom; relative frequency; random experiment; random event

1. Introduction Probability theory is nowadays an important tool in physics and information theory, engineering and industry. A great discovery of twentieth century physics was the probabilistic nature of physical phenomena at microscopic scales, described in quantum mechanics and quantum field theory. At present there is a firm consensus among the physicists that probability theory is necessary to describe quantum phenomena.

At the same time, problems of physics brought physicists to the necessity to use not only classical probability but also negative probability. Negative probabilities emerged in physics in 1930s when Dirac (1930) and Heisenberg (1931) introduced probability distributions with negative values within the context of quantum theory. However, both physicists missed its significance and possibility to take negative values, using this distribution as an approximation to the full quantum description of a system such as the atom. Wigner (1932) came to the conclusion that quantum corrections often lead to negative probabilities while he was supplanting the wavefunction from Schrödinger's equation with a probability distribution in phase space. To do this, he introduced a function, which looked like a conventional probability distribution and has later been better known as the Wigner quasi-probability distribution because in contrast to conventional probability distributions, it took negative values, which could not be eliminated or made nonnegative. Dirac (1942) not only supported Wigner’s

**approach but also introduced the physical concept of negative energy. He wrote:**

“Negative energies and probabilities should not be considered as nonsense.

They are well-defined concepts mathematically, like a negative of money."

Richard Feynman in his keynote talk on Simulating Physics with Computers

**said:**

“The only difference between a probabilistic classical world and the equations of the quantum world is that somehow or other it appears as if the probabilities would have to go negative … “ Feyman (1987) studied negative probability and discussed differentexamples demonstrating how negative probabilities naturally exist in physics and beyond.

Based on mathematical ideas from classical statistics and modern ideas from information theory, Lowe (2004/2007) also argues that the use of non-positive probabilities is both inevitable and natural.

After this, negative probabilities a little by little have become a popular although questionable technique in physics. Many physicists have used negative probabilities to solve physical problems (cf., for example, (Dirac, 1930; 1942;

Heisenberg, 1931; Wigner, 1932; Sokolovski and Connor, 1991; Youssef, 1994;

1995; 2001; Scully, et al, 1994; Khrennikov, 1995; 1997; Han, et al, 1996;

Curtright and Zachos, 2001; Sokolovski, 2007; Bednorz and Belzig, 2009;

Hofmann, 2009)).

It is necessary to remark that are also used in machine learning Lowe, D.

(2004/2007) and mathematical finance (Haug, 2007).

Mathematical problems with negative probabilities were also studied. Bartlett (1945) worked out the mathematical and logical consistency of negative probabilities. However, he did not establish rigorous foundation for negative probability utilization. Khrennikov (2009) developed mathematical theory of negative probabilities in the framework of p-adic analysis. This is adequate not for the conventional physics in which the majority of physicists work but only for the so-called p-adic physics.

The mathematical grounding for the negative probability in the real number domain was developed by Burgin (2009) who constructed a Kolmogorov type axiom system building a mathematical theory of extended probability as a probability function, which is defined for random events and can take both positive and negative real values. As a result, extended probabilities include negative probabilities. It was also demonstrated that the classical probability is a positive section (fragment) of extended probability.

At the same time, there are many problems with interpretations of the conventional concept of probability. The goal of this paper is to build the frequency interpretation of extended probability because the frequency interpretation reflects the most popular approach to probability treatment in science, in particular, in physics, and in other applications of probability. In Section 2, going after Introduction, traditional approaches to interpretations of probability are discussed. In Section 3, the frequency interpretation of extended probability is constructed. It is called the frequency interpretation probability. In Section 4, it is proved that the frequency interpretation probability satisfies all axioms from (Burgin, 2009).

** 2. Traditional approaches to interpretations of probability**

An interpretation of the concept of probability is a choice of some class of events (or statements) and an assignment of some meaning to probability claims about those events (or statements). Usually researchers concentrate on three main interpretations of the probability: the frequency interpretation, the belief interpretation, and the support interpretation. There are also other interpretations, such as the logical interpretation or the propensity interpretation.

**All these interpretations belong to three groups:**

- Objective probability is defined (exists) as a numerical property of sequences of frequencies associated with an event in natural, social or technical phenomena.

- Subjective probability is defined (exists) as a belief or measure of confidence in an outcome of a certain event.

- Combined probability is defined (exists) as a belief supported by both observational and experimental evidence or measure of confidence in an outcome of some event.

** Example 2.1.**

When there are 10 blue balls, 5 green balls and 5 red balls in a box, and these balls are well mixed, then the objective probability to draw a blue ball from this box is ½.

** Example 2.2.**

When we believe that the ratio of blue balls to all balls in a box is ½, then the subjective probability of drawing a blue ball from this box is ½.

** Example 2.3.**

When we conjecture that the ratio of blue balls to all balls in a box is ½, we tested this hypothesis and got some experimental evidence in support of it, then the combined probability of drawing a blue ball from this box is ½.

Each of these three groups can be divided into several subclasses.

It is possible to interpret objective probability in three main ways: as the actual finite relative frequency (the actual frequency interpretation), in a form of a limit (or limiting behavior) of hypothetical infinite relative frequencies (the potential frequency interpretation) or as a propensity (the propensity interpretation).

It is possible to interpret subjective probability in three main ways: as an actual belief, or system of actual beliefs, (the belief interpretation), as an idealized belief, system of idealized beliefs, based on the Bayes theorem (the Bayesian or personalist interpretation) or in a form of a logical system (the logical interpretation).

It is possible to interpret combined probability in three main ways: as an actual belief, or system of actual beliefs, supported by observational or experimental evidence (the supported belief interpretation), as an idealized belief, system of idealized beliefs, based on the Bayes theorem and supported by observational or experimental evidence (the supported Bayesian interpretation) or in a form of a logical system supported by observational or experimental evidence (the supported logical interpretation).

It is important to distinguish two types of probabilities: ensemble probabilities and sequential (or cumulative) probabilities. A special case of sequential (or cumulative) probabilities is temporal probability when the sequence of events is consequential, i.e., events happen one after another.

In the actual frequency interpretation, the probability p(A) of an event A is taken equal to the long-run relative frequency with which A occurs in identical repeats of an experiment or observation.

In the potential frequency interpretation, the probability p(A) of an event A is taken equal to the limit of relative frequency with which A occurs in identical repeats of an experiment or observation.

A long-run propensity theory is one in which propensities are associated with repeatable conditions, and are regarded as propensities to produce in a long series of repetitions of these conditions. In this theory, frequencies are approximately equal to the probabilities. A single-case propensity theory is one in which propensities are regarded as propensities to produce a particular result on a specific occasion.

In the Bayesian interpretation, the probability p(A/B) of an event (proposition/hypothesis) A, given (conditional on) the happening (truth of) the event (proposition/hypothesis) B is a measure of the plausibility of the event (proposition/hypothesis) A, given (conditional on) the happening (truth of) the event (proposition/hypothesis) B. Bayesian inference uses probability distribution as an encoding of our uncertainty about some model parameter or set of competing theories, based on our current state of information.

The supported Bayesian approach in the style of de Finetti (1937) recognizes

**no rational constraints on subjective probabilities beyond:**

1. conformity to the probability calculus (coherence);

2. a rule for updating probabilities in the face of new evidence (conditioning).

Conditioning means that an agent with probability function P1, who becomes certain of a piece of evidence E, should shift to a new probability function P2

**related to P1 by:**

(Conditioning) P2(X) = P1(X | E) (provided P1(E) 0).

Frequency approach was mathematically grounded and further developed in algorithmic information theory (Kolmogorov, 1965; Martin-Löf, 1970).

Considering different applications of probability theory, we come to the fundamental question: what interpretation of probability is appropriate for scientific practice? One of the most outstanding philosophers of science, Rudolf Carnap suggested that both the objective (frequency) and subjective (for Carnap (1950), logical) interpretations are needed for representing different aspects of probability usage.

However, reality shows that there is no easy compromise. Indeed, most scientists (and some philosophers) support the frequency approach and do not trust the Bayesian approach oriented to subjective evaluations. At the same time, many philosophers (and some scientists) are subjectivists, supporting Bayesian approach and many of them do not believe in objective probabilities.

All approaches have their caveats. The frequency interpretation contains the term ‘‘identical repeats.’’ Of course the repeated experiments can never be identical in all respects. The Bayesian definition of probability involves the rather vague sounding term ‘‘plausibility,’’ which must be given a precise meaning for the theory to provide quantitative results.

Problems with probability interpretations and necessity to have sound mathematical foundations brought forth an axiomatic approach in probability theory. Based on ideas of Fréchet and following the axiomatic mainstream in mathematics, Kolmogorov developed his famous axiomatic exposition of probability theory (1933).

Here we consider only the frequency interpretation as the most popular approach to probability in physics because negative probability comes from physics and is more and more use by physicists.

3. Frequency interpretations of extended probability

Taking an event { ui } with ui ∈, we denote by N+ the number times that events with the same sign as the event ui occur during a sequence of N trials, by N- the number times that events with the opposite to the event ui sign occur during a sequence of N trials, by ni the number of times that the event ui occurs during a sequence of N trials, and by mi the number of times that the event -ui occurs during the same sequence of trials. Let vN(ui ) = (ni )/ N+ - (mi )/ N-. Then we define the extended frequency probability of the event ui as

Consequently, we have p(wi ) = - p(-wi ) In a general case of random events, we have a random event A = {wi1, wi2, wi3, …, wik, -wj1, -wj2, -wj3, …, -wjt }, the number N+ of times that positive events occur during a sequence of N trials, the number N- of times that negative events occur during a sequence of N trials, the number npos of times that positive events wir from A occur during a sequence of N trials and the number nneg of times that negative events -wiq occur during the same sequence of trials.

Then we define vN(A ) = (ni )/ N+ - (mi )/ N- and the extended frequency probability of the random event A as

We assume that this limit exists for random events, i.e., for all events from the set F.