Interactive Statistics: The Project

Aims and Hypotheses

As noticed in the abstract, an interactive domain is each research domain in which the observed system is intrinsically influenced by the observation itself. Important examples of these domains are quantum mechanics3, the study of psychological decision processes 4 and the analysis of ultra-weak signals 5, for instance in tomography and remote-sensing. As shown in 4 ,these research domains cannot be investigated by means of classical statistical procedures because these methods do not model an interaction between the observer and the system under investigation. Thus, the aim of this project is the construction of a statistical method that can handle this kind of situation and a corresponding analysis to compare the practical behaviour of interactive systems. As a working hypothesis, that can be immediately put in a mathematical procedure, we consider the construction of a more general statistics by supposing that we have an incomplete knowledge of the way we collect the information, or, otherwise, of the interaction between the observer and the observed. This is in sharp contrast with the usual statistical hypothesis of an incomplete knowledge of the system we consider.

Material and Methods

The above hypothesis has been studied in the two-dimensional case (that is, questions with only two possible answers) with very interesting results4. The only knowledge we assume to posses is the magnitude of the influence of the observation on the observed system. We have called this magnitude of the interaction epsilon and the accompanying model the epsilon-model. Epsilon can vary between 0 and 1: in the case of epsilon equals one we have a maximal interaction; for epsilon equal to 0 there is no interaction. It seems now that the case epsilon=1 is isomorphic to a quantum mechanical experiment on a two-dimensional Hilbert space and that the case epsilon=0 agrees with classical (or kolmogorovian) statistics. For the other values of epsilon we found a new kind of statistics not isomorphic with either of the two former cases. In other words, this is the first statistics mentioned in literature that describes the continuous transition of a microscopical (quantum mechanics)

to a macroscopical (classical mechanics) situation. However, it follows from our hypothesis that this statistics in the case of epsilon=1 is not only appropriate for the microscopical world, but in the region of social sciences too. To check this idea one has to extend the exising model to a more dimensional model. Mathematically, this means that the computations for the two-dimensional case should be extended to the more dimensional model. The idea is sufficiently precise to be implemented in a relative short time (9 to 15 months) on a personal computer. Because explicit expressions are easier to manipulate, we want to investigate which results can be analytically solved. The results we obtain will be tested in three ways:

1) Concrete experimental situations.

2) The two limit cases (of maximal and minimal interaction) should agree with a quantum mechanical description and a classical description respectively.

3) Results should be consistent with the results for the two-dimensional model.

Applications

1. Quantum mechanics

The above hypothesis allows us to reinterprete the origin of probabilities (ontological or epistemological) in quantum mechanics. Although many researchers have the feeling that they have an epistemological origin, it is generally accepted they are of an ontological nature. The reason for this discord was that it seemed impossible to construct a statistical theory consistent with quantum mechanics from the hypothesis of a lack of knowledge on the observed entities. A group of six researchers (Dirk Aerts, Thomas Durt, Bob Coecke, Frank Valckenborgh, Sven Aerts, Bart D'Hooghe) investigates the possibility to find back the quantum mechanical probabilities by replacing the hypothesis just above by the aforementioned working hypothesis. If the results of this research continues as to-day, the philosophical background of quantum theory should be reconsidered.

2. Psychological Decision Procedures

In psychological research one regularly works with questionnaires where the participants have to choose between several options. If the choice of the test person is not contained within the different options (maybe he thinks the question to be badly posed) he'll have to make a choice at the very moment itself. It is clear that his choice will have a contextual component: who asks this question, why, in what circumstances, ... The fluctuations that enter the measurement in this way are due to the measurement process (and not of the people that are questioned) and as such they have their moment of action during the measurement. In other words: one and the same person can can give a different answer to same question merely because the context is different. It is exactly this kind of fluctuations we wish to model because it are these fluctuations that give rise to a statistical structure that is much closer to quantum statistics than to classical statistics. In a further phase of the project we want to apply this approach to complex problems that involve correlations between different measurements, because it is known from quantum theory that it is in these situations that one encounters the most profound differences between the two types of statistics. More specifically, we wish to consider the applicability of this approach in the item-response theory and multi-dimensional scaling models[6] . It will be clear that if the introduction of quantum statistics in the field of psychology is succesful then this would constitute a promising new methodology.

3. Analysis of ultra-weak signals

In a review article by Malley and Hornstein [7] it is shown that the use of quantum statistical methods in the area of weak-signal analysis leads to a risc-curve that is lower everywhere than the usual Bayes-inference risc-curve.

The method they use (named quantum Statistical Inference) was developed by engineers who wanted to use the advantages of LASER in communicationsystems. The fluctuations in the workings of the LASER (whoes basic principles are dictated by quantum rules) demanded of a new statistical analysing method, capable of depicting transmission efficiency. The same methods are now applied to as diverse fields as tomography, optical-imaging, remote-sensing,... The use of these new methods has lead to an improvement of the signal-noise ratio by as much as 20dB. Because these are all areas where one formerly used to apply classical statistics, our conjecture is that these areas do not posses a maximal fluctuations structure for the measurement, that is, we believe that the more appropriate statistical model for these situations is neither a quantum mechanical nor a classical one but an intermediate one. If this is the case then it must be possible to improve the signal-noise ratio even further by using the probabilities of the epsilon-model.

4. Principia Cybernetica

This project also has the following two connections with the international project Principia Cybernetica ( VUB representants are F. Heylighen and J. Bollen).

1) From cybernetics it is well-known that a subsystem never has complete knowledge of the system it belongs to. This is because there is a feedback (interaction) between the two systems.

2) The evolution of a knowledge system is not only modelled in Principia Cybernetica: an adaptive semantical network has been implemented on the Internet. The evolution of this semantical network is being examined. This should be accompanied by an analysis. An important question that arises is the following: to what extent is the self-organising process that is now going on on the Internet influenced by the initial choices we gave the users of the Internet? This question can only be answered in a framework that takes into account the interaction.

As such, this project is a test for the development of the semantical network as much as it is an experimental test for the statistical techniques here proposed.

Timing

In the first year of the project the two-dimensional model will be elaborated to the three-dimensional case. This has the advantage that one doesn't need yet to overcome the full complexity of the n-dimensional case while it can be shown that the three-dimensional case already displays all the particularities of the n-dimensional one. The second project year will be dedicated to the analysis of the data from experimental set-ups where interactivity plays an important role.

Computer analyses will be made of this data taking into account the working hypothesis. A start for the n-dimensional case will be made. Parameters that are evidently important in the experiments will play a role in the development of the n-dimensional case. In the third year of the project we will consider how the statistical techniques can be adapted to the more dimensional epsilon-model and to what extent this constitutes an improvement to the techniques that are being used now. The last year serves as an evaluation year where a case-study shall be made.