On Universality of Classical Probability with Contextually Labeled Random Variables: Response to A. Khrennikov
Ehtibar N. Dzhafarov, Maria Kon
TL;DR
This paper defends the universality of classical probability theory and argues against Khrennikov's criticisms, emphasizing that contexts are observable and that non-uniqueness of couplings is not problematic.
Contribution
It clarifies the applicability of classical probability to empirical phenomena and counters critiques of the Contextuality-by-Default theory, reaffirming its robustness.
Findings
Classical probability effectively describes empirical phenomena with observable contexts.
Non-uniqueness of couplings is not a flaw but a feature of modeling non-commutative systems.
Quantum formalisms are special cases with additional assumptions.
Abstract
In his constructive and well-informed commentary, Andrei Khrennikov acknowledges a privileged status of classical probability theory with respect to statistical analysis. He also sees advantages offered by the Contextuality-by-Default theory, notably, that it `demystifies quantum mechanics by highlighting the role of contextuality,' and that it can detect and measure contextuality in inconsistently connected systems. He argues, however, that classical probability theory may have difficulties in describing empirical phenomena if they are described entirely in terms of observable events. We disagree: contexts in which random variables are recorded are as observable as the variables' values. Khrennikov also argues that the Contextuality-by-Default theory suffers the problem of non-uniqueness of couplings. We disagree that this is a problem: couplings are all possible ways of imposing…
Peer Reviews
No public reviews on file for this paper yet. If you reviewed it on a platform where reviews are public (OpenReview, ICLR, NeurIPS, ICML), you can paste yours below so the community can read it here.
Videos
No videos yet. Explain this paper in a talk, walkthrough, or lecture? Add one.
Taxonomy
TopicsQuantum Mechanics and Applications · Philosophy and History of Science · Statistical Mechanics and Entropy
††thanks: Corresponding author: Ehtibar Dzhafarov, Purdue University, Department of Psychological Sciences, 703 Third Street West Lafayette, IN 47907, USA. email: [email protected].
On Universality of Classical Probability with Contextually Labeled
Random Variables: Response to A. Khrennikov
Ehtibar N. Dzhafarov
Purdue University
Maria Kon
Purdue University
Abstract
In his constructive and well-informed commentary, Andrei Khrennikov acknowledges a privileged status of classical probability theory with respect to statistical analysis. He also sees advantages offered by the Contextuality-by-Default theory, notably, that it “demystifies quantum mechanics by highlighting the role of contextuality,” and that it can detect and measure contextuality in inconsistently connected systems. He argues, however, that classical probability theory may have difficulties in describing empirical phenomena if they are described entirely in terms of observable events. We disagree: contexts in which random variables are recorded are as observable as the variables’ values. Khrennikov also argues that the Contextuality-by-Default theory suffers the problem of non-uniqueness of couplings. We disagree that this is a problem: couplings are all possible ways of imposing counterfactual joint distributions on random variables that de facto are not jointly distributed. The uniqueness of modeling experiments by means of quantum formalisms brought up by Khrennikov is achieved for the price of additional, substantive assumptions. This is consistent with our view of quantum theory as a special-purpose generator of classical probabilities. Khrennikov raises the issue of “mental signaling,” by which he means inconsistent connectedness in behavioral systems. Our position is that it is as inherent to behavioral systems as their stochasticity.
KEYWORDS: classical probability, contextuality, contextual labeling, quantum formalisms, random variables.
In our target paper (Dzhafarov & Kon, 2018) we argued that classical probability theory (CPT) is a universally applicable mathematical language, unfalsifiable by empirical phenomena or substantive theories. We considered several phenomena claimed to pose difficulties for CPT, and showed that these difficulties only arise if one misidentifies the random variables used to describe these phenomena. We pointed out a general way of preventing the possibility of such misidentifications: contextual labeling of random variables. We contrasted the universality of CPT with substantive, special-purpose theories aimed at deriving probability distributions for specific empirical situations. We view quantum formalisms as such a special-purpose theory (or set of models, if applied outside quantum mechanics). Khrennikov’s (2019) commentary further strengthens our view of CPT by mentioning that statistical analysis of data is “fundamentally based” on CPT. In the target paper we mention this too: “Any prediction derived in a special-purpose theory […] can be (and always is, eventually) fully expressed in the language of CPT, making it possible to relate the prediction to empirical data.”
1 Is CPT inadequate if one sticks to observable events?
However, Khrennikov still thinks that CPT has problems with certain empirical situations, such as the double-slit experiment in quantum physics. He asserts “the impossibility of defining a Kolmogorov probability space which is based solely on experimentally verifiable events.” Thus, he says, in the double-slit experiment CPT can describe the situation in each of the contexts defined by which slit is open and which is closed, but “it is impossible to represent observables with respect to these contexts in the same Kolmogorov probability space without introducing unobservable events.” Perhaps this comment is prompted by the fact that in the target paper (and, in greater detail, in Dzhafarov & Kujala, 2018) we presented a contextuality analysis of the double-slit experiment that used unobservable events, such as “particle hits the detector having passed through left open slit.” This, however, is an optional analysis that in no way replaces the basic CPT description of this empirical situation. In this basic description, one considers the random variable
[TABLE]
with both values being entirely observable, and their probabilities being empirically estimable. By the general principle we advocated in the target paper, this random variable should be labeled by its context, which is also entirely observable:
[TABLE]
We have therefore four distinct random variables (), stochastically unrelated to each other. Their distributions cannot be specified unless we make physical assumptions, and it is these physical assumptions that can agree or disagree with the experiment. Thus, one can make the physical assumption that
[TABLE]
i.e., the particle cannot hit the detector if both slits are closed. It is, as we know, empirically correct, but this is a physical prediction rather than one of CPT. One can make another physical assumption,
[TABLE]
and this one is known to be empirically false. This fact was interpreted by Feynman (1951) as a violation of the additivity law for probabilities, and Khrennikov seems to concur. We consider this a mistake, because CPT in no way constrains the probability distributions of in (4). The issue, once again, is one of correctly identifying random variables, and contextual notation (involving, we emphasize, only observable contexts) is a general way to prevent mistaken identification.
2 In CPT, contextual labeling cannot be circumvented
Khrennikov presents our Contextuality-by-Default (CbD) theory as just one possible approach within the framework of CPT. We cannot but agree, if one takes the CbD theory in its entirety. We think, however, that the departure point of CbD, that all random variables should be labeled not only by what they measure or respond to but also by the contexts in which they are recorded, is a logical necessity.
Khrennikov’s exposition of the CbD theory is preceded by an analysis that contravenes this requirement. He considers a set of random variables such that we have joint distributions for pairs (), but not for all four of them.111The reader having difficulties in reconciling the notation used in the present note with that in Khrennikov’s commentary (and with the notation used in the target article) may consult Appendix, where the correspondences are spelled out. Regrettably, Khrennikov did not keep his mathematical notation close to one in the target paper, forcing us here to adopt “bridging” notation, both resembling Khrennikov’s and acceptable in the CbD theory, even if not optimal. He calls these random variables “observables,” but it is clear from Footnote 2 of Khrennikov’s commentary that by an “observable” Khrennikov means a random variable in the normative CPT sense (as opposed to, say, Hermitian operators or POVMs in quantum theory, that are also called “observables”). However, if are classical random variables, the situation considered by Khrennikov (following many other authors) is logically impossible.
In CPT, random variables are jointly distributed if and only if they are defined on the same domain probability space. If and are defined on a probability space , and if and are defined on a probability space , then . Otherwise, we will have a mathematical impossibility of one and the same random variable (here, ) defined on two different domains. But then are all defined on the same probability space and are, therefore, jointly distributed. Applying this simple reasoning to , if any three of the pairs
[TABLE]
are jointly distributed, then all four random variables are jointly distributed.
Suppose, however, that Alice records a joint distribution of , Bob records a joint distribution of , and Charlie, who gets the results from both Alice and Bob, finds from some trustworthy theory that cannot be jointly distributed. What should Charlie conclude? The only way of dealing with this situation is to treat it as a reductio ad absurdum case: Charlie has arrived at a mathematical impossibility, whence there should be a mistake made in the initial assumptions. If there can be no doubt that the joint distributions of and exist, and if the “trustworthy theory” in question cannot be doubted either, the only remaining possibility is that it was a mistake to assume that in Alice’s and in Bob’s is one and the same random variable. Since and are measured separately (in different contexts), it is perfectly possible that we have in fact and with and being distinct random variables. This would be obvious to Charlie if and measured different things or responded to different questions (in the CbD terminology, if they had different contents). This would also be obvious to Charlie if the distributions of and were different. However, Charlie can be “fooled” into confusing with if they have the same content and the same distribution: in this case the contradiction arrived at by Charlie is the only guide in deciding that and are distinct random variables.
In the CbD theory, the possibility of erroneously assuming is precluded by “automatically” defining the identity of a random variable by both its content and its context. So we have for Alice, two random variables in the same context , distinguished from each other by their different contents ( and ); and we have for Bob, provided his has the same content () as Alice’s . Now there is no possibility of a mistake, as and are not the same random variable. Charlie can meaningfully ask, however, whether and can be coupled in a special way, in particular, if there is a coupling in which with probability
- The affirmative answer to this question would mean that the naive substitution of and for the rigorously defined and is relatively innocuous: it can be viewed as an informal simplification that, with some care, leads to no confusion. If, however, a coupling with does not exist, then the naive treatment is misleading and confusing.
Returning to the CHSH-related situation focused on by Khrennikov, in the CbD theory we “automatically” have to describe it by eight random variables grouped into four jointly distributed pairs,
[TABLE]
Consider the situation when this system is consistently connected, i.e., when the distributions of the content-sharing random variables (those denoted by the same letter and the same subscript, e.g., and ) are identical. Rather than ignoring the contexts (superscripts), which, as we know, may lead to a contradiction, in the CbD theory we consider this system’s couplings
[TABLE]
defined as sets of jointly distributed random variables that respect all the probabilities that characterize the correspondingly labeled random variables in (6). The question we ask in the CbD theory is whether among these couplings we can find ones in which, with probability 1,
[TABLE]
Such a coupling is called maximally connected, and if it exists, the system is noncontextual. Otherwise it is contextual.
3 Is non-uniqueness of couplings a problem?
Khrennikov’s account of the CbD theory, using the example (6) above, is fairly accurate, except for his assertion that the random variables in (6) are jointly distributed. In the CbD theory, the random variables belonging to different contexts are stochastically unrelated, i.e., they have no joint distribution. It is the couplings (7) subject to (8) that impose all possible joint distributions on the eight random variables in (6). For all practical purposes, however, Khrennikov’s imprecision here is relatively innocuous. Instead of speaking of four stochastically unrelated pairs of random variables, Khrennikov speaks of eight jointly distributed variables about which we only know joint distributions within same-context pairs. The overall joint distribution then should be treated as unknown (and unknowable, because different contexts are mutually exclusive). Instead of different couplings, then, one would speak of different possibilities for this unknown joint distribution.
The only problem with this language is that the existence of one “true” but unknown joint distribution is disconcerting. What could this distribution be? Khrennikov indeed finds fault with the CbD theory in this respect. “How can one select the ‘right coupling’?” he asks, and suggests a substantive interpretation for our inability to find one, in the spirit of the Copenhagen school of quantum mechanics: perhaps, he says, it is due to the fact that “real physical (or psychological) contextuality is determined not only by semantically defined observables, but also by apparatuses used for their measurement.”
We think that the answer is much simpler: there is no “right coupling” because couplings are not part of a system of random variables being coupled. Couplings are imposed on a system like (6) rather than found in it. They formalize the answer to the following counterfactual question: while in reality the different pairs in (6) are stochastically unrelated, what could their joint distribution be if they were jointly distributed? The set of all possible couplings of a system characterizes this system, in the same way as the set of all factors of an integer characterizes this integer. We do not think of the multitude of the factors as a deficiency of number theory or our ignorance of one “true” factor. One should also keep in mind that in CbD analysis we are interested not in all possible couplings but only in special, maximally connected ones. By definition, if a system is contextual no such couplings exist. So, the non-uniqueness problem here does not arise.
Khrennikov contrasts the non-uniqueness of couplings in the CbD theory with the uniqueness of quantum formalisms: “[Quantum probability] does not suffer the non-uniqueness problem. There is one fixed quantum state given by a normalized vector , or generally by a density operator ; and there is the unique representation of observables by Hermitian operators.” This is consistent with our view of quantum formalisms as special-purpose computations aimed at generating probability distributions, in the classical sense of the latter term. The uniqueness in question is achieved at the price of making additional assumptions about a system being studied, leading to empirically falsifiable predictions.
The following example, a modification of one given in a conference paper by Robert Griffiths (Griffiths, 2018), may serve as an illustration. Consider the system
[TABLE]
Irrespective of the row-wise distributions, this system is noncontextual, and it generally has an infinity of maximally connected couplings. Suppose now that that are generated by quantum measurements of in some fixed quantum state. In particular, invoking the no-disturbance principle, and are identically distributed. Then either the joint distribution of the four random variables with exists uniquely (if the Hermitian matrices generating and commute), or it does not exist (if they do not commute). We have no multitude of possible joint distributions here, but only because we have made additional assumptions about the origins of the random variables.
4 On “mental signaling”
Khrennikov correctly points out that the main advantage of the CbD conceptualization is its natural extendibility to inconsistently connected systems, i.e., those in which content-sharing random variables in different contexts may have different distributions. In relation to the system (6), the extended definition is that the system is noncontextual if it has a coupling (7) in which each of the equations in (8) holds with maximal possible probability (constrained by the marginal distributions of the random variables). This makes the CbD theory especially applicable to behavioral systems, most if not all of which are inconsistently connected. Khrennikov uses the term “mental signaling” to refer to this inconsistency, because “signaling” is the technical term used for inconsistent connectedness in some physical systems. He points out that inconsistently connected systems are also abundant in quantum physics, for a variety of reasons. (Khrennikov, in fact, was instrumental in bringing this issue to the attention of experimental physicists.) However, most physicists would agree that inconsistent connectedness can be greatly reduced or altogether eliminated by better experimental designs and greater control over experimental arrangements. In this relation Khrennikov poses a question of whether or not the same may be true for the ubiquitous inconsistent connectedness of behavioral systems. He writes: “In psychology the situation is more complicated. There are no theoretical reasons to expect no signaling. Therefore, it is not obvious whether signaling is a technicality or a fundamental feature of cognition. […] it may be that mental signaling is really a fundamental feature of cognition.”
We think the latter is definitely the case. Psychological experiments share with quantum physical experiments the fundamental presence of stochasticity in responses to stimuli (measurements of properties). Regardless of how strictly these stimuli (properties) are controlled, stochasticity in the responses (measurements) cannot be reduced by nontrivial amounts. In addition, however, psychological experiments exhibit the property of non-selectivity in responses to several stimuli: when the task is to respond to stimulus and to respond to stimulus , then almost always, if not always, influences the distribution of responses to , and vice versa (Cervantes & Dzhafarov, 2018; Dzhafarov, Zhang, & Kujala, 2015). In this respect behavioral systems differ from some quantum physical systems, those in which quantum theory includes “no-signaling” or “no-disturbance” constraints. These constraints, however, do not apply to all quantum physical systems (see, e.g., a CbD analysis of the Leggett-Garg-type systems in Bacciagaluppi, 2016, and a CbD analysis of the double-slit experiment in Dzhafarov & Kujala, 2018).
5 Conclusion
Khrennikov’s commentary is a constructive and even sympathetic analysis of our views on CPT with contextual labeling of random variables. The sympathy in the commentary can be easily understood: Khrennikov has himself advocated variants of contextual labeling of random variables, with some publications well preceding the CbD theory (Khrennikov, 2009a, 2009b, 2015a, 2015b). In this note we have discussed those aspects of Khrennikov’s commentary that allowed us to further elucidate our positions on CPT and the conceptual advantages of the CbD theory. Khrennikov defends some of the CPT inadequacy claims by bringing up the observability-of-events restriction. We have disagreed: contexts are observable too. Khrennikov uses noncontextual indexation of random variables and introduces contextuality as non-existence of a global distribution of a set of random variables whose overlapping subsets are jointly distributed. We have argued that this widespread description leads to contradictions that can only be avoided by labeling all random variables contextually. Khrennikov points out a problem of non-uniqueness faced by the CbD theory in finding couplings for systems of random variables. We have argued that this is not a problem because the various joint distributions imposed by couplings do not exist in the same sense as the random variables being coupled. Rather, the couplings are counterfactual, describing all possible ways in which variables that are de facto stochastically unrelated could be related to each other if they were jointly distributed. Finally, we have supported Khrennikov’s assumption that inconsistent connectedness in behavioral systems is inherent and irreducible.
REFERENCES
Bacciagaluppi, G. (2016). Einstein, Bohm and Leggett-Garg. In E. Dzhafarov, S. Jordan, R. Zhang & V. Cervantes (Eds.), Contextuality from Quantum Physics to Psychology (pp. 63-76). New Jersey: World Scientific.
Cervantes, V.H., & Dzhafarov, E.N. (2018). Snow Queen is evil and beautiful: Experimental evidence for probabilistic contextuality in human choices. Decision, 5, 193-204.
Dzhafarov, E.N., & Kon, M. (2018). On universality of classical probability with contextually labeled random variables. Journal of Mathematical Psychology, 85, 17-24
Dzhafarov, E.N., & Kujala, J.V. (2018). Contextuality analysis of the double slit experiment (with a glimpse into three slits). Entropy, 20, 278; doi:10.3390/e20040278.
Dzhafarov, E.N., Zhang, R., & Kujala, J.V. (2015). Is there contextuality in behavioral and social systems? Philosophical Transactions of the Royal Society A, 374: 20150099.
Feynman, R.P. (1951). The concept of probability in quantum mechanics. In J. Neyman (Ed.), Proceedings of the Second Berkeley Symposium on Mathematical Statistics and Probability (pp. 533-541). Berkeley: University of California Press.
Griffiths, R.B. (2018, November 9). Quantum measurements and contextuality. Paper presented at the Purdue Winer Memorial Lectures 2018: Probability and Contextuality. https://www.purdue.edu/hhs/psy/conferences/pwml/slides/PWML18_Griffiths.pdf.
Khrennikov, A. (2009a). Contextual Approach to Quantum Formalism. Dordrecht: Springer.
Khrennikov, A. (2009b). Bell’s inequality: Physics meets probability. Information Science, 179, 492-504.
Khrennikov, A. (2015a). CHSH inequality: Quantum probabilities as classical conditional probabilities. Foundations of Physics, 45, 711-725.
Khrennikov, A. (2015b). Two-slit experiment: Quantum and classical probabilities. Physica Scripta, 90, 1-9.
Khrennikov, A. (2019). Classical versus quantum probability: Comments on the paper “On universality of classical probability with contextually labeled random variables” by E. Dzhafarov and M. Kon.
Appendix: Notation
In the target paper we use the usual CbD notation for a random variable measuring, or responding to in context . Thus, , e.g., is a random variable with content in context , and this uniquely identifies it within a system of random variables being considered. In this note will be replaced with and , so that the content of a random variable is now identified by both its subscript and the letter or . This is done to bring the notation closer to Khrennikov’s use of and , when he refers to the CbD theory, to designate random variables recorded in context that he denotes by . Khrennikov’s own (noncontextual) notation for random variables is (not to be confused with the use of the same symbols when he refers to the CbD theory: there are not random variables). This may be confusing, but the choice was not ours. In this note, when we speak of noncontextual notation (considered incorrect in the CbD theory), we use , .
Khrennikov focuses on systems of random variables that in the CbD theory are called cyclic systems of rank 4. Their notation choices here are as shown:
[TABLE]
[TABLE]
[TABLE]