Causal completeness of probability theories – results and open problems Miklós Rédei Department of Philosophy, Logic and Scientific Method London School of Economics and Political Science Houghton Street London WC2A 2AE, UK e-mail: [email protected] Balázs Gyenis Department of History and Philosophy of Science University of Pittsburgh 1017 Cathedral of Learning Pittsburgh, PA 15260, U.S.A. e-mail: [email protected] Forthcoming in “Causality in the Sciences”, ed. by F. Russo and J. Williamson 1 The Common Cause Principle and causal completeness informally The aim of this paper is to give a review of the known results and open problems concerning causal completeness (closedness) of classical and non-classical (quantum) probability spaces. Causal closedness of a probabilistic theory means that the theory is causally rich enough to be able to explain causally all the correlations it predicts. It is natural to ask whether probabilistic theories are causally closed if one assumes that Reichenbach’s Common Cause Principle holds: this principle states that if two events A and B are probabilistically correlated then either the correlation is due to a causal interaction between A and B, or, if A and B are causally independent, Rind (A, B), then there exists a third event C, a so-called common cause that explains the correlation by being related to A and B probabilistically in a specific way (see Definition 1). Causal closedness of a probabilistic theory is intended to express that the theory complies with the Common Cause Principle; accordingly, and more precisely, a probability theory is defined to be causally closed with respect to a causal independence relation 1 Rind defined between pairs of random events if for any correlation between elements A and B such that Rind (A, B) holds, there exists a common cause C in that theory of the correlation between A and B (Definition 3). The problem is then under what conditions on the probability space and on Rind is the probability space causally closed. We will see that causal closedness is non-trivial and not impossible in classical probability spaces, not even if the probability space contains a finite number of random events only – but causal completeness is not typical either. There does not seem to be any regular and easily characterizable behavior of probability spaces from the perspective of causal closedness; one has to check in each and every case by brute force whether the causal closedness holds. The notion of causal closedness of classical probability spaces was introduced in [1] but the notion of common cause can be naturally defined in non-classical (quantum) probability spaces as well; hence the notion of causal closedness also makes perfect sense for such probability theories. Very little known about causal closedness of such general probability spaces. The known results are summarized in Section 4. 2 The notion of common cause and some terminology In what follows (X, S, p) denotes a classical (Kolmogorovian) probability space with Boolean algebra S of subsets of a set X (with respect to the set theoretic operations ∩, ∪ and A⊥ = X \ A as Boolean algebra operations) and with the probability measure p on S. To simplify notation, occasionally we write (S, p) instead of (X, S, p), when the precise nature of X is not important. For instance, when S has a final number of elements, then S is the set of all subsets of a finite set having n < ∞ elements, in which case we write (Sn , p). Given (S, X, p), the quantity Corrp (A, B) defined by Corrp (A, B) ≡ p(A ∩ B) − p(A)p(B) (1) is called the correlation of A, B in p. Events A and B are said to be positively correlated if Corrp (A, B) > 0. A correlation Corrp (A, B) 6= 0 is called non-degenerate (and (A, B) a non-degenerate correlated pair) if A 6= B. A correlation is called maximal if p(A|B) = p(B|A) = 1 (2) The next definition specifies the notion of common cause. Since the definition was first given by Reichenbach in [13], this type of common cause is called “Reichenbachian”; however, since in this paper only Reichenbachian common causes feature, the qualifier “Reichenbachian” will be omitted. 2 Definition 1. C is a common cause of the correlation (1) if the following (independent) conditions hold: p(A ∩ B|C) = p(A|C)p(B|C) p(A ∩ B|C ⊥ ) = p(A|C ⊥ )p(B|C ⊥ ) p(A|C) > p(A|C ⊥ ) p(B|C) > p(B|C ⊥ ) where (3) (4) (5) (6) p(X ∩ Y ) p(Y ) p(X|Y ) = denotes the conditional probability of X on condition Y , C ⊥ denotes the complement of C and it is assumed that none of the probabilities p(X) (X = A, B, C, C ⊥ ) is equal to zero. Since the notion of common cause is a measure theoretic one, measure zero sets have to be dealt with. The next definition takes care of this and summarizes some terminology used later. Definition 2. 1. Let ∆(X, Y ) ≡ (X \ Y ) ∪ (Y \ X) be the symmetric difference of sets X, Y . A common cause C of the correlation between A, B is called proper if p (∆(B, C)) 6= 0 6= p (∆(A, C)) (7) That is to say, a common cause C of the correlation Corrp (A, B) > 0 is proper if the common cause differs from the correlated events by more than a measure zero event. Otherwise C is called improper. 2. It can happen that, in addition to being a probabilistic common cause, the common cause event C logically implies both A and B, i.e. C ⊆ A ∩ B. If this is the case then we call C a strong common cause. If C is a common cause such that C 6⊆ A and C 6⊆ B then C is called a genuinely probabilistic common cause. 3. A common cause C will be called deterministic if p(A|C) = ⊥ p(A|C ) = 3 1 = p(B|C) ⊥ 0 = p(B|C ) (8) (9) 3 Causal closedness of classical probability theories Given the notion of common cause one can define the concept of common cause closedness in a very natural manner: Definition 3. Let (X, S, p) be a probability space and Rind be a two-place causal independence relation between elements of S. The probability space (X, S, p) is called common cause closed with respect to Rind , if for every correlation Corrp (A, B) > 0 with A ∈ S and B ∈ S such that Rind (A, B) holds, there exists a common cause C in S. If there are no elements A, B in S that are positively correlated, then (X, S, p) is called trivially common cause closed. Proposition 1 ([1]). Let (Sn , p) be a finite probability space. If Rind contains all the pairs of events A, B in Sn that are correlated in p, then (Sn , p) is not non-trivially common cause closed with respect to Rind . Proposition 1 shows that a probability space containing a finite number of random events contains more correlations than it can account for exclusively in terms of common causes. But this is not surprising because common cause closedness with respect to a causal independence relation that leaves no room for causal dependence is unreasonably strong. One can of course make a probability space (X, S, p) causally closed by stipulating that Rind (A, B) does not hold (i.e. that A and B are causally related) whenever A and B are correlated but there exists no common cause C ∈ S of the correlation. But this is unacceptable in general since this move makes the notion of causal closedness trivial and the causal dependence so defined (and the causal independence relation Rind so defined) may turn out not to have reasonable features. One needs a disciplined, independent definition of the causal independence relation. Intuitively, causal independence of A and B should imply that from the presence or absence of A one should not be able to infer either the occurrence or non-occurrence of B, and conversely: presence or absence of B should not entail occurrence or non-occurrence of A. Taking, as it is common, the partial ordering ⊆ in the Boolean algebra S as the implication relation between events (equivalently: between propositions that the corresponding events occur), this requirement about Rind can be expressed by the demand that Rind (A, B) should imply all of the following relations A 6⊆ B, B 6⊆ A, A⊥ 6⊆ B B ⊥ 6⊆ A , A 6⊆ B ⊥ , , B⊆ 6 A⊥ , A⊥ 6⊆ B ⊥ B ⊥ 6⊆ A⊥ This requirement can be expressed compactly by saying that Rind (A, B) implies that A and B are logically independent; equivalently, that {∅, A, A⊥ , X} and {∅, B, B ⊥ , X} are logically independent Boolean subalgebras of S in the sense of the following Definition 4: 4 Definition 4. Two Boolean subalgebras L1 , L2 of the Boolean algebra S are called logically independent if A ∩ B 6= ∅ whenever ∅, X 6= A ∈ L1 and ∅, X 6= B ∈ L2 (10) The pair (L01 , L02 ) of Boolean subalgebras of Boolean algebra S is called a maximal logically independent pair, if logical independence of Boolean subalgebras L1 and L2 containing respectively L01 and L02 as Boolean subalgebras implies L01 = L1 and L02 = L2 . For later purposes we also need the following notions: Definition 5. The pair (A, B) is called logically independent modulo zero probability if there exist A0 , B 0 such that p(A0 ) = p(B 0 ) = 0 (11) and (A \ A0 ) and (B \ B 0 ) are logically independent. This motivates the following definition, which formulates a natural notion of causal closedness. Definition 6. (X, S, p) is called common cause closed with respect to the pair (L1 , L2 ) of logically independent Boolean subalgebras of S, if for every A ∈ L1 and B ∈ L2 that are correlated in p, there exists a common cause C in S of the correlation between A and B. Proposition 2 ([1]). Let (S5 , pu ) be the probability space with the Boolean algebra S5 generated by 5 atoms and with pu being the probability measure defined by the uniform distribution on atoms of S5 . Then (S5 , pu ) is common cause closed with respect to every pair of logically independent Boolean subalgebras (L1 , L2 ) of S5 . The next proposition shows that the behavior of the probability space (S5 , pu ) described in Proposition 2 is exceptional. Proposition 3 ([1]). If the probability space (Sn , p) is not (S5 , pu ), then it is not non-trivially common cause closed with respect to every pair of logically independent Boolean subalgebras. But causally not closed probability spaces can be extended in such a manner that the extension contains common causes of a finite number of correlations in a given pair of logically independent sublattices: Proposition 4 ([1],[5]). If (X, S, p) with finite S is not common cause closed with respect to a logically independent pair (L1 , L2 ), then it can be extended into a (X 0 , S 0 , p0 ), with S 0 being also finite, in such a manner that (X 0 , S 0 , p0 ) is common cause closed with respect to the logically independent pair (h(L1 ), h(L2 )), where h(Li ) is the homomorphic image in S 0 of Li (i = 1, 2). 5 By an extension is meant here that there exists a Boolean algebra embedding h of S into S 0 that preserves the probability in the sense that p(X) = p0 (h(X)) for all X ∈ S. Note that the images of L1 , L2 under h will not necessarily be maximally logically independent, not even if L1 , L2 is a maximally logically independent pair; so we have the following Problem, which is open: Problem 1. Does Proposition 4 remain true if “logically independent” means maximal logically independent? The following problem is also open: Problem 2. Let (S, p) be a probability space with S having an infinite number of elements and assume that (L1 , L2 ) is a logically independent pair of Boolean subalgebras of S such that there exist an infinite number of pairs (Ai , Bi ) of events Ai ∈ L1 and Bi ∈ L2 that are correlated in p. Does there exist and extension (S 0 , p0 ) of (S, p) such that (S 0 , p0 ) is common cause closed with respect to (h(L1 ), h(L2 ))? Does there exist such an extension so that (h(L1 ), h(L2 )) is a maximal pair of logically independent sub-Boolean algebras in S 0 ? Probability spaces with infinite Boolean algebras can however be causally closed as the next proposition shows. Before stating the proposition recall that a probability space (X, S, p) is called atomless if for any A ∈ S, p(A) 6= 0 there exists B ⊆ A, B ∈ S such that 0 < p(B) < p(A). Proposition 5 ([1]). If (X, S, p) is an atomless probability space, then it contains uncountably many proper common causes of every non-degenerate correlation in it. Moreover if A and B are correlated, logically independent modulo measure zero events, then S contains both uncountably many strong and uncountably many genuinely probabilistic common causes of the correlation between A and B. The notion of common cause can be naturally generalized to cover the case when the correlation is not explainable by a single common cause but by system of common cause like events. One such generalization was given in [3, 4]: Definition 7. Let (X, S, p) be a probability space and A, B be two events in S. The partition {Ci }i∈I of S is said to be a Reichenbachian common cause system (RCCS for short) for the pair (A, B) if the following two conditions are satisfied p(A ∩ B|Ci ) = p(A|Ci )p(B|Ci ) for all i ∈ I [p(A|Ci ) − p(A|Cj )][p(B|Ci ) − p(B|Cj )] > 0 (i 6= j) (12) (13) The cardinality of the index set I (i.e. the number of events in the partition) is called the size of the RCCS. Since C, C ⊥ with a Reichenbachian common cause C is a RCCS of size 2, we call a RCCS proper if its size is greater than 2. It was shown in [3, 4] that Reichenbachian Common Cause systems of arbitrary finite size exist for any non-maximal correlation in the sense that for any such correlation in any probability space there exists an extension of that 6 probability space that contains a Reichenbachian Common Cause system of the prescribed size. In view of this, a natural refinement of the definition of causal closedness of (X, S, p) is obtained if one replaces the notion of common cause with the concept of common cause system: Definition 8. (X, S, p) is called causally N -closed with respect to a causal independence relation Rind if for any correlation Corrp (A, B) > 0 such that Rind (A, B) holds, there exists in (X, S, p) a Reichenbachian common cause system of size N for the correlation. There are a number of questions one can ask in connection with causal N closedness: Problem 3. On what condition on (X, S, p) and Rind is (X, S, p) causally N closed for a fixed N ? We have seen that probability spaces may or may not be causally 2-closed – causal 2-closedness depends sensitively on how Rind is defined. This leads to the following open problems: Problem 4. Can a probability space which is not causally 2-closed be causally N -closed for some fixed N > 2 (with respect to some non-trivial causal independence relation Rind )? Problem 5. Can a probability space be causally N -closed for every N (with respect to some non-trivial causal independence relation Rind )? We conjecture that atomless probability spaces are causally N -closed for every N with respect to every pair of logically independent Boolean subalgebras. 4 Causal closedness of non-classical probability spaces The notion of common cause can be defined in non-classical (quantum) probability spaces (L, φ), where an orthomodular lattice L takes the place of the Boolean algebra and φ is an additive (or σ additive) map from L into [0, 1] (generalized probability measure), replacing a classical probability measure. Special examples of such spaces are the quantum probability spaces (P(N ), φ) where P(N ) is the projection lattice of a von Neumann algebra N and φ is a (normal) state on N (see [21], [6] for the theory of von Neumann algebras). An even more specific example of the latter is the probability space when P(N ) is the von Neumann lattice of all projections on a Hilbert space H (in this case we write (H, P(H), φ); this latter non-classical probability space describes standard, nonrelativistic quantum systems. Two elements A, B ∈ L are called compatible if they belong to the same distributive sublattice of L. This condition is equivalent to A = (A ∧ B) ∨ (A ∨ B ⊥ ) 7 If A, B are compatible and Corrφ (A, B) ≡ φ(A ∧ B) − φ(A)φ(B) > 0 (14) then A and B are called (positively) correlated with respect to the state φ. Definition 9. If A and B are positively correlated, then C ∈ L is called a common cause of the correlation (14) if C is compatible with both A and B and the following conditions hold. φ(A ∧ B|C) φ(A ∧ B|C ⊥ ) φ(A|C) φ(B|C) = = > > φ(A|C)φ(B|C) φ(A|C ⊥ )φ(B|C ⊥ ) φ(A|C ⊥ ) φ(B|C ⊥ ) where φ(X|Y ) = (15) (16) (17) (18) φ(X ∧ Y ) φ(Y ) denotes the conditional probability of X on condition Y and it is assumed that none of the probabilities φ(X), (X = A, B, C, C ⊥ ) is equal to zero. Extension of (L, φ), logical independence of events in L and causal independence relation Rind on L can all be defined in complete analogy with the classical definitions, which makes it possible to define causal completeness as well in complete analogy with the classical case: Definition 10. Let (L, φ) be a non-classical probability space and Rind be a two-place causal independence relation between elements of L. The probability space (L, φ) is called common cause closed with respect to Rind , if for every correlation Corrφ (A, B) > 0 with A ∈ L and B ∈ L such that Rind (A, B) holds, there exists a common cause C in L (in the sense of Definition 9). If there are no compatible elements A, B in L that are positively correlated, then (L, p) is called trivially common cause closed. Problem 6. On what conditions on (L, φ) and Rind is the probability space (L, φ) common cause closed with respect to Rind ? This problem is largely open in this generality. The only general result known is Proposition 6 ([7]). If L is an atomless, complete, orthomodular lattice and φ is a faithful state then (L, φ) is causally closed with respect to every pair of logically independent sublattices. The above result is the quantum counterpart of Proposition 5; and the key fact that it rests on is that if L is an atomless lattice and φ is a faithful state on L then (L, φ) is atomless as a measure space in the sense that for any 0 6= A ∈ L, and for any real number 0 6= r < p(A) there exists B ≤ A, B ∈ L 8 such that p(B) = r. This latter fact was proved in [11] (see also [12]) for the specific quantum probability space (P(N ), φ) where N is a type III von Neumann algebra and φ is a faithful normal state on N , and Kitajima [7] showed that the proof of can be carried over from the von Neumann algebra framework to more general non-classical probability spaces. It is known that the projection lattices of type II von Neumann algebras are also atomless; so one has as a specific case of Proposition 6 the following Proposition 7. Let (P(N ), φ) be a quantum probability space with N as a type III or type II von Neumann algebra and φ as a faithful normal state on N . Then (P(N ), φ) is causally closed with respect to every pair of logically independent sublattices. Note that the lattice P(H) of all projections on a Hilbert space H is not atomless (it is atomic) irrespective of the dimensionality of the Hilbert space H [9]; moreover, the quantum probability spaces (H, P(H), φ) are not atomless in a measure theoretic sense; consequently, Propositions 6 and 7 do not say anything about the causal closedness of the quantum probability space (H, P(H), φ) and it is not known under what conditions such quantum probability spaces are causally closed (with respect to some Rind ). Just like in the classical case, the notion of a (Reichenbachian) common cause system also can be formulated in a non-classical probability space, and one can define naturally the more general notion of causal N -closedness of a non-classical probability space: The set {Ci , i ∈ J} of elements (J being an index set, Ci ∈ L) is called a partition in L if ∨i Ci = I and Ci and Cj are orthogonal whenever i 6= j; i.e. Ci ≤ Cj⊥ for i 6= j. Definition 11. A partition {Ci , i ∈ J} is a (Reichenbachian) common cause system for the correlation (14) between compatible elements A and B if Ci is compatible with both A and B for every i ∈ J and the following conditions (analogous to (12)-(13)) hold φ(A ∧ B|Ci ) = φ(A|Ci )φ(B|Ci ) for all i ∈ J [φ(A|Ci ) − φ(A|Cj )][φ(B|Ci ) − φ(B|Cj )] > 0 (i 6= j) (19) (20) The cardinality of the index set J is called the size of the common cause system. Definition 12. The probability space (L, φ) is called causally N -closed (with respect to some causal independence relation Rind ) if for any correlation between elements that stand in the causal independence relation there exists in (L, φ) a Reichenbachian common cause system of size N . There are a number of open problems in connection with Reichenbachian common cause systems in non-classical probability spaces and causal N -closedness of such probability theories: Problem 7. Given a correlation in a general probability space (L, φ) that does not have a common cause system of a given size N > 2 of the correlation, does there exist an extension (L0 , φ0 ) of (L, φ) such that there exists a Reichenbachian common cause system of size N of the correlation in the extension (L0 , φ0 )? 9 We conjecture a positive answer to the above question. Problem 8. 1. Do there exist non-classical probability spaces that are causally N -closed for some fixed N (with respect to some nontrivial Rind )? 2. Do there exist non-classical probability spaces that are causally N -closed for every N (with respect to some decent Rind )? 3. Do there exist non-classical probability spaces that are causally closed (with respect to some non-trivial causal independence relation Rind ) in such a way that every correlation in the space has a common cause system of countably infinite size? These questions have not been investigated. 5 Closing comments on causal closedness Further generalization can be achieved by treating the specific form the correlation measure Corr takes as a variable of the notion of the common cause. By allowing Corr to measure correlation between pairs of ordered partitions it becomes possible to handle the case of common cause-type explanations of correlating variables, not just that of events. This allows a more detailed analysis of causal closedness and of falsification attempts against the Common Cause Principle. It can be shown that Reichenbachian common cause systems are special cases of the resulting notion of a generalized Reichenbachian common cause. By imposing mild conditions on Corr extension theorems analogous to Proposition 4 can be proven. However the question of common cause closedness of general probability spaces with respect to generalized Reichenbachian common causes is still open. For further details the Reader is referred to [2]. One can strengthen Reichenbach’s notion of common cause by requiring the common cause to satisfy some additional conditions. The additional conditions can be motivated by physics: after all, the probability measure spaces in terms of which the concept of common cause is formulated are not just abstract mathematical entities in physics but physically interpreted structures. Being organic parts of specific physical theories, these measure spaces offer means to express a possibly large variety of physical facts and principles. Two of such important principles are locality and causality. Both locality and causality are rich and many-layered concepts and there is no unique way of expressing them in terms of probability measure spaces. But it can happen that a physical theory entails both some additional conditions as necessary for the common cause C of a correlation between A and B to be “local” and a causal dependence relation between random events. In such a situation the problem of causal closedness should be reformulated by taking into account these further restrictions. This happens in local, algebraic, relativistic quantum field theory (AQFT). The theory predicts correlations between localized, causally independent (specelike separated) observables [16, 17, 18, 19], [20, 14, 15], and the common causes 10 of these correlations have to be localized [8]. It turns out that there is no unique way of defining locality of the common cause and, consequently, causal closedness of AQFT can also be specified in different ways [12]: strong, weak and right (”desirable”) localizability of common causes lead to the concepts of “strong”, “weak” and “desirable” causal closedness of AQFT. To decide which of these causal closedness hold for AQFT is a difficult matter. While it is easy to see that strong causal closedness is violated in AQFT [12], and it could be shown that AQFT is weakly causally closed [11] (see also the review [10]), it remains an open problem whether AQFT is causally closed in the most desirable sense. Acknowledgement: Work supported in part by the Hungarian Scientific Research Found (OTKA), contract number: K68043. References [1] B. Gyenis and M. Rédei. When can statistical theories be causally closed? Foundations of Physics, 34:1285–1303, 2004. [2] B. Gyenis and M. Rédei. Causal completeness of general probability theories. In M. Suarez, editor, Probabilities, causes and propensities in physics, Synthese Library. Springer, 2008. forthcoming. [3] G. Hofer-Szabó and M. Rédei. Reichenbachian common cause systems. International Journal of Theoretical Physics, 43:1819–1826, 2004. [4] G. Hofer-Szabó and M. Rédei. Reichenbachian common cause systems of arbitrary finite size exist. Foundations of Physics Letters, 35:745–746, 2006. [5] G. Hofer-Szabó, M. Rédei, and L.E. Szabó. On Reichenbach’s Common Cause Principle and Reichenbach’s notion of common cause. The British Journal for the Philosophy of Science, 50:377–398, 1999. [6] R.V. Kadison and J.R. Ringrose. Fundamentals of the Theory of Operator Algebras, volume I. and II. Academic Press, Orlando, 1986. [7] Yuichiro Kitajima. Reichenbach’s Common Cause in an atomless and complete orthomodular lattice. International Journal of Theoretical Physics, 47:511–519, 2008. [8] M. Rédei. Reichenbach’s Common Cause Principle and quantum field theory. Foundations of Physics, 27:1309–1321, 1997. [9] M. Rédei. Quantum Logic in Algebraic Approach, volume 91 of Fundamental Theories of Physics. Kluwer Academic Publisher, 1998. [10] M. Rédei. Reichenbach’s common cause principle and quantum correlations. In T. Placek and J. Butterfield, editors, Modality, Probability and Bell’s Theorems, volume 64 of NATO Science Series, II., pages 259–270. Kluwer Academic Publishers, Dordrecht, Boston, London, 2002. 11 [11] M. Rédei and S.J. Summers. Local primitive causality and the common cause principle in quantum field theory. Foundations of Physics, 32:335– 355, 2002. [12] M. Rédei and S.J. Summers. Remarks on causality in relativistic quantum field theory. International Journal of Theoretical Physics, 44:10291039, 2005. [13] H. Reichenbach. The Direction of Time. University of California Press, Los Angeles, 1956. [14] S.J. Summers. Bell’s inequalities and quantum field theory. In Quantum Probability and Applications V., volume 1441 of Lecture Notes in Mathematics, pages 393–413. Springer, 1990. [15] S.J. Summers. On the independence of local algebras in quantum field theory. Reviews in Mathematical Physics, 2:201–247, 1990. [16] S.J. Summers and R. Werner. The vacuum violates Bell’s inequalities. Physics Letters A, 110:257–279, 1985. [17] S.J. Summers and R. Werner. Bell’s inequalities and quantum field theory, I. General setting. Journal of Mathematical Physics, 28:2440–2447, 1987. [18] S.J. Summers and R. Werner. Bell’s inequalities and quantum field theory, II. Bell’s inequalities are maximally violated in the vacuum. Journal of Mathematical Physics, 28:2448–2456, 1987. [19] S.J. Summers and R. Werner. Maximal violation of Bell’s inequalities is generic in quantum field theory. Commununications in Mathematical Physics, 110:247–259, 1987. [20] S.J. Summers and R. Werner. Maximal violation of Bell’s inequalities for algebras of observables in tangent spacetime regions. Annales de l’Institut Henri Poincaré – Physique théorique, 49:215–243, 1988. [21] M. Takesaki. Theory of Operator Algebras, volume I. Springer Verlag, New York, 1979. 12
© Copyright 2026 Paperzz