A system's wave function is uniquely determined by its underlying physical state

We address the question of whether the quantum-mechanical wave function $\Psi$ of a system is uniquely determined by any complete description $\Lambda$ of the system's physical state. We show that this is the case if the latter satisfies a notion of"free choice". This notion requires that certain experimental parameters---those that according to quantum theory can be chosen independently of other variables---retain this property in the presence of $\Lambda$. An implication of this result is that, among all possible descriptions $\Lambda$ of a system's state compatible with free choice, the wave function $\Psi$ is as objective as $\Lambda$.


I. INTRODUCTION
The quantum-mechanical wave function, Ψ, has a clear operational meaning, specified by the Born rule [1]. It asserts that the outcome X of a measurement, defined by a family of projectors {Π x }, follows a distribution P X given by P X (x) = Ψ|Π x |Ψ , and hence links the wave function Ψ to observations. However, the link is probabilistic; even if Ψ is known to arbitrary precision, we cannot in general predict X with certainty.
In classical physics, such indeterministic predictions are always a sign of incomplete knowledge. 1 This raises the question of whether the wave function Ψ associated to a system corresponds to an objective property of the system, or whether it should instead be interpreted subjectively, i.e., as a representation of our (incomplete) knowledge about certain underlying objective attributes. Another alternative is to deny the existence of the latter, i.e., to give up the idea of an underlying reality completely.
Despite its long history, no consensus about the interpretation of the wave function has been reached. A subjective interpretation was, for instance, supported by the famous argument of Einstein, Podolsky and Rosen [2] (see also [3]) and, more recently, by information-theoretic considerations [4][5][6]. The opposite (objective) point of view was taken, for instance, by Schrödinger (at least initially), von Neumann, Dirac, and Popper [7][8][9].
To turn this debate into a more technical question, one may consider the following gedankenexperiment: Assume you are provided with a set of variables Λ that are intended to describe the physical state of a system. Suppose, furthermore, that the set Λ is complete, i.e., there is nothing that can be added to Λ to increase the accuracy of any predictions about the outcomes of measurements * roger.colbeck@york.ac.uk † renner@phys.ethz.ch 1 For example, when we assign a probability distribution P to the outcomes of a die roll, P is not an objective property but rather a representation of our incomplete knowledge. Indeed, if we had complete knowledge, including for instance the precise movement of the thrower's hand, the outcome would be deterministic.
on the system. If you were now asked to specify the wave function Ψ of the system, would your answer be unique? If so then Ψ is a function of the variables Λ and hence as objective as Λ. The model defined by Λ would then be called Ψ-ontic [10]. Conversely, the existence of a complete set of variables Λ that does not determine the wave function Ψ would mean that Ψ cannot be interpreted as an objective property. Λ would then be called Ψ-epistemic. 2 In a seminal paper [14], Pusey, Barrett and Rudolph showed that any complete model Λ is Ψ-ontic if it satisfies an assumption, termed "preparation independence". It demands that Λ consists of separate variables for each subsystem, e.g., Λ = (Λ A , Λ B ) for two subsystems S A and S B , and that these are statistically independent, i.e., P Λ A Λ B = P Λ A ×P Λ B , whenever the joint wave function Ψ of the total system has product form, i.e., Ψ = Ψ A ⊗ Ψ B .
Here we show that the same conclusion can be reached without imposing any internal structure on Λ. More precisely, we prove that Ψ is a function of any complete set of variables that are compatible with a notion of "free choice" (Corollary 1). This captures the idea that experimental parameters, e.g., which state to prepare or which measurement to carry out, can be chosen independently of all other information (relevant to the experiment), except for information that is created after the choice is made, e.g., measurement outcomes. While this notion is implicit to quantum theory, we demand that it also holds in the presence of Λ. 3 The proof of our result is inspired by our earlier work [15] in which we observed that the wave function Ψ is uniquely determined by any complete set of variables Λ, provided that Ψ is itself complete (in the 2 Note that the existence or non-existence of Ψ-epistemic theories is also relevant in the context of simulating quantum systems. Here Λ can be thought of as the internal state of a computer performing the simulation, and one would ideally like that storing Λ requires significantly fewer resources than would be required to store Ψ. However, a number of existing results already cast doubt on this possibility (see, for example, [11][12][13]). 3 Free choice of certain variables is also implied by the preparation independence assumption used in [14], as discussed below.
sense described above). Furthermore, in another previous work [16], we showed that completeness of Ψ holds within any framework compatible with free choice, if one makes the additional assumption that any quantum measurement on a system corresponds to a unitary evolution of an extended system. Hence, in this case, Ψ is determined by Λ [15]. The argument we provide in the present work shows that this conclusion is true more generally, even if the unitarity assumption does not hold. 4

II. THE UNIQUENESS THEOREM
Our argument refers to an experimental setup where a particle emitted by a source decays into two, each of which is directed towards one of two measurement devices (see Fig. 1). The measurements that are performed depend on parameters A and B, and their respective outcomes are denoted X and Y .
Quantum theory allows us to make predictions about these outcomes based on a description of the initial state of the system, the evolution it undergoes and the measurement settings. For our purposes, we assume that the quantum state of each particle emitted by the source is pure, and hence specified by a wave function 5 . As we 4 Note, however, that the argument we give in this work doesn't allow us to conclude that Ψ is complete. 5 We consider it uncontroversial that a mixed state can be thought will consider different choices for this wave function, we model it as a random variable Ψ that takes as values unit vectors ψ in a complex Hilbert space H. Furthermore, we take the decay to act like an isometry, denoted U , from H to a product space H A ⊗ H B . Finally, for any choices a and b of the parameters A and B, the measurements are given by families of projectors {Π a x } x∈X and {Π b y } y∈Y on H A and H B , respectively. The Born rule, applied to this setting, now asserts that the joint probability distribution of X and Y , conditioned on the relevant parameters, is given by To model the system's "physical state", we introduce an additional random variable Λ. We do not impose any structure on Λ (in particular, Λ could be a list of values). We will consider predictions P XY |ABΛ (x, y|a, b, λ) conditioned on any particular value λ of Λ, analogously to the predictions based on Ψ according to the Born rule (1).
To define the notions of free choice and completeness, as introduced informally in the introduction, we use the fact that any experiment takes place in spacetime and therefore has a chronological structure. 6 For example, the measurement setting A is chosen before the measurement outcome X is obtained. This may be modelled mathematically by a preorder relation 7 , denoted , on the relevant set of random variables. While our technical claim does not depend on how the chronological structure is interpreted physically, it is intuitive to imagine it being compatible with relativistic spacetime. In this case, A X would mean that the spacetime point where X is accessible lies in the future light cone of the spacetime point where the choice A is made.
For our argument we consider the chronological structure defined by the transitive completion of the relations of as a state of knowledge.
(cf. Fig. 2). This reflects, for instance, that Ψ is chosen at the very beginning of the experiment, and that A and B are chosen later, right before the two measurements are carried out. Note, furthermore, that A Y and B X. With the aforementioned interpretation of the relation in relativistic spacetime, this would mean that the two measurements are carried out at spacelike separation.
Using the notion of a chronological structure, we can now specify mathematically what we mean by free choices and by completeness. We note that the two definitions below should be understood as necessary (but not necessarily sufficient) conditions characterising these concepts. Since they appear in the assumptions of our main theorem, our result also applies to any more restrictive definitions. We remark furthermore that the definitions are generic, i.e., they can be applied to any set of variables equipped with a preorder relation. 8 Definition 1. When we say that a variable A is a free choice from a set A (w.r.t. a chronological structure) this means that the support of P A contains A and that P A|A ↑ = P A where A ↑ is the set of all random variables Z (within the chronological structure) such that A Z.
In other words, a choice A is free if it is uncorrelated with any other variables, except those that lie in the future of A in the chronological structure. For a further discussion and motivation of this notion we refer to Bell's work [17] as well as to [18].
Crucially, we note that Definition 1 is compatible with the usual understanding of free choices within quantum theory. For example, if we consider our experimental setup (cf. Fig. 1) in ordinary quantum theory (i.e., where there is no Λ), the initial state Ψ as well as the measurement settings A and B can be taken to be free choices w.r.t. Ψ A, Ψ B, A X, B Y (which is the chronological structure defined by Eq. 2 with Λ removed).
Definition 2. When we say that a variable Λ is complete (w.r.t. a chronological structure) this means that 9 where Λ ↑ and Λ ↓ denote the sets of random variables Z (within the chronological structure) such that Λ Z and Z Λ, respectively.
Completeness of Λ thus implies that predictions based on Λ about future values Λ ↑ cannot be improved by taking into account additional information Λ ↓ available in the past. 10 Recall that this is meant as a necessary criterion for completeness and that our conclusions hold for any more restrictive definition. For example, one may replace the set Λ ↑ by the set of all values that are not in the past of Λ.
We are now ready to formulate our main result as a theorem. Note that, the assumptions to the theorem as well as its claim correspond to properties of the joint probability distribution of X, Y , A, B, Ψ and Λ. Theorem 1. Let Λ and Ψ be random variables and assume that the support of Ψ contains two wave functions, ψ and ψ , with | ψ|ψ | < 1. If for any isometry U and measurements {Π a x } x and {Π b y } y , parameterised by a ∈ A and b ∈ B, there exist random variables A, B, X and Y such that The theorem asserts that, assuming validity of the Born rule and freedom of choice, the values taken by any complete variable Λ are different for different choices of the wave function Ψ. This implies that Ψ is indeed a function of Λ.
To formulate this implication as a technical statement, we consider an arbitrary countable 11 set S of wave functions such that | ψ|ψ | < 1 for any distinct elements ψ, ψ ∈ S. Corollary 1. Let Λ and Ψ be random variables with Ψ taking values from the set S of wave functions. If the conditions of Theorem 1 are satisfied then there exists a function f such that Ψ = f (Λ) holds almost surely.
The proof of this corollary is given in Appendix A.

III. PROOF OF THE UNIQUENESS THEOREM
The argument relies on specific wave functions, which depend on parameters d, k ∈ N and ξ ∈ [0, 1], with k < d. They are defined as unit vectors on a product space H A ⊗ H B , where H A and H B are (d + 1)dimensional Hilbert spaces equipped with an orthonor- Lemma 1. For any 0 ≤ α < 1 there exist k, d ∈ N with k < d and ξ ∈ [0, 1] such that the vectors φ and φ defined by (3) and (4) have overlap φ|φ = α.
We also set Π a d = Π b d = |d d|. The outcomes X and Y will generally be correlated. To quantify these correlations, we define 13 The next lemma shows that I n,d gives an upper bound on the distance of the distribution P X|AΛ from a uniform distribution over {0, . . . , d − 1}. The bound holds for any random variable Λ, provided the joint distribution P XY Λ|AB satisfies certain conditions. 12 We use here the abbreviation |j |j for |j ⊗ |j . 13 Note that the first sum corresponds to the probability that X ⊕ 1 = Y , conditioned on A = 0 and B = 2n − 1. The terms in the second sum can be interpreted analogously.
Lemma 2. Let P XY ABΛ be a distribution that satisfies P XΛ|AB = P XΛ|A , P Y Λ|AB = P Y Λ|B and P ABΛ = P A × P B ×P Λ with supp(P A ) ⊇ A n and supp(P B ) ⊇ B n . Then The proof of this lemma is given in Appendix C. It generalises an argument described in [16], which is in turn based on work related to chained Bell inequalities [19,20] (see also [21,22]).
We have now everything ready to prove the uniqueness theorem.
Proof of Theorem 1. Let α, γ ∈ R such that e iγ α = ψ|ψ . Furthermore, let k, d, ξ be as defined by Lemma 1, so that φ|φ = α. Then there exists an isometry U such that U ψ = φ and U ψ = e iγ φ (see Lemma 3 of Appendix D). 14 Now let n ∈ N and let A, B, X and Y be random variables that satisfy the three conditions of the theorem for the isometry U and for the projective measurements defined by (5) and (6), which are parameterised by a ∈ A n and b ∈ B n , respectively. According to the Born rule (Condition 1), the distribution P XY |ABψ ≡ P XY |ABΨ (·, ·|·, ·, ψ) conditioned on the choice of initial state Ψ = ψ corresponds to the one considered in (7), i.e., Note that P A|BΨ P Y Λ|ABΨ = P AY Λ|BΨ = P A|BY ΛΨ P Y Λ|BΨ . Freedom of choice (Condition 2) implies that P A|BΨ = P A|BY ΛΨ .
It follows that P Y Λ|ABΨ = P Y Λ|BΨ . By a similar reasoning, we also have P XΛ|ABΨ = P XΛ|AΨ . The freedom of choice condition also ensures that P ABΛ|Ψ = P A × P B × P Λ|Ψ with suppP A ⊇ A n and suppP B ⊇ B n . We can thus apply Lemma 2 to give, with (8), Considering only the term x = k (recall that k < d) and noting that the left hand side does not depend on n, we have dP Λ|ψ (λ) P X|AΛΨ (k|0, λ, ψ) − 1 d = 0 (otherwise, by taking n sufficiently large, we will get a contradiction with the above). Let L be the set of all elements λ from the range of Λ for which P X|AΛΨ (k|0, λ, ψ) 14 If H has a larger dimension than H A ⊗H B (e.g., because H is infinite dimensional) then we can consider an (infinite dimensional) extension of H B , keeping the same notation for convenience.
is defined and equal to 1 d . The above implies that P Λ|Ψ (L|ψ) = 1.

IV. DISCUSSION
It is interesting to compare Theorem 1 to the result of [14], which we briefly described in the introduction. The latter is based on a different experimental setup, where n particles with wave functions Ψ 1 , . . . , Ψ n , each chosen from a set {ψ, ψ }, are prepared independently at n remote locations. The n particles are then directed to a device where they undergo a joint measurement with outcome Z.
The main result of [14] is that, for any variable Λ that satisfies certain assumptions, the wave functions Ψ 1 , . . . , Ψ n are determined by Λ. One of these assumptions is that Λ consists of n parts, Λ 1 , . . . , Λ n , one for each particle. To state the other assumptions and compare them to ours, it is useful to consider the chronological structure defined by the transitive completion of the relations 15 It is then easily verified that the assumptions of [14] imply the following: 1. P Z|Ψ1···Ψn satisfies the Born rule; 15 Note that this chronological structure captures the aforementioned experimental setup. In particular, we have Ψ i Λ j for i = j, reflecting the idea that the n particles are prepared in separate isolated devices.
These conditions are essentially in one-to-one correspondence with the assumptions of Theorem 1. 16 The main difference thus concerns the modelling of the physical state Λ, which in the approach of [14] is assumed to have an internal structure. A main goal of the present work was to avoid using this assumption (see also [23,24] for alternative arguments).
We conclude by noting that the assumptions to Theorem 1 and Corollary 1 may be weakened. For example, the independence condition that is implied by free choice may be replaced by a partial independence condition along the lines considered in [25]. An analogous weakening was given in [26,27] regarding the argument of [14]. More generally, recall that all our assumptions are properties of the probability distribution P XY ABΨΛ . One may therefore replace them by relaxed properties that need only be satisfied for distributions that are εclose (in total variation distance) to P XY ABΨΛ . (For example, the Born rule may only hold approximately.) It is relatively straightforward to verify that the proof still goes through, leading to the claim that Ψ = f (Λ) holds with probability at least 1 − δ, with δ → 0 in the limit where ε → 0.
Nevertheless, none of the three assumptions of Theorem 1 can be dropped without replacement. Indeed, without the Born rule, the wave function Ψ has no meaning and could be taken to be independent of the measurement outcomes X. Furthermore, a recent impossibility result [28] implies that the freedom of choice assumption cannot be omitted. It also implies that the statement of Theorem 1 cannot hold for a setting with only one single measurement. This means that there exist Ψ-epistemic theories compatible with the remaining assumptions. However, in this case, it is still possible to exclude a certain subclass of such theories, called maximally Ψ-epistemic theories [29] (see also [30]). Finally, completeness of Λ is necessary because, without it, Λ could be set to a constant, in which case it clearly cannot determine Ψ. For any distinct ψ, ψ ∈ S, let L ψ,ψ be the set defined by Theorem 1, i.e., P Λ|Ψ (L ψ,ψ |ψ) = 1 P Λ|Ψ (L ψ,ψ |ψ ) = 0 , and for any ψ ∈ S define the (countable) intersection L ψ ≡ ψ ∈S\{ψ} L ψ,ψ . This satisfies P Λ|Ψ (L ψ |ψ ) = 1 if ψ = ψ 0 otherwise.
(Here we have used that for any probability distribution P and for any events L, L , P (L) = P (L ) = 1 implies that P (L ∩ L ) = 1.) To define the function f , we specify the inverse sets The function f is well defined on ψ∈S f −1 (ψ) because, by construction, the sets f −1 (ψ) are disjoint for different ψ ∈ S. Furthermore, it follows from the above that for any ψ ∈ S P Λ|Ψ (f −1 (ψ)|ψ) = 1 .
This implies that f (Λ) = Ψ holds with probability 1 conditioned on Ψ = ψ. The assertion of the corollary then follows because this is true for any ψ ∈ S. dP Λ (λ)D(P X|a0λ , 1/d) ≤ d 4 I n,d (P XY |AB ) .
For this, we consider the distribution P X⊕1|aλ , which corresponds to the distribution of X if its values are shifted by one (modulo d). According to Lemma 5 and using 1 d d 2 The assertion then follows with where we have set b 0 ≡ 2n − 1; the first inequality follows from Lemma 4; the second is obtained with P X|abλ = P X|aλ and P Y |abλ = P Y |bλ (which are implied by the conditions stated in the lemma) as well as the triangle inequality for D(·, ·).