Shifting the Quantum-Classical Boundary: Theory and Experiment for Statistically Classical Optical Fields

The growing recognition that entanglement is not exclusively a quantum property, and does not even originate with Schr\"odinger's famous remark about it [Proc. Camb. Phil. Soc. 31, 555 (1935)], prompts examination of its role in marking the quantum-classical boundary. We have done this by subjecting correlations of classical optical fields to new Bell-analysis experiments, and report here values of the Bell parameter greater than ${\cal B} = 2.54$. This is many standard deviations outside the limit ${\cal B} = 2$ established by the Clauser-Horne-Shimony-Holt (CHSH) Bell inequality [Phys. Rev. Lett. 23, 880 (1969)], in agreement with our theoretical classical prediction, and not far from the Tsirelson limit ${\cal B} = 2.828...$. These results cast a new light on the standard quantum-classical boundary description, and suggest a reinterpretation of it.

Introduction: For many decades the term "entanglement" has been attached to the world of quantum mechanics [1]. However, it is true that non-quantum optical entanglement can exist (realized very early by Spreeuw [2]) and its applications have concrete consequences. These are based on entanglements between two, or more than two, degrees of freedom, which are easily avalable classically [2][3][4][5][6]. Multi-entanglements of the same kind are also being explored quantum mechanically [7]. Applications in the classical domain have included, for example, resolution of a long-standing issue concerning Mueller matrices [8], an alternative interpretation of the degree of polarization [9], introduction of the Bell measure as a new index of coherence in optics [10], and innovations in polarization metrology [11]. Here we present theoretical and experimental results extending these results by showing that probabilistic classical optical fields can exhibit violations of the Clauser-Horne-Shimony-Holt (CHSH) Bell inequality [12] of quantum strength. This is evidence of a new kind that asks for reconsideration of the common understanding that Bell violation signals quantum physics. We emphasize that our discussion focuses on non-quantum entanglement of nondeterministic classical optical fields, and does not engage issues such as non-locality that are important for some applications in quantum information.
The observations and applications of non-quantum wave entanglement noted above [2][3][4][5][6][8][9][10][11] exploited nonseparable correlations among two or more modes or degrees of freedom (DOF) of optical wave fields. Nonseparable correlations among modes are an example of entanglement [13], but are not enough for our present purpose. In addition, we want to conform to three criteria that Shimony has identified for Bell tests [1], facts of quantum Nature that must be satisfied when examining possible tests of the quantum-classical border. Fortuitously, the ergodic stochastic optical fields of the classical theory of partial coherence and partial polarization (see Wolf [15]) satisfy these criteria fully (see Suppl. Materials [16]), and we have used such fields as our test bed. Background Theory: We will deal here only with the simplest suitable example, the theory of completely unpolarized classical light, and have explained elsewhere (see [16], [3]) the generalizations needed to treat partially polarized fields, which lead to the same conclusions. In all cases there are only two degrees of freedom (DOF) to deal with, namely the direction of polarization and the temporal amplitude of the optical field. In both classical and quantum theories these are fundamentally independent attributes. An electric field, for a beam travelling in the z direction, is written In the classical theory of unpolarized light [18] the optical field's two amplitudes E x and E y are statistically completely uncorrelated, and are treated as vectors in a stochastic function space. A scalar product of the vectors in this space corresponds physically to observable correlation functions such as E x E y . For unpolarized light we have E x E x = E y E y , and E x E y = 0. Now it is possible to talk of entanglement of the classical field. This is because entangled states are superpositions of products of vectors from different vector spaces, whenever the superpositions can't be rearranged into a single product that separates the two spaces [1]. Looking again at (1), we see that this is the case because we've taken E to be unpolarized. That is, by the definition of unpolarized light, there is no directionû of polarization that captures the total intensity, so E(t) can't, for any directionû, be written in the form E(t) =ûF (t), which would factorably separate the polarization and amplitude DOF [19].
Beyond its probabilistic indeterminacy, the E in (1) has other quantum-like attributes -it has the same form as a quantum state superposition and can be called a pure state in the same sense, more precisely a two-party state living in two vector spaces at once, polarization space forx andŷ, and infinitely continuous stochastic function space for E x and E y .
The Bell inequality most commonly used for correlation tests is due to Clauser, Horne, Shimony and Holt (CHSH) [12]. It deals with correlations between two different DOF when each is two-dimensional. The Schmidt Theorem of analytic function theory [20] ensures two-dimensionality, by guaranteeing that among the infinitely many dimensions available to the amplitudes in (1), only two dimensions are active. This is a consequence arising just from the fact that the partner polarization vectorsx andŷ live in a two-dimensional space.
For convenience, we introduce e, the field normalized to the intensity I = E x E x + E y E y : where now e · e = e x e x + e y e y = 1. For some simplification in writing, we will use Dirac notation for the vectors without, of course, imparting any quantum character to the fields. The unit polarization vectorsx andŷ will be renamed asx → |u 1 andŷ → |u 2 and the unit amplitudes will be rewritten e x → |f 1 and e y → |f 2 . If desired, the Dirac notation can be discarded at any point and the vector signs and hats re-installed. For the case of unpolarized light we have u 1 |u 2 = 0 and f 1 |f 2 = 0. Unit projectors in the two spaces take the form 1 = |u 1 u 1 | + |u 2 u 2 | and 1 = |f 1 f 1 | + |f 2 f 2 |. In this notation, and in the original notation for comparison, the field takes the form In this notation the field actually looks like what it is, a two-party superposition of products in independent vector spaces, i.e., an entangled two-party state (actually a Bell state). Here the two parties are the independent polarization and amplitude DOF. The notation for a CHSH correlation coefficient C(a, b) implies that arbitrary rotations of the unit vectors |u j and |f k (j, k = 1, 2) through angles a and b can be managed independently in the two spaces. An arbitrary rotation through angle a of the polarization vectors |u 1 and |u 2 takes the form |u a 1 = cos a|u 1 − sin a|u 2 and |u a 2 = sin a|u 1 + cos a|u 2 .
For function space rotations we have |f b 1 and |f b 2 defined similarly: where the rotation angles a and b are unrelated. Next, the correlation between the polarization (u) and function (f ) degrees of freedom is given by the standard average where Z is shorthand for the difference projection: Z u (a) ≡ |u a 1 u a 1 | − |u a 2 u a 2 |, analogous to a σ z spin operation. C(a, b) is thus a combination of four joint projections such as: This is all classical and all of the correlation projections P jk (a, b) with j, k = 1, 2 have familiar roles in classical optical polarization theory [18]. Gisin [5] observed that any quantum state entangled in the same way as the classical pure state (2) will lead to violation of the CHSH inequality, which takes the form B ≤ 2, where The same result will be found here, as one uses only DOF independence and properties of positive functions and normed vectors to arrive at it (see details in Suppl. Matls. [16]). We note again that the issue of entanglement itself is pertinent to the discussion, but the usefulness of entanglement as a resource for particular applications is not. Thus we have reached the main goal of our theoretical background sketch. This was to demonstrate the existence of a purely classical field theory that can exhibit a violation of the CHSH Bell inequality. Experimental Testing: The remaining task is to show that experimental observation confirms this theoretical prediction, in effect shifting one's interpretation of tests of the quantum-classical border by showing that, along with quantum fields, classical fields conforming to the Shimony Bell-test criteria are capable of Bell violation. In order to make such a demonstration, a classical field source must be used. This means one producing a field that is quantum mechanical (since we believe all light fileds are intrinsically quantum), but a field whose quantum statistics are not distinguishable from classical statistics. This is only necessary up to second order in the field because the CHSH procedure engages no higher order statistics. Such sources are easily available. Since the earliest testing of laser light it has been known that a laser operated below threshold has statistical character not distinguishable from classical thermal statistics. So in our experiments we have used a broadband laser diode operated below threshold. Our experiment repeatedly records the correlation function C(a, b) defined in (6) for four different angles in order to construct the value of the Bell parameter B. This is done through measurements of the joint projections P jk (a, b). We will describe explicitly only the recording of P 11 (a, b), identified in (7), but the others are done similarly in an obvious way. In the classical context that we are examining, the optical field is macroscopic and correlation detection is essentially calorimetric (i.e., not requiring or employing individual photon recognition). Polarization Tomography: The first step is to determine tomographically the polarization state of the test field. A polarization tomography setup is shown in Fig. 1. Using a polarizing beam splitter and half and quarter wave plates to project onto circular and diagonal bases, the Stokes parameters (S 1 , S 2 , S 3 ), relative to S 0 = 1, are found to be (−0.0827, −0.0920, −0.0158), providing a small non-zero degree of polarization equal to 0.125. This departure from zero requires a slight modification of the theory presented above (see the Supplemental Material) and reduces the maximum possible value of B able to be achieved for our specific experimental field to B = 2.817, below but close to B = 2 √ 2 = 2.828..., the theoretical maximum for completely unpolarized light. Experimental Bell Test: The experimental test has two major components, as shown in Fig. 1: a source of the light to be measured, and a Mach-Zehnder (MZ) interferometer. The source utilizes a 780 nm laser diode, operated in the multi-mode region below threshold, giving it a short coherence length on the order of 1 mm. The beam is assumed statistically ergodic, stable and stationary, as commonly delivered from such a multimode below-threshold diode. It is incident on a 50:50 beam splitter and recombined on a polarizing beam splitter after adequate delay so that the light to be studied is an incoherent mix of horizontal and vertical polarizations before being sent to the measurement area via a single mode fiber. A half wave plate in one arm controls the relative power, and thus the degree of polarization (DOP). Quarter and half wave plates help correct for polarization changes introduced by the fiber.
In Fig. 1 the partially polarized beam entering the MZ is separated by a 50:50 beam splitter into primary test beam |E and auxiliary beam |Ē . The two beams inherit the same statistical properties from their mother beam and thus both can be expressed as in Eq. (3), with intensities I andĪ. The phase of the auxiliary beam |Ē is shifted by an unimportant factor i at the beam splitter.
To determine the joint projection P 11 (a, b) of the test beam |E , the first step is to project the field to obtain |E a 1 ≡ |u a 1 u a 1 |E . This can be realized by the polarizer labelled a on the bottom arm of the MZ. The transmitted beam retains both |f components in function space: where I a 1 is the intensity, and c 11 and c 12 are normalized amplitude coefficients with |c 11 | 2 + |c 12 | 2 = 1. Here c 11 relates to P 11 in an obvious way: P 11 (a, b) = I a 1 |c 11 | 2 /I. One sees that the intensities I and I a 1 can be measured directly, but not the coefficient c 11 .
For P 11 (a, b) our aim is to produce a field that combines a projection onto |f b 1 in function-space with the |u a 1 projection in polarization space. The challenge of overcoming the lack of polarizers for projection of a non-deterministic field onto an arbitrary direction in its independent infinite-dimensional function space is managed by a "stripping" technique [16] applied to the auxiliaryĒ field in the left arm. We passĒ through a polarizer rotated from the initial |u 1 − |u 2 basis by a specially chosen angle s, so that the statistical component |f b 2 is stripped off. The transmitted beam |Ē s 1 then has only the |f b 1 component, as desired: |Ē s 1 = i Īs 1 |u s 1 |f b 1 . HereĪ s 1 is the corresponding intensity and the special stripping angle s is given by tan s = (κ 1 /κ 2 ) tan b (see [3,16]).
The function-space-oriented beam |Ē s 1 is then sent through another polarizer a to become |Ē a 1 = |u a 1 u a 1 |Ē a 1 = i Īa 1 |u a 1 |f b 1 , whereĪ a 1 is the corresponding intensity. Finally, the beams |E a 1 and |Ē a 1 are combined by a 50:50 beam splitter which yields the outcome beam |E T 1 = (|Ē a 1 + i|E a 1 )/ √ 2. The total intensity I T 1 of this outcome beam can be easily expressed in terms of the needed coefficient c 11 .
Some simple arithmetic will immediately provide the joint projection P 11 (a, b) in terms of various measurable intensities: Other P jk (a, b) values can be obtained similarly by rotations of polarizers a and s. To make our measurements, polarizers a were simultaneously rotated using motorized mounts, while the third polarizer s was fixed at different values in a sequence of runs. Results: For each angle, measurements were made at detector D1 for the total intensity I T , and the separate intensities from each arm I a andĪ a by closing the shutters S alternately. In this way, the measurements of the polarization space and statistical amplitude space are carried out separately. From these measurements the needed correlations C(a, b) were determined and Eq. (8) used to evaluate the CHSH parameter B. Fig. 2 shows C(a, b) obtained by measuring the joint projections P jk (a, b) for a complete rotation of polarizer a, with different curves corresponding to b (and thus s) fixed at different values. It is apparent from the near-identity of the curves that, to good approximation, the correlations are a function of the difference in angles, i.e. C(a, b) = C(a − b). The maximum value for B can then be found straightforwardly from any one of the curves in Fig. 2. Among them the smallest and largest values of B (obtained for curves 1 and 4), are 2.548 ± 0.004 and 2.679 ± 0.007.
To be careful, we note that in our experiments the field was almost but not quite completely unpolarized, thus not quite the same field sketched in the Background Theory paragraphs. Thus we couldn't expect to get the maximum quantum result B = 2 √ 2 = 2.828... for the Bell parameter, but the values achieved also present a strong violation. The background theory is mildly more complicated for partially polarized rather than unpolarized light, but when worked out for the degree of polarization of our light beams (see [3] and the Suppl. Matl. [16]) it supports the values we observed. Summary: In summary, we first sketched the purely classical theory of optical beam fields (1) that satisfy the Bell-test criteria of Shimony [1,16]. Their bipartite pure state form shows the entanglement of their two independent degrees of freedom [22]. The classical theory defines them as dynamically probabilistic fields, meaning that individual field measurements yield values that cannot be predicted except in an average sense, which is another feature shared with quantum systems but also associated for more than 50 years with the well-understood and well-tested optical theory of partial coherence [18]. Our theoretical sketch for the simplest case, unpolarized light, indicated that such fields or states are predicted to possess a range of correlation strengths equal to that of two-party quantum systems, that is, outside the bound B ≤ 2 of the CHSH Bell inequality and potentially as great as B = 2 √ 2. In our experimental test we used light whose statistical behavior (field second-order statistics) is indistinguishable from classical, viz., the light from a broadband laser diode operating below threshold. Our detections of whole-beam intensity are free of the heralding requirements familiar in paired-photon CHSH experiments. Repeated tests confirmed that such a field can strongly violate the CHSH Bell inequality and can attain Bell-violating levels of correlation similar to those found in tests of maximally entangled quantum systems.
One naturally asks, how are these results possible? We know that a field with classically random statistics is a local real field, and we also know that Bell inequalities prevent local physics from containing correlations as strong as what quantum states provide. But the experimental results directly contradict this. The resolution of the apparent contradiction is not complicated but does mandate a shift in the conventional understanding of the role of Bell inequalities, particularly as markers of a classical-quantum border. Bell himself came close to addressing this point. He pointed out [2] that even adding classical indeterminism still wouldn't be enough for any type of hidden variable system to overcome the restriction imposed by his inequalities. This is correct as far as it goes but fails to engage the point that local fields can be statistically classical and exhibit entanglement at the same time. For the fields under study, the entanglement is a strong correlation that is intrinsically present between the amplitude and polarization degrees of freedom, and it is embedded in the field from the start (as it also is embedded ab initio in any quantum states that violate a Bell inequality). The possibility of such pre-existing structural correlation is bypassed in a CHSH derivation. Thus one sees that Bell violation has less to do with quantum theory than previously thought, but everything to do with entanglement.

Shimony's Bell-Test Conditions:
In his extended analyses [1] of Bell inequalities and their testing, Shimony recognized that in order to deserve serious attention, an alternative non-quantum theory entering what is considered a quantum domain (in our case, the domain of Bell test violation) needs to embrace in some way aspects of Nature that appear completely random, i.e., purely probabilistic. These aspects are dealt with by quantum theory in well known ways, and this is the reason Bell once raised the issue of classical indeterminism [2], but without going as far as classical entanglement. Shimony summarized these considerations by naming three key features, all of which should be considered, in his words, "... as established parts of physical theory: (I) In any state of a physical system S there are some eventualities which have indefinite truth values. (II) If an operation is performed which forces an eventuality with indefinite truth value to achieve definiteness ... the outcome is a matter of chance. (III) There are 'entangled systems' (in Schrödinger's phrase) which have the property that they constitute a composite system in a pure state, while neither of them separately is in a pure state." Here by eventualities Shimony means measurement outcomes. We have identified the electric field as it is dealt with in the standard classical theory of partial coherence and partial polarization as a physical system satisfying all three conditions. Partially Polarized Fields: Classical statistical light of any degree of polarization can be treated exactly as the unpolarized light in the text, with a small change. One needs to insert the parameters that allow the two orthogonal components provided by the Schmidt decomposition to have different intensities. These parameters, κ 1 and κ 2 , are the (real) Schmidt eigenvalues. They satisfy κ 2 1 + κ 2 2 = 1, and both equal 1/ √ 2 in the completely unpolarized case treated in the text. Any intensity-normalized partially polarized field can then be written as in text Eqn. (3), but with κ 1 and κ 2 attached [3]: It is clear that the field is entangled between polarization and amplitude unless one of the κ's is zero, in which case |e is plainly separable. All of the formulas for unpolarized light will change, but only to the extent that the presence of the κ's requires. The conventional degree of polarization P is given in terms of the κ's by P = |κ 2 1 − κ 2 2 | [4], and this can be used to find the Schmidt coefficients. For the experimental classical statistical optical light field we obtained κ 1 , κ 2 = 0.750, 0.661. Another interesting formula is the full result for the partially polarized Bell parameter: One sees two uncomplicated limits: when either κ is zero (no entanglement) no result higher than B = 2 can be achieved, and when κ 1 = κ 2 = 1/ √ 2 (maximal entanglement) the Tsirelson bound can be reached: B = 2 √ 2. Also, if one follows Gisin's approach [5] by choosing the rotation angles as a = 0, a ′ = π/4, and cos2b = −cos2b ′ = 1/ 1 + 4|κ 1 κ 2 | 2 , the Bell parameter becomes Apparently, there will be a Bell violation (B¿2) as long as κ 1 κ 2 = 0, i.e., when there is non-zero entanglement. Basis stripping, rotation, and projection in amplitude function space: To observe correlation, it is essential to be able to access and measure both polarization and amplitude function degrees of freedom. Unfortunately, unlike the polarization degree of freedom, there is no systematic technology working directly in the infinite-dimensional amplitude function space that can project a non-deterministic field onto an arbitrary basis |f b 1 in that space. This requires an innovation using an indirect measurement such as incorporated in the experimental setup sketched in Fig.  1 of the text. In our setup we employ an auxiliary beam that contains only the |f b 1 basis to interfere with the primary test beam (which in general contains both |f b 1 and |f b 2 ). By interference measurements one is able to obtain the |f b 1 information from the test beam, and finally determine the needed c 11 value given in equation (9) of the main text. This section shows specifically how to "strip" off and rotate a basis in the statistical amplitude function space of the auxiliary field |Ē , which shares the same statistical properties as the primary test beam. For generality, we take the auxiliary beam as initially in the form of Eq. (11) with arbitrary κ 1 , κ 2 . Such a beam can always be rewritten in the rotated amplitude function space basis |f b 1 , |f b 2 , i.e, |Ē = Ī (κ 1 cos b|u 1 − κ 2 sin b|u 2 ) |f b 1 + (κ 1 sin b|u 1 + κ 2 cos b|u 2 ) |f b 2 .
One notes from the second term of the equation that a properly chosen polarizer that blocks completely the polarization-space component κ 1 sin b|u 1 + κ 2 cos b|u 2 will effectively strip off the amplitude function space basis vector |f b 2 . Such a stripping polarizer |u s 1 is oriented with respect to the polarization space basis |u 1 , |u 2 by a specific angle we have called s, i.e., |u s 1 = cos s|u 1 − sin s|u 2 .
Then the stripping condition is directly given as u s 1 | (κ 1 sin b|u 1 + κ 2 cos b|u 2 ) = 0, which specifies the rotation angle s according to so s is determined by the values of κ 1 and κ 2 for any choice of rotation angle b in the amplitude function space. As a result of this stripping polarizer, the beam (14) becomes where the amplitude function space component |f b 2 of the transmitted beam is completely stripped off. One notes immediately that arbitrary rotations of the function space basis |f b 1 (i.e., arbitrary change of angle b) for the auxiliary beam can be effectively realized by the change of angle s through Eq. (17). Consequently, as described in the main text, by further interference of this auxiliary beam with the primary test beam, one is able to project the test beam onto any function space basis |f b 1 with an arbitrary angle b. This is exactly how we make measurements in the amplitude function space of the statistical light beam.