God Has Not Been Playing Dice All Along: Einstein's Determinism as a Statement about Operational Structure, and Probability as a Feature of Its Description
T.O.
Einstein’s conviction that nature does not play dice is usually reported as a dispute that he lost. This paper separates what he said from what is attached to it, and asks in which sense he was right. Part I compiles his statements, from the letter to Born of December 1926 through the article of 1948 to the last appendix of The Meaning of Relativity (1955), each with the status of its source. Part II comments on them in the vocabulary of Operatiology (Japanese: Sousa Genron, 操作原論), an independent line of work in which a Tier-2 layer (states, operations, classes of states) is distinguished from a Tier-3 layer of description (spacetime, Hilbert space, measure, probability, the time parameter, outcomes). We argue that his diagnosis, that probability is not a primitive property of nature but a feature of an incomplete and indirect description, was right in direction, within that vocabulary and under bridge premises, stated as such in the text, that connect an elementary structure to quantum mechanics; and we show in the elementary case how a deterministic law yields probabilities once the situation is recorded by a class of states and not by a state, and a class is carried into more than one class. We argue that his limits are two. The first is historical: he saw the limit of the continuum and called for a purely algebraic theory, but stated that nobody knew how to obtain its basis. The second is structural: he tied separability, a premise stated in terms of spatially separated regions and therefore belonging to the description layer, to physical reality; this rests on a further bridge premise about the form of his own formulation. The paper does not refute Bell’s theorem and does not decide whether the quantum state describes nature completely. Numerical verification. The Python code below (Code A1, Appendix A of the paper) checks the mathematical facts on a composite system that the separability part relies on. For random pure states of 2×2, 2×3 and 3×3 systems it checks that a state is a product state if and only if its Schmidt rank is one, if and only if its reduced state is pure, if and only if its best overlap with a product vector is one. For the Bell state it checks that the reduced state is maximally mixed and that no pure state of one part reproduces the statistics of all observables on that part. It also checks that the Bell state and the product state |+>|+> have the same marginals and different joint distributions. import numpy as np # Code A1: separability of a pure bipartite state, reduced states, and marginals. rng = np.random.default_rng(20261031) def schmidt(psi, dA, dB): s = np.linalg.svd(psi.reshape(dA, dB), compute_uv=False) return s # Schmidt coefficients sqrt(lambda_i), descending def reduced_A(psi, dA, dB): M = psi.reshape(dA, dB) return M @ M.conj().T def purity(rho): return float(np.real(np.trace(rho @ rho))) def max_product_overlap(psi, dA, dB): # max over unit vectors a, b of |<a (x) b | psi>|^2 equals the largest squared Schmidt coefficient. return float(schmidt(psi, dA, dB)[0] ** 2) ok = True # (1) Random states: product iff Schmidt rank 1 iff reduced state pure. for (dA, dB) in [(2, 2), (2, 3), (3, 3)]: for trial in range(200): if trial % 2 == 0: # a product state a = rng.normal(size=dA) + 1j * rng.normal(size=dA); a /= np.linalg.norm(a) b = rng.normal(size=dB) + 1j * rng.normal(size=dB); b /= np.linalg.norm(b) psi = np.kron(a, b) else: # a generic state psi = rng.normal(size=dA * dB) + 1j * rng.normal(size=dA * dB); psi /= np.linalg.norm(psi) s = schmidt(psi, dA, dB) rank = int(np.sum(s > 1e-10)) pur = purity(reduced_A(psi, dA, dB)) ovl = max_product_overlap(psi, dA, dB) is_prod = (rank == 1) ok &= (is_prod == (abs(pur - 1.0) < 1e-10)) and (is_prod == (abs(ovl - 1.0) < 1e-10)) if trial % 2 == 0: ok &= is_prod print("(1) product <=> Schmidt rank 1 <=> pure reduced state <=> overlap 1:", ok) # (2) Bell state: reduced state is maximally mixed, best product overlap is 1/2. bell = np.array([1, 0, 0, 1], dtype=complex) / np.sqrt(2) rhoA = reduced_A(bell, 2, 2) print("(2) Bell reduced state:\n", np.round(rhoA.real, 6)) print(" purity of reduced state: %.6f ; best product overlap: %.6f" % (purity(rhoA), max_product_overlap(bell, 2, 2))) ok &= abs(purity(rhoA) - 0.5) < 1e-12 and abs(max_product_overlap(bell, 2, 2) - 0.5) < 1e-12 # (3) No unit vector phi on A reproduces the statistics of all observables on A: # expectation of X in |phi> equals tr(rhoA X) for all Hermitian X iff rhoA = |phi><phi|. best = 0.0 for _ in range(20000): phi = rng.normal(size=2) + 1j * rng.normal(size=2); phi /= np.linalg.norm(phi) d = np.linalg.norm(np.outer(phi, phi.conj()) - rhoA) # Frobenius distance of density matrices best = d if best == 0.0 else min(best, d) print("(3) smallest distance between rhoA and any sampled pure state: %.4f (positive; exact value %.4f)" % (best, np.sqrt(0.5))) ok &= best > 0.5 # (4) Same marginals, different joint distributions (computational basis): plus = np.array([1, 1], dtype=complex) / np.sqrt(2) prod = np.kron(plus, plus) def joint(psi): return (np.abs(psi) ** 2).reshape(2, 2) for name, psi in [("Bell", bell), ("|+>|+>", prod)]: J = joint(psi) print("(4) %-7s joint:" % name, np.round(J.flatten(), 4), "marginal A:", np.round(J.sum(1), 4), "marginal B:", np.round(J.sum(0), 4)) Jb, Jp = joint(bell), joint(prod) same_marg = np.allclose(Jb.sum(1), Jp.sum(1)) and np.allclose(Jb.sum(0), Jp.sum(0)) diff_joint = not np.allclose(Jb, Jp) print(" same marginals:", same_marg, "; different joint distributions:", diff_joint) ok &= same_marg and diff_joint print("ALL CHECKS PASSED:", bool(ok))
•Jan 1, 2026•0 citations