Milburn and Shrapnel (Entropy 20, 687 (2018)) modeled a causal intervention — the operation that gives causal models empirical meaning — as a physical process consisting of a measurement followed by conditional control, and showed in two specific models (a classical free particle localized by a Gaussian momentum measurement; a quantum harmonic oscillator with a binary (left/right) position measurement and parity feedback) that Pearl-style surgical do-interventions are physically impossible: accuracy is limited by the low entropy of the apparatus, interventions cost work, and the work extractable after an intervention is bounded by the erasure cost of the intervention memory. Two gaps were left open and define this problem: general constraints on physically realizable interventions from thermodynamics and quantum mechanics, valid for broad classes of systems and causal models rather than two examples, have not been derived; and it remains conjectural whether causal asymmetry derives from the act of intervention by a local low-entropy agent, or whether causal structure is an agent-independent property contingent on local entropy gradients between cause and effect. The problem asks for the general constraint theory, in classical and quantum settings, and for the adjudication of the entropy-gradient conjecture.

Background

Causal models are the standard formal device for reasoning about cause and effect: a structural causal model represents each variable as a function of its direct causes (its parents) plus noise, and supports predictions about what would happen if a variable were forcibly set to a chosen value. Such a forced assignment is an intervention, written do(X = x) in Pearl's do-calculus; an atomic or surgical intervention cuts the variable off from its parents while leaving all other mechanisms untouched. Interventions are what give causal models empirical content, since they formalize the experiments from which causal structure is learned, and quantum causal models generalize this apparatus to quantum systems, where the intervened objects are quantum operations (the quantum analogues of stochastic maps, acting on the system's density operator). The physical status of the intervention operation itself, however, is not part of the formal theory. Milburn and Shrapnel (2018) filled part of this gap by modeling an intervention as a measurement followed by a conditional control operation and analyzing it with stochastic thermodynamics, the extension of thermodynamics to fluctuating, individually measured open systems. There, measurement-and-feedback protocols obey generalized second laws (Sagawa and Ueda 2008; Ito and Sagawa 2013): the work extractable using feedback is bounded by the information acquired, and erasing the resulting memory costs at least kBTln⁡2k_B T\ln 2 per bit (Landauer erasure). In their two worked examples — a classical free particle whose momentum is localized by a Gaussian measurement plus an impulsive kick, and a quantum harmonic oscillator driven by a binary (left/right) position measurement followed by a parity feedback (a conditional unitary implementing the oscillator's parity operation) — they showed that perfect surgical interventions are unattainable: the accuracy of an intervention is limited by the entropy of the apparatus (which must be prepared below the entropy of the system), every intervention costs work, and the work extractable after an intervention is bounded by the erasure cost of the intervention memory; in the quantum case entanglement with the apparatus and the uncertainty principle enforce the same limits. Two questions were left open and define the present problem. First, the analysis is model-specific: general constraints on physically realizable interventions — theorems valid for broad classes of classical and quantum systems and causal models, rather than two illustrations — were not derived. Second, the discussion poses the interpretive question the examples illuminate: is the source of causal asymmetry (the directionality of cause-effect relations in a world whose microscopic laws are time-reversal invariant) the act of intervention itself, which requires a local low-entropy agent? Or is causal structure an agent-independent property of the world, contingent on a local entropy gradient between cause and effect that exists prior to any intervention? The authors state the entropy-gradient claim explicitly as a conjecture.

Problem Statement

Deliver a physical theory of causal interventions that closes the two gaps explicitly left open by Milburn and Shrapnel (2018). First, derive general constraints from thermodynamics and quantum mechanics on physically realizable interventions — where an intervention is a physical operation consisting of a measurement of the target variable followed by a conditional control operation — valid for broad classes of classical and quantum systems and causal models, not only the Gaussian-momentum-measurement model of a free particle and the binary (left/right) position measurement with parity feedback model of a harmonic oscillator. The general theory should characterize which ideal interventions, including Pearl surgical do-interventions, can be approximated by physical operations and at what minimal thermodynamic cost (work, heat, and memory-erasure cost), and how quantum features — measurement invasiveness, uncertainty relations, entanglement — modify these constraints relative to the classical case; it should recover the two source models and the known feedback second laws as special cases. Second, adjudicate the paper conjecture: does the directionality of causal relations derive from the act of intervention by a local low-entropy agent (agent-dependent causation), or is causal structure an agent-independent property contingent on local entropy gradients — specifically, must causal relations built on reversible physical interactions require a local entropy gradient between cause and effect prior to any intervention? Both questions are open in the source; a complete answer consists of the general theorems (or a no-go theorem showing the claimed generality is unattainable), together with a proof, refutation, or sharp delimitation of the entropy-gradient conjecture. For the first part, the minimum required generality is one general class of classical systems together with its quantum generalization, with the characterization of which ideal interventions in that class are approximable and their minimal work, heat, and memory-erasure costs supplied as part of the same result; recovering the two source models and the known feedback second laws is a consistency check on that result, not a separately required deliverable.

References

  1. G. J. Milburn and S. Shrapnel, Classical and Quantum Causal Interventions, Entropy 20(9), 687 (2018). DOI: 10.3390/e20090687, https://doi.org/10.3390/e20090687 (preprint: arXiv:1809.03191, https://arxiv.org/abs/1809.03191)
  2. J. Pearl, Causality: Models, Reasoning, and Inference, 2nd ed., Cambridge University Press (2009). DOI: 10.1017/CBO9780511803161, https://doi.org/10.1017/CBO9780511803161
  3. T. Sagawa and M. Ueda, Second Law of Thermodynamics with Discrete Quantum Feedback Control, Physical Review Letters 100, 080403 (2008). DOI: 10.1103/PhysRevLett.100.080403, https://doi.org/10.1103/PhysRevLett.100.080403
  4. S. Ito and T. Sagawa, Information Thermodynamics on Causal Networks, Physical Review Letters 111, 180603 (2013). DOI: 10.1103/PhysRevLett.111.180603, https://doi.org/10.1103/PhysRevLett.111.180603
  5. G. E. Crooks and S. Still, Marginal and conditional second laws of thermodynamics, EPL 125, 40005 (2019). DOI: 10.1209/0295-5075/125/40005, https://doi.org/10.1209/0295-5075/125/40005
  6. P. Strasberg and A. Winter, Stochastic thermodynamics with arbitrary interventions, Physical Review E 100, 022135 (2019). DOI: 10.1103/PhysRevE.100.022135, https://doi.org/10.1103/PhysRevE.100.022135 (preprint: arXiv:1905.07990, https://arxiv.org/abs/1905.07990)
  7. P. Strasberg, Thermodynamics of Quantum Causal Models: An Inclusive, Hamiltonian Approach, Quantum 4, 240 (2020). DOI: 10.22331/q-2020-03-02-240, https://doi.org/10.22331/q-2020-03-02-240 (preprint: arXiv:1911.01730, https://arxiv.org/abs/1911.01730)
  8. J. Thompson, A. J. P. Garner, J. R. Mahoney, J. P. Crutchfield, V. Vedral, and M. Gu, Causal Asymmetry in a Quantum World, Physical Review X 8, 031013 (2018). DOI: 10.1103/PhysRevX.8.031013, https://doi.org/10.1103/PhysRevX.8.031013 (preprint: arXiv:1712.02368, https://arxiv.org/abs/1712.02368)
  9. C. Rovelli, Agency in Physics (2020). arXiv:2007.05300, https://arxiv.org/abs/2007.05300
  10. L. Escolà-Farràs and D. Braun, Quantifying causal influence in quantum mechanics, Physical Review A 106, 062415 (2022). DOI: 10.1103/PhysRevA.106.062415, https://doi.org/10.1103/PhysRevA.106.062415 (preprint: arXiv:2105.08197, https://arxiv.org/abs/2105.08197)
  11. M. A. Javidian, V. Aggarwal, F. Bao, and Z. Jacob, Quantum Entropic Causal Inference, in Quantum Information and Measurement VI (QIM 2021), paper F2C.3. DOI: 10.1364/qim.2021.f2c.3, https://doi.org/10.1364/qim.2021.f2c.3 (preprint: arXiv:2102.11764, https://arxiv.org/abs/2102.11764)
  12. L. Hutter, R. Chaves, R. Nery, G. Moreno, and D. J. Brod, Quantifying quantum causal influences, Physical Review A 108, 022222 (2023). DOI: 10.1103/PhysRevA.108.022222, https://doi.org/10.1103/PhysRevA.108.022222 (preprint: arXiv:2210.04306, https://arxiv.org/abs/2210.04306)
  13. S. Kechrimparis, M. Gu, and H. Kwon, Causal asymmetry of classical and quantum autonomous agents, npj Quantum Information (2026). DOI: 10.1038/s41534-026-01335-7, https://doi.org/10.1038/s41534-026-01335-7 (preprint: arXiv:2309.13572, https://arxiv.org/abs/2309.13572)
  14. D. Rancati, M. Welling, and F. Locatello, Reconciling Causality and Non-Equilibrium Thermodynamics with Hamiltonian Causal Models (2026). arXiv:2606.04822, https://arxiv.org/abs/2606.04822
  15. M. V. Churilov, The Locality Gap: A Thermodynamic Law for Objective Facts (2026). arXiv:2608.05753, https://arxiv.org/abs/2608.05753
  16. G. J. Milburn and S. Basiri-Esfahani, The physics of learning machines, Contemporary Physics 63, 34–60 (2022). DOI: 10.1080/00107514.2022.2135672, https://doi.org/10.1080/00107514.2022.2135672 (preprint: arXiv:2209.11954, https://arxiv.org/abs/2209.11954)
  17. N. Ares, A. N. Pearson, and G. A. D. Briggs, Eight Oxford Questions: Quantum Mechanics Under a New Light, in Do Wave Functions Jump? (Fundamental Theories of Physics, Vol. 198, Springer, 2020), pp. 361–384. DOI: 10.1007/978-3-030-46777-7_25, https://doi.org/10.1007/978-3-030-46777-7_25