PASCAL - Pattern Analysis, Statistical Modelling and Computational Learning

Relating the Thermodynamic Arrow of Time to the Causal Arrow.
Armen Allahverdyan and Dominik Janzing
Journal of Statistical Mechanics Volume P04001, pp. 1-21, 2008.


Consider a Hamiltonian system that consists of a slow subsystem S and a fast subsystem F. The autonomous dynamics of S is driven by an effective Hamiltonian, but its thermodynamics is unexpected. We show that a well-defined thermodynamic arrow of time (second law) emerges for S whenever there is a well-defined causal arrow from S to F and the back-action is negligible. This is because the back-action of F on S is described by a non-globally Hamiltonian Born–Oppenheimer term that violates the Liouville theorem, and makes the second law inapplicable to S. If S and F are mixing, under the causal arrow condition they are described by microcanonical distributions P(S) and P(S|F). Their structure supports a causal inference principle proposed recently in machine learning.

EPrint Type:Article
Project Keyword:Project Keyword UNSPECIFIED
Subjects:Computational, Information-Theoretic Learning with Statistics
ID Code:5373
Deposited By:Dominik Janzing
Deposited On:31 March 2009