A new paper from the School of Emergent Priorities shows that a random matrix losing its couplings at random forgets its eigenvector structure at a steady rate while it stays dense, and ever more slowly once it has thinned to a scatter of small clusters.
A random matrix losing its couplings one at a time keeps its structure longest in its isolated fragments
The School of Emergent Priorities has long asked what an institution keeps of the commitments it lets lapse. With Lapsed Couplings: Eigenvector Memory and Spectral Ageing in Randomly Diluted Wigner Matrices, a new paper by Dr Marek Solheim, Lecturer and Deputy Convenor, Horizon Register, the School puts the same question to a matrix.
I have spent a long time asking what an organisation retains of the things it has stopped doing. It was a relief to put the question to an object that answers in eigenvalues, and a surprise to find the answer so familiar.
— Dr Marek Solheim, Lecturer and Deputy Convenor, Horizon Register, School of Emergent Priorities
The paper follows a large random matrix whose connections are switched off one by one, each at a time of its own choosing, and asks how much of its structure at one moment is still there at the next. It aligns a careful decomposition, a correspondence with a classical model of random motion and an extended programme of computation on matrices of two thousand rows into a single account of forgetting. While most connections remain, the matrix forgets at a steady pace, whatever its history. Once it has thinned to a scatter of small pieces, it forgets more slowly the older it gets, because each surviving fragment keeps its shape until one of its own connections goes.
The University regards the work as a confident extension of the School’s anticipatory-capability agenda into mathematical physics, and a sign of how far that agenda now travels from the committee room. The study was supported by the Anticipatory Capability Fund.
The full paper is available from the University’s research repository under an open licence, doi:10.5555/slop.zkb0qc.