Info
This seminar covered the results in Machine Learning Catalysis of quantum tunneling.
Abstract: The so-called second quantum technological revolution is evolving at a rapid pace and promises significant impacts not only on science but also within the industrial sector. Progress in this field critically relies on efficient methods for controlling the quantum properties of systems and their dynamics. In this talk we are going to give two illustrative examples of how, using a key algorithm in modern machine learning, automatic differentiation, we can control the properties of interest of a quantum system. Our initial case study focuses on the control of the tunnelling probability of particles in a two-mode system. We show that when the quantum system is coupled to an ancilla, one can learn the optimal ancillary component and the optimal coupling, such that the tunnelling probability/time can be controlled. The subsequent example addresses the mitigation of decoherence within a quantum system with noise. Employing a similar methodology, we show how we can learn an ancillary system and its corresponding noise parameters to counteract and diminish the impact of system noise.
by Prof. Fabio Anselmi (Department of Mathematics, Informatics, and Geo-science, University of Trieste)
Key concepts
- Autodiff: computes exact derivatives of a program by applying the chain rule to its elementary operations. Reverse mode gives the gradient of a scalar loss with respect to all parameters at a cost of a small constant multiple of one forward evaluation, which is what makes optimising many control and coupling parameters of a simulated quantum system feasible.
- Quantum noise: uncontrolled coupling to an environment, which makes the reduced dynamics of the system non-unitary (decoherence, dissipation), modelled by completely positive trace preserving maps or a Lindblad master equation.
- Quantum tunnelling: here, the coherent transfer of particles between the two modes (wells) through the barrier, driven by the term below; the tunnelling probability and time are the quantities being controlled.
- Activation Layers: elementwise nonlinear functions (ReLU, tanh, sigmoid) between the linear layers of a neural network; without them a deep network collapses to a single linear map.
- ancillary system: an auxiliary system coupled to the system through . Tracing out gives an effective, generally non-unitary, dynamics on , and the talk learns and so that this effective dynamics has the desired tunnelling or noise properties.
- Bose-Hubbard Model, two-mode version: (Particle Interaction + Tunnelling + Asymmetry), with the Schwinger operators , . For bosons this is a spin system: is the on-site interaction, the hopping between the wells, and an energy offset between them.
- Basin of attraction (machine learning): the set of initial parameters from which the optimiser (gradient descent) converges to a given local minimum of the loss.
- Noon state: , which acquires a relative phase and so reaches Heisenberg limited phase sensitivity , but loses its coherence entirely if a single particle is lost.
- Stochastic resonance: in a nonlinear (typically bistable) system driven by a weak periodic signal, adding noise of the right intensity enhances the response to the signal, the optimum being where the noise-activated (Kramers) hopping rate matches the signal frequency. It is an example of noise helping rather than hurting, as in the second part of the talk.
- Characterisation of positive maps vs Completely Positive maps: a linear map is positive if for all , and completely positive if is positive for all . By Choi’s theorem, on is completely positive iff its Choi matrix is positive semidefinite, iff it has a Kraus form . The transpose is positive but not completely positive. Physical channels must be completely positive since the system may be entangled with another one, and positive but not completely positive maps act as entanglement witnesses (the Peres-Horodecki criterion uses the partial transpose).