Skip to main content
Back to timeline
arXivSource publication:

Topological quantum reservoir computing: protected edge modes hold memory while bulk dynamics compute

Synopsis

Topological quantum reservoir computing moves quantum reservoir computing onto a one-dimensional Kitaev chain: topologically protected Majorana edge modes retain information over long times while faster bulk dynamics process the input, and only a linear readout is trained. Without interactions, edge retention grows exponentially with chain length while local write and read amplitudes are exponentially suppressed; a matched ancillary fermion makes the edge a programmable memory, with stronger writing shortening retention. Interactions let information load autonomously from bulk to edge at the cost of reduced retention, and this channel improves long-timescale temporal integration.

Source-provided article image: Topological Quantum Reservoir Computing
arXiv

Interpretation

Quantum reservoir computing had been implemented on spin systems, photonic and bosonic systems, superconducting circuits and neutral-atom arrays, where the system's intrinsic evolution maps temporal inputs into high-dimensional features and only a linear readout is trained. Topological quantum reservoir computing splits storage and processing across two kinds of degrees of freedom in one many-body system: protected boundary modes hold information while extended bulk modes process it quickly. Earlier quantum reservoir computing platforms did not use topologically protected boundary modes as long-term memory; storage and processing fell to the same set of modes. In the topological phase of a one-dimensional Kitaev chain, Majorana edge modes give a long-lived parity memory while the bulk keeps a faster response, both within the same chain.

The noninteracting Kitaev chain forms an exact linear reservoir in covariance space, so the stability condition and the memory function can be written analytically. In the topological phase the edge mode's retention time grows exponentially with chain length, while local write and read amplitudes decay exponentially with chain length, producing a protection–access tradeoff. In the trivial phase the slowest mode is an extended standing wave whose retention grows only algebraically with chain length, in contrast to the exponential growth of the topological phase. Numerical results confirm exponential and algebraic size scalings in the topological and trivial phases respectively, and show how retention changes across the topological transition.

Adding an ancillary fermion that is reinitialized at each update and couples directly to the nonlocal edge fermion turns the edge mode into a directly writable long-term memory register, with a write coefficient that stays finite as chain length grows. Write and retention coefficients satisfy a fixed relation: stronger writing stores a larger fraction of each new input but accelerates decay of stored information. With a local interface the edge write amplitude vanishes exponentially with chain length; the matched interface removes this local-access penalty. The update equation for edge parity yields the retention and write coefficients, and numerical results show the write–retention tradeoff curve.

Turning on nearest-neighbor density interactions breaks the exact covariance-space description, letting information transfer autonomously from bulk to edge while edge retention drops, a loading–retention competition. On a temporal integration task, an augmented readout that includes the edge coordinate performs similarly to the local readout at short target timescales and substantially better at long ones. Without interactions there is no such autonomous bulk-to-edge transfer channel, and long-timescale memory rests on local features alone. Within the finite-size topological superconducting regime of the interacting chain, bulk-to-edge loading shows a broad maximum at intermediate interaction and intermediate parameters, while edge retention is largest for weak interactions deep in that regime.

Perspective

The scope is the specific one-dimensional Kitaev chain model: the linear covariance-space reservoir in the noninteracting limit, programmable edge memory with a matched ancillary fermion interface, and bulk-to-edge loading once nearest-neighbor density interactions are added. The intended readers are researchers in quantum computing, quantum machine learning and condensed-matter theory, and the conclusions rest on analytical derivations and numerical calculations.

Results such as exponential growth of edge retention with chain length and exponential suppression of local access come from the noninteracting limit, while the loading–retention competition under interactions relies on finite-size numerical calculations. The authors list further control of interfaces and interaction structure, and extension to other topological platforms, as future directions without giving concrete results for those extensions.