Memory Separations from a Collective-Spin Quantum Register

Technical Summary: Memory Separations from a Collective-Spin Quantum Register

Problem Statement
The paper investigates the efficiency of quantum versus classical memory in generating stochastic processes, specifically focusing on renewal processes. While it is established that quantum hidden Markov models (qHMMs) can utilize non-orthogonal states to reduce memory requirements compared to classical hidden Markov models (HMMs), this study addresses a specific, experimentally relevant constraint: the memory register is driven solely by collective operations. These operations act identically on every qubit in the register, depending only on the total spin. This restriction models physical systems like atomic ensembles or spin-squeezed samples, where the state remains within the permutation-symmetric subspace (dimension $n_m + 1$ rather than $2^{n_m}$). The central question is whether such a restricted quantum system can still generate processes that require an unbounded or significantly larger number of classical states to reproduce.

Methodology
The author models the memory as a sequential Born machine with a coherent ancilla, where a work register is prepared, rotated, coupled to the memory via collective spin operators ($J_\alpha$ and $Z_k \otimes J_z$), measured, and reset. The memory itself is never measured directly. The study focuses on two specific constructions within this framework:

1. Forbidden Interevent Time: A process where the probability of an event occurring at a specific time interval $n^*$ is exactly zero ($d_{n^*} = 0$), while all other intervals have non-zero probability. This is achieved by setting a specific relative phase of $\pi$ between two generator states.
2. Quadratic Channel Generation: A generalization using $K$ amplitudes to generate a renewal process where the interevent distribution is a sum of $K(K+1)/2$ decay channels. This includes $K$ "population" channels and $K(K-1)/2$ "interference" channels arising from cross-terms.

The author analyzes the classical cost (minimum number of hidden states) required to reproduce these processes using Propositions derived from the properties of non-negative matrices, Hankel ranks, and the Wielandt bound. They also verify the realizability of these processes using a depth-three quantum circuit whose parameter count is independent of the register size.

Key Contributions and Results

• Unbounded Separation via a Single Phase:
  The paper demonstrates that a single spin-1/2 (a 2-dimensional quantum memory) can generate a renewal process where a specific interevent time $n^*$ is forbidden.

  • Result: Any classical HMM reproducing this process requires at least $1 + \lceil \sqrt{n^*} \rceil$ states.
  • Significance: This establishes an unbounded separation between quantum and classical memory costs. The quantum resource is minimal: a single relative phase of $\pi$ and one amplitude ratio. The classical cost grows without bound as $n^*$ increases, despite the quantum system being a fixed, small dimension.

• Quadratic Separation via Interference:
  By utilizing $K$ amplitudes, the quantum memory generates a process with $K(K+1)/2$ distinct decay channels.

  • Result: For generic decay rates and positive weights, the minimal classical HMM requires exactly $R = K(K+1)/2$ states to reproduce the process. The quantum memory requires only dimension $K$.
  • Mechanism: The interference terms in the quantum superposition create "off-diagonal" channels with rates equal to the pairwise averages of the base rates. Crucially, the weights of these interference channels can be negative (destructive interference), a feature impossible in classical mixtures.
  • Significance: The paper provides a family of processes where the classical cost is exactly $K(K+1)/2$, not just bounded below by it. This contrasts with previous work that often compared lower bounds; here, the attainable costs themselves exhibit a quadratic separation (scaling as $K^2$ vs. $K$), rather than an exponential one.

• Realization with Collective Operations:
  The author shows that these processes are realizable using a depth-three circuit acting on a collective-spin register.

  • Result: The dynamical Lie algebra of the proposed circuit is the full $su(D)$, ensuring reachability of any unitary on the joint register. Numerical verification confirms that a depth-three block with twenty parameters can reproduce the target Kraus operators for the $n^*=5$ case to machine precision, whereas depth-two fails.
  • Constraint: The number of parameters does not grow with the register size ($n_m$), making the construction scalable.

Significance and Claims
The paper claims to demonstrate that even under the severe restriction of collective operations—where the memory dimension scales linearly with the number of qubits rather than exponentially—a quantum system can generate processes that are significantly more efficient to simulate classically in terms of state count.

Specifically, the author highlights that:

1. Resource Efficiency: The entire resource for the unbounded separation is a single relative phase of $\pi$.
2. Exact Attainment: Unlike previous studies that compared theoretical lower bounds, this work exhibits families of processes where the actual minimal classical cost equals $K(K+1)/2$, while the quantum cost is $K$.
3. Role of Coherence: The separation is attributed to the ability of quantum coherence to generate negative weights in the decay channel decomposition (destructive interference), which classical probability theory cannot replicate without increasing the state space.
4. Experimental Relevance: The construction relies on operations naturally available in atomic ensembles (total-spin operators), suggesting that these theoretical advantages are accessible in current experimental platforms.

The paper does not propose new experimental setups beyond the described circuit realization nor does it speculate on future applications beyond the demonstration of memory separation. The focus remains strictly on the theoretical characterization of memory costs in restricted quantum registers.