Why redundant records, not just decoherence, matter
Decoherence suppresses interference between the components of a superposition and selects stable pointer states, but on its own it does not explain why separate observers end up agreeing on the same pointer value. Quantum Darwinism answers that question with redundancy: the environment stores many copies of the pointer information, so different observers can read disjoint fragments and infer the same value without measuring the system directly.
The paper asks whether those copies must be written by direct system-environment contact. In many measurement settings information reaches observers through an intermediate device, so the authors build a model in which the system first correlates with an apparatus, and the environment then interacts only with that apparatus. The route for pointer information becomes S → A → E rather than S → E.
Two circuits that differ only in who controls the rotation
Both architectures are qubit circuits with the same environment of N qubits, initially all in |0⟩, and the same controlled rotation Ry(θ) applied to each environment qubit. In system monitoring, the system qubit itself controls that rotation on every environment qubit. In apparatus monitoring, gates first correlate the system with the apparatus, and the apparatus then controls the same rotation on the environment.
The interaction is non-demolition: it preserves the control's pointer states. Each environment record is either |0⟩ or |ε⟩ = Ry(θ)|0⟩, with overlap γ = cos(θ/2). At θ = 0 the environment stays uncorrelated with the control; at θ = π the records are orthogonal. Because the two architectures produce the same conditional environmental records with the pointer labels exchanged, their record distinguishability and probabilities are unchanged, which lets the authors isolate what the apparatus itself does.
The apparatus changes total correlations, not readable records
The paper separates two quantities that are easy to conflate. Quantum mutual information I(S:Fm) measures all correlations between the system and a fragment of m environment qubits. The Holevo quantity χ(ΠS:Fm) bounds the pointer information an observer can actually read from that fragment. Their difference is the pointer-basis discord D = I − χ.
In system monitoring, the mutual information depends on the overlaps of the fragment, its complement, and the full environment, and for the complete environment it reaches 2H(ρS). In apparatus monitoring, tracing out the orthogonal apparatus states removes coherence between the two system branches, so the mutual information reduces to the fragment entropy alone and never exceeds H(ΠS), even when the whole environment is observed.
The accessible information is nevertheless identical. Both architectures give χ = h2((1 + |cos(θ/2)|^m)/2) for every fragment size m, because their conditional records differ only by an exchange of pointer labels. Apparatus monitoring has zero discord, and the system-monitoring discord exactly accounts for the gap between the two mutual-information curves. Since redundancy is defined through the Holevo threshold, the minimum fragment size and the redundancy are equal in the two noiseless models.
Encoding moves correlations from the apparatus to the environment
To track the transfer, the authors replace the Hadamard with Ry(φ) on the system, tuning the initial system-apparatus entanglement. Before encoding, the pure S–A state carries mutual information 2H(ΠS) while the environment is uncorrelated with the system. After all N environment qubits have interacted with the apparatus at θ = π, the S–A mutual information falls to H(ΠS) and the S–E mutual information rises to the same value.
The bookkeeping is exact: the sum of the S–A and S–E mutual informations is 2H(ΠS) both before and after encoding. At φ = π/2, the initial S–A value is two bits and both final values are one bit. The decrease in system-apparatus correlation is balanced by the increase in system-environment correlation, which is what makes the apparatus a conduit rather than a sink.
Noise location decides whether more records help
The equivalence is a noiseless result, and the paper's noise analysis is where the two architectures diverge. Bit flips on the apparatus before each encoding interaction introduce errors shared by successive records. If the earliest qubit in a fragment has index i1, later records add nothing about the original pointer, so χ = 1 − h2(q_i1) with q_k = (1 − (1 − 2p)^k)/2. For the complete environment only the first record matters, giving χ = 1 − h2(p), so strong apparatus noise keeps even the full environment below the 90% threshold.
Independent bit flips on environment qubits after encoding behave differently. Each error affects only one record, so larger fragments recover pointer information: at p = 0.10, a fragment of m = 4 qubits (f = 0.2) already exceeds the 90% threshold, and for fixed p ≠ 1/2 the Holevo information tends to 1 as m grows. At p = 1/2 each record is independent of the pointer and the information vanishes.
Dephasing in the pointer basis leaves the conditional record populations unchanged, so χ = 1 for any m ≥ 1 and redundancy stays at N at either noise location. Depolarizing noise acts as a bit flip with probability p/2, so its curves follow the bit-flip results at half the error probability. The paper reports that apparatus noise reduces redundancy more strongly than independent environmental noise in the cases considered.
Where the apparatus route is most fragile
The decision-relevant comparison is not whether indirect recording works, but which noise channel threatens it. With δ = 0.1 and H(ΠS) = 1, redundancy is set by the smallest fragment reaching 90% of the pointer information. Dephasing leaves that number at N regardless of location. Bit flips and depolarizing noise lower it, and the asymmetry between locations is the paper's central practical point: an apparatus flip is copied into every later record, whereas an environment flip damages a single record that other fragments can outvote.
This also explains why the mutual-information comparison alone would mislead. System monitoring shows larger total S–Fm correlations, yet the accessible pointer information and redundancy are the same as in apparatus monitoring without noise. The extra correlation is discord, not readable record content, so a reader comparing architectures should look at the Holevo quantity and the fragment threshold rather than at mutual information.
An exact model, not a claim about every apparatus
The conclusions come from exact analytical solutions of qubit circuit models with a fixed environment size N, a single interaction angle θ, and specific noise channels applied at defined locations. The dephasing result is explicitly limited to the model considered, and the depolarizing correspondence to bit flips at p/2 holds for these pointer records. The redundancy criterion uses an averaged Holevo threshold with δ = 0.1 and assumes H(ΠS) > 0.
Within that scope, the paper's priority is clear: indirect apparatus-mediated recording is sufficient for redundant pointer records when noise is absent, and the reliability question then turns on where noise enters. Apparatus bit flips are the binding constraint because they create shared errors that no amount of additional environmental records can remove, while independent environment noise is recoverable by combining more records.
Sources used in this answer
Quantum Darwinism through apparatus-mediated environmental recording
Exact qubit-circuit solutions show that an environment interacting only with an intermediate apparatus acquires the same accessible pointer information and redundancy as direct system-environment monitoring without noise, while apparatus bit flips create shared errors that limit recoverable information and independent environment noise can be mitigated by larger fragments.