Quantum Homogenizer: Collision State Transformation

• Quantum Homogenizer is a device that uses sequential collisions between a system qubit and identical reservoir qubits to drive any initial state toward a desired target state with high precision.
• It operates via both coherent (partial-SWAP) and incoherent (controlled-SWAP) mechanisms, achieving exponential state convergence and universal state transformation under tight error bounds.
• The concept underpins applications in thermalization, quantum erasure, and reservoir computing, with implementations demonstrated in NMR and photonic systems.

A quantum homogenizer is a collision-based device in which a single system qubit repeatedly interacts with a large reservoir of identical qubits, so that the joint transformation

$$(\rho\otimes\xi\otimes\cdots\otimes\xi)\mapsto \approx \xi\otimes\xi\otimes\cdots\otimes\xi$$

is realized with arbitrarily small error: after $N$ sequential interactions the system state $\rho_N$ satisfies $D(\rho_N,\xi)\le\delta$, while each reservoir qubit’s final state $\xi_j$ satisfies $D(\xi_j,\xi)\le\delta$, where $D$ may be the trace distance or infidelity. In this sense the homogenizer models thermalization by collisions, while also functioning as a universal state-transformation machine that drives a qubit from any initial state to any target reservoir state $\xi$ with negligible disturbance to the reservoir (Beever et al., 2023).

1. Formal definition and operational setting

In its standard formulation, the quantum homogenizer consists of a system qubit $S$ and $N$ identical reservoir qubits, all initially prepared in the same state $N$0. One cycle initializes $N$1 in an arbitrary state $N$2, applies the same interaction sequentially between $N$3 and each reservoir qubit, and leaves $N$4 approximately in state $N$5 while each reservoir qubit is only slightly perturbed. This operational picture underlies both the thermalization interpretation and the machine-like interpretation in which the same reservoir may be reused if deterioration remains small (Violaris et al., 2022).

The homogenization conditions are stronger than simple convergence of the system alone. The protocol must work for arbitrary input $N$6, arbitrary target $N$7, and any desired $N$8, while preserving each reservoir unit up to the same tolerance. This universality is central: the homogenizer is not merely a relaxation channel toward a fixed bath state, but a tunable mechanism for approximate realization of the map $N$9 for any qubit pair (Beever et al., 2023).

A common parametrization writes qubit states in Bloch form. In the single-axis setting used in analyses of repeated machine use, the reservoir and system states can be written as

$\rho_N$0

with the interaction strength encoded by $\rho_N$1 and $\rho_N$2. This parametrization makes the contraction toward the reservoir state explicit and is also used in fidelity-based analyses of accuracy, robustness, and deterioration under repeated cycling (Violaris et al., 2022).

2. Coherent partial-SWAP homogenizer

The original coherent homogenizer is built from the partial-swap unitary

$\rho_N$3

where $\rho_N$4 is the full SWAP and $\rho_N$5 controls the coupling strength. This gate is distinguished by a strong structural result: the partial-swap $\rho_N$6 is the only two-qubit unitary that satisfies the universal homogenization conditions for all $\rho_N$7 (Beever et al., 2023).

Because $\rho_N$8 is fully coherent, successive interactions generate multi-qubit superpositions and entanglement among the system and the reservoir qubits it has touched. Reservoir qubits that have not directly interacted can still interfere via the system acting as a quantum bus. In single-collision Bloch-vector form, this coherent character appears through an additional cross-product term: $\rho_N$9 so the trajectory is not purely relaxational (Karpat et al., 27 Jan 2025).

The convergence theorem is nevertheless simple. Writing the initial system and reservoir Bloch vectors as $D(\rho_N,\xi)\le\delta$0 and $D(\rho_N,\xi)\le\delta$1, and $D(\rho_N,\xi)\le\delta$2, one has after $D(\rho_N,\xi)\le\delta$3 interactions

$D(\rho_N,\xi)\le\delta$4

Hence

$D(\rho_N,\xi)\le\delta$5

If $D(\rho_N,\xi)\le\delta$6, then achieving $D(\rho_N,\xi)\le\delta$7 requires

$D(\rho_N,\xi)\le\delta$8

Meanwhile each reservoir qubit interacts only once, and its post-interaction distance from $D(\rho_N,\xi)\le\delta$9 is $\xi_j$0, so demanding $\xi_j$1 fixes $\xi_j$2. In the worst case $\xi_j$3, this yields

$\xi_j$4

The key consequence is exponential convergence in $\xi_j$5, with the reservoir remaining nearly intact for small $\xi_j$6 (Beever et al., 2023).

3. Incoherent controlled-SWAP homogenizer

An alternative homogenizer replaces coherent system–reservoir interference by a controlled-SWAP mechanism. Here an ancillary control qubit is prepared in

$\xi_j$7

and a three-qubit Fredkin gate is applied: $\xi_j$8 The control is traced out after each use, so system–reservoir coherence is not propagated beyond that step (Beever et al., 2023).

The resulting single-step update is linear and lacks interference terms. If the initial Bloch vectors are $\xi_j$9 and $D(\xi_j,\xi)\le\delta$0, then after one interaction

$D(\xi_j,\xi)\le\delta$1

The absence of $D(\xi_j,\xi)\le\delta$2 terms is the defining feature of the incoherent model. Iteration immediately gives

$D(\xi_j,\xi)\le\delta$3

so

$D(\xi_j,\xi)\le\delta$4

with the same convergence bound $D(\xi_j,\xi)\le\delta$5 as in the coherent case. The most disturbed reservoir qubit is again the first one, for which $D(\xi_j,\xi)\le\delta$6, and choosing $D(\xi_j,\xi)\le\delta$7 ensures all reservoir qubits remain within $D(\xi_j,\xi)\le\delta$8 of $D(\xi_j,\xi)\le\delta$9 (Beever et al., 2023).

This equivalence of asymptotic convergence is the principal structural result of the coherent–incoherent comparison: the convergence properties that matter for modeling thermalization are not dependent on coherence between qubits in the homogenization protocol. The incoherent and coherent machines also share the same single-use resource scaling,

$D$0

which in the worst case $D$1 gives the well-known $D$2 behavior. At the level of fidelity, the infidelity difference $D$3 is at most $D$4 at each step and vanishes as $D$5 (Beever et al., 2023).

The equivalence is only asymptotic. In transient Bloch-sphere dynamics, CSWAP produces straight-line relaxation from $D$6 to $D$7, while PSWAP produces spiraling trajectories because of its coherent back-action. The continuous-time weak-coupling limit of CSWAP is

$D$8

whereas PSWAP carries an additional coherent commutator term (Karpat et al., 27 Jan 2025).

4. Non-Markovian extensions, memory, and transient phenomena

Markovian collision models assume that each ancilla interacts with the system once and is discarded. Non-Markovian variants introduce ancilla–ancilla collisions, so that memory from older ancillas is transferred into ancillas that the system meets later. In the qubit model with partial swap used for both system–bath and bath–bath collisions,

$D$9

numerical analysis shows that homogenization is achieved irrespective of the initial states of the system or bath units. This preserves universality, but the homogenization rate is slower than in the Markovian counterpart (Saha et al., 2022).

In the Bloch-vector description, the non-Markovian evolution is governed by coupled recurrences for the system vector $\xi$0, the ancilla state immediately before system contact $\xi$1, and the ancilla state immediately after it $\xi$2. The Markovian contraction bound

$\xi$3

gives $\xi$4, while the non-Markovian case retains the same asymptotic target but displays a slower approach when ancilla–ancilla coupling $\xi$5 (Saha et al., 2022).

Changing the bath–bath interaction can reverse that slowdown, but only by sacrificing universality. For a generalized bath–bath unitary

$\xi$6

homogenization can become faster than in the Markovian model, but then the initial ancilla state must be restricted to the $\xi$7-axis; coherent superpositions in the bath are no longer faithfully imposed on the system (Saha et al., 2022).

Transient structure is even more sharply differentiated in coherent versus incoherent collision models. With no ancilla–ancilla coupling, both PSWAP and CSWAP realizations are Markovian. With intra-environment collisions, PSWAP–PSWAP becomes non-Markovian only once the environmental coupling is large enough, PSWAP–CSWAP can generate a higher degree of BLP non-Markovianity even for moderate coupling, while CSWAP–CSWAP and CSWAP–PSWAP remain Markovian in the sense that the trace distance decays monotonically for any intra-environment coupling (Karpat et al., 27 Jan 2025). The discrete BLP measure is

$\xi$8

A further extension uses Fredkin-mediated intra-ancilla interactions to ask whether the memory required by the reduced dynamics is classical or quantum. In that setting, product reservoirs yield classical memory, a GHZ reservoir still requires only classical memory because its two- and three-qubit marginals are separable, whereas a Bell-pair reservoir activates quantum memory for sufficiently large coupling. The criterion

$\xi$9

is necessary and sufficient for the emergence of quantum memory. Numerically, for the Bell-state reservoir the quantity $S$0 becomes negative at $S$1 rad, while perturbed GHZ states cross into the quantum-memory regime at smaller $S$2 as the perturbation strengthens (Yosifov et al., 29 Jul 2025).

5. Reusability, erasure, and irreversibility

The homogenizer can be analyzed not only as a one-shot state transformer but also as a reusable machine. For repeated homogenization of $S$3 different system qubits with the same reservoir, the first reservoir qubit drifts the most, and its Bloch distance from the original state becomes $S$4, equivalently $S$5. Meanwhile each new system approaches the degraded reservoir state with an error bound that leads to the explicit constraint

$S$6

together with $S$7. For any finite $S$8, both coherent and incoherent homogenizers remain universal: by taking $S$9 and $N$0, one can homogenize $N$1 qubits to arbitrary accuracy at the cost of more reservoir qubits and weaker coupling (Beever et al., 2023).

This reusability problem becomes thermodynamically significant in the context of erasure. In the constructor-theoretic analysis of Szilard-engine purification, the homogenizer is used as an approximate catalytic machine for transforming a qubit from any state to any other state while remaining approximately unchanged. The system error after $N$2 cycles with $N$3 reservoir qubits is defined as

$N$4

while the machine deterioration is approximated by

$N$5

Their ratio

$N$6

is the central diagnostic. In the double limit $N$7, $N$8 means the task can be made arbitrarily accurate and reusable, whereas $N$9 means it cannot be performed indefinitely in a cycle (Violaris et al., 2022).

The resulting asymmetry is sharp. Pure$N$00mixed randomization is a possible task, with $N$01. Mixed$N$02pure erasure is not: $N$03, so the homogenizer deteriorates faster than the error can be suppressed. In the language of Szilard’s engine, erasure via homogenization alone cannot run indefinitely in a cycle; maintaining accurate erasure requires refreshing or replacing the reservoir, which implies an additional irreversibility cost beyond the Landauer bound $N$04 (Violaris et al., 2022).

A common misconception is to identify this cyclic asymmetry with asymmetry in single-use state evolution. The NMR experiment on a finite four-qubit homogenizer instead found experimental results consistent with the theoretical symmetry in how the qubit states evolve in pure-to-mixed and mixed-to-pure cases after accounting for decoherence. The apparent tension is resolved by distinguishing finite-horizon state evolution from indefinite catalytic reusability (Violaris et al., 2020).

6. Implementations and broader adaptations

A finite quantum homogenizer has been implemented in nuclear magnetic resonance using fully $N$05C-labelled crotonic acid dissolved in $N$06-acetone, with four active $N$07C spins labeled $N$08. Two spins served as the system register and two as the reservoir register. The internal Hamiltonian in the rotating frame was

$N$09

with chemical shifts $N$10, $N$11, $N$12, $N$13 Hz and couplings $N$14, $N$15, $N$16 Hz, plus weak long-range $N$17 Hz. Partial swaps were implemented by GRAPE-optimized shaped pulses for ten values $N$18. The measured polarizations tracked the theoretical curves

$N$19

$N$20

and the entropy data showed $N$21 rising from $N$22 bit while $N$23 fell from $N$24, with the sum $N$25 exceeding $N$26 bits at intermediate $N$27, indicating information scrambling in the four-qubit state (Violaris et al., 2020).

A distinct development recasts homogenization as a platform for temporal information processing. The disordered quantum homogenizer uses randomized partial-SWAP collisions, with each collision equivalent to a mixing channel

$N$28

The iterated system state is then

$N$29

which proves asymptotic stability toward a unique steady state $N$30. This model satisfies the stability and contractivity conditions required for quantum reservoir computing and has been proposed as physically implementable in either NMR ensembles or photonic systems (Yosifov et al., 2024).

In photonic quantum information processing, the term also appears in “frequency auto-homogenization,” where a $N$31 quantum frequency conversion process satisfying the group-velocity-matching condition $N$32 maps spectrally distinct inputs onto the same output mode. In that setting the output amplitude

$N$33

is rendered nearly independent of the input center frequency by a horizontal phase-matching function and broadband pump. Proof-of-principle measurements in a 2.5 mm periodically poled Rb:KTP waveguide yielded a Schmidt number $N$34, purity $N$35, interferometric visibility above $N$36 over $N$37 nm of input detuning around $N$38 nm, and conversion efficiency $N$39 with a transform-limited $N$40 nm pump, with simulations predicting up to $N$41 after chirping the pump to $N$42 ps (Heberle et al., 4 Feb 2025).

Across these formulations, the central theme remains the same: homogenization is the controlled suppression of distinguishability through repeated weak interactions. In the qubit collision model, that suppression is universal, exponentially convergent, and compatible with both coherent and incoherent realizations. The deeper differences emerge not in the fixed point but in transient structure, memory, reusability, and thermodynamic cost.