Amplitude-Phase Separation toward Optimal and Fast-Forwardable Simulation of Non-Unitary Dynamics

Quantum simulation of the linear non-unitary dynamics is crucial in scientific computing. In this work, we establish a generic framework, referred to as the Amplitude-Phase Separation (APS) methods, which formulates any non-unitary evolution into separate simulation of a unitary operator and a Hermitian operator, thus allow one to take best advantage of, and to even improve existing algorithms, developed for unitary or Hermitian evolution respectively. We utilize two techniques: the first achieves a provably optimal query complexity via a shifted Dyson series; the second breaks the conventional linear dependency, achieving fast-forwarding by exhibiting a square-root dependence on the norm of the dissipative part. Furthermore, one can derive existing methods such as the LCHS (Linear Combination of Hamiltonian Simulation) and the NDME (Non-Diagonal Density Matrix Encoding) methods from APS. The APS provides an effective and generic pathway for developing efficient quantum algorithms for general non-unitary dynamics to achieve either optimal query complexity or fast-forwarding property, outperforming the existing algorithms for the same problems.

💡 Research Summary

The paper introduces a novel framework called Amplitude‑Phase Separation (APS) for quantum simulation of general linear non‑unitary dynamics, a class of problems that arise in scientific computing (e.g., PDEs, ODEs) but cannot be directly simulated on a quantum computer because they are not governed by unitary evolution. APS decomposes any non‑unitary propagator (T\exp!\big(-\int_{0}^{t}A(s)ds\big)) into a product of a unitary (phase) operator and a Hermitian (amplitude) operator. Two concrete decompositions are defined:

1. Phase‑Driven APS: First apply the anti‑Hermitian part (A_{2}(t)) as a unitary flow (U_{p}(t)=T\exp!\big(-i\int_{0}^{t}A_{2}(s)ds\big)). The remaining Hermitian part is transformed to (A_{p}(t)=U_{p}^{\dagger}(t)A_{1}(t)U_{p}(t)) and simulated as a dissipative evolution.

2. Amplitude‑Driven APS: First apply the Hermitian part (A_{1}(t)) as a dissipative flow (U_{a}(t)=T\exp!\big(-\int_{0}^{t}A_{1}(s)ds\big)). The residual anti‑Hermitian part becomes (A_{a}(t)=U_{a}^{\dagger}(t)A_{2}(t)U_{a}(t)) and is simulated as a unitary flow.

The key technical contributions are:

• Shifted Dyson Series – For the Hermitian operator in Phase‑Driven APS, the authors shift the spectrum by the maximal norm (A_{\max}) and expand the resulting operator around 0 instead of the usual expansion point. This “shifted” Dyson series reduces the number of required terms to
  (M = O!\big(A_{\max}t + \log! \epsilon^{-1}\log!\log! \epsilon^{-1}\big)),
  dramatically improving over the naïve (O(A_{\max}t/\epsilon)) scaling.

• Quantum Singular Value Transformation (QSVT) – The unitary pieces (e^{iA_{2}h_j}) are implemented via QSVT, achieving optimal query complexity for the unitary part. By composing at most (\lceil\log N_m\rceil) such pieces, the full Hermitian flow is realized with only (O(\log! \epsilon^{-1})) gate overhead.

• Fast‑Forwarding for Dissipative Dynamics – In the amplitude‑driven regime where the Hermitian part dominates ((A_{1,\max}\gg A_{2,\max})), the authors combine Lindbladian simulation techniques with a “fast‑forwarding” construction that yields a square‑root dependence on evolution time: the query complexity scales as
  (O!\big(p,A_{1,\max}t + A_{2,\max}t,\log! \epsilon^{-1}\log!\log! \epsilon^{-1}\big))
  (with (p) the order of the series), breaking the linear‑in‑time barrier that is unavoidable for unitary Hamiltonian simulation.

• Unified View of Existing Methods – The paper shows that Linear Combination of Hamiltonian Simulation (LCHS) and Non‑Diagonal Density Matrix Encoding (NDME) are special cases of APS: LCHS corresponds to Phase‑Driven APS combined with a Fourier transform of the scalar exponential, while NDME corresponds to Amplitude‑Driven APS together with the interaction‑picture Lindbladian representation.

The authors provide rigorous proofs (Appendices A–D) that the APS‑based algorithms achieve:

• Non‑Unitary Optimality – For time‑independent non‑unitary dynamics, the query complexity to the Hamiltonian oracle is
  (O!\big(|u_{0}||u(t)|,(A_{\max}t + \log! \epsilon^{-1}\log!\log! \epsilon^{-1})\big)),
  matching the lower bound for unitary Hamiltonian simulation (linear in (t) plus additive (\log! \epsilon^{-1}) error term) and improving over previous non‑unitary algorithms that suffered multiplicative (\log! \epsilon^{-1}) penalties.

• Fast‑Forwarding – For purely dissipative or dissipation‑dominated systems, the APS algorithm attains a √t scaling, the first known result that applies to general time‑dependent non‑unitary evolutions.

• Time‑Dependent Extensions – The framework naturally extends to time‑dependent matrices (A(t)). By using a direct‑access model for the time‑ordered exponential, the authors obtain a query complexity of
  (O!\big(|u_{0}||u(t)|,(A_{\max}t\log! \epsilon^{-1}\log!\log! \epsilon^{-1})\big)),
  improving prior work by at least a factor of (\log! \epsilon^{-1}).

• Handling Inhomogeneous Terms – When a source term (b(t)) appears in the differential equation, the algorithm adapts by adjusting the number of repetitions and the queries to the (u_{0}) oracle, yielding a complexity proportional to (|u_{0}| + b_{\max}t|u(t)|).

A comprehensive comparison table (Table I) lists APS alongside Hamiltonian simulation, Lindbladian simulation, LCHS, NDME, and various recent non‑unitary algorithms, highlighting that APS consistently offers lower query counts, additive error dependence, and, when appropriate, √t fast‑forwarding.

In summary, the paper delivers a powerful, generic toolkit for quantum simulation of non‑unitary dynamics. By cleanly separating amplitude (dissipative) and phase (unitary) components and leveraging state‑of‑the‑art quantum algorithmic primitives (QSVT, shifted Dyson series, fast‑forwardable Lindbladians), APS achieves both optimal query complexity (linear (t) + additive (\log! \epsilon^{-1})) and fast‑forwarding (√t) in regimes where previous methods fell short. This advances the frontier of quantum scientific computing, opening the door to efficient quantum treatment of a broad class of differential equations and other linear non‑unitary processes.