---
title: Hierarchical Adiabatic Elimination
url: https://www.emergentmind.com/topics/hierarchical-adiabatic-elimination
type: topic
---

# Hierarchical Adiabatic Elimination

Searching arXiv for recent and foundational work on hierarchical adiabatic elimination and closely related adiabatic-elimination formalisms.
arXiv search query: "hierarchical adiabatic elimination open quantum systems fast unitary dynamics"
Hierarchical adiabatic elimination denotes a family of reduction procedures for quantum dynamics with separated timescales in which fast degrees of freedom, fast subspaces, or rapidly oscillating transitions are removed recursively or order by order, while an effective dynamics is retained for the slow sector. Across closed and open quantum settings, the common structure is a decomposition into relevant and irrelevant components—often written as \(P\) and \(Q\) sectors, or as slow and fast subsystems—followed by perturbative, projection-based, geometric, resolvent, Heisenberg, or diagrammatic constructions of reduced generators and embedding maps. In the supplied literature, the term is associated both with adiabatic control of leakage in driven finite-dimensional systems and with systematic elimination in Lindbladian models having nested or multipartite timescale separation [1611.05054].

## 1. Conceptual scope and hierarchical structure

Hierarchical adiabatic elimination appears in several closely related senses. In one sense, it refers to **isolating subspaces (adiabatic/hierarchical elimination) in quantum dynamics** using Leakage Elimination Operators (LEOs), with the target subspace identified as an instantaneous eigenspace or other chosen adiabatic sector. In another, it refers to **iterative or recursive treatment** of multiple timescales in open quantum systems, where one eliminates the fastest component first and then repeats the procedure on the remaining slower dynamics. A further usage concerns **order-by-order hierarchies** of corrections, where zeroth-order adiabatic elimination is improved by first-, second-, or higher-order approximations [1611.05054].

A standard decomposition is
\[
\mathcal{H} = P\mathcal{H} \oplus Q\mathcal{H},
\]
or, in bipartite settings, a tensor-product split between a slow subsystem and a fast subsystem. In the LEO formulation, the total Hamiltonian is partitioned as
\[
H = H_P + H_Q + H_L,
\]
where \(H_P\) and \(H_Q\) act within the \(P\) and \(Q\) subspaces, and \(H_L\) generates unwanted transitions (leakage) between them. In open-system formulations, the corresponding object is a Lindbladian or GKSL generator split into fast and slow pieces, often with a small parameter \(\epsilon\) controlling the perturbation [1611.05054].

The hierarchical aspect is explicit in several forms. One summary states that if the center manifold itself has a "secondary" fast-slow structure, or if there are **multiple nested timescale separations**, the same operator formalism applies recursively. Another states that in systems with several well-separated timescales, the same formalism may be applied **iteratively or hierarchically**, identifying the slowest remaining subspace at each step. A related multipartite version decomposes the environment into \(K\) components and exploits that decomposition so that each environment subsystem’s contribution is computed independently in its own smaller Hilbert space, avoiding the exponential scaling associated with a monolithic elimination [2404.01802].

This suggests that “hierarchical” is not a single algorithmic prescription but a structural attribute: the reduction scheme must remain valid when the retained dynamics is itself nontrivial, when further layers of fast variables remain, or when successive corrections are organized into a controlled expansion.

## 2. Closed-system formulations: leakage suppression and higher-order adiabatic deviations

In adiabatic control problems, the principal difficulty is that slow driving is never literally infinitesimal, so leakage out of an instantaneous eigenstate manifold can accumulate. One formulation addresses this by embedding a Leakage Elimination Operator in the **adiabatic frame**. For the ground eigenstate \( |E_0(t)\rangle \), the added term is
\[
H_{LEO}(t) = f(t) |E_0(t)\rangle\langle E_0(t)|,
\]
with \(f(t)\) a control function, typically a sequence of fast, strong pulses. The LEO operator \(R_L\) is defined by
\[
\{R_L,H_L\} = 0, \quad [R_L, P] = [R_L, Q] = 0,
\]
and repeated pulsing yields
\[
\lim_{m\rightarrow\infty}\left[e^{-iHt/m} R_L^\dag e^{-iHt/m} R_L\right]^m = e^{-iH_P t/m} e^{-iH_Q t/m},
\]
which effectively eliminates \(P \leftrightarrow Q\) transitions. In the associated \(PQ\)-partitioning analysis, the target-subspace amplitude satisfies
\[
\dot{p}(t) = \int_0^t g'(t, s) p(s) ds,
\]
and for the two-level example
\[
g'(t, s) = -\frac{\omega^2}{4} \exp\left\{ i \int_s^t [f(s') - \omega_1] ds' \right\}.
\]
A central claim is that the effectiveness of the control depends on the **average control frequency**
\[
\langle \omega_2(s, t) \rangle = \frac{1}{t-s}\int_s^t f(s') ds',
\]
rather than the detailed pulse shape; regular, random, or noisy pulses are described as equally effective provided they deliver sufficient average control frequency [1611.05054].

A different hierarchical closed-system picture treats deviations from the quantum adiabatic theorem itself. For a slowly varying, non-degenerate Hamiltonian \(\hat{H}_0(R(t))\), a classical-mechanics framework is constructed in which the zeroth-order dynamics follows the instantaneous eigenstate, while first-, second-, and higher-order deviations are governed by a hierarchy of effective Hamiltonians \(H_1,H_2,\dots,H_k\). In that framework, the \(k\)th-order deviations depend on \(R,\dot R,\ddot R,\ldots,d^kR/dt^k\), and **adiabaticity at order \(k\) fails if the \(k\)-th time derivative becomes large, even if all lower derivatives remain small**. For a two-level mapping with \(p=\arg(c_2)-\arg(c_1)\) and \(q=|c_2|^2\), the first-order shift is determined from a linearized equation involving the Hessian \(\Gamma_0(R)\), and the dynamics can be generated by a quadratic effective Hamiltonian centered at the first-order shifted fixed point. The same paper states that this construction exposes a deep connection between classical adiabatic theory and quantum adiabatic theory [1402.6431].

Taken together, these results separate two notions sometimes conflated in the literature. One is **elimination of leakage by control** within an adiabatic subspace; the other is **hierarchical description of nonadiabatic corrections** order by order. The two are compatible but conceptually distinct.

## 3. Open quantum systems: projection, resolvent, Sylvester, and Heisenberg constructions

In open systems, adiabatic elimination is usually formulated at the level of density operators and Lindbladians. A common exact starting point is the projected resolvent identity
\[
P G(z)P = (z - \mathcal{L}_\text{eff}(z))^{-1},
\]
with
\[
\mathcal{L}_\text{eff}(z) = P \mathcal{L} P + P \mathcal{L} Q (z - Q\mathcal{L} Q)^{-1} Q \mathcal{L} P,
\]
or, in equivalent notation,
\[
\mathcal{L}_{\text{eff}(z)} = PLP + PLQ [z - QLQ]^{-1} QLP.
\]
Expanding around \(z=0\) gives
\[
L_0 = P \mathcal{L} P - P\mathcal{L}Q (Q\mathcal{L}Q)^{-1} Q\mathcal{L}P,
\qquad
L_1 = -P\mathcal{L}Q (Q\mathcal{L}Q)^{-2} Q\mathcal{L}P.
\]
One formulation emphasizes that \(L_0\) is trace-preserving within the slow subspace only approximately, and introduces the normalization correction
\[
\alpha = \frac{1}{\text{tr}\big( - L_1 \bar{\rho} \big)},
\]
where \(\bar\rho\) is the stationary state of \(L_0\). The same source interprets the correction as a detailed balance equation and states that discrete and continuous fast subspaces converge for very large dissipation and at coherent population trapping points [1909.04211].

A perturbative alternative uses an embedding map \(\mathcal{K}\) and reduced generator \(\mathcal{L}_s\),
\[
\mathcal{L}_s(\rho_s) = \sum_{j=0}^\infty \epsilon^j \mathcal{L}_{s,j}(\rho_s),\qquad
\mathcal{K}(\rho_s) = \sum_{j=0}^\infty \epsilon^j \mathcal{K}_j(\rho_s),
\]
subject to
\[
\mathcal{K}(\mathcal{L}_s(\rho_s)) = \mathcal{L}(\mathcal{K}(\rho_s)).
\]
For fast unitary center dynamics, the first-order correction obeys a **Sylvester equation** of the form \(AX+XB=C\), with explicit solution
\[
X = -\int_0^\infty e^{tA} C e^{tB} dt.
\]
The first-order embedding and second-order reduced dynamics are then expressed as explicit time integrals involving backward evolution \(A_k^-(t)=e^{-itH_A}A_k e^{itH_A}\) and Heisenberg-adjoint bath evolution \(B_l(t)=e^{t\mathcal{L}_B^*}(B_l)\). This approach is described as systematic at all orders, and the same recursive structure is said to apply when the center manifold itself has nested timescale separation [2404.01802].

A Heisenberg-picture construction instead parameterizes the slow manifold by invariant operators \(J_a\) satisfying \(\mathcal{L}_0^\dagger(J_a)=0\) and coordinates \(x_a=\mathrm{Tr}(J_a\rho)\). The slow basis and reduced dynamics are expanded as
\[
S_a(\epsilon)=S_a^{(0)}+\epsilon S_a^{(1)}+\epsilon^2 S_a^{(2)}+\cdots,
\qquad
F_{a,a'}(\epsilon)=\epsilon F^{(1)}_{a,a'}+\epsilon^2 F^{(2)}_{a,a'}+\cdots,
\]
with
\[
\frac{dx_a}{dt} = \sum_{a'} F_{a,a'}(\epsilon)\,x_{a'}.
\]
The pseudo-inverse
\[
\mathcal{R}_0(W) = \int_0^\infty e^{s \mathcal{L}_0} (W - K_0(W))\,ds
\]
generates first- and second-order corrections, and the second-order propagator is stated to preserve trace and complete positivity up to second-order terms [2303.17308].

These formulations are mathematically different but structurally aligned: all of them define a slow manifold or projected sector, compute an effective generator there, and provide a map back to the full space.

## 4. Multipartite, bipartite, Gaussian, and measured settings

Several supplied works generalize adiabatic elimination from a single fast subspace to composite fast environments. In a multipartite model with a target system on \(\mathcal{H}_B\) and an environment decomposed as \(\mathcal{H}_A=\bigotimes_k\mathcal{H}_A^{(k)}\), the full dynamics is
\[
\frac{d \rho}{dt} = \sum_{k}{ \Big( \mathcal{L}_A^{(k)}(\rho) + \varepsilon \mathcal{L}_{int}^{(k)}(\rho) \Big) } + \varepsilon \mathcal{L}_B(\rho) - i \left[ \tilde{\bm{H}_B}, \rho \right].
\]
The reduced model is expanded as
\[
\mathcal{K}(\rho_s) = \mathcal{K}_0(\rho_s) + \varepsilon \mathcal{K}_1(\rho_s) + \varepsilon^2 \mathcal{K}_2(\rho_s) + \dots ,
\qquad
\mathcal{L}_s(\rho_s) = \mathcal{L}_{s,0}(\rho_s) + \varepsilon \mathcal{L}_{s,1}(\rho_s) + \varepsilon^2 \mathcal{L}_{s,2}(\rho_s) + \dots .
\]
At zeroth order,
\[
\mathcal{K}_0(\rho_s) = \left(\bigotimes_{k} \bar{\rho}_A^{(k)}\right) \otimes \rho_s,
\qquad
\mathcal{L}_{s,0}(\rho_s) = -i [ \tilde{\bm{H}_B}, \rho_s ].
\]
The first-order correction is a sum of individual contributions from each environment component, and the second-order reduced Lindbladian is given explicitly in Lindblad form. The same source emphasizes that the decomposition into \(K\) components enables efficient treatment and avoids the quantum curse of dimension, while preserving Kraus form for the embedding map and Lindblad form for the reduced dynamics [1803.07810].

In a bipartite open system \(H=H^{(A)}\otimes H^{(B)}\), the projection may be taken as
\[
P \rho^{(AB)} = \operatorname{tr}_B[\rho^{(AB)}] \otimes \rho_b,
\]
with \(\rho_b\) the fast subsystem steady state. The long-time reduced generator is again based on
\[
L_0 = P\mathcal{L}P - P\mathcal{L}Q(Q\mathcal{L}Q)^{-1}Q\mathcal{L}P,
\]
and the method is explicitly described as extensible to **nesting projectors** for multiple subsystems or subspaces [2006.07528].

A specialized but important instance is elimination of Gaussian bosonic transducers under continuous measurement. There the global stochastic master equation has the form
\[
\dot{\rho} = \mathcal{L}_S \rho + \mathcal{L}_T \rho + \mathcal{L}_{\text{int}}\rho + \sum_m \mathcal{H}[\lambda_m]\rho \cdot dW_m,
\]
with linear transducer quadratures \(\mathbf r\), first moments \(\mathbf x\), and covariance matrix \(\Gamma\). The transducer dynamics is specified by
\[
\dot{\mathbf{x}} = A\mathbf{x} + \sum_m (\Gamma c_m - \sigma m_m) dW_m,
\]
\[
\dot{\Gamma} = A\Gamma + \Gamma A^T + 2N - 2 \sum_m (\Gamma c_m - \sigma m_m)(\Gamma c_m - \sigma m_m)^T.
\]
Elimination yields an effective SME
\[
d\rho_S = \mathcal{L}_{\text{eff}} \rho_S \, dt + \sum_m \mathcal{H}[i \Lambda_m^T \mathbf{s}] \rho_S \, dW_m,
\]
with
\[
\mathcal{L}_{\text{eff}} \rho_S = \frac{1}{2} (A^{-1} \Gamma)_{ij} \left[ s_i, [s_j, \rho_S] \right] + \frac{i}{2} (A^{-1} \sigma)_{ij} \left[ s_i, \{ s_j, \rho_S \} \right].
\]
This formulation explicitly handles arbitrary numbers of bosonic modes and finite-temperature transducers, and it applies both to unconditional and conditional dynamics [1503.07484].

A related cavity-emitter reduction eliminates an ensemble \(B\) and derives effective equations for a subsystem \(\mathcal S\) consisting of a cavity and a single emitter \(A\). The effective master equation is
\[
\partial_t v = -i [H_0^{(\mathrm{eff})} + H_{\rm JC}^{(\mathrm{eff})}, v] + \mathcal{L}^{(\mathrm{eff})} v,
\]
with effective parameters determined by \(\vec G\), \(\vec V\), and \(\mathbf M^{-1}\), and with an off-diagonal dissipative rate
\[
\mu = \mathrm{Im} [\vec{G}^T {\bf M}^{-1} \vec{V} ].
\]
The validity condition is stated as
\[
|\lambda_j| \gg \big\{ |g_A|, |\vec{G}|, |\vec{V}|, |\Delta^{(\mathrm{eff})}_{A,c}|, \gamma_A^{(\mathrm{eff})}, \kappa^{(\mathrm{eff})}, |g_A^{(\mathrm{eff})}|, |\mu| \big\},
\]
and, for a single \(B\) emitter,
\[
\max(|\Omega_{AB}|, |\gamma_{AB}|) \ll |\Delta_B - i \gamma_B|.
\]
That work notes that the same projection technique can in principle be applied hierarchically when \(B\) contains sub-ensembles with very different timescales [1912.12703].

## 5. Recursive approximation hierarchies beyond standard adiabatic elimination

A recurring criticism of standard adiabatic elimination is that it is often only a lowest-order approximation and can be ambiguous or difficult to improve systematically. Two supplied works build explicit approximation hierarchies from integro-differential equations of Lippmann–Schwinger type.

For multi-level, multi-photon processes with relevant amplitudes \(\psi(t)\) and irrelevant amplitudes \(\epsilon(t)\), the interaction-picture Hamiltonian is written as
\[
H_\mathrm{I} = \hbar \begin{pmatrix} \boldsymbol{\omega} & \frac{1}{2}\boldsymbol{\Omega} \\
\frac{1}{2}\boldsymbol{\Omega}^\dagger & \boldsymbol{\Delta} \end{pmatrix},
\]
and the exact elimination of \(\epsilon(t)\) gives
\[
i \frac{\partial \psi}{\partial t} = \boldsymbol{\omega} \psi - \frac{i}{4} \boldsymbol{\Omega} \int_0^t dt' ~ e^{-i \boldsymbol{\Delta} (t-t')} \boldsymbol{\Omega}^\dagger \psi(t').
\]
The **zeroth-order Markov approximation** yields
\[
H_\mathrm{eff}^{(0)} = \hbar \left(\boldsymbol{\omega} - \boldsymbol{\Omega} \frac{1}{4\boldsymbol{\Delta}}\boldsymbol{\Omega}^\dagger \right),
\]
which is identified with standard adiabatic elimination. The **first-order Markov approximation** gives
\[
i \frac{\partial \psi}{\partial t}
=
\Big[1 + \boldsymbol{\Omega} \frac{1}{4\boldsymbol{\Delta}^2} \boldsymbol{\Omega}^\dagger \Big]^{-1}
\left[ \boldsymbol{\omega} - \boldsymbol{\Omega} \frac{1}{4\boldsymbol{\Delta}} \boldsymbol{\Omega}^\dagger \right]\psi,
\]
and higher orders follow by further Taylor expansion of the memory term. The procedure also motivates criteria for optimizing the interaction picture, such as trace-centering \(\mathrm{tr}(\boldsymbol{\omega}+\widetilde\omega)=0\) [1209.6568].

For Raman transitions, a related strategy avoids eliminating the intermediate state entirely. The interaction-picture Hamiltonian is cast as a \(2+1\) block system, and after squaring the Hamiltonian one writes
\[
(H_{\mathrm{I}}/\hbar)^2 = M_0^2 + \epsilon.
\]
The exact evolution is recast into an integro-differential equation
\[
U(t) = U_0(t) - \int_0^t dt' \frac{\sin[M_0 (t-t')]}{M_0} \epsilon \, U(t'),
\]
which generates a hierarchy
\[
U_k^{(R)}(t) = U_{k-1}^{(R)}(t) - \int_0^t dt' \frac{\sin[M_0(t - t')]}{M_0} \epsilon U_{k-1}^{(R)}(t').
\]
The supplied summary states that very accurate results are already obtained in the lowest order, that all state populations—including the intermediate state—remain accessible, and that the method avoids the interaction-picture ambiguity of standard adiabatic elimination [1209.6569].

These results show a second sense of hierarchy: not successive removal of different subsystems, but successive improvement of a single reduction by controlled memory corrections.

## 6. Reformulations, equivalences, and transition-space projection

Recent formulations make the equivalence between previously separate adiabatic-elimination frameworks explicit. One such reformulation uses the time-convolutionless (TCL) master equation for
\[
\frac{d}{dt} \rho(t) = ( \mathcal{L}_0 + \epsilon \mathcal{L}_1 ) \rho(t).
\]
With a projection onto surviving modes,
\[
\mathcal{P}_{\rm inv} = \sum_s | r_s \rangle \langle l_s |,
\]
the TCL equation introduces
\[
\Sigma(t) = \epsilon \int_{0}^{t} d\tau\, e^{\mathcal{Q}\mathcal{L}\mathcal{Q} \tau} \mathcal{Q}\mathcal{L}_1 \mathcal{P} e^{- \mathcal{L} \tau},
\]
\[
\mathcal{J}(t) = [1 - \Sigma(t)]^{-1} e^{\mathcal{Q}\mathcal{L}\mathcal{Q} t} \mathcal{Q},
\qquad
\mathcal{P}(t) = [1 - \Sigma(t)]^{-1} \mathcal{P},
\]
and, in the long-time limit, a corrected invariant projector
\[
\mathcal{P}^{(\epsilon)}_{\rm inv} = \mathcal{P}_{\rm inv} + \epsilon \mathcal{P}_1 + \epsilon^2 \mathcal{P}_2 + \dots,
\]
with
\[
\mathcal{P}_1 = \int_0^\infty d\tau\, e^{\mathcal{L}_0 \tau} \mathcal{Q}_{\rm inv} \mathcal{L}_1 \mathcal{P}_{\rm inv} e^{-\mathcal{L}_0 \tau}.
\]
The reduced and embedding maps are
\[
\mathcal{K}^{(\epsilon)}_{\rm TCL} = \mathcal{P}^{(\epsilon)}_{\rm inv} \chi_R,
\qquad
\mathcal{F}^{(\epsilon)}_{\rm TCL} = \chi_L^\dagger \mathcal{L} \mathcal{P}^{(\epsilon)}_{\rm inv} \chi_R,
\]
and they satisfy the same invariance equation as geometric adiabatic elimination,
\[
\mathcal{K}^{(\epsilon)} \mathcal{F}^{(\epsilon)} = \mathcal{L} \mathcal{K}^{(\epsilon)}.
\]
The supplied summary states that the TCL formulation yields results equivalent to those of the geometric formulation, while also handling transients and complex cases that are challenging within the geometric approach [2409.08332].

A more recent development formulates adiabatic elimination in the dispersive regime directly in transition-operator space. The free Liouvillian satisfies
\[
\mathcal{L}_{\text{free}} \hat{\xi} = \Delta_{\hat{\xi}} \hat{\xi},
\]
and the resolvent expansion is written as
\[
\mathcal{G}(s) = \left(s+i \mathcal{L}_{\text{free}}\right)^{-1} \sum_{n=0}^\infty \left[ -i\,\mathcal{L}_{\text{int}} \left(s + i \mathcal{L}_{\text{free}}\right)^{-1} \right]^n.
\]
Adiabatic elimination is then implemented by a projection
\[
\mathcal{P}_T \hat{\xi} = \frac{1}{T} \int_0^T dt\, \hat{\xi}(t),
\]
which keeps only transitions with small cumulative detuning on the timescale \(T\). At order \(n\), the diagram amplitude is
\[
v_n(t) = (-i)^n \int_0^t d\tau_n \ldots \int_0^{\tau_2} d\tau_1 \,
\exp\left( -i \sum_{k=0}^n (\Delta_k - i\Theta_k) (\tau_{k+1}-\tau_k) \right),
\]
or equivalently as a sum over poles. The summary explicitly describes the controlled projections as being applied repeatedly at each order so that only slow or resonant transitions are retained, which is a particularly literal realization of hierarchical elimination in transition space [2605.14100].

A plausible implication is that the current landscape is less a competition among incompatible methods than a set of equivalent or complementary parameterizations—state-space, operator-space, resolvent, center-manifold, or transition-space—chosen according to spectral structure and computational convenience.

## 7. Interpretation, validity, and recurrent misconceptions

A recurrent misconception is that adiabatic elimination always means literal removal of variables with no trace of the eliminated sector. Several supplied works contradict that simplification. In the Raman and multi-photon hierarchies, eliminated or nominally irrelevant levels can still have computable population, and higher-order corrections encode memory and normalization effects rather than simply enforcing a slaving relation [1209.6569]. In open systems, another misconception is that the reduced slow dynamics is automatically trace-preserving without correction; one supplied resolvent treatment introduces an explicit trace correction factor and interprets it as detailed balance [1909.04211].

A second misconception is that the slow manifold must itself be dynamically slow in a naive sense. One 2024 formulation explicitly treats the case where the center manifold follows **fast unitary dynamics instead of just being slow**, and one earlier multipartite formulation extends reduction to cases where the target component is subject to Hamiltonian evolution at the fast timescale [2404.01802]. This broadens the meaning of “adiabatic” in open-system reduction: the essential requirement is timescale separation relative to the eliminated sector, not absolute slowness of the retained one.

A third misconception is that idealized pulse shapes or perfect regularity are necessary in adiabatic leakage control. The LEO-based analysis states that the performance depends only on the average value of \(f(t)\), not on its detailed form, and that regular, random, or noisy pulses are all equally effective provided they deliver sufficient average control frequency [1611.05054].

Validity conditions remain method-dependent. Projection-based and perturbative reductions require a spectral gap or a regime where couplings are small compared with fast decay or detuning scales. Representative statements in the supplied literature include \(\epsilon=g/\kappa \ll 1\), the requirement that bath eigenvalues dominate all system-bath couplings and effective rates, and conditions such as
\[
\max(|\Omega_{AB}|, |\gamma_{AB}|) \ll |\Delta_B - i \gamma_B|.
\]
Several summaries also note breakdown when timescales become comparable or when memory effects become important [1912.12703].

The literature additionally distinguishes between preserving mathematical structure and merely obtaining approximate rates. Some frameworks are designed so that reduced dynamics remains in Lindblad form and the embedding remains in Kraus form; others prove trace preservation and complete positivity up to a specified order; still others emphasize full-state reconstruction from the reduced manifold. This suggests that hierarchical adiabatic elimination is best understood not only as model reduction but as **structure-preserving reduction**.

In aggregate, the supplied works portray hierarchical adiabatic elimination as a broad research program rather than a single theorem: it includes \(PQ\)-partitioning with LEO control, resolvent and frequency-domain effective operators, Sylvester-equation perturbation theory with adjoint dynamics, Heisenberg slow-manifold expansions, Gaussian and measurement-conditioned elimination, multipartite and bipartite projector methods, TCL reformulations, and transition-space diagrammatic projections. What unifies these strands is the systematic exploitation of timescale separation, together with explicit rules for retaining the correct slow physics while controlling leakage, normalization, positivity, and higher-order corrections [2409.08332].

Source: https://www.emergentmind.com/topics/hierarchical-adiabatic-elimination