If a bath correlation is known, how should it drive an electronic density matrix? Two common answers are the second-order time-convolutionless equation (TCL2) and the second-order Nakajima–Zwanzig equation (NZ2). One evolves the current density; the other integrates its history. Twelve examples below show why that difference does not establish an accuracy ranking.

Part I explained how Gaussian HEOM stores bath memory in auxiliary density operators. This part makes a controlled reduction in ambition: truncate a generator or a memory kernel at second order in the system–bath coupling. We derive both equations, use a resonance approximation to anticipate their behavior, and then read the numerical examples. The analytic interpretation is a consistency check against existing results, not an independent blind prediction.

1. What information enters the reduced equation?

Set \(\hbar=1\), write the coupling strength explicitly as \(g\), and prepare a thermal bath independently of the electronic state:

\[ \begin{aligned} H&=H_S+H_B+gV,\qquad V=A\otimes B,\\ \rho_{SB}(0)&=\rho_S(0)\otimes\rho_B,\qquad \rho_B=Z_B^{-1}e^{-\beta H_B},\qquad \langle B\rangle_B=0. \end{aligned} \]

The system observable is obtained from the reduced density, \(\rho_S(t)=\operatorname{Tr}_B\rho_{SB}(t)\). The two-point input is the unsymmetrized correlation with the later operator on the left:

\[ \begin{aligned} C(\tau)&=\operatorname{Tr}_B[B(\tau)B(0)\rho_B],\\ B(\tau)&=e^{iH_B\tau}Be^{-iH_B\tau},\qquad C(-\tau)=C(\tau)^*. \end{aligned} \]

Its real and imaginary parts retain fluctuations and response. Both are used below. The correlation-function note explains the physical meaning and thermal weights. When comparing conventions, note that the reverse product \(\langle B(0)B(\tau)\rangle_B\) is the complex conjugate of the convention used here.

A two-point correlation completely specifies a Gaussian bath influence under the assumptions in Part I. It does not make every second-order electronic equation exact. For a non-Gaussian bath, higher bath correlations contain additional information even if this \(C(\tau)\) is known exactly.

2. One second-order kernel, two equations

In the interaction picture of \(H_S+H_B\), let \(\mathcal L(t)X=-i[V_I(t),X]\). The full equation is \(\dot\rho_{SB,I}=g\mathcal L(t)\rho_{SB,I}\). Expanding its propagator and tracing the bath defines a reduced map:

\[ \begin{aligned} \rho_I(t)&=\Lambda(t)\rho_I(0),\\ \Lambda(t)&=I+g^2M_2(t)+O(g^3),\\ M_2(t)X&=\int_0^t du\int_0^u ds\, \operatorname{Tr}_B\{\mathcal L(u)\mathcal L(s) [X\otimes\rho_B]\}. \end{aligned} \]

The first-order term vanishes because the bath coupling is centered. Wherever the exact reduced map is invertible, its exact time-local generator is \(\dot\Lambda\Lambda^{-1}\). Expanding that generator gives \(g^2\dot M_2+O(g^3)\). Keeping only the second-order term gives TCL2.

Define the common kernel

\[ \mathcal K_2(t,s)X =-\operatorname{Tr}_B [V_I(t),[V_I(s),X\otimes\rho_B]]. \]

The two approximations can now be compared without a change of picture:

\[ \begin{aligned} \mathrm{TCL2}:\quad \dot\rho_I(t)&=g^2\int_0^t ds\, \mathcal K_2(t,s)\rho_I(t),\\ \mathrm{NZ2}:\quad \dot\rho_I(t)&=g^2\int_0^t ds\, \mathcal K_2(t,s)\rho_I(s). \end{aligned} \]

The difference is the density on which the same kernel acts. TCL2 is time-local in the unknown density, but its coefficients retain the bath correlation over the entire interval. Neither equation here takes a Markov limit or a secular approximation. The current-versus-history distinction agrees with Babu and co-workers, Sections II.1–II.2.

How eliminating the projected bath gives NZ2

Use \(\mathcal PX=\operatorname{Tr}_B X\otimes\rho_B\), \(\mathcal Q=I-\mathcal P\), and \(R=\rho_{SB,I}\). Centering gives \(\mathcal P\mathcal L(t)\mathcal P=0\), while the preparation gives \(\mathcal QR(0)=0\). Thus

\[ \begin{aligned} \mathcal P\dot R&=g\mathcal P\mathcal L(t)\mathcal QR,\\ \mathcal Q\dot R&=g\mathcal Q\mathcal L(t)\mathcal QR +g\mathcal Q\mathcal L(t)\mathcal PR. \end{aligned} \]

Let \(\mathcal G_Q(t,s)\) propagate the homogeneous equation in the \(\mathcal Q\) subspace. Integrate the second line and substitute into the first:

\[ \mathcal P\dot R(t)=g^2\int_0^t ds\, \mathcal P\mathcal L(t)\mathcal G_Q(t,s) \mathcal Q\mathcal L(s)\mathcal PR(s). \]

This is an exact memory representation. Replacing \(\mathcal G_Q=I_Q+O(g)\) inside its kernel gives NZ2. It truncates propagation in the eliminated sector; it does not retain the full system–bath dynamics merely because the history remains.

At fixed finite time, \(\rho_I(s)-\rho_I(t)=O(g^2)\). Multiplying this difference by the outer \(g^2\) explains why TCL2 and NZ2 first differ at \(O(g^4)\). This is not a uniform long-time error bound. In general the exact solution can also contain a third-order contribution that both methods omit. The even oscillator potentials used below have a stronger symmetry, \(B\mapsto-B\), so their odd coupling orders vanish and the first general omitted order is fourth.

What does a second-order truncation omit?

Both methods omit a fourth-order object; they omit different objects. For the parity-symmetric models below, expand the exact reduced map one order further:

\[ \begin{aligned} \Lambda(t)&=I+g^2M_2(t)+g^4M_4(t)+O(g^6),\\ \Lambda(t)^{-1}&=I-g^2M_2(t) +g^4[M_2(t)^2-M_4(t)]+O(g^6),\\ \dot\Lambda(t)\Lambda(t)^{-1} &=g^2\mathcal R_2(t)+g^4\mathcal R_4(t)+O(g^6),\\ \mathcal R_2(t)&=\dot M_2(t),\\ \mathcal R_4(t)&=\dot M_4(t)-\dot M_2(t)M_2(t). \end{aligned} \]

TCL2 retains only \(g^2\mathcal R_2\). Here \(\mathcal R_2(t)\) is a time-local generator, whereas \(\mathcal K_2(t,s)\) above is a two-time memory kernel. The exact NZ equation instead expands as

\[ \dot\rho_I(t)=\int_0^t ds\, [g^2\mathcal K_2(t,s)+g^4\mathcal K_4(t,s)+\cdots] \rho_I(s). \]

NZ2 retains only \(g^2\mathcal K_2\): it also drops \(\mathcal K_4\) and higher kernels. The exact TCL and NZ representations describe the same dynamics wherever the time-local generator exists. Truncating them at the same order need not give the same approximation, and neither truncation contains all the physics of the other.

Why solving a second-order equation still generates fourth-order terms

Let \(\Lambda_X=I+g^2M_2+g^4M_{4,X}+\cdots\) be the map obtained by solving method \(X\). Iterating each approximate equation gives

\[ \begin{aligned} \dot M_{4,\mathrm{TCL2}}(t) &=\mathcal R_2(t)M_2(t),\\ \dot M_{4,\mathrm{NZ2}}(t) &=\int_0^t ds\,\mathcal K_2(t,s)M_2(s). \end{aligned} \]

For the exact map, these become

\[ \begin{aligned} \dot M_4(t)&=\mathcal R_2(t)M_2(t)+\mathcal R_4(t),\\ \dot M_4(t)&=\int_0^t ds\, [\mathcal K_2(t,s)M_2(s)+\mathcal K_4(t,s)]. \end{aligned} \]

Thus both solutions resum selected higher powers of \(g\), but their generated fourth-order terms differ. Neither includes the complete exact fourth-order coefficient in general. Keeping the past density does not supply the missing \(\mathcal K_4\), and replacing it by the current density does not mean simply deleting the entire fourth-order term of NZ2.

Time locality is also not equivalent to absence of memory. Even an exact time-local generator can encode coherent return through its time-dependent coefficients. The plateau errors below concern the second-order generator for these models, not a prohibition on revivals in all time-local equations.

For a harmonic bath with linear coupling, Gaussian factorization reduces higher bath moments to products of two-point correlations. It does not generally set \(\mathcal R_4\) or \(\mathcal K_4\) to zero: the time-ordered system operators need not commute. The commuting Gaussian pure-dephasing limit is a special exception for TCL2, not a general harmonic-bath result.

An analytic counterexample to a universal NZ2 advantage

As a separate pure-dephasing control, take static zero-mean Gaussian noise \(\xi\) with variance \(\sigma^2\), and \(H_\xi=g\xi\sigma_z/2\). This is a commuting Gaussian-noise model, not the finite-frequency population-transfer model plotted below. Its normalized coherence is

\[ f(t)=\langle e^{-ig\xi t}\rangle =e^{-at^2},\qquad a=\frac{g^2\sigma^2}{2}. \]

The common second-order kernel is constant. The two equations give

\[ \begin{aligned} \mathrm{TCL2}:\quad \dot f(t)&=-2at f(t), & f_{\mathrm{TCL2}}(t)&=e^{-at^2},\\ \mathrm{NZ2}:\quad \dot f(t)&=-2a\int_0^t f(s)\,ds, & f_{\mathrm{NZ2}}(t)&=\cos(\sqrt{2a}\,t). \end{aligned} \]

TCL2 is exact here. Both answers begin with \(1-at^2\); their next terms are \(a^2t^4/2\) and \(a^2t^4/6\), respectively. Since \(a\propto g^2\), they first differ at fourth coupling order. A negative coherence is not a negative population: this counterexample demonstrates an accuracy difference, not a positivity violation of the NZ2 dephasing map.

The equation actually integrated

For the single Hermitian channel, tracing the double commutator gives

\[ \begin{aligned} \mathcal K_2(t,s)X={}& -C(t-s)[A_I(t),A_I(s)X]\\ &-C(s-t)[XA_I(s),A_I(t)]. \end{aligned} \]

For reproducibility in the Schrödinger picture, define \(\mathcal L_0X=-i[H_S,X]\), \(\mathcal G_\tau=e^{\mathcal L_0\tau}\), and

\[ F(t)=\int_0^t C(\tau)\mathcal G_\tau A\,d\tau. \]

The implemented TCL2 equation is

\[ \dot\rho=\mathcal L_0\rho -g^2[A,F(t)\rho-\rho F(t)^\dagger]. \]

The implemented NZ2 equation is

\[ \begin{aligned} \dot\rho(t)=\mathcal L_0\rho(t)-g^2\int_0^t d\tau\, \Big[A,\,&C(\tau)\mathcal G_\tau(A\rho(t-\tau))\\ &-C(\tau)^*\mathcal G_\tau(\rho(t-\tau)A)\Big]. \end{aligned} \]

The free propagation \(\mathcal G_\tau\) is essential: replacing the current Schrödinger-picture density by a bare historical density in the TCL2 formula would give a different equation.

Replacing the NZ2 history by auxiliary matrices

If \(C(\tau)=\sum_jc_je^{-\nu_j\tau}\), introduce

\[ \begin{aligned} Q_j(t)&=\int_0^t c_je^{-\nu_j\tau} \mathcal G_\tau(A\rho(t-\tau))\,d\tau,\\ R_j(t)&=\int_0^t c_j^*e^{-\nu_j^*\tau} \mathcal G_\tau(\rho(t-\tau)A)\,d\tau. \end{aligned} \]

Differentiating these convolutions gives

\[ \begin{aligned} \dot Q_j&=(\mathcal L_0-\nu_j)Q_j+c_jA\rho,\\ \dot R_j&=(\mathcal L_0-\nu_j^*)R_j+c_j^*\rho A,\\ \dot\rho&=\mathcal L_0\rho -g^2\sum_j([A,Q_j]+[R_j,A]), \qquad Q_j(0)=R_j(0)=0. \end{aligned} \]

These equations exactly localize the specified NZ2 convolution. They do not recover the higher-order feedback omitted by NZ2. For Hermitian physical densities \(R_j=Q_j^\dagger\), but independent variables are needed when applying the map to arbitrary non-Hermitian matrix units. The frozen runner uses an equivalent paired-frequency combination of these auxiliaries.

3. Predicting the population curves

To connect the formal equations to the examples, take \(H_S=\Delta\sigma_z/2\) and \(A=\sigma_x\). For the initially occupied upper state, the population sector closes. Define the kernels and their finite-time integrals:

\[ \begin{aligned} k_{\downarrow,\uparrow}(\tau) &=2g^2\operatorname{Re}[C(\tau)e^{\pm i\Delta\tau}],\\ r_{\downarrow,\uparrow}(t) &=\int_0^t k_{\downarrow,\uparrow}(\tau)\,d\tau . \end{aligned} \]

TCL2 uses \(r_\downarrow(t)\) and \(r_\uparrow(t)\) in an instantaneous population equation. NZ2 convolves the corresponding kernels with past populations:

\[ \begin{aligned} \dot P_e^{\mathrm{TCL2}}(t) &=-r_\downarrow(t)P_e(t)+r_\uparrow(t)[1-P_e(t)],\\ \dot P_e^{\mathrm{NZ2}}(t) &=\int_0^t d\tau\,\{ -k_\downarrow(\tau)P_e(t-\tau) +k_\uparrow(\tau)[1-P_e(t-\tau)]\}. \end{aligned} \]

Now keep a dominant resonant pair of bath lines, an additional approximation used only for interpretation:

\[ C_{\mathrm{res}}(\tau)=a e^{-i\Delta\tau}+b e^{i\Delta\tau}, \qquad W=a+b,\qquad p_*=\frac{b}{W}. \]

Dropping the fast counter-rotating terms gives \(k_\downarrow\simeq2g^2a\) and \(k_\uparrow\simeq2g^2b\). With \(P_e(0)=1\), integration for TCL2 and one further differentiation for NZ2 give

\[ \begin{aligned} P_e^{\mathrm{TCL2}}(t) &\simeq p_*+(1-p_*)e^{-g^2Wt^2},\\ P_e^{\mathrm{NZ2}}(t) &\simeq p_*+(1-p_*)\cos(g\sqrt{2W}\,t). \end{aligned} \]

This predicts a decay envelope for TCL2 and an undamped oscillation for NZ2. The latter can reach \(2p_*-1<0\), so keeping a memory convolution does not by itself preserve physical probabilities. The apparent TCL2 plateau is also not proof of thermalization: the reference system contains one undamped mode and can return population coherently.

For the normalized harmonic bath at \(\beta=2\), \(W=1\) and \(p_*=(e^2+1)^{-1}\). The approximation predicts the first NZ2 zero at

\[ t_0=\frac{\arccos(-e^{-2})}{\sqrt2\,g} \simeq\frac{1.20671}{g}. \]

Thus \(g=0.05\) predicts a zero near \(t=24.13\), inside the displayed interval, even though the coupling is small. At \(g=0.01\), the predicted zero is outside that interval. The important secular variable is \(gt\), not coupling strength alone.

The quantum reference has a different structure. Under the same harmonic rotating-wave approximation it is a thermal average of coherent Rabi exchanges:

\[ \begin{aligned} P_e^{\mathrm{JC}}(t) &=\sum_{n=0}^{\infty}(1-r)r^n \cos^2\!\left(gt\sqrt{\frac{n+1}{2\bar n+1}}\right),\\ r&=e^{-2},\qquad \bar n=(e^2-1)^{-1}. \end{aligned} \]

The thermal sum contains nonnegative exchanges at occupation-dependent frequencies. Neither simple second-order expression reproduces this thermal frequency structure. The full numerical calculations below retain counter-rotating terms and all retained bath transitions; they do not use these rotating-wave formulas.

4. The model and numerical workflow

All twelve plots use the same preparation and readout, with four bath/gap choices and three couplings. The bath and coupling operator are

\[ H_B=\frac{p^2+q^2}{2}+\kappa q^4,\qquad B=\frac{q}{\sqrt{\langle q^2\rangle_\beta}},\qquad \langle B^2\rangle_\beta=1,\qquad \beta=2. \]

The odd coordinate has zero thermal mean by parity. Unit variance makes \(g\) the root-mean-square coupling energy. Time is in \(\omega_{\mathrm{ref}}^{-1}\), and energies are in \(\hbar\omega_{\mathrm{ref}}\), with \(\omega_{\mathrm{ref}}=1\). These dimensionless couplings are not material derivatives in eV per unit displacement.

ExamplesBathElectronic gapCouplings
A–C\(\kappa=0\)\(\Delta=1\), resonant\(0.05,\ 0.2,\ 0.6\)
D–F\(\kappa=0\)\(\Delta=1.6\), detunedsame
G–I\(\kappa=0.01\)\(E_1-E_0\simeq1.02839\)same
J–L\(\kappa=0.5\)\(E_1-E_0\simeq1.62823\)same
  1. Prepare the bath. Diagonalize \(H_B\) in a harmonic-oscillator basis and assign its Gibbs weights. There is no trajectory sampling or stochastic hopping.
  2. Construct the correlation. In the bath eigenbasis, evaluate
    \[ C(t)=\sum_{mn}p_m|B_{mn}|^2e^{-i(E_n-E_m)t}. \]
    Both approximate methods use this same spectral input. Discarded spectral pairs have total positive weight below \(2\times10^{-13}\).
  3. Propagate three electronic preparations. Use the bare upper state, bare lower state, and \((|e\rangle+|g\rangle)/\sqrt2\), each with the same factorized thermal bath.
  4. Obtain the reference independently. Diagonalize the full electron–bath Hamiltonian, propagate the joint density spectrally, and trace out the bath. This retains bath response and electron–bath entanglement without a second-order closure.
  5. Integrate TCL2 and NZ2. Use adaptive DOP853, finite-time TCL2 coefficients and an equivalent auxiliary form of NZ2. No Markov, secular, empirical thermal, positivity-repair or population-clipping step is added.
  6. Check numerical errors separately. Compare oscillator bases 40 and 64, and tighten integration tolerances while halving the maximum internal step from 0.1 to 0.05. The plotted output interval is 0.0125; it is not the adaptive integration timestep.

The largest integration-control difference in the full study is \(1.22\times10^{-10}\); the largest basis-control difference is \(2.25\times10^{-8}\), measured by half the trace norm of the density difference. Both are below the declared numerical tolerance \(2\times10^{-6}\). The reference is therefore a numerically converged finite-time result for these models, not an unrestricted claim of exactness for arbitrary baths.

All twelve panels are collected here before their interpretation. A–C use harmonic resonance; D–F use harmonic detuning; G–I use weak anharmonicity; J–L use strong anharmonicity. Within each group, the coupling increases from 0.05 to 0.2 to 0.6. Select a panel to open its full-resolution image.

Common axes and legend. Every plot shows only the initially excited preparation and \(P_e(t)=\operatorname{Re}\rho_{ee}(t)\) in the same fixed bare electronic basis. Black solid is the full quantum reference, orange dashed is TCL2, and blue dotted is NZ2. Vertical ranges vary between panels; negative values and values above one remain visible. These are twelve examples from a wider seven-coupling scan. Open the original four-by-three overview.

Example A: resonant harmonic bath at g equals 0.05; TCL2 remains above exact and NZ2 becomes negative after about time 24
A: \(\kappa=0,\Delta=1,g=0.05\). Horizontal: \(t\omega_{\mathrm{ref}}\); vertical: dimensionless upper-state population. Black solid: exact; orange dashed: TCL2; blue dotted: NZ2.
Example B: resonant harmonic bath at g equals 0.2; exact revives, TCL2 approaches a low plateau, and NZ2 has a large negative oscillation
B: \(\kappa=0,\Delta=1,g=0.2\). Time and population use the units of A. Black solid: exact; orange dashed: TCL2; blue dotted: NZ2.
Example C: resonant harmonic bath at g equals 0.6; both approximate dynamics miss the exact amplitude and phase
C: \(\kappa=0,\Delta=1,g=0.6\). Time and population use the units of A. Black solid: exact; orange dashed: TCL2; blue dotted: NZ2.
Example D: detuned harmonic bath at g equals 0.05; all methods predict only a small loss of upper-state population
D: \(\kappa=0,\Delta=1.6,g=0.05\). Time and population use the units of A. Black solid: exact; orange dashed: TCL2; blue dotted: NZ2.
Example E: detuned harmonic bath at g equals 0.2; oscillations show increasing phase and amplitude errors
E: \(\kappa=0,\Delta=1.6,g=0.2\). Time and population use the units of A. Black solid: exact; orange dashed: TCL2; blue dotted: NZ2.
Example F: detuned harmonic bath at g equals 0.6; TCL2 exceeds population one and NZ2 becomes negative
F: \(\kappa=0,\Delta=1.6,g=0.6\). Time and population use the units of A. Black solid: exact; orange dashed: TCL2; blue dotted: NZ2.
Example G: weak quartic bath at g equals 0.05; the long-time separation closely resembles harmonic resonance
G: \(\kappa=0.01,\Delta\simeq1.02839,g=0.05\). Time and population use the units of A. Black solid: exact; orange dashed: TCL2; blue dotted: NZ2.
Example H: weak quartic bath at g equals 0.2; exact revives, TCL2 misses the return, and NZ2 becomes negative
H: \(\kappa=0.01,\Delta\simeq1.02839,g=0.2\). Time and population use the units of A. Black solid: exact; orange dashed: TCL2; blue dotted: NZ2.
Example I: weak quartic bath at g equals 0.6; neither approximation follows the quantum reference
I: \(\kappa=0.01,\Delta\simeq1.02839,g=0.6\). Time and population use the units of A. Black solid: exact; orange dashed: TCL2; blue dotted: NZ2.
Example J: strong quartic bath at g equals 0.05; TCL2 remains above exact and NZ2 becomes negative near time 23.48
J: \(\kappa=0.5,\Delta\simeq1.62823,g=0.05\). Time and population use the units of A. Black solid: exact; orange dashed: TCL2; blue dotted: NZ2.
Example K: strong quartic bath at g equals 0.2; exact population returns almost completely, TCL2 misses it and NZ2 has a large negative valley
K: \(\kappa=0.5,\Delta\simeq1.62823,g=0.2\). Time and population use the units of A. Black solid: exact; orange dashed: TCL2; blue dotted: NZ2.
Example L: strong quartic bath at g equals 0.6; both methods fail, NZ2 has deep negative oscillations and TCL2 also briefly becomes negative
L: \(\kappa=0.5,\Delta\simeq1.62823,g=0.6\). Time and population use the units of A. Black solid: exact; orange dashed: TCL2; blue dotted: NZ2.

6. Examples A–C: harmonic resonance

The bath is Gaussian, but the coupling is transverse: \([H_S,A]\ne0\). Gaussian statistics alone do not eliminate higher-order reduced-system contributions. A pure-dephasing problem can make TCL2 exact; this population-transfer problem does not have that commuting structure.

A. Weak coupling can still accumulate a large long-time error

The actual zero lies between 24.1375 and 24.1500. At \(t=30\), the resonance estimate gives TCL2 \(0.2120\) and NZ2 \(-0.3416\); the full equations give \(0.2128\) and \(-0.3406\). The reference is \(0.0768\). The agreement with the estimates identifies a closure error, not a failure to compute the bath spectrum.

B. A low plateau misses coherent return

The orange curve is smoother and stays positive here, but it loses the main physical event: population returns from the mode to the electronic system. NZ2 preserves oscillatory structure without the correct quantum exchange. Neither visual smoothness nor the mere presence of oscillations establishes accuracy.

C. Strong coupling lies outside the simple resonance estimate

Here \(g/\Delta\) is not a small parameter. The rotating-wave estimate can help recognize a qualitative mechanism, but its detailed frequency and zero times are no longer controlled predictions of the full equations.

7. Examples D–F: harmonic detuning

Keep the same bath and change only the electronic gap to 1.6. A detuned oscillatory kernel can cancel rather than accumulate. This control tests resonance directly, without changing Gaussianity or introducing anharmonicity.

D. Weak coupling and detuning protect this finite window

The maximum full-density errors over all three preparations are \(0.00376\) for TCL2 and \(0.00265\) for NZ2. Both satisfy the study's 0.01 absolute-error criterion. NZ2 is slightly better here, so even this small set already rules out a universal TCL2-over-NZ2 ranking.

E. An acceptable-looking population does not certify the density matrix

For the coherent initial preparation, NZ2 has a minimum density eigenvalue of about \(-0.00512\). That violation is invisible in this initially excited population panel. A single population between zero and one cannot replace a full-density positivity check.

F. Detuning is not protection against arbitrary coupling

The dominant detuning is 0.6, now comparable to \(g\). The small ratio that protected D is gone. The orange curve demonstrates that TCL2 has no general positivity guarantee.

8. Examples G–I: weak anharmonicity

For \(\kappa=0.01\), the first-order level spacing is \(E_{n+1}-E_n=1+3\kappa(n+1)+O(\kappa^2)\). Tuning to the lowest transition leaves nearby thermally occupied transitions only weakly detuned. Their phase difference is of order 0.9 over \(t=30\) in this estimate. Weak anharmonicity therefore does not establish a rapidly decaying bath correlation.

G. The harmonic failure pattern survives

At \(t=30\), TCL2 gives 0.2092, NZ2 gives −0.3243, and the reference gives 0.0752. Those values are close to A. Retaining only the ground-to-first-excited bath line would miss important nearby thermal lines and is not a reliable quantitative reduction here.

H. A weak quartic term does not remove the reference revival

Changing the spectrum and introducing non-Gaussian statistics modifies the curves, but does not repair the underlying long-time closure error. A bath with slightly unequal level spacings is still a confined discrete quantum system.

I. Strong electronic coupling remains the dominant challenge

The word “weak” describes the quartic coefficient, not the electronic coupling. These are independent approximation parameters. Small \(\kappa\) cannot justify truncating at small \(g\) when \(g\) is actually large.

9. Examples J–L: strong anharmonicity

For \(\kappa=0.5\), the low-energy gap increases and higher bath levels become less thermally occupied. The \(0\leftrightarrow1\) line carries about 0.93952 of the normalized correlation weight. Its thermal ratio gives \(p_*\simeq0.03710\). Unlike the weak-quartic case, this dominant-line reduction is a useful small-coupling estimate, although the bath is not literally a two-level system.

J. A dominant line predicts the negative crossing

The estimate gives \(P_e(30)\simeq0.1534\) for TCL2 and \(-0.4121\) for NZ2. The full second-order equations give \(0.1536\) and \(-0.4083\); the reference is \(0.0506\). Strong anharmonicity does not prevent an accurate explanation of how the approximate equations fail.

K. Strong anharmonicity does not imply short memory

A low-temperature dominant transition can make the reference exchange especially clear. Non-Gaussian statistics and long memory are separate properties. Confinement and discrete lines remain even when the potential is strongly quartic.

L. Neither closure is reliable at strong coupling

The upper-state TCL2 population reaches about \(-0.00356\); the minimum eigenvalue across all three initial preparations is about \(-0.01378\). Neither value should be clipped or dismissed because the orange curve looks comparatively stable.

10. What the comparison establishes

Numerical convergence and physical accuracy are different questions. The reference, TCL2 and NZ2 can all be numerically converged while the latter two solve inaccurate reduced equations. The raw negative values are retained because they diagnose the approximation.

The wider comparison uses the maximum full-density error over three preparations and the whole observation window:

\[ \epsilon(T)=\max_{i,\ 0\le t\le T} \frac12\|\rho_i^{\mathrm{method}}(t)-\rho_i^{\mathrm{exact}}(t)\|_1. \]

The numerical maximum is evaluated on the saved time grid, with the separate refinement controls described above. The declared accuracy criterion is \(\epsilon\le0.01\), together with density eigenvalues no smaller than \(-10^{-7}\), on those preparations. This is an absolute density error, not a relative error in a small transfer probability. Once an approximate density is nonpositive, half the trace norm of its difference remains a useful matrix diagnostic, but need not have the operational interpretation or upper bound of trace distance between physical states.

Bath/gap\(g\)TCL2 \(\epsilon(30)\)NZ2 \(\epsilon(30)\)Interpretation
Harmonic, resonant0.050.13600.4175Both inaccurate; NZ2 nonpositive
Harmonic, detuned0.050.003760.00265Both pass; NZ2 slightly closer
Weak quartic, resonant0.050.13510.3996Weak anharmonicity does not repair the closure
Strong quartic, resonant0.050.12830.4589TCL2 worst error occurs for the coherent initial state
Strong quartic, resonant0.60.98101.8433Both severely inaccurate and nonpositive

Three lessons follow. First, retaining the density history is not a higher-order approximation: NZ2 is not systematically above TCL2. Second, exact two-point input cannot remove the neglected reduced-system orders, or the additional higher bath correlations of a non-Gaussian environment. Third, increasing anharmonicity does not universally increase the error: at \(g=0.05\), the TCL2 error decreases from 0.1360 through 0.1351 to 0.1283, while the NZ2 error first decreases and then increases.

Changing \(\kappa\) also changes the spectrum, matrix elements and thermal occupations, and the gap is retuned to each bath's lowest transition. It is not an experiment that varies only “non-Gaussianity.” Nor can a large \(\kappa\) turn one undamped mode into a macroscopic short-memory reservoir.

The special exactness of TCL2 for Gaussian pure dephasing is useful as a control, not a population-transfer theorem; see Fetherolf and Berkelbach, Section II. Conversely, an auxiliary representation of NZ2 remains NZ2. For Gaussian baths, converging the higher HEOM tiers from Part I can recover contributions absent from either second-order equation. Feeding only \(C(t)\) from a genuinely non-Gaussian bath into a Gaussian hierarchy does not recover that bath's missing higher correlations.

How to read the limitations

The shared limitation is a perturbative treatment of the coupling, not the presence or absence of an explicit history integral. At a fixed finite time, the parity-symmetric models have a leading generic omitted order of \(g^4\). Its coefficient can grow with time or accumulate near resonance. Consequently, small \(g\) alone does not control the error over an arbitrarily long window.

For the harmonic bath, exact \(C(t)\) contains all Gaussian bath information, but either second-order closure can still miss repeated electronic–bath exchanges. For an anharmonic, non-Gaussian bath, there is an additional loss: the connected four-point and higher bath correlations are not fixed by \(C(t)\). A strongly quartic potential does not by itself make either loss larger, nor does it imply short memory.

The appropriate checks are therefore finite-window errors against a suitable reference, the relevant resonances and correlation times, and positivity for the preparations of interest. Trace and Hermiticity preservation alone are insufficient. Neither unmodified equation has a general positivity or complete-positivity guarantee, and the thermal symmetry of the input correlation alone does not guarantee the exact stationary reduced state.

For applications, method choice must therefore be checked against coupling, resonance, bath memory, the preparation, the observable and the required time window. These examples isolate closure error using spectral correlations. They do not validate RPMD correlations for anharmonic materials, or establish a correction to a particular mixed quantum–classical production workflow.

11. Reproducible code and data

The public bundle contains the frozen benchmark runner and configuration, twelve lossless CSV exports, the plotting script, the full comparison table and provenance checksums. It depends on NumPy, SciPy and Matplotlib. The plotter uses every saved sample without smoothing, interpolation or probability repair.

The displayed 36 curves pass the original numerical smoothness thresholds. Integration and basis checks are separate from that display check. The broader study contains 42 parameter points and six model families, including pure-dephasing controls; this article selects the twelve transfer examples above.

References

  1. A. P. Babu, S. Alipour, A. T. Rezakhani and T. Ala-Nissila, “Unfolding system-environment correlation in open quantum systems: Revisiting master equations and the Born approximation.” Open manuscript, Sections II.1–II.2.
  2. J. H. Fetherolf and T. C. Berkelbach, “Linear and nonlinear spectroscopy from quantum master equations.” Open manuscript.

Open Quantum Systems series

  1. I. HEOM: Gaussian bath influence, hierarchy, and quantum-reference tests.
  2. II. TCL2 and NZ2: second-order closures and twelve population examples.