Interactions with the environment shape energy transport and photochemical processes in molecular materials and biological systems [1 , 2 , 3 ] . Dissipation and dephasing also complicate the connection between measured optical spectra and the underlying resonances [2 ] . When the relevant optical coherences admit an effective non-Hermitian description, complex eigenvalues encode resonance energies and decay rates [4 , 5 , 6 ] . The contribution of each resonance to the optical response also depends on a residue involving its left and right eigenvectors [7 ] . These contributions can interfere to produce two spectral peaks even when the real pole positions coincide [7 , 8 ] . Moreover, as an exceptional point (EP) is approached, eigenvalues and eigenvectors coalesce [9 , 10 ] , making the decomposition into individual resonance contributions ill-conditioned [5 , 11 ] . Connecting non-Hermitian modes to optical spectra therefore requires their complex poles and biorthogonal residues.
Quantum algorithms have expanded access to non-Hermitian spectral properties through variational eigenstate preparation [12 ] , eigenvalue estimation and transformation [13 , 14 ] , pseudospectral analysis [15 ] , and oblique eigenprojection [16 ] . These approaches address complementary tasks. Eigenvalue and singular-value estimates identify spectral structure, but do not by themselves determine the complex residues of a chosen optical channel. Variational schemes can evaluate biorthogonal matrix elements using separately optimised left and right states [12 ] , while quantum linear-algebra methods provide access to resolvent matrix elements [17 ] . Differences between expectation values evaluated in left and right eigenstates can also signal EPs [18 ] , illustrating how the relevant eigenstate information depends on the observable. In chemical spectroscopy of open molecular aggregates and lossy molecule-cavity systems [1 , 2 , 7 ] , the task is to resolve each mode’s contribution to a chosen optical channel, including modes with the same resonance frequency but different decay rates. Yet the real-frequency peak-ratio construction of generalised quantum computational spectroscopy does not directly separate such modes [19 ] . We show how a nonvariational algorithm built from unitary evolutions can extract both complex poles and the residues of a chosen optical channel.
Here we develop a quantum algorithm by embedding the shifted operator H z = H − z I H_z = H - z\mathbb{I} H z = H − z I , for an effective non-Hermitian Hamiltonian H H H and complex trial energy z z z , in a Hermitian dilation. At an eigenvalue, the dilation’s zero-energy subspace contains separate embeddings of the corresponding left and right eigenvectors. Gaussian filtering targets this subspace, and randomised sampling of unitary evolutions estimates the matrix elements needed to reconstruct the optical residues. We demonstrate the method numerically for molecular excitations coupled to plasmonic cavity modes, an established setting for strong-coupling spectroscopy [20 , 21 , 22 ] and molecular polaritonics [23 , 24 , 25 , 26 , 27 , 28 ] , including efforts to modify chemical dynamics [29 , 30 , 31 ] and energy transport [32 ] via strong coupling. For models with second- and third-order EPs, comparisons with exact diagonalisation validate the reconstructed complex poles and probe-dependent optical spectra. These results distinguish interference-induced doublets from genuine splitting of the resonance energies.
Chemical non-Hermitian Hamiltonian. —We consider N N N identical molecules uniformly coupled to one plasmon mode. The optically accessible single-excitation sector is spanned by ∣ P ⟩ = a † ∣ G ⟩ |P\rangle= a^\dagger|G\rangle ∣ P ⟩ = a † ∣ G ⟩ and the collective molecular bright state ∣ X ⟩ = N − 1 / 2 ∑ n σ n + ∣ G ⟩ |X\rangle= N^{-1/2}\sum_n \sigma_n^+|G\rangle ∣ X ⟩ = N − 1/2 ∑ n σ n + ∣ G ⟩ , where ∣ G ⟩ |G\rangle ∣ G ⟩ is the joint ground state. The resulting 2 × 2 2 \times2 2 × 2 model describes a collective bright-sector projection. The N − 1 N-1 N − 1 orthogonal molecular states are dark under the uniform-coupling and uniform-dipole assumptions. In the basis { ∣ P ⟩ , ∣ X ⟩ } \{|P\rangle, |X\rangle\} { ∣ P ⟩ , ∣ X ⟩} , the non-Hermitian Hamiltonian is
H e f f = ( ω ~ P g g ω ~ X ) , (1) H_{\mathrm{eff}} = \begin{pmatrix} \widetilde{\omega}_P & g \\ g & \widetilde{\omega}_X \end{pmatrix}, \tag*{(1)} H eff = ( ω P g g ω X ) , (1)
where ω ~ P = ω p − i κ 2 \widetilde{\omega}_P = \omega_p-i\frac{\kappa}{2} ω P = ω p − i 2 κ and ω ~ X = ω x − i γ 2 \widetilde{\omega}_X = \omega_x-i\frac{\gamma}{2} ω X = ω x − i 2 γ denote the complex bare energies incorporating the transition frequencies ω p \omega_p ω p and ω x \omega_x ω x alongside their respective decay linewidths κ \kappa κ and γ \gamma γ . Here g g g is the collective plasmon-exciton coupling, and κ / 2 \kappa/2 κ /2 and γ / 2 \gamma/2 γ /2 are the total optical-coherence decay rates. As shown in SM Sec. II, Lindblad recycling terms vanish on the optical-coherence block, making this effective Hamiltonian sufficient for linear absorption within the specified model. Diagonalisation of H e f f H_{\mathrm{eff}} H eff yields the exact complex eigenvalues
z ± = ω ~ P + ω ~ X 2 ± ( ω ~ P − ω ~ X 2 ) 2 + g 2 . (2) z_{\pm} = \frac{\widetilde{\omega}_P+\widetilde{\omega}_X}{2} \pm\sqrt{\left(\frac{\widetilde{\omega}_P-\widetilde{\omega}_X}{2}\right)^2+g^2}. \tag*{(2)} z ± = 2 ω P + ω X ± ( 2 ω P − ω X ) 2 + g 2 . (2)
On resonance ω p = ω x ≡ ω ‾ \omega_p=\omega_x\equiv\overline{\omega} ω p = ω x ≡ ω , this simplifies to
z ± = ω ‾ − i κ + γ 4 ± g 2 − ( κ − γ ) 2 16 . (3) z_{\pm} = \overline{\omega}-i\frac{\kappa+\gamma}{4}\pm\sqrt{g^2-\frac{(\kappa-\gamma)^2}{16}}. \tag*{(3)} z ± = ω − i 4 κ + γ ± g 2 − 16 ( κ − γ ) 2 . (3)
This reveals an EP at critical collective coupling g E P = ∣ κ − γ ∣ 4 g_{\mathrm{EP}}=\frac{|\kappa-\gamma|}{4} g EP = 4 ∣ κ − γ ∣ . Below this threshold (weak coupling), the bright poles have the same real part but different linewidths; above it (strong coupling), their real parts split while their linewidths coincide in the symmetric model. The EP is thereby controlled jointly by collective coherent coupling and dissipation, rather than by the resonance of an isolated molecule with an isolated plasmon.
Hermitian dilation and biorthogonal eigenspaces. —The chemical model shows that both complex poles and the associated left and right eigenvectors are needed. We now formulate a quantum algorithm for accessing them. Let H ≡ H e f f H\equiv H_{\mathrm{eff}} H ≡ H eff , with right and left eigenvectors defined as H ∣ r α ⟩ = z α ∣ r α ⟩ H|r_\alpha\rangle=z_\alpha|r_\alpha\rangle H ∣ r α ⟩ = z α ∣ r α ⟩ and H † ∣ l α ⟩ = z α ∗ ∣ l α ⟩ H^\dagger|l_\alpha\rangle=z_\alpha^*|l_\alpha\rangle H † ∣ l α ⟩ = z α ∗ ∣ l α ⟩ , respectively. For a complex trial energy z z z and shifted operator H z H_z H z , the smallest-singular-value landscape is f ( z ) ≡ σ min ( H z ) f(z)\equiv\sigma_{\min}(H_z) f ( z ) ≡ σ m i n ( H z ) [33 ] . The condition f ( z ) = 0 f(z)=0 f ( z ) = 0 holds iff z z z is an eigenvalue of H H H (Proposition 1). Moreover, under H → H + Δ H\to H+\Delta H → H + Δ , Weyl’s inequality gives ∣ σ min ( H + Δ − z I ) − σ min ( H − z I ) ∣ ≤ ∥ Δ ∥ |\sigma_{\min}(H+\Delta-zI)-\sigma_{\min}(H-zI)|\leq\|\Delta\| ∣ σ m i n ( H + Δ − z I ) − σ m i n ( H − z I ) ∣ ≤ ∥Δ∥ , where ∥ ⋅ ∥ \|\cdot\| ∥ ⋅ ∥ is the operator norm. Thus, finite-threshold regions of f ( z ) f(z) f ( z ) provide a perturbation-controlled measure of near-singular spectral structure, including near an EP. The corresponding Lipschitz-continuity result and stability bounds for the detection regions are proved in Proposition 2 and Corollary 1.
To access this landscape through unitary evolution, we introduce the Hermitian dilation
P ( z ) = ( 0 H z H z † 0 ) , \mathcal{P}(z)=\begin{pmatrix} 0 & H_z \\ H_z^\dagger& 0 \end{pmatrix}, P ( z ) = ( 0 H z † H z 0 ) ,
which acts on one additional two-level register and the original chemical Hilbert space. Its spectrum is spec P ( z ) = ± σ k ( H z ) k \operatorname{spec}\mathcal{P}(z)=\pm\sigma_k(H_z)_k spec P ( z ) = ± σ k ( H z ) k , so locating an eigenvalue of H H H is equivalent to locating a zero-energy eigenvalue of P ( z ) \mathcal{P}(z) P ( z ) (Proposition 3). The dilation also contains the eigenvectors. At z = z α z=z_\alpha z = z α , the embedded states ( ∣ l α ⟩ 0 ) \begin{pmatrix}|l_\alpha\rangle\\0\end{pmatrix} ( ∣ l α ⟩ 0 ) and ( 0 ∣ r α ⟩ ) \begin{pmatrix}0\\|r_\alpha\rangle\end{pmatrix} ( 0 ∣ r α ⟩ ) are zero modes of P ( z α ) \mathcal{P}(z_\alpha) P ( z α ) (Proposition 4). So for a simple eigenvalue, the corresponding kernel is ker P ( z α ) = span { ( ∣ l α ⟩ 0 ) , ( 0 ∣ r α ⟩ ) } \ker\mathcal{P}(z_\alpha)=\operatorname{span}\left\{\begin{pmatrix}|l_\alpha\rangle\\0\end{pmatrix},\begin{pmatrix}0\\|r_\alpha\rangle\end{pmatrix}\right\} ker P ( z α ) = span { ( ∣ l α ⟩ 0 ) , ( 0 ∣ r α ⟩ ) } . The dilation thus provides an orthogonal zero-energy subspace of a Hermitian operator containing both embedded members of the biorthogonal pair. This is the structure used below to reconstruct chemical observables. At the exact EP, the zero-mode construction remains well defined, but the simple-pole residue formula no longer applies; the response must be treated using a Jordan chain.
Gaussian filtering with unitary operations. —The zero-energy subspace of P ( z ) \mathcal{P}(z) P ( z ) can be isolated with a Gaussian spectral filter
G τ ( P ( z ) ) = exp [ − τ 2 2 P ( z ) 2 ] , (4) G_\tau(\mathcal{P}(z))=\exp\left[-\frac{\tau^2}{2}\mathcal{P}(z)^2\right], \tag*{(4)} G τ ( P ( z )) = exp [ − 2 τ 2 P ( z ) 2 ] , (4)
where τ \tau τ controls the spectral resolution (Lemma 1 and Proposition 5). Acting on an initial state ∣ Φ 0 ⟩ |\Phi_0\rangle ∣ Φ 0 ⟩ gives the filtered state ∣ Φ τ ( z ) ⟩ = G τ ( P ( z ) ) ∣ Φ 0 ⟩ |\Phi_\tau(z)\rangle=G_\tau(\mathcal{P}(z))|\Phi_0\rangle ∣ Φ τ ( z )⟩ = G τ ( P ( z )) ∣ Φ 0 ⟩ . Since P ( z ) \mathcal{P}(z) P ( z ) is Hermitian, the filter suppresses components with nonzero dilation energies and preserves the zero-energy subspace. For an isolated zero-energy subspace, the filtering error decreases as exp [ − τ 2 Δ σ 2 / 2 ] \exp[-\tau^2\Delta_\sigma^2/2] exp [ − τ 2 Δ σ 2 /2 ] , where Δ σ \Delta_\sigma Δ σ is the smallest nonzero singular value of H z H_z H z . Eq. (4) can be implemented using only unitary operations. Its Fourier representation is
G τ ( P ( z ) ) = ∫ − ∞ ∞ d x 2 π e − x 2 / 2 e − i τ x P ( z ) , (5) G_\tau(\mathcal{P}(z))=\int_{-\infty}^{\infty}\frac{dx}{\sqrt{2\pi}}e^{-x^2/2}e^{-i\tau x\mathcal{P}(z)}, \tag*{(5)} G τ ( P ( z )) = ∫ − ∞ ∞ 2 π d x e − x 2 /2 e − i τ x P ( z ) , (5)
where a finite quadrature gives G τ ( P ( z ) ) ≃ ∑ m w m e − i τ x m P ( z ) G_\tau(\mathcal{P}(z))\simeq\sum_m w_m e^{-i\tau x_m\mathcal{P}(z)} G τ ( P ( z )) ≃ ∑ m w m e − i τ x m P ( z ) (proof in SM Sec. III). Hence, randomised sampling of the quadrature points estimates observables associated with the discretised Gaussian filter through repeated unitary experiments, without non-unitary evolution under $H_z ({xref}Proposition 9`).
To locate the complex poles, we scan z z z over the complex plane and evaluate the filtered signal S τ ( z ) = ⟨ Φ τ ( z ) ∣ Φ τ ( z ) ⟩ S_\tau(z)=\langle\Phi_\tau(z)|\Phi_\tau(z)\rangle S τ ( z ) = ⟨ Φ τ ( z ) ∣ Φ τ ( z )⟩ . The signal is enhanced when H z H_z H z approaches singularity, and we use the maxima of S τ ( z ) S_\tau(z) S τ ( z ) as numerical estimates of the complex eigenvalues. In practice, several initial states can be used to ensure adequate overlap with each zero-energy sector (Figure 1 (a-Figure 1 c)). The filter isolates the left-right eigenspace; the embedded measurements we introduce below then separate the two components and convert the filtered subspace into biorthogonal observables.
Figure 1. Validation of the Gaussian-filtering algorithm around the EP. (a)–(c) Filtered signal S τ ( z ) S_{\tau}(z) S τ ( z ) in the complex plane (τ = 40 \tau= 40 τ = 40 , 250 samples) across three regimes: (a) weak coupling, (b) the EP, and (c) strong coupling, with exact eigenvalues marked by red crosses. (d) Re ( E ) \operatorname{Re}(E) Re ( E ) and (e) Im ( E ) \operatorname{Im}(E) Im ( E ) energy spectra versus g g g , highlighting Rabi splitting in the strong-coupling regime. Exact eigenvalues (blue) are compared against Gaussian (red squares), MC Gaussian (green crosses), and Trotterisation (orange triangles) reconstructions. (f) Average peak distance error versus g g g , calculated over 15 random initial states.
Biorthogonal reconstruction of optical response. —Locating the complex poles is only the first step: their contributions to an optical channel are determined by the associated residues. Away from the EP, where H H H is diagonalisable and ⟨ l α ∣ r α ⟩ ≠ 0 \langle l_{\alpha}|r_{\alpha}\rangle\ne0 ⟨ l α ∣ r α ⟩ = 0 , its resolvent has the expansion ( ω − H ) − 1 = ∑ α ∣ r α ⟩ ⟨ l α ∣ / [ ( ω − z α ) ⟨ l α ∣ r α ⟩ ] (\omega-H)^{-1} = \sum_{\alpha}|r_{\alpha}\rangle\langle l_{\alpha}|/[(\omega-z_{\alpha})\langle l_{\alpha}|r_{\alpha}\rangle] ( ω − H ) − 1 = ∑ α ∣ r α ⟩ ⟨ l α ∣/ [( ω − z α ) ⟨ l α ∣ r α ⟩] . For a preparation state ∣ s ⟩ |s\rangle ∣ s ⟩ and a detection state ⟨ d ∣ \langle d| ⟨ d ∣ , the residue associated with z α z_{\alpha} z α is hence
f α = ⟨ d ∣ r α ⟩ ⟨ l α ∣ s ⟩ ⟨ l α ∣ r α ⟩ , (6) f_{\alpha} = \frac{\langle d|r_{\alpha}\rangle\langle l_{\alpha}|s\rangle}{\langle l_{\alpha}|r_{\alpha}\rangle}, \tag*{(6)} f α = ⟨ l α ∣ r α ⟩ ⟨ d ∣ r α ⟩ ⟨ l α ∣ s ⟩ , (6)
where f α f_{\alpha} f α is generally complex and depends on the mode and the probe channel. For linear absorption, ∣ s ⟩ |s\rangle ∣ s ⟩ and ⟨ d ∣ \langle d| ⟨ d ∣ are generated by the transition dipole, with ∣ s ⟩ = μ ^ ∣ 0 ⟩ |s\rangle= \hat{\mu}|0\rangle ∣ s ⟩ = μ ^ ∣0 ⟩ and ⟨ d ∣ = ⟨ 0 ∣ μ ^ † \langle d| = \langle0|\hat{\mu}^{\dagger} ⟨ d ∣ = ⟨ 0∣ μ ^ † . The response is then reconstructed from the filtered poles and residues as A ( ω ) ∝ − Im ∑ α f α / ( ω − z α ) A(\omega) \propto-\operatorname{Im}\sum_{\alpha} f_{\alpha}/(\omega-z_{\alpha}) A ( ω ) ∝ − Im ∑ α f α / ( ω − z α ) (SM Sec. II). This pole-residue form makes the origin of interference explicit: two modes can produce a split or asymmetric line shape even when their real pole positions are not separated. For the matched dipole channel ⟨ d ∣ = ⟨ s ∣ \langle d| = \langle s| ⟨ d ∣ = ⟨ s ∣ , the optical residue generally differs from the right-state weight ∣ ⟨ r α ∣ s ⟩ ∣ 2 / ⟨ r α ∣ r α ⟩ |\langle r_{\alpha}|s\rangle|^{2}/\langle r_{\alpha}|r_{\alpha}\rangle ∣ ⟨ r α ∣ s ⟩ ∣ 2 / ⟨ r α ∣ r α ⟩ . The latter is an ordinary expectation value in a prepared right eigenstate, whereas the residue also contains the dual-basis excitation coefficient. These quantities coincide for an orthonormal eigenbasis. Although the complete right-eigenvector matrix determines the dual basis through its inverse, the dilation gives direct access to the required left-right matrix elements.
For a chemical operator O O O , suppose the off-diagonal embedded operators O = ( 0 O 0 0 ) O = \begin{pmatrix} 0 & O \\ 0 & 0 \end{pmatrix} O = ( 0 0 O 0 ) and I = ( 0 I 0 0 ) \mathcal{I} = \begin{pmatrix} 0 & \mathcal{I} \\ 0 & 0 \end{pmatrix} I = ( 0 0 I 0 ) , where I \mathcal{I} I is the off-diagonal embedding of the chemical identity.
identity. If the filtered state has nonzero overlap with both embedded components, their common preparation amplitudes cancel in the ratio
⟨ O ⟩ ⟨ I ⟩ = ⟨ l α ∣ O ∣ r α ⟩ ⟨ l α ∣ r α ⟩ . (7) \frac{\langle O\rangle}{\langle\mathcal{I}\rangle} = \frac{\langle l_{\alpha}|O|r_{\alpha}\rangle}{\langle l_{\alpha}|r_{\alpha}\rangle}. \tag*{(7)} ⟨ I ⟩ ⟨ O ⟩ = ⟨ l α ∣ r α ⟩ ⟨ l α ∣ O ∣ r α ⟩ . (7)
To make that explicit, consider the zero-energy component of the filtered state at z = z α z = z_{\alpha} z = z α , written as ∣ Ψ α , 0 ⟩ = a α ( ∣ l α , 0 ⟩ ) T + b α ( 0 , ∣ r α ⟩ ) T |\Psi_{\alpha,0}\rangle= a_{\alpha}(|l_{\alpha},0\rangle)^{\mathsf{T}} + b_{\alpha}(0,|r_{\alpha}\rangle)^{\mathsf{T}} ∣ Ψ α , 0 ⟩ = a α ( ∣ l α , 0 ⟩ ) T + b α ( 0 , ∣ r α ⟩ ) T , with a α ∗ b α ≠ 0 a_{\alpha}^{*}b_{\alpha} \ne0 a α ∗ b α = 0 . For the embedded operators O \mathcal{O} O and I \mathcal{I} I , direct block multiplication gives ⟨ Ψ α , 0 ∣ O ∣ Ψ α , 0 ⟩ = a α ∗ b α ⟨ l α ∣ O ∣ r α ⟩ \langle\Psi_{\alpha,0}|\mathcal{O}|\Psi_{\alpha,0}\rangle= a_{\alpha}^{*}b_{\alpha}\langle l_{\alpha}|O|r_{\alpha}\rangle ⟨ Ψ α , 0 ∣ O ∣ Ψ α , 0 ⟩ = a α ∗ b α ⟨ l α ∣ O ∣ r α ⟩ and ⟨ Ψ α , 0 ∣ I ∣ Ψ α , 0 ⟩ = a α ∗ b α ⟨ l α ∣ r α ⟩ \langle\Psi_{\alpha,0}|\mathcal{I}|\Psi_{\alpha,0}\rangle= a_{\alpha}^{*}b_{\alpha}\langle l_{\alpha}|r_{\alpha}\rangle ⟨ Ψ α , 0 ∣ I ∣ Ψ α , 0 ⟩ = a α ∗ b α ⟨ l α ∣ r α ⟩ . Their ratio hence gives Eq. (7) . Choosing O = ∣ s ⟩ ⟨ d ∣ O = |s\rangle\langle d| O = ∣ s ⟩ ⟨ d ∣ we get Eq. (6) , the pole residue entering the optical response. The off-diagonal operators are decomposed into Hermitian components using the dilation-register Pauli operators, so their real and imaginary parts can be measured with standard protocols. This converts the filtered zero-energy subspace into the biorthogonal observables required for non-Hermitian spectroscopy. Direct Gaussian filtering of a non-Hermitian operator is not an orthogonal spectral projection. The Fourier representation of the direct Gaussian filter is generally not realisable as a RLCU constructed from unitary evolutions.
Interference splitting versus Rabi splitting.
We now test the complete procedure by varying the molecular-plasmon coupling g g g through the EP of the plasmon-molecular excitation model. For each value of g g g , we obtain reference poles by direct diagonalisation and independently scan z z z using the filtered signal S τ ( z ) S_{\tau}(z) S τ ( z ) . The proposed pole estimates are obtained from the maxima of S τ ( z ) S_{\tau}(z) S τ ( z ) . These estimates reproduce both the real and imaginary parts of the exact poles [Figure 1 (d,Figure 1 e)], with average peak-distance errors shown in [Figure 1 (f)].
The pole trajectories establish the distinction between two spectroscopically different regimes. Below the EP, the two bright poles have the same real but different imaginary parts. Their different linewidths and complex residues can nevertheless interfere to produce two apparent maxima in the optical response. The resulting doublet is hence an interference feature rather than evidence of two distinct resonance energies. Above the EP, the real parts of the poles split by 2 g 2 − g E P 2 2\sqrt{g^2-g_{\rm EP}^2} 2 g 2 − g EP 2 while their linewidths become equal in the symmetric model. In this regime, the optical splitting is accompanied by eigenvalue separation and corresponds to Rabi-like splitting.
The response also depends on how the system is prepared and measured. Mixed, molecular- and plasmon-only probes produce different residues for the same pole structure. Figure 2 compares the reconstructed responses for different dipole channels, while Figure 3 summarizes the coupling dependence of the spectra and reconstruction errors. This channel dependence shows why eigenvalue detection alone cannot reproduce non-Hermitian spectroscopy. The combined pole and residue analysis provides a stricter interpretation of spectral splitting in dissipative molecular systems. A doublet should be identified as genuine Rabi splitting only when it is accompanied by separated real poles in the complex spectrum; otherwise, it may result from interference between modes with different decay rates. The Gaussian-filtering procedure supplies both pieces of information within a unitary-compatible framework.
Figure 2. Reconstructed effective absorption spectra A ( ω ) A(\omega) A ( ω ) under distinct coupling regimes and dipole configurations. (a),(b) Weak-coupling regime g = 0.023 g = 0.023 g = 0.023 for plasmon-dominant (μ P = 1 , μ X = 0 \mu_{\mathrm{P}} = 1,\mu_{\mathrm{X}} = 0 μ P = 1 , μ X = 0 ) and exciton-dominant (μ P = 0 , μ X = 1 \mu_{\mathrm{P}} = 0,\mu_{\mathrm{X}} = 1 μ P = 0 , μ X = 1 ) dipoles. (c),(d) Strong-coupling regime g = 0.043 g = 0.043 g = 0.043 displaying Rabi splitting for plasmon-dominant dipole. Spectra obtained via exact Hamiltonian diagonalisation are compared against exact Gaussian filtering, noiseless circuit MC simulation (τ = 35 \tau= 35 τ = 35 , 30 samples), and IBM quantum hardware results. Vertical dotted lines denote the real parts of the complex resonance energies.
Figure 3. Optical spectra and average reconstruction errors across coupling strength g g g averaged over 15 initial states. (a) Effective optical response A ( ω ) A(\omega) A ( ω ) at g = 0.023 g = 0.023 g = 0.023 (solid) and g = 0.043 g = 0.043 g = 0.043 (dashed). (b)–(d) Average reconstruction error as a function of g g g for filter times (b) τ = 40 \tau= 40 τ = 40 , (c) τ = 80 \tau= 80 τ = 80 , and (d) τ = 160 \tau= 160 τ = 160 using Gaussian process regression fits with 95% confidence bands, showing sharp error peaks near the EP.
At the EP, it is the decomposition into separate modes that becomes singular, not the damped optical response. As the two poles coalesce, individual residues can diverge while their contributions cancel in the total polarizability. Writing H = z E P I + J H = z_{\rm EP}I + J H = z EP I + J , with J 2 = 0 J^2 = 0 J 2 = 0 on the two-dimensional bright sector gives
α + , E P ( ω ) = − ⟨ s ∣ s ⟩ ω − z E P − ⟨ s ∣ J ∣ s ⟩ ( ω − z E P ) 2 , (8) \alpha_{+,\mathrm{EP}}(\omega) = -\frac{\langle s|s\rangle}{\omega-z_{\mathrm{EP}}} - \frac{\langle s|J|s\rangle}{(\omega-z_{\mathrm{EP}})^2}, \tag*{(8)} α + , EP ( ω ) = − ω − z EP ⟨ s ∣ s ⟩ − ( ω − z EP ) 2 ⟨ s ∣ J ∣ s ⟩ , (8)
where z E P = ω ˉ − i ( κ + γ ) / 4 z_{\rm EP} = \bar{\omega} - i(\kappa+\gamma)/4 z EP = ω ˉ − i ( κ + γ ) /4 in the linewidth convention. This expression is finite at real probe frequencies for positive damping and is equivalent to SM Eq. (48), where the rates denote coherence decay rates. The dipole-dependent coefficients show why an EP fixes neither the number of spectral maxima nor a universal squared-Lorentzian line shape. The cancellation also explains why reconstruction becomes sensitive near the EP, as reflected in Figure 3 . The dilation kernel remains well defined there, but ordinary zero-mode filtering alone does not supply the higher-order resolvent coefficient.
Beyond the two-mode model. —The construction is not restricted to a single bright molecular mode or a second-order EP. As a higher-order example Figure 4 , consider two plasmon modes with no direct intermode coupling and one collective molecular excitation (i.e., EP3) [34 ] . In the basis { ∣ P 1 ⟩ , ∣ X ⟩ , ∣ P 2 ⟩ } \{|P_1\rangle,|X\rangle,|P_2\rangle\} { ∣ P 1 ⟩ , ∣ X ⟩ , ∣ P 2 ⟩} , let
Figure 4. Higher-order validation of the Gaussian-filtering algorithm around the EP3. Extension of Figure 1 to Eq. (9) . Using the same filtering and reconstruction parameters, (a)–(c) show S τ ( z ) S_{\tau}(z) S τ ( z ) , respectively, in the weak-coupling, EP3, and strong-coupling regimes. (d), (e) The real and imaginary pole trajectories, including three-branch coalescence and bifurcation. (f) The average peak-distance error over 15 initial states.
H 3 = ( ω 0 − i ( Γ 0 − η ) g 0 g ω 0 + δ − i Γ 0 g 0 g ω 0 − i ( Γ 0 + η ) ) , (9) H_3 =
\begin{pmatrix}
\omega_0-i(\Gamma_0-\eta) & g & 0 \\
g & \omega_0+\delta-i\Gamma_0 & g \\
0 & g & \omega_0-i(\Gamma_0+\eta)
\end{pmatrix}, \tag*{(9)} H 3 = ω 0 − i ( Γ 0 − η ) g 0 g ω 0 + δ − i Γ 0 g 0 g ω 0 − i ( Γ 0 + η ) , (9)
where g = N g 0 g=\sqrt{N}g_0 g = N g 0 and Γ 0 > η > 0 \Gamma_0>\eta>0 Γ 0 > η > 0 . Here δ \delta δ is the molecular detuning and η \eta η is the linewidth imbalance between the two plasmon modes. Subtracting the common complex energy ω 0 − i Γ 0 \omega_0-i\Gamma_0 ω 0 − i Γ 0 gives the shifted core H 3 − ( ω 0 − i Γ 0 ) I 3 = H_3-(\omega_0-i\Gamma_0)I_3 = H 3 − ( ω 0 − i Γ 0 ) I 3 =
( i n g 0 g δ g 0 g − i n ) \begin{pmatrix} in & g & 0 \\ g & \delta& g \\ 0 & g & -in \end{pmatrix} in g 0 g δ g 0 g − in
, which is P T \mathcal{PT} P T -symmetric for real δ \delta δ , while the uniform shift ensures the physical system remains passive.
At resonance δ = 0 \delta= 0 δ = 0 , the poles are z 0 = ω 0 − i Γ 0 z_0 = \omega_0 - i\Gamma_0 z 0 = ω 0 − i Γ 0 and z ± = ω 0 − i Γ 0 ± 2 g 2 − η 2 z_{\pm} = \omega_0 - i\Gamma_0 \pm\sqrt{2g^2-\eta^2} z ± = ω 0 − i Γ 0 ± 2 g 2 − η 2 . Consequently, EP3 occurs at g E P 3 = η / 2 g_{\mathrm{EP3}} = \eta/\sqrt{2} g EP3 = η / 2 , where all three poles and their right eigenvectors coalesce at z E P 3 = ω 0 − i Γ 0 z_{\mathrm{EP3}} = \omega_0 - i\Gamma_0 z EP3 = ω 0 − i Γ 0 . The coalesced right eigenvector is proportional to ∣ P 1 ⟩ − i 2 ∣ X ⟩ = ∣ P 2 ⟩ |P_1\rangle- i\sqrt{2}|X\rangle= |P_2\rangle ∣ P 1 ⟩ − i 2 ∣ X ⟩ = ∣ P 2 ⟩ . For g < g E P 3 g < g_{\mathrm{EP3}} g < g EP3 , the three poles have three real part but distinct decay rates; for g > g E P 3 g > g_{\mathrm{EP3}} g > g EP3 , the pair z ± z_{\pm} z ± separates in real energy. At the EP3, the resolvent can contain poles of order up to three [35 ] . Away from EP3, the Hermitian-dilation, filtering, and residue-reconstruction procedures are unchanged. As shown in Figure 4 , the Gaussian-filtering algorithm remains numerically stable across the EP3, accurately recovering the complex spectrum without singular divergence near the coalescence point. Furthermore, the reconstructed response functions show that interference-induced line-shape splitting persists in the weak-coupling regime g < g E P 3 g < g_{\mathrm{EP3}} g < g EP3 prior to genuine real-energy pole separation, Figure 5 .
Figure 5. Extension of Figure 2 to EP3. (a),(b) Weak-coupling spectra with interference-induced doublets. (c),(d) Strong-coupling spectra with genuine pole splitting and channel-dependent line shapes. Vertical dotted lines mark the real parts of the complex poles.
Discussion. Spectroscopy in dissipative molecular systems requires distinguishing changes in resonance energies from changes in how modes contribute to an optical channel. Our protocol makes both quantities accessible through unitary quantum experiments by estimating complex poles and biorthogonal residues. The plasmon-molecule examples show why this matters: overlapping modes can produce a doublet without real-energy pole splitting, and changing the probe can reshape the spectrum without changing the poles. The reconstructed response thus connects non-Hermitian mode structure to observable line shapes. The second- and third-order EP examples show that this distinction also applies at higher-order coalescence. At an EP, the total damped response remains finite even as the individual residues are ill-conditioned. Operationally, the Hermitian dilation turns the non-Hermitian spectral problem into filtering and measurement with unitary evolutions. The filtered zero modes encode the left-right matrix elements needed to calculate optical residues, which combine with the poles to reconstruct channel-resolved spectra. Our circuit and calculations on IBM hardware verify this. Applying the protocol to realistic aggregates can support model-based spectral interpretation and probe design.
Acknowledgments. —P.M. is supported by the Grant Agency of the Czech Republic (Grant No. 24-10992S). A.Y. is supported by the UKRI Future Leaders Fellowship (Grant No. 10128920). We acknowledge the use of IBM Quantum services for this work. The views expressed are those of the authors and do not reflect the official policy or position of IBM or the IBM Quantum team.
Data availability. —The data are available upon reasonable request from the authors.
Supplemental Material for “Quantum computational spectroscopy of complex resonances and optical residues”
Mathematical Formulation of the Non-Hermitian Spectral Problem
The spectrum of a non-Hermitian operator is generally not a stable object: arbitrarily small perturbations may induce large spectral displacements for non-normal operators. A perturbation-stable characterisation is instead provided by the pseudospectrum, which can be characterised through the smallest singular value of the shifted operator H z = H − z I H_z = H - zI H z = H − z I . Specifically, the zeros of the smallest singular value recover the spectrum, while its sublevel sets define the pseudospectrum. Consequently, the non-Hermitian spectral problem can be reformulated as the construction of the singular-value landscape over the complex plane. This reformulation provides the mathematical foundation of the quantum algorithm developed in this work.
Spectrum and Pseudospectrum
The eigenvalues of a Hermitian operator are stable under small perturbations. In contrast, the spectrum of a non-Hermitian operator may change dramatically under arbitrarily small perturbations, particularly for non-normal matrices. Consequently, the spectrum alone does not provide a robust characterisation of the spectral properties of a non-Hermitian operator. To account for this sensitivity, one introduces the ε \varepsilon ε -pseudospectrum, defined as
Λ ε ( H ) = ⋃ ∥ Δ ∥ ≤ ε spec ( H + Δ ) , (10) \Lambda_{\varepsilon}(H) = \bigcup_{\lVert\Delta\rVert\le\varepsilon} \operatorname{spec}(H+\Delta), \tag*{(10)} Λ ε ( H ) = ∥ Δ ∥ ≤ ε ⋃ spec ( H + Δ ) , (10)
where the union is taken over all perturbations satisfying ∥ Δ ∥ ≤ ε \lVert\Delta\rVert\le\varepsilon ∥ Δ ∥ ≤ ε . Thus, Λ ε ( H ) \Lambda_{\varepsilon}(H) Λ ε ( H ) consists of all points in the complex plane that become eigenvalues after perturbations of size at most ε \varepsilon ε . The spectrum is recovered as the limiting case Λ 0 ( H ) = spec ( H ) \Lambda_0(H) = \operatorname{spec}(H) Λ 0 ( H ) = spec ( H ) . Although Eq. (10) provides a robust mathematical definition, it is not directly suitable for computation since it requires considering infinitely many perturbations. The following result shows that the entire pseudospectrum can instead be characterised by a single scalar quantity.
Proposition 1 (Singular-value characterisation). For every z ∈ C z \in\mathbb{C} z ∈ C
z ∈ Λ ε ( H ) ⟺ σ min ( H z ) ≤ ε , (11) z \in\Lambda_{\varepsilon}(H) \Longleftrightarrow\sigma_{\min}(H_z) \le\varepsilon, \tag*{(11)} z ∈ Λ ε ( H ) ⟺ σ m i n ( H z ) ≤ ε , (11) where σ min ( ⋅ ) \sigma_{\min}(\cdot) σ m i n ( ⋅ ) denotes the smallest singular value. Consequently
z ∈ spec ( H ) ⟺ σ min ( H z ) = 0. (12) z \in\operatorname{spec}(H) \Longleftrightarrow\sigma_{\min}(H_z) = 0. \tag*{(12)} z ∈ spec ( H ) ⟺ σ m i n ( H z ) = 0. (12)
Proof. For any matrix A A A , the Eckart-Young theorem states that the distance from A A A to the nearest singular matrix equals its smallest singular value σ min ( A ) = min det ( A + Δ ) = 0 ∥ Δ ∥ \sigma_{\min}(A) = \min_{\det(A+\Delta)=0} \lVert\Delta\rVert σ m i n ( A ) = min d e t ( A + Δ ) = 0 ∥ Δ ∥ . Applying this identity to H z H_z H z gives σ min ( H z ) = min det ( H z + Δ ) = 0 ∥ Δ ∥ \sigma_{\min}(H_z) = \min_{\det(H_z+\Delta)=0} \lVert\Delta\rVert σ m i n ( H z ) = min d e t ( H z + Δ ) = 0 ∥ Δ ∥ . Therefore, σ min ( H z ) ≤ ε \sigma_{\min}(H_z) \le\varepsilon σ m i n ( H z ) ≤ ε iff there exists a perturbation satisfying ∥ Δ ∥ ≤ ε \lVert\Delta\rVert\le\varepsilon ∥ Δ ∥ ≤ ε such that det ( H z + Δ ) = 0 \det(H_z + \Delta) = 0 det ( H z + Δ ) = 0 , which is precisely the definition of z ∈ Λ ε ( H ) z \in\Lambda_{\varepsilon}(H) z ∈ Λ ε ( H ) . Setting ε = 0 \varepsilon= 0 ε = 0 immediately yields Eq. (12) . □
Proposition 1 shows that both the spectrum and the pseudospectrum are completely determined by the smallest singular value of H z H_z H z . Consequently, the non-Hermitian eigenvalue problem can be reformulated as the construction of the scalar function f H ( z ) = σ min ( H z ) f_H(z) = \sigma_{\min}(H_z) f H ( z ) = σ m i n ( H z ) over the complex plane. The eigenvalues of H H H correspond to the zeros of f H f_H f H , while its sublevel sets represent the pseudospectrum. Therefore, rather than searching directly for the roots of det ( H z ) \det(H_z) det ( H z ) , the remainder of this work is devoted to constructing and measuring the singular-value landscape f H ( z ) f_H(z) f H ( z ) .
Stability of the Singular-Value Landscape
For the landscape f H ( z ) f_H(z) f H ( z ) to serve as the basis of a robust quantum algorithm, it must vary continuously under perturbations of the underlying non-Hermitian operator. The following establishes this stability.
Proposition 2 (Lipschitz continuity). Let H ′ = H + Δ H' = H + \Delta H ′ = H + Δ , where ∥ Δ ∥ ≤ δ \lVert\Delta\rVert\le\delta ∥ Δ ∥ ≤ δ . Then for every z ∈ C z \in\mathbb{C} z ∈ C
∣ f H ′ ( z ) − f H ( z ) ∣ ≤ δ . (13) \lvert f_{H'}(z) - f_H(z)\rvert\le\delta. \tag*{(13)} ∣ f H ′ ( z ) − f H ( z )∣ ≤ δ . (13)
Proof. By definition, f H ( z ) = σ min ( H z ) f_H(z) = \sigma_{\min}(H_z) f H ( z ) = σ m i n ( H z ) and f H ′ ( z ) = σ min ( H z + Δ ) f_{H'}(z) = \sigma_{\min}(H_z + \Delta) f H ′ ( z ) = σ m i n ( H z + Δ ) . Weyl’s perturbation theorem for singular values states that ∣ σ j ( A ) − σ j ( B ) ∣ ≤ ∥ A − B ∥ \lvert\sigma_j(A)-\sigma_j(B)\rvert\le\lVert A-B\rVert ∣ σ j ( A ) − σ j ( B )∣ ≤ ∥ A − B ∥ for every singular value. Choosing A = H z A = H_z A = H z and B = H z + Δ B = H_z + \Delta B = H z + Δ gives A − B = − Δ A-B = -\Delta A − B = − Δ . Therefore, ∣ f H ′ ( z ) − f H ( z ) ∣ = ∣ σ min ( A ) − σ min ( B ) ∣ ≤ ∥ Δ ∥ \lvert f_{H'}(z) - f_H(z)\rvert= \lvert\sigma_{\min}(A) - \sigma_{\min}(B)\rvert\le\lVert\Delta\rVert ∣ f H ′ ( z ) − f H ( z )∣ = ∣ σ m i n ( A ) − σ m i n ( B )∣ ≤ ∥ Δ ∥ , which proves Eq. (13) .
Proposition 2 shows that the singular-value landscape depends continuously on perturbations of the underlying operator. Consequently, small perturbations cannot produce large pointwise distortions of f H ( z ) f_H(z) f H ( z ) . In particular, if z ∗ z^* z ∗ is an isolated zero of f H f_H f H , then every sufficiently small perturbation produces a nearby minimum of the perturbed landscape. Hence, the localisation of eigenvalues can be formulated as the problem of tracking the minima of a stable scalar function over the complex plane. Unlike the eigenvalues themselves, which may exhibit arbitrarily large displacements near defective points, the singular-value landscape remains uniformly Lipschitz continuous with respect to operator perturbations. As a result, the quantum algorithm constructs a perturbation-stable object even when the spectrum is highly ill-conditioned.
Corollary 1 (Stability of detection regions). For ϵ ≥ 0 \epsilon\ge0 ϵ ≥ 0 , define the detection region
Z ϵ ( H ) = { z ∈ C : f H ( z ) ≤ ϵ } , \mathcal{Z}_{\epsilon}(H) = \{z \in\mathbb{C} : f_H(z) \le\epsilon\}, Z ϵ ( H ) = { z ∈ C : f H ( z ) ≤ ϵ } , where f H ( z ) = σ min ( H z ) f_H(z) = \sigma_{\min}(H_z) f H ( z ) = σ m i n ( H z ) . If ∥ Δ ∥ ≤ δ \lVert\Delta\rVert\le\delta ∥ Δ ∥ ≤ δ , then Z ϵ ( H ) ⊆ Z ϵ + δ ( H + Δ ) \mathcal{Z}_{\epsilon}(H) \subseteq\mathcal{Z}_{\epsilon+\delta}(H + \Delta) Z ϵ ( H ) ⊆ Z ϵ + δ ( H + Δ ) and Z ϵ ( H + Δ ) ⊆ Z ϵ + δ ( H ) \mathcal{Z}_{\epsilon}(H + \Delta) \subseteq\mathcal{Z}_{\epsilon+\delta}(H) Z ϵ ( H + Δ ) ⊆ Z ϵ + δ ( H ) .
Proof. Let z ∈ Z ϵ ( H ) z \in\mathcal{Z}_{\epsilon}(H) z ∈ Z ϵ ( H ) . Then f H ( z ) ≤ ϵ f_H(z) \le\epsilon f H ( z ) ≤ ϵ . By Proposition 2
f H + Δ ( z ) ≤ f H ( z ) + δ ≤ ϵ + δ , f_{H+\Delta}(z) \le f_H(z) + \delta\le\epsilon+ \delta, f H + Δ ( z ) ≤ f H ( z ) + δ ≤ ϵ + δ , which implies z ∈ Z ϵ + δ ( H + Δ ) z \in\mathcal{Z}_{\epsilon+\delta}(H + \Delta) z ∈ Z ϵ + δ ( H + Δ ) . Hence, Z ϵ ( H ) ⊆ Z ϵ + δ ( H + Δ ) \mathcal{Z}_{\epsilon}(H) \subseteq\mathcal{Z}_{\epsilon+\delta}(H + \Delta) Z ϵ ( H ) ⊆ Z ϵ + δ ( H + Δ ) . Interchanging the roles of H H H and H + Δ H + \Delta H + Δ yields Z ϵ ( H + Δ ) ⊆ Z ϵ + δ ( H ) \mathcal{Z}_{\epsilon}(H + \Delta) \subseteq\mathcal{Z}_{\epsilon+\delta}(H) Z ϵ ( H + Δ ) ⊆ Z ϵ + δ ( H ) , which completes the proof.
Hence, perturbations of size at most δ \delta δ modify the near-zero regions of the singular-value landscape only through an additive increase of the detection threshold. Although the eigenvalues of a non-normal operator may undergo large displacements under small perturbations, the singular-value landscape itself remains uniformly stable. Consequently, the quantum algorithm constructs a perturbation-stable detection landscape whose near-zero regions provide a robust characterisation of the spectrum and the pseudospectrum. Together, Proposition 1, Proposition 2 , and Corollary 1 establish that the non-Hermitian spectral problem can be reformulated as the construction of a perturbation-stable singular-value landscape. The remaining task is to develop an efficient quantum procedure for evaluating f H ( z ) = σ min ( H z ) f_H(z) = \sigma_{\min}(H_z) f H ( z ) = σ m i n ( H z ) .
The remaining challenge is that singular values are not directly accessible through unitary quantum evolution. Our strategy is thus to embed the shifted operator H z H_z H z into a Hermitian operator whose energy spectrum reproduces its singular values. Once this is achieved, the smallest singular value can be obtained by detecting the lowest energy of a Hermitian Hamiltonian. For each complex parameter z z z , define A ( z ) = H z − z I A(z) = H_z - zI A ( z ) = H z − z I , and introduce the Hermitian dilation
P ( z ) = ( 0 A ( z ) A ( z ) † 0 ) , \mathcal{P}(z) =
\begin{pmatrix}
0 & A(z) \\
A(z)^\dagger& 0
\end{pmatrix}, P ( z ) = ( 0 A ( z ) † A ( z ) 0 ) ,
which acts on the enlarged Hilbert space H ⊕ H \mathcal{H} \oplus\mathcal{H} H ⊕ H . Since P ( z ) = P ( z ) † \mathcal{P}(z) = \mathcal{P}(z)^\dagger P ( z ) = P ( z ) † , the family U ( t ) = exp ( − i t P ( z ) ) U(t) = \exp\left(-it\mathcal{P}(z)\right) U ( t ) = exp ( − i t P ( z ) ) is unitary. Consequently, the non-Hermitian spectral problem has been converted into a Hamiltonian-simulation problem for a Hermitian operator. The key property of this construction is that the spectrum of P ( z ) \mathcal{P}(z) P ( z ) is determined entirely by the singular values of A ( z ) A(z) A ( z ) .
Proposition 3 (Hermitian dilation). For every z ∈ C z \in\mathbb{C} z ∈ C
spec ( P ( z ) ) = { ± σ k ( H z ) } k . (14) \operatorname{spec}(\mathcal{P}(z)) = \{\pm\sigma_k(H_z)\}_k. \tag*{(14)} spec ( P ( z )) = { ± σ k ( H z ) } k . (14) As a result
min λ ∈ spec ( P ( z ) ) ∣ λ ∣ = σ min ( H z ) . (15) \min_{\lambda\in\operatorname{spec}(\mathcal{P}(z))} \lvert\lambda\rvert= \sigma_{\min}(H_z). \tag*{(15)} λ ∈ spec ( P ( z )) min ∣ λ ∣ = σ m i n ( H z ) . (15)
and thus
z ∈ spec ( H ) ⟺ P ( z ) possesses a zero-energy eigenstate . (16) z \in\operatorname{spec}(H) \Longleftrightarrow\mathcal{P}(z)\text{ possesses a zero-energy eigenstate}. \tag*{(16)} z ∈ spec ( H ) ⟺ P ( z ) possesses a zero-energy eigenstate . (16)
Proof. Let
P ( z ) ( u v ) = λ ( u v ) , \mathcal{P}(z)\begin{pmatrix}u\\v\end{pmatrix}=\lambda\begin{pmatrix}u\\v\end{pmatrix}, P ( z ) ( u v ) = λ ( u v ) ,
where λ \lambda λ is an eigenvalue of P ( z ) \mathcal{P}(z) P ( z ) . From the definition of P ( z ) \mathcal{P}(z) P ( z ) , the eigenvalue equation is equivalent to the coupled system A ( z ) v = λ u A(z)v=\lambda u A ( z ) v = λ u , A ( z ) † u = λ v A(z)^\dagger u=\lambda v A ( z ) † u = λ v . Applying A ( z ) † A(z)^\dagger A ( z ) † to the first equation and substituting the second yields A ( z ) † A ( z ) v = λ A ( z ) † u = λ 2 v A(z)^\dagger A(z)v=\lambda A(z)^\dagger u=\lambda^2v A ( z ) † A ( z ) v = λ A ( z ) † u = λ 2 v . Thus, λ 2 \lambda^2 λ 2 is an eigenvalue of A ( z ) † A ( z ) A(z)^\dagger A(z) A ( z ) † A ( z ) . By definition, the singular values of A ( z ) A(z) A ( z ) are the positive square roots of the eigenvalues of A ( z ) † A ( z ) A(z)^\dagger A(z) A ( z ) † A ( z ) . Hence, λ 2 = σ k ( A ( z ) ) 2 \lambda^2=\sigma_k(A(z))^2 λ 2 = σ k ( A ( z ) ) 2 , which implies λ = ± σ k ( A ( z ) ) \lambda=\pm\sigma_k(A(z)) λ = ± σ k ( A ( z )) . This proves Eq. (14) . Eq. (15) follows immediately, since the eigenvalues of P ( z ) \mathcal{P}(z) P ( z ) occur in symmetric pairs ± σ k ( A ( z ) ) \pm\sigma_k(A(z)) ± σ k ( A ( z )) , and so the smallest absolute eigenvalue equals σ min ( A ( z ) ) = σ min ( H z ) \sigma_{\min}(A(z))=\sigma_{\min}(H_z) σ m i n ( A ( z )) = σ m i n ( H z ) . Finally
σ min ( H z ) = 0 ⟺ H z is singular ⟺ z ∈ spec ( H ) , \sigma_{\min}(H_z)=0 \Longleftrightarrow H_z\text{ is singular} \Longleftrightarrow z\in\operatorname{spec}(H), σ m i n ( H z ) = 0 ⟺ H z is singular ⟺ z ∈ spec ( H ) ,
which is equivalent to the existence of a zero-energy eigenstate of P ( z ) \mathcal{P}(z) P ( z ) by Eq. (15) . This establishes Eq. (16) .
Proposition 3 converts the original non-Hermitian eigenvalue problem into the detection of zero-energy states of the Hermitian family P ( z ) \mathcal{P}(z) P ( z ) . Constructing the singular-value landscape is thereby equivalent to measuring the lowest-energy branch of a Hermitian Hamiltonian, enabling the use of unitary quantum simulation techniques.
Recovery of the Biorthogonal Eigenspace
Locating the eigenvalues alone, however, is insufficient for evaluating physical observables. Unlike the Hermitian case, expectation values are determined by the biorthogonal expression
⟨ O ⟩ j = ⟨ L j ∣ O ∣ R j ⟩ ⟨ L j ∣ R j ⟩ , (17) \langle O\rangle_j=\frac{\langle L_j|O|R_j\rangle}{\langle L_j|R_j\rangle}, \tag*{(17)} ⟨ O ⟩ j = ⟨ L j ∣ R j ⟩ ⟨ L j ∣ O ∣ R j ⟩ , (17)
which depends simultaneously on the left and right eigenvectors. The following proposition shows that these vectors are naturally embedded in the zero-energy subspace of the Hermitian dilation.
Proposition 4 (Zero modes and biorthogonal eigenvectors). Let λ j \lambda_j λ j be an eigenvalue of H H H , and define A ( λ j ) = H − λ j I A(\lambda_j)=H-\lambda_j I A ( λ j ) = H − λ j I . Then the embedded states
∣ χ j L ⟩ = ( ∣ L j ⟩ 0 ) , ∣ χ j R ⟩ = ( 0 ∣ R j ⟩ ) , (18) |\chi_j^L\rangle=\begin{pmatrix}|L_j\rangle\\0\end{pmatrix},\qquad
|\chi_j^R\rangle=\begin{pmatrix}0\\|R_j\rangle\end{pmatrix}, \tag*{(18)} ∣ χ j L ⟩ = ( ∣ L j ⟩ 0 ) , ∣ χ j R ⟩ = ( 0 ∣ R j ⟩ ) , (18) span the kernel of the Hermitian dilation
ker P ( λ j ) = span { ∣ χ j L ⟩ , ∣ χ j R ⟩ } . (19) \operatorname{ker}\mathcal{P}(\lambda_j)=\operatorname{span}\left\{|\chi_j^L\rangle,|\chi_j^R\rangle\right\}. \tag*{(19)} ker P ( λ j ) = span { ∣ χ j L ⟩ , ∣ χ j R ⟩ } . (19) Proof. Since H ∣ R j ⟩ = λ j ∣ R j ⟩ H|R_j\rangle=\lambda_j|R_j\rangle H ∣ R j ⟩ = λ j ∣ R j ⟩ and ⟨ L j ∣ H = λ j ⟨ L j ∣ \langle L_j|H=\lambda_j\langle L_j| ⟨ L j ∣ H = λ j ⟨ L j ∣ , the shifted operator satisfies A ( λ j ) ∣ R j ⟩ = 0 A(\lambda_j)|R_j\rangle=0 A ( λ j ) ∣ R j ⟩ = 0 , A ( λ j ) † ∣ L j ⟩ = 0 A(\lambda_j)^\dagger|L_j\rangle=0 A ( λ j ) † ∣ L j ⟩ = 0 . Substituting these relations into the definition of the Hermitian dilation gives
P ( λ j ) ∣ χ j R ⟩ = ( A ( λ j ) ∣ R j ⟩ 0 ) = 0 , \mathcal{P}(\lambda_j)|\chi_j^R\rangle=
\begin{pmatrix}A(\lambda_j)|R_j\rangle\\0\end{pmatrix}=0, P ( λ j ) ∣ χ j R ⟩ = ( A ( λ j ) ∣ R j ⟩ 0 ) = 0 , and
P ( λ j ) ∣ χ j L ⟩ = ( 0 A ( λ j ) † ∣ L j ⟩ ) = 0 , \mathcal{P}(\lambda_j)|\chi_j^L\rangle=
\begin{pmatrix}0\\A(\lambda_j)^\dagger|L_j\rangle\end{pmatrix}=0, P ( λ j ) ∣ χ j L ⟩ = ( 0 A ( λ j ) † ∣ L j ⟩ ) = 0 , establishing that both embedded vectors belong to the zero-energy subspace. Conversely, every zero mode of P ( λ j ) \mathcal{P}(\lambda_j) P ( λ j ) satisfies A ( λ j ) ∣ v ⟩ = 0 A(\lambda_j)|v\rangle=0 A ( λ j ) ∣ v ⟩ = 0 and A ( λ j ) † ∣ u ⟩ = 0 A(\lambda_j)^\dagger|u\rangle=0 A ( λ j ) † ∣ u ⟩ = 0 , and therefore is a linear combination of the embedded right and left eigenvectors. Hence ker P ( λ j ) = span { ∣ χ j L ⟩ , ∣ χ j R ⟩ } \operatorname{ker}\mathcal{P}(\lambda_j)=\operatorname{span}\left\{|\chi_j^L\rangle,|\chi_j^R\rangle\right\} ker P ( λ j ) = span { ∣ χ j L ⟩ , ∣ χ j R ⟩ } .
Proposition 4 establishes that the Hermitian dilation preserves not only the spectral information encoded by the singular-value landscape, but also the complete biorthogonal eigenspace of the original non-Hermitian operator. Consequently, isolating the zero-energy subspace of P ( λ j ) \mathcal{P}(\lambda_j) P ( λ j ) simultaneously identifies the target eigenvalue and provides direct access to the left and right eigenvectors required for evaluating Eq. (17) . So the Hermitian dilation preserves not only the eigenvalue information encoded in the singular-value landscape, but also the complete biorthogonal eigenspace required for evaluating non-Hermitian expectation values. This observation forms the basis of the Gaussian spectral-filtering algorithm developed below.
Linear Absorption Spectroscopy
We consider a single plasmon mode coupled to N N N identical two-level molecules, described by the Tavis–Cummings (TC) Hamiltonian under the rotating-wave approximation
H T C = ω p a † a + ω x ∑ n = 1 N σ n + σ n − + g 0 ∑ n = 1 N ( a † σ n − + a σ n + ) , (20) H_{\mathrm{TC}}=\omega_{\mathrm{p}}a^\dagger a+\omega_{\mathrm{x}}\sum_{n=1}^{N}\sigma_n^+\sigma_n^-+g_0\sum_{n=1}^{N}\left(a^\dagger\sigma_n^-+a\sigma_n^+\right), \tag*{(20)} H TC = ω p a † a + ω x n = 1 ∑ N σ n + σ n − + g 0 n = 1 ∑ N ( a † σ n − + a σ n + ) , (20)
where a † a^\dagger a † creates a plasmon, σ n + \sigma_n^+ σ n + excites molecule n n n , and g 0 g_0 g 0 denotes the uniform, real molecule-plasmon coupling strength. For a fixed linear polarisation, the transition-dipole operator u = u + + u − u=u^+ + u^- u = u + + u − is given by
u + = u p a † + u 0 ∑ n = 1 N σ n + , u − = ( u + ) † , (21) u^+=u_{\mathrm{p}}a^\dagger+u_0\sum_{n=1}^{N}\sigma_n^+,\qquad u^-=(u^+)^\dagger, \tag*{(21)} u + = u p a † + u 0 n = 1 ∑ N σ n + , u − = ( u + ) † , (21)
where u p u_{\mathrm{p}} u p and u 0 u_0 u 0 are the transition-dipole amplitudes of the plasmon mode and each molecule, respectively, projected onto the polarisation direction. The weak probe field E ( t ) E(t) E ( t ) couples to the transition dipoles through H i n t ( t ) = − u E ( t ) H_{\mathrm{int}}(t)=-uE(t) H int ( t ) = − u E ( t ) . The system density matrix evolves by
ρ ˙ ( t ) = L ρ ( t ) + i E ( t ) [ u , ρ ( t ) ] , (22) \dot{\rho}(t)=\mathcal{L}\rho(t)+iE(t)[u,\rho(t)], \tag*{(22)} ρ ˙ ( t ) = L ρ ( t ) + i E ( t ) [ u , ρ ( t )] , (22)
where L \mathcal{L} L is the field-free Liouvillian with Markovian dissipation
L ρ = − i [ H T C , ρ ] + D [ L p , p ] ρ + D [ L ϕ , p ] ρ + ∑ n = 1 N ( D [ L p , n ] ρ + D [ L ϕ , n ] ρ ) , (23) \mathcal{L}\rho=-i[H_{\mathrm{TC}},\rho]+\mathcal{D}[L_{\mathrm{p,p}}]\rho+\mathcal{D}[L_{\phi,\mathrm{p}}]\rho+\sum_{n=1}^{N}\left(\mathcal{D}[L_{\mathrm{p,n}}]\rho+\mathcal{D}[L_{\phi,\mathrm{n}}]\rho\right), \tag*{(23)} L ρ = − i [ H TC , ρ ] + D [ L p , p ] ρ + D [ L ϕ , p ] ρ + n = 1 ∑ N ( D [ L p , n ] ρ + D [ L ϕ , n ] ρ ) , (23)
where D [ L ] ρ = L ρ L † − 1 2 { L † L , ρ } \mathcal{D}[L]\rho=L\rho L^\dagger-\frac{1}{2}\{L^\dagger L,\rho\} D [ L ] ρ = L ρ L † − 2 1 { L † L , ρ } is the dissipator. The corresponding Lindblad operators are
L p , p = κ p a , L p , n = γ p σ n − , (24) L_{\mathrm{p,p}}=\sqrt{\kappa_{\mathrm{p}}}\,a,\qquad L_{\mathrm{p,n}}=\sqrt{\gamma_{\mathrm{p}}}\,\sigma_n^-, \tag*{(24)} L p , p = κ p a , L p , n = γ p σ n − , (24)
L ϕ , p = 2 κ ϕ a † a , L ϕ , n = 2 γ ϕ σ n + σ n − . (25) L_{\phi,\mathrm{p}}=\sqrt{2\kappa_{\phi}}\,a^\dagger a,\qquad L_{\phi,\mathrm{n}}=\sqrt{2\gamma_{\phi}}\,\sigma_n^+\sigma_n^-. \tag*{(25)} L ϕ , p = 2 κ ϕ a † a , L ϕ , n = 2 γ ϕ σ n + σ n − . (25)
The rates κ p \kappa_{\mathrm{p}} κ p and γ p \gamma_{\mathrm{p}} γ p describe plasmon and molecular population decay through radiative and nonradiative relaxation. The rates κ ϕ \kappa_{\phi} κ ϕ and γ ϕ \gamma_{\phi} γ ϕ describe pure dephasing of the corresponding ground-to-single-excitation coherences. Pure dephasing does not directly change populations in the local excitation basis.
In the weak-field limit, we expand the density matrix as ρ ( t ) = ρ 0 + ρ ( 1 ) ( t ) + O ( E 2 ) \rho(t)=\rho_0+\rho^{(1)}(t)+O(E^2) ρ ( t ) = ρ 0 + ρ ( 1 ) ( t ) + O ( E 2 ) , where the stationary reference ρ 0 \rho_0 ρ 0 is the collective ground state by neglecting thermal excitation, ρ 0 = ∣ G ⟩ ⟨ G ∣ \rho_0=|G\rangle\langle G| ρ 0 = ∣ G ⟩ ⟨ G ∣ with ∣ G ⟩ = ∣ 0 P , g 1 , g 2 , … , g N ⟩ |G\rangle=|0_{\mathrm{P}},g_1,g_2,\ldots,g_N\rangle ∣ G ⟩ = ∣ 0 P , g 1 , g 2 , … , g N ⟩ . The first-order correction is
ρ ( 1 ) ( t ) = i ∫ − ∞ t d t ′ E ( t ′ ) G ( t − t ′ ) [ u , ρ 0 ] , G ( t ) = e L t . (26) \rho^{(1)}(t)=i\int_{-\infty}^{t}dt'\,E(t')\,\mathcal{G}(t-t')[u,\rho_0],\qquad\mathcal{G}(t)=e^{\mathcal{L}t}. \tag*{(26)} ρ ( 1 ) ( t ) = i ∫ − ∞ t d t ′ E ( t ′ ) G ( t − t ′ ) [ u , ρ 0 ] , G ( t ) = e L t . (26)
Here the field acts at time t ′ t' t ′ , and the resulting perturbation evolves under G ( t − t ′ ) \mathcal{G}(t-t') G ( t − t ′ ) until the observation time t t t . The induced dipole of the coupled plasmon-molecule system is
p ( 1 ) ( t ) = Tr { μ ρ ( 1 ) ( t ) } = ∫ − ∞ t d t ′ R ( 1 ) ( t − t ′ ) E ( t ′ ) = ∫ 0 ∞ d τ R ( 1 ) ( τ ) E ( t − τ ) , (27) p^{(1)}(t)=\operatorname{Tr}\{\mu\rho^{(1)}(t)\}=\int_{-\infty}^{t}dt'\,R^{(1)}(t-t')E(t')=\int_{0}^{\infty}d\tau\,R^{(1)}(\tau)E(t-\tau), \tag*{(27)} p ( 1 ) ( t ) = Tr { μ ρ ( 1 ) ( t )} = ∫ − ∞ t d t ′ R ( 1 ) ( t − t ′ ) E ( t ′ ) = ∫ 0 ∞ d τ R ( 1 ) ( τ ) E ( t − τ ) , (27)
in which the causal linear response function
R ( 1 ) ( t ) = i θ ( t ) Tr { μ G ( t ) [ μ , ρ 0 ] } , (28) R^{(1)}(t)=i\theta(t)\operatorname{Tr}\{\mu\mathcal{G}(t)[\mu,\rho_0]\}, \tag*{(28)} R ( 1 ) ( t ) = i θ ( t ) Tr { μ G ( t ) [ μ , ρ 0 ]} , (28)
with θ ( t ) \theta(t) θ ( t ) as the Heaviside step function describes the dipole response at a time t t t after a field interaction. The commutator describes the creation of an optical coherence by the probe, the propagator describes its subsequent evolution, and the final dipole operator determines its contribution to the induced dipole. Write the linear response in terms of dipole-dipole correlation function R ( 1 ) ( t ) = i θ ( t ) ( C ( t ) − C ( t ) ∗ ) R^{(1)}(t)=i\theta(t)(C(t)-C(t)^*) R ( 1 ) ( t ) = i θ ( t ) ( C ( t ) − C ( t ) ∗ ) , in which
C ( t ) = ⟨ μ ( t ) μ ( 0 ) ⟩ = Tr { μ − G ( t ) [ μ + ρ 0 ] } . (29) C(t)=\langle\mu(t)\mu(0)\rangle=\operatorname{Tr}\{\mu^-\mathcal{G}(t)[\mu^+\rho_0]\}. \tag*{(29)} C ( t ) = ⟨ μ ( t ) μ ( 0 )⟩ = Tr { μ − G ( t ) [ μ + ρ 0 ]} . (29)
In the frequency domain
p ( 1 ) ( ω ) = ∫ − ∞ ∞ d t e i ω t p ( 1 ) ( t ) = α ( ω ) E ( ω ) . (30) p^{(1)}(\omega)=\int_{-\infty}^{\infty}dt\,e^{i\omega t}p^{(1)}(t)=\alpha(\omega)E(\omega). \tag*{(30)} p ( 1 ) ( ω ) = ∫ − ∞ ∞ d t e iω t p ( 1 ) ( t ) = α ( ω ) E ( ω ) . (30)
where E ( ω ) E(\omega) E ( ω ) is the Fourier transform of the applied electric field and α ( ω ) \alpha(\omega) α ( ω ) is the complex polarizability that determines the amplitude and phase of the induced dipole relative to the applied field
α ( ω ) = ∫ 0 ∞ d τ e i ω τ R ( 1 ) ( τ ) . (31) \alpha(\omega)=\int_{0}^{\infty}d\tau\,e^{i\omega\tau}R^{(1)}(\tau). \tag*{(31)} α ( ω ) = ∫ 0 ∞ d τ e iω τ R ( 1 ) ( τ ) . (31)
For a monochromatic field E ( t ) = Re [ E ω e − i ω t ] E(t)=\operatorname{Re}[E_{\omega}e^{-i\omega t}] E ( t ) = Re [ E ω e − iω t ] , the cycle-averaged power transferred from the field to the system is
P ‾ ( ω ) = ω 2 ∣ E ω ∣ 2 Im α ( ω ) . (32) \overline{P}(\omega)=\frac{\omega}{2}|E_{\omega}|^{2}\operatorname{Im}\alpha(\omega). \tag*{(32)} P ( ω ) = 2 ω ∣ E ω ∣ 2 Im α ( ω ) . (32)
We report the reduced absorption line shape A ( ω ) = Im α ( ω ) A(\omega)=\operatorname{Im}\alpha(\omega) A ( ω ) = Im α ( ω ) . Near positive-frequency optical resonances, the resonant contribution α + ( ω ) = i ∫ 0 ∞ d t e i ω t C ( t ) \alpha_{+}(\omega)=i\int_{0}^{\infty}dt\,e^{i\omega t}C(t) α + ( ω ) = i ∫ 0 ∞ d t e iω t C ( t ) gives
A ( ω ) ≃ Im α + ( ω ) = Re ∫ 0 ∞ d t e i ω t C ( t ) , (33) A(\omega)\simeq\operatorname{Im}\alpha_{+}(\omega)=\operatorname{Re}\int_{0}^{\infty}dt\,e^{i\omega t}C(t), \tag*{(33)} A ( ω ) ≃ Im α + ( ω ) = Re ∫ 0 ∞ d t e iω t C ( t ) , (33)
where the approximation neglects the counter-rotating contribution. To evaluate C ( t ) C(t) C ( t ) , we define the effective Hamiltonian
H e f f = H T C − i 2 ∑ ν L ν † L ν , (34) H_{\mathrm{eff}}=H_{\mathrm{TC}}-\frac{i}{2}\sum_{\nu}L_{\nu}^{\dagger}L_{\nu}, \tag*{(34)} H eff = H TC − 2 i ν ∑ L ν † L ν , (34)
where ν \nu ν runs over all Lindblad operators. Here, H e f f H_{\mathrm{eff}} H eff acts on the full system Hilbert space. H e f f H_{\mathrm{eff}} H eff does not by itself generate the complete dynamics, because the recycling terms L ν ρ L ν † L_{\nu}\rho L_{\nu}^{\dagger} L ν ρ L ν † play an important role in excited-state population. However, these recycling terms vanish exactly for the optical coherences entering C ( t ) C(t) C ( t ) , i.e., L ν μ + ρ 0 L ν † = 0 L_{\nu}\mu_{+}\rho_{0}L_{\nu}^{\dagger}=0 L ν μ + ρ 0 L ν † = 0 . Thus, the effective Hamiltonian propagates the optical coherence exactly within the Lindblad model.
We define ∣ P ⟩ = a † ∣ G ⟩ |P\rangle=a^{\dagger}|G\rangle ∣ P ⟩ = a † ∣ G ⟩ as one plasmon excitation, and ∣ n ⟩ = σ n + ∣ G ⟩ |n\rangle=\sigma_{n}^{+}|G\rangle ∣ n ⟩ = σ n + ∣ G ⟩ as one excitation localised on molecule n n n . Thus molecular bright state is ∣ X ⟩ = 1 N ∑ n = 1 N ∣ n ⟩ |X\rangle=\frac{1}{\sqrt{N}}\sum_{n=1}^{N}|n\rangle ∣ X ⟩ = N 1 ∑ n = 1 N ∣ n ⟩ . The remaining N − 1 N-1 N − 1 molecular dark states are orthogonal to ∣ X ⟩ |X\rangle ∣ X ⟩ , decouple from the plasmon, and are not directly excited by the uniform molecular dipole under identical-molecule assumption. The restriction of H e f f H_{\mathrm{eff}} H eff to this optically active single-excitation subspace is
H e f f ( e ) = ( ω p − i κ / 2 g g ω x − i γ / 2 ) , (35) H_{\mathrm{eff}}^{(e)}=
\begin{pmatrix}
\omega_{p}-i\kappa/2 & g\\
g & \omega_{x}-i\gamma/2
\end{pmatrix}, \tag*{(35)} H eff ( e ) = ( ω p − iκ /2 g g ω x − iγ /2 ) , (35)
where g = N g 0 g=\sqrt{N}g_{0} g = N g 0 is the collective plasmon-exciton coupling, κ = κ p + 2 κ ϕ \kappa=\kappa_{p}+2\kappa_{\phi} κ = κ p + 2 κ ϕ , and γ = γ p + 2 γ ϕ \gamma=\gamma_{p}+2\gamma_{\phi} γ = γ p + 2 γ ϕ . Thus the optical-coherence decay rates are κ / 2 \kappa/2 κ /2 and γ / 2 \gamma/2 γ /2 , as in the main text. The transition dipole is rewritten
μ + = μ P ∣ P ⟩ ⟨ G ∣ + μ X ∣ X ⟩ ⟨ G ∣ , (36) \mu^{+}=\mu_{P}|P\rangle\langle G|+\mu_{X}|X\rangle\langle G|, \tag*{(36)} μ + = μ P ∣ P ⟩ ⟨ G ∣ + μ X ∣ X ⟩ ⟨ G ∣ , (36)
where μ X = N u 0 \mu_{X}=\sqrt{N}u_{0} μ X = N u 0 . Thus, the dipole correlation function and the resonant optical response reduce to
C ( t ) = ⟨ G ∣ μ − e − i H e f f ( e ) t μ + ∣ G ⟩ , α + ( ω ) = − ⟨ G ∣ μ − ( ω I e − H e f f ( e ) ) − 1 μ + ∣ G ⟩ , (37) C(t)=\langle G|\mu^{-}e^{-iH_{\mathrm{eff}}^{(e)}t}\mu^{+}|G\rangle,\qquad
\alpha_{+}(\omega)=-\langle G|\mu^{-}\left(\omega I_{e}-H_{\mathrm{eff}}^{(e)}\right)^{-1}\mu^{+}|G\rangle, \tag*{(37)} C ( t ) = ⟨ G ∣ μ − e − i H eff ( e ) t μ + ∣ G ⟩ , α + ( ω ) = − ⟨ G ∣ μ − ( ω I e − H eff ( e ) ) − 1 μ + ∣ G ⟩ , (37)
where I e I_{e} I e is the identity in the plasmon–bright-state subspace. This expression does not require H e f f ( e ) H_{\mathrm{eff}}^{(e)} H eff ( e ) to be diagonalisable. For distinct eigenvalues λ ± \lambda_{\pm} λ ± , we consider the spectral decomposition of H e f f ( e ) H_{\mathrm{eff}}^{(e)} H eff ( e )
H e f f ( e ) ∣ r α ⟩ = λ α ∣ r α ⟩ , ⟨ l α ∣ H e f f ( e ) = λ α ⟨ l α ∣ . (38) H_{\mathrm{eff}}^{(e)}|r_{\alpha}\rangle=\lambda_{\alpha}|r_{\alpha}\rangle,\qquad
\langle l_{\alpha}|H_{\mathrm{eff}}^{(e)}=\lambda_{\alpha}\langle l_{\alpha}|. \tag*{(38)} H eff ( e ) ∣ r α ⟩ = λ α ∣ r α ⟩ , ⟨ l α ∣ H eff ( e ) = λ α ⟨ l α ∣. (38)
The left eigenvectors form the dual basis and are not generally the Hermitian adjoints of the right eigenvectors. We choose the biorthogonal normalisation
⟨ l α ∣ r β ⟩ = δ α β , ∑ α ∣ r α ⟩ ⟨ l α ∣ = I e . (39) \langle l_{\alpha}|r_{\beta}\rangle=\delta_{\alpha\beta},\qquad
\sum_{\alpha}|r_{\alpha}\rangle\langle l_{\alpha}|=I_{e}. \tag*{(39)} ⟨ l α ∣ r β ⟩ = δ α β , α ∑ ∣ r α ⟩ ⟨ l α ∣ = I e . (39)
The response can then be decomposed as
C ( t ) = ∑ α = ± f α e − i λ α t , α + ( ω ) = − ∑ α = ± f α ω − λ α , (40) C(t)=\sum_{\alpha=\pm}f_{\alpha}e^{-i\lambda_{\alpha}t},\qquad
\alpha_{+}(\omega)=-\sum_{\alpha=\pm}\frac{f_{\alpha}}{\omega-\lambda_{\alpha}}, \tag*{(40)} C ( t ) = α = ± ∑ f α e − i λ α t , α + ( ω ) = − α = ± ∑ ω − λ α f α , (40)
where
f α = a α b α ∗ (41) f_\alpha= a_\alpha b_\alpha^* \tag*{(41)} f α = a α b α ∗ (41)
is the optical residue of the dipole-projected resolvent as a product of the excitation amplitude a α = ⟨ l α ∣ μ + ∣ G ⟩ a_\alpha= \langle l_\alpha|\mu^+|G\rangle a α = ⟨ l α ∣ μ + ∣ G ⟩ and the detection amplitude b α ∗ = ⟨ G ∣ μ − ∣ r α ⟩ b_\alpha^* = \langle G|\mu^-|r_\alpha\rangle b α ∗ = ⟨ G ∣ μ − ∣ r α ⟩ . The eigenvalues locate the resolvent poles, while the residues determine their contribution to the optical response. The excitation amplitude a α a_\alpha a α is determined by the left eigenvector ⟨ l α ∣ \langle l_\alpha| ⟨ l α ∣ , whereas the overlap with ∣ r α ⟩ |r_\alpha\rangle ∣ r α ⟩ generally includes contributions from other modes
b α = ⟨ r α ∣ μ + ∣ G ⟩ = ∑ β ⟨ r α ∣ r β ⟩ a β . (42) b_\alpha= \langle r_\alpha|\mu^+|G\rangle= \sum_\beta\langle r_\alpha|r_\beta\rangle a_\beta. \tag*{(42)} b α = ⟨ r α ∣ μ + ∣ G ⟩ = β ∑ ⟨ r α ∣ r β ⟩ a β . (42)
Thus, b α b_\alpha b α generally differs from a α a_\alpha a α , and b α b α ∗ = ⟨ r α ∣ μ + ∣ G ⟩ ⟨ G ∣ μ − ∣ r α ⟩ b_\alpha b_\alpha^* = \langle r_\alpha|\mu^+|G\rangle\langle G|\mu^-|r_\alpha\rangle b α b α ∗ = ⟨ r α ∣ μ + ∣ G ⟩ ⟨ G ∣ μ − ∣ r α ⟩ cannot replace the optical residue f α = a α b α ∗ f_\alpha= a_\alpha b_\alpha^* f α = a α b α ∗ . Only when κ = γ \kappa= \gamma κ = γ , the eigenvectors can be chosen orthonormal, giving a α = b α a_\alpha= b_\alpha a α = b α and f α = b α b α ∗ f_\alpha= b_\alpha b_\alpha^* f α = b α b α ∗ . In general, however, f α f_\alpha f α is complex; its real and imaginary parts contribute Lorentzian and dispersive terms, respectively, to the spectrum. Completeness fixes the total residue
∑ α f α = ⟨ G ∣ μ − μ + ∣ G ⟩ = ∣ μ P ∣ 2 + ∣ μ X ∣ 2 . (43) \sum_\alpha f_\alpha= \langle G|\mu^-\mu^+|G\rangle= |\mu_P|^2 + |\mu_X|^2. \tag*{(43)} α ∑ f α = ⟨ G ∣ μ − μ + ∣ G ⟩ = ∣ μ P ∣ 2 + ∣ μ X ∣ 2 . (43)
This decomposition into two separate modes breaks down at an exceptional point (EP), where both the eigenvalues and their eigenvectors coalesce. At zero detuning, ω P = ω X = ω 0 \omega_P = \omega_X = \omega_0 ω P = ω X = ω 0 , the eigenvalues are
λ ± = ω 0 − i Γ ± g 2 − ( κ − γ ) 2 4 , Γ = κ + γ 2 . (44) \lambda_\pm= \omega_0 - i\Gamma\pm\sqrt{g^2 - \frac{(\kappa-\gamma)^2}{4}}, \qquad\Gamma= \frac{\kappa+\gamma}{2}. \tag*{(44)} λ ± = ω 0 − i Γ ± g 2 − 4 ( κ − γ ) 2 , Γ = 2 κ + γ . (44)
For ∣ g ∣ > ∣ κ − γ ∣ / 2 |g| > |\kappa-\gamma|/2 ∣ g ∣ > ∣ κ − γ ∣/2 , the modes have different resonance frequencies Re λ ± \operatorname{Re}\lambda_\pm Re λ ± and the same optical-coherence decay rate Γ \Gamma Γ . For ∣ g ∣ < ∣ κ − γ ∣ / 2 |g| < |\kappa-\gamma|/2 ∣ g ∣ < ∣ κ − γ ∣/2 , their resonance frequencies coincide, while their coherence decay rates − Im λ ± -\operatorname{Im}\lambda_\pm − Im λ ± differ. The EP occurs at
2 ∣ g ∣ = ∣ κ − γ ∣ > 0 , λ E P = ω 0 − i Γ . (45) 2|g| = |\kappa-\gamma| > 0, \qquad\lambda_{\mathrm{EP}} = \omega_0 - i\Gamma. \tag*{(45)} 2∣ g ∣ = ∣ κ − γ ∣ > 0 , λ EP = ω 0 − i Γ. (45)
It therefore marks the boundary between frequency splitting and decay-rate splitting at zero detuning. The condition involves the difference between the two coherence decay rates; the mean decay rate Γ \Gamma Γ remains finite at the EP. At this point, H e f f ( e ) H_{\mathrm{eff}}^{(\mathrm{e})} H eff ( e ) has only one independent eigenvector, so a complete two-mode biorthogonal basis no longer exists. The singularity of the separate-mode description is explicit in the spectral projectors
∣ r ± ⟩ ⟨ l ± ∣ = H e f f ( e ) − λ ∓ I e λ ± − λ ∓ , (46) |r_\pm\rangle\langle l_\pm| = \frac{H_{\mathrm{eff}}^{(\mathrm{e})} - \lambda_\mp I_{\mathrm{e}}}{\lambda_\pm-\lambda_\mp}, \tag*{(46)} ∣ r ± ⟩ ⟨ l ± ∣ = λ ± − λ ∓ H eff ( e ) − λ ∓ I e , (46)
which give the equivalent residue expression
f ± = ⟨ G ∣ μ − ( H e f f ( e ) − λ ∓ I e ) μ + ∣ G ⟩ λ ± − λ ∓ . (47) f_\pm= \frac{\langle G|\mu^- \left(H_{\mathrm{eff}}^{(\mathrm{e})} - \lambda_\mp I_{\mathrm{e}}\right)\mu^+|G\rangle}{\lambda_\pm-\lambda_\mp}. \tag*{(47)} f ± = λ ± − λ ∓ ⟨ G ∣ μ − ( H eff ( e ) − λ ∓ I e ) μ + ∣ G ⟩ . (47)
As the eigenvalue separation approaches zero, the individual residues can diverge. Their divergent contributions cancel in the combined polarizability. The singularity therefore concerns the decomposition into separate modes and does not imply an infinite optical response. The response at the EP can be evaluated directly, giving
α + , E P ( ω ) = − ∣ μ P ∣ 2 ( δ ω + i γ ) + ∣ μ X ∣ 2 ( δ ω + i κ ) + 2 g Re μ P ∗ μ X ( δ ω + i Γ ) 2 , δ ω = ω − ω 0 . (48) \alpha_{+,{\rm EP}}(\omega) = -\frac{|\mu_P|^2(\delta\omega+i\gamma)+|\mu_X|^2(\delta\omega+i\kappa)+2g\,\operatorname{Re}\mu_P^*\mu_X}{(\delta\omega+i\Gamma)^2}, \qquad\delta\omega=\omega-\omega_0. \tag*{(48)} α + , EP ( ω ) = − ( δ ω + i Γ ) 2 ∣ μ P ∣ 2 ( δ ω + iγ ) + ∣ μ X ∣ 2 ( δ ω + iκ ) + 2 g Re μ P ∗ μ X , δ ω = ω − ω 0 . (48)
For real probe frequencies and Γ > 0 \Gamma> 0 Γ > 0 , ∣ δ ω + i Γ ∣ 2 = δ ω 2 + Γ 2 > 0 |\delta\omega+i\Gamma|^2 = \delta\omega^2+\Gamma^2 > 0 ∣ δ ω + i Γ ∣ 2 = δ ω 2 + Γ 2 > 0 . Thus, the pole remains below the real-frequency axis, and the absorption A ( ω ) ≃ Im α + , E P ( ω ) A(\omega)\simeq\operatorname{Im}\alpha_{+,{\rm EP}}(\omega) A ( ω ) ≃ Im α + , EP ( ω ) is finite. The line shape depends on the transition dipoles as well as the coalesced eigenvalue. Its frequency-dependent numerator can produce a single peak, an asymmetric profile, or two maxima. The EP condition therefore does not by itself determine the number of absorption peaks or imply a squared-Lorentzian line shape.
The same response-function framework extends to nonlinear spectroscopy, including pump-probe and two-dimensional electronic spectroscopy. Their leading signals involve the third-order response, which for nonnegative delays can be written as
R ( 3 ) ( t 3 , T 2 , t 1 ) = i 3 Tr { μ G ( t 3 ) μ × G ( T 2 ) μ × G ( t 1 ) μ × ρ 0 } , μ × A = [ μ , A ] . (49) R^{(3)}(t_3,T_2,t_1) = i^3\operatorname{Tr}\{\mu\,\mathcal{G}(t_3)\mu^\times\mathcal{G}(T_2)\mu^\times\mathcal{G}(t_1)\mu^\times\rho_0\}, \qquad\mu^\times A = [\mu,A]. \tag*{(49)} R ( 3 ) ( t 3 , T 2 , t 1 ) = i 3 Tr { μ G ( t 3 ) μ × G ( T 2 ) μ × G ( t 1 ) μ × ρ 0 } , μ × A = [ μ , A ] . (49)
Here μ \mu μ is the original TC dipole operator, retained between all excitation manifolds involved in the response. Unlike linear absorption, these signals depend on excited-state populations and coherences during the waiting time T 2 T_2 T 2 . Excited-state absorption further involves transitions between the single- and double-excitation manifolds. Consequently, the single-excitation poles and residues f α f_\alpha f α do not determine the nonlinear spectra by themselves. Additional dipole matrix elements and the evolution of the excited-state density matrix are required. Recycling terms must be retained in sectors where they do not vanish; for example, local molecular dephasing transfers population between bright and dark states. A future extension could use the left and right eigenoperators of the full Liouvillian, supplemented by generalised eigenoperators at its EPs, to construct the propagators in R ( 3 ) ( t 3 , T 2 , t 1 ) R^{(3)}(t_3,T_2,t_1) R ( 3 ) ( t 3 , T 2 , t 1 ) , retaining both the relevant excitation manifolds and these population-transfer processes.
Spectral Filter Approach and Randomised LCU Construction for Non-Hermitian Spectroscopic Dilation
We now develop a quantum algorithm for constructing the corresponding singular-value landscape by combining Gaussian spectral filtering with a randomised linear-combination-of-unitaries (RLCU) implementation.
Continuous Gaussian Spectral Filtering
Having mapped the non-Hermitian eigenvalue problem to the zero-mode detection problem of the Hermitian dilation P ( z ) \mathcal{P}(z) P ( z ) , we introduce the Gaussian spectral filter
G τ ( P ( z ) ) = exp ( − τ 2 2 P ( z ) 2 ) , (50) G_\tau(\mathcal{P}(z))=\exp\left(-\frac{\tau^2}{2}\mathcal{P}(z)^2\right), \tag*{(50)} G τ ( P ( z )) = exp ( − 2 τ 2 P ( z ) 2 ) , (50)
where τ > 0 \tau>0 τ > 0 controls the spectral resolution. The following proposition is established for isolated, diagonalisable eigenvalues of H H H . EPs require a separate analysis because the spectral projection must be generalised to the corresponding Jordan invariant subspace.
Lemma 1 (Gaussian spectral projection). Since P ( z ) \mathcal{P}(z) P ( z ) is Hermitian, it admits the spectral decomposition P ( z ) = ∑ k ϵ k ∣ χ k ⟩ ⟨ χ k ∣ \mathcal{P}(z)=\sum_k\epsilon_k|\chi_k\rangle\langle\chi_k| P ( z ) = ∑ k ϵ k ∣ χ k ⟩ ⟨ χ k ∣ , where ϵ k = ± σ k ( H z ) \epsilon_k=\pm\sigma_k(H_z) ϵ k = ± σ k ( H z ) . Consequently
G τ ( P ( z ) ) = ∑ k e − τ 2 ϵ k 2 / 2 ∣ χ k ⟩ ⟨ χ k ∣ . G_\tau(\mathcal{P}(z))=\sum_k e^{-\tau^2\epsilon_k^2/2}|\chi_k\rangle\langle\chi_k|. G τ ( P ( z )) = k ∑ e − τ 2 ϵ k 2 /2 ∣ χ k ⟩ ⟨ χ k ∣. If z = λ j z=\lambda_j z = λ j is an isolated, diagonalisable eigenvalue of H H H , then σ min ( H − λ j I ) = 0 \sigma_{\min}(H-\lambda_j\mathbb{I})=0 σ m i n ( H − λ j I ) = 0 , and therefore the kernel of P ( λ j ) \mathcal{P}(\lambda_j) P ( λ j ) is left invariant while every nonzero singular-value component is exponentially suppressed. In the limit lim τ → ∞ G τ ( P ( λ j ) ) = Π 0 \lim_{\tau\to\infty}G_\tau(\mathcal{P}(\lambda_j))=\Pi_0 lim τ → ∞ G τ ( P ( λ j )) = Π 0 , where Π 0 \Pi_0 Π 0 denotes the orthogonal projector onto ker ( P ( λ j ) ) \ker(\mathcal{P}(\lambda_j)) ker ( P ( λ j )) .
Proof. Since P ( λ j ) \mathcal{P}(\lambda_j) P ( λ j ) is Hermitian, the spectral theorem applies. Hence, G τ ( P ( λ j ) ) = ∑ k e − τ 2 ϵ k 2 / 2 ∣ χ k ⟩ ⟨ χ k ∣ G_\tau(\mathcal{P}(\lambda_j))=\sum_k e^{-\tau^2\epsilon_k^2/2}|\chi_k\rangle\langle\chi_k| G τ ( P ( λ j )) = ∑ k e − τ 2 ϵ k 2 /2 ∣ χ k ⟩ ⟨ χ k ∣ . For every nonzero eigenvalue e − τ 2 ϵ k 2 / 2 ⟶ 0 e^{-\tau^2\epsilon_k^2/2}\longrightarrow0 e − τ 2 ϵ k 2 /2 ⟶ 0 (τ → ∞ \tau\to\infty τ → ∞ ), whereas for every zero mode e − τ 2 ϵ k 2 / 2 = 1 e^{-\tau^2\epsilon_k^2/2}=1 e − τ 2 ϵ k 2 /2 = 1 . Therefore, G τ ( P ( λ j ) ) ⟶ Π 0 G_\tau(\mathcal{P}(\lambda_j))\longrightarrow\Pi_0 G τ ( P ( λ j )) ⟶ Π 0 . Furthermore, let Δ σ = min ϵ k ≠ 0 ∣ ϵ k ∣ \Delta_\sigma=\min_{\epsilon_k\ne0}|\epsilon_k| Δ σ = min ϵ k = 0 ∣ ϵ k ∣ denote the smallest nonzero singular-value gap of P ( λ j ) \mathcal{P}(\lambda_j) P ( λ j ) . Then
∣ G τ ( P ( λ j ) ) − Π 0 ∣ = ∣ ∑ ϵ k ≠ 0 e − τ 2 ϵ k 2 / 2 ∣ χ k ⟩ ⟨ χ k ∣ ∣ = max ϵ k ≠ 0 e − τ 2 ϵ k 2 / 2 ≤ e − τ 2 Δ σ 2 / 2 . \left|G_\tau(\mathcal{P}(\lambda_j))-\Pi_0\right|=\left|\sum_{\epsilon_k\ne0}e^{-\tau^2\epsilon_k^2/2}|\chi_k\rangle\langle\chi_k|\right|=\max_{\epsilon_k\ne0}e^{-\tau^2\epsilon_k^2/2}\le e^{-\tau^2\Delta_\sigma^2/2}. ∣ G τ ( P ( λ j )) − Π 0 ∣ = ϵ k = 0 ∑ e − τ 2 ϵ k 2 /2 ∣ χ k ⟩ ⟨ χ k ∣ = ϵ k = 0 max e − τ 2 ϵ k 2 /2 ≤ e − τ 2 Δ σ 2 /2 .
Hence, the Gaussian filter converges exponentially fast (in operator norm) to the orthogonal projector onto the kernel of the Hermitian dilation, with convergence rate determined by the smallest nonzero singular value. ▫
Corollary 2 (Overlap weight). Let the initial state be expanded in the eigenbasis of the Hermitian dilation ∣ Φ 0 ⟩ = ∑ k c k ∣ χ k ⟩ |\Phi_0\rangle=\sum_k c_k|\chi_k\rangle ∣ Φ 0 ⟩ = ∑ k c k ∣ χ k ⟩ , where c k = ⟨ χ k ∣ Φ 0 ⟩ c_k=\langle\chi_k|\Phi_0\rangle c k = ⟨ χ k ∣ Φ 0 ⟩ . Under the assumptions of Lemma 1 , lim τ → ∞ G τ ( P ( λ j ) ) ∣ Φ 0 ⟩ = Π 0 ∣ Φ 0 ⟩ \lim_{\tau\to\infty}G_\tau(\mathcal{P}(\lambda_j))|\Phi_0\rangle=\Pi_0|\Phi_0\rangle lim τ → ∞ G τ ( P ( λ j )) ∣ Φ 0 ⟩ = Π 0 ∣ Φ 0 ⟩ , where Π 0 \Pi_0 Π 0 denotes the orthogonal projector onto ker ( P ( λ j ) ) \ker(\mathcal{P}(\lambda_j)) ker ( P ( λ j )) . If the zero eigenspace is one-dimensional, Π 0 = ∣ χ 0 ⟩ ⟨ χ 0 ∣ \Pi_0=|\chi_0\rangle\langle\chi_0| Π 0 = ∣ χ 0 ⟩ ⟨ χ 0 ∣ , and thereby lim τ → ∞ G τ ( P ( λ j ) ) ∣ Φ 0 ⟩ = c 0 ∣ χ 0 ⟩ \lim_{\tau\to\infty}G_\tau(\mathcal{P}(\lambda_j))|\Phi_0\rangle=c_0|\chi_0\rangle lim τ → ∞ G τ ( P ( λ j )) ∣ Φ 0 ⟩ = c 0 ∣ χ 0 ⟩ , with c 0 = ⟨ χ 0 ∣ Φ 0 ⟩ c_0=\langle\chi_0|\Phi_0\rangle c 0 = ⟨ χ 0 ∣ Φ 0 ⟩ . Consequently, ∣ G τ ( P ( λ j ) ) ∣ Φ 0 ⟩ ∣ 2 = ∣ c 0 ∣ 2 \left|G_\tau(\mathcal{P}(\lambda_j))|\Phi_0\rangle\right|^2=|c_0|^2 ∣ G τ ( P ( λ j )) ∣ Φ 0 ⟩ ∣ 2 = ∣ c 0 ∣ 2 .
The quantity ∣ c 0 ∣ 2 |c_0|^2 ∣ c 0 ∣ 2 therefore represents the overlap weight of the initial state with the target zero-mode of the Hermitian dilation and serves as the non-Hermitian analogue of the overlap parameter appearing in Hermitian Gaussian spectral filtering. To implement this non-unitary operator using unitary time evolution, we employ the Fourier representation of the Gaussian filter
G τ ( P ( z ) ) = ∫ − ∞ ∞ p ( x ) e − i τ x P ( z ) d x , (51) G_{\tau}(\mathcal{P}(z)) = \int_{-\infty}^{\infty} p(x)e^{-i\tau x\mathcal{P}(z)}\,dx, \tag*{(51)} G τ ( P ( z )) = ∫ − ∞ ∞ p ( x ) e − i τ x P ( z ) d x , (51)
where p ( x ) = 1 2 π e − x 2 / 2 p(x) = \frac{1}{\sqrt{2\pi}}e^{-x^2/2} p ( x ) = 2 π 1 e − x 2 /2 is the standard normal distribution. The integration variable x x x is dimensionless and sampled from a fixed probability distribution independent of the target spectral resolution. All dependence on the filtering strength is encoded through the evolution time t = τ x t = \tau x t = τ x .
Discrete Realisation of the Gaussian Spectral Filter
The continuous Gaussian filter provides an exact representation of the non-unitary operator G τ ( P ( z ) ) G_{\tau}(\mathcal{P}(z)) G τ ( P ( z )) in terms of unitary real-time evolutions generated by P ( z ) \mathcal{P}(z) P ( z ) . To realize this operation on a digital quantum processor, the continuous integral must be converted into a finite linear combination of experimentally accessible unitary operations. This compilation procedure introduces two controllable approximation parameters: a finite integration cutoff x c x_c x c and a finite quadrature resolution M M M . These parameters determine the range and density of sampled evolution times while preserving the underlying spectral-filtering mechanism.
Finite Integration Domain
Because the Gaussian weight in Eq. (51) rapidly suppresses contributions from large values of ∣ x ∣ |x| ∣ x ∣ , the integral can be restricted to a finite interval [ − x c , x c ] [-x_c, x_c] [ − x c , x c ] without significantly affecting the filtered state. We can hence introduce the truncated filter
G τ , x c ( P ( z ) ) = c ∫ − x c x c p ( x ) e − i τ x P ( z ) d x . (52) G_{\tau,x_c}(\mathcal{P}(z)) = c\int_{-x_c}^{x_c} p(x)e^{-i\tau x\mathcal{P}(z)}\,dx. \tag*{(52)} G τ , x c ( P ( z )) = c ∫ − x c x c p ( x ) e − i τ x P ( z ) d x . (52)
The cutoff parameter x c x_c x c determines the maximum evolution time explored during the filtering procedure t max = τ x c t_{\max} = \tau x_c t m a x = τ x c . Consequently, increasing x c x_c x c systematically improves the approximation to the ideal Gaussian filter while requiring access to longer real-time evolutions. In practice, x c x_c x c is chosen so that this cutoff error is below the target ϵ c \epsilon_c ϵ c .
Discrete Quadrature Representation
To obtain a finite LCU representation, the truncated integral is discretised over a symmetric grid x m = m Δ x x_m = m\Delta x x m = m Δ x , where m = − M , … , M m = -M,\ldots,M m = − M , … , M , with grid spacing Δ x = x c M \Delta x = \frac{x_c}{M} Δ x = M x c . The resulting discrete approximation takes the form
G ~ τ ( P ( z ) ) = c ∑ m = − M M w m e − i τ x m P ( z ) , (53) \widetilde{G}_{\tau}(\mathcal{P}(z)) = c\sum_{m=-M}^{M} w_m e^{-i\tau x_m\mathcal{P}(z)}, \tag*{(53)} G τ ( P ( z )) = c m = − M ∑ M w m e − i τ x m P ( z ) , (53)
where w m w_m w m are quadrature weights determined by the chosen integration rule. Where for a uniform discretisation we have w m = Δ x p ( x m ) w_m = \Delta x\,p(x_m) w m = Δ x p ( x m ) . Each term in the expansion corresponds to a unitary evolution generated by the Hermitian dilation P ( z ) \mathcal{P}(z) P ( z ) for an evolution time t m = τ x m t_m = \tau x_m t m = τ x m . The Gaussian filter is therefore realised as a weighted linear combination of real-time evolutions with different durations. The approximation becomes systematically more accurate as x c x_c x c and M M M increase. Importantly, although the original eigenvalue problem is non-Hermitian, all quantum evolutions appearing in the implementation are generated by the Hermitian dilation P ( z ) \mathcal{P}(z) P ( z ) . Thus, the resulting algorithm inherits the operational advantages of standard Hermitian Hamiltonian simulation while retaining access to the complex spectrum and biorthogonal eigenstates of the underlying non-Hermitian operator.
Randomised LCU Realisation of the Filtered Observable
The discretised Gaussian filter is naturally expressed as a linear combination of unitary (LCU) evolutions generated by the Hermitian dilation P ( z ) \mathcal{P}(z) P ( z ) . A direct coherent implementation of this LCU operator would require additional ancilla qubits and controlled-select operations. Instead, we adopt RLCU strategy in which individual unitary components are sampled according to their normalised quadrature weights. The discretised filter can be written as
G ~ τ ( P ( z ) ) = ∑ m = − M M w m U m , U m = e − i τ x m P ( z ) . (54) \widetilde{G}_{\tau}(\mathcal{P}(z)) = \sum_{m=-M}^{M} w_m U_m,\qquad U_m = e^{-i\tau x_m\mathcal{P}(z)}. \tag*{(54)} G τ ( P ( z )) = m = − M ∑ M w m U m , U m = e − i τ x m P ( z ) . (54)
Defining the normalisation factor as μ = ∑ m = − M M ∣ w m ∣ \mu= \sum_{m=-M}^{M} |w_m| μ = ∑ m = − M M ∣ w m ∣ , and the corresponding probability distribution p ~ m = w m μ \widetilde{p}_m = \frac{w_m}{\mu} p m = μ w m , the filter admits the stochastic representation
G ~ τ ( P ( z ) ) = μ ∑ m = − M M p ~ m U m = μ E m ∼ p ~ m [ U m ] . (55) \widetilde{G}_{\tau}(\mathcal{P}(z)) = \mu\sum_{m=-M}^{M} \widetilde{p}_m U_m = \mu\mathbb{E}_{m\sim\widetilde{p}_m}[U_m]. \tag*{(55)} G τ ( P ( z )) = μ m = − M ∑ M p m U m = μ E m ∼ p m [ U m ] . (55)
The objective is to estimate the biorthogonal expectation value ⟨ O ⟩ j = ⟨ L j ∣ O ∣ R j ⟩ ⟨ L j ∣ R j ⟩ \langle O\rangle_j = \frac{\langle L_j|O|R_j\rangle}{\langle L_j|R_j\rangle} ⟨ O ⟩ j = ⟨ L j ∣ R j ⟩ ⟨ L j ∣ O ∣ R j ⟩ through filtered observables defined on the dilated Hilbert space. Let M ∈ O , I \mathcal{M} \in\mathcal{O},\mathcal{I} M ∈ O , I denote either the dilated observable operator or the dilated normalisation operator. The corresponding filtered quantity is
N G ~ ( M ) = Tr [ G ~ τ ( P ( z ) ) ρ G ~ τ ( P ( z ) ) † M ] . (56) N_{\widetilde{G}}(\mathcal{M}) = \operatorname{Tr}\left[\widetilde{G}_{\tau}(\mathcal{P}(z))\rho\widetilde{G}_{\tau}(\mathcal{P}(z))^\dagger\mathcal{M}\right]. \tag*{(56)} N G ( M ) = Tr [ G τ ( P ( z )) ρ G τ ( P ( z ) ) † M ] . (56)
Substituting the RLCU representation yields
N G ~ ( M ) = μ 2 ∑ m , m ′ = − M M p ~ m p ~ m ′ ⟨ Φ 0 ∣ U m ′ † M U m ∣ Φ 0 ⟩ . (57) N_{\widetilde{G}}(\mathcal{M}) = \mu^2 \sum_{m,m'=-M}^{M} \widetilde{p}_m\widetilde{p}_{m'}\langle\Phi_0|U_{m'}^\dagger\mathcal{M}U_m|\Phi_0\rangle. \tag*{(57)} N G ( M ) = μ 2 m , m ′ = − M ∑ M p m p m ′ ⟨ Φ 0 ∣ U m ′ † M U m ∣ Φ 0 ⟩ . (57)
This expression admits a stochastic interpretation in which the indices m m m and m ′ m' m ′ are sampled independently from the probability distribution p ~ m \widetilde{p}_m p m , while the corresponding matrix element is evaluated on quantum hardware. The resulting estimator avoids coherent implementation of the full LCU operator and replaces it with repeated evaluations of shallow real-time evolution circuits. The randomised implementation therefore provides an unbiased Monte Carlo estimator of the filtered expectation value. Specifically, if ( m a , m b ) (m_a,m_b) ( m a , m b ) are sampled independently from the distribution p ~ m \widetilde{p}_m p m , then A M = ⟨ Φ 0 ∣ U m b † M U m a ∣ Φ 0 ⟩ A_{\mathcal{M}}=\langle\Phi_0|U_{m_b}^\dagger\mathcal{M}U_{m_a}|\Phi_0\rangle A M = ⟨ Φ 0 ∣ U m b † M U m a ∣ Φ 0 ⟩ satisfies N G ~ ( M ) = μ 2 E [ A M ] N_{\widetilde{G}}(\mathcal{M})=\mu^2\mathbb{E}[A_{\mathcal{M}}] N G ( M ) = μ 2 E [ A M ] . In practice, E [ A M ] \mathbb{E}[A_{\mathcal{M}}] E [ A M ] is estimated from a finite number of randomised circuit executions, while each transition amplitude is evaluated by a Hadamard-test measurement. Although the target spectral problem is non-Hermitian, every sampled circuit instance involves only the unitary evolution U m = e − i τ x m P ( z ) U_m=e^{-i\tau x_m\mathcal{P}(z)} U m = e − i τ x m P ( z ) generated by the Hermitian dilation P ( z ) \mathcal{P}(z) P ( z ) . Consequently, the quantum implementation reduces to standard Hermitian Hamiltonian simulation while retaining access to the complex spectrum and biorthogonal observables of the original non-Hermitian Hamiltonian.
High-Order Simulation of the Hermitian Dilation
Each realisation of the randomised Gaussian filter requires the implementation of the unitary evolution U ( t ) = e − i t P ( z ) U(t)=e^{-it\mathcal{P}(z)} U ( t ) = e − i t P ( z ) , where t = τ x t=\tau x t = τ x , generated by P ( z ) \mathcal{P}(z) P ( z ) . Since P ( z ) \mathcal{P}(z) P ( z ) is Hermitian, the simulation reduces to a standard Hamiltonian-simulation problem on the enlarged Hilbert space. After decomposing the dilation into Hermitian Pauli operators P ( z ) = ∑ α = 0 Γ h α P α \mathcal{P}(z)=\sum_{\alpha=0}^{\Gamma}h_\alpha P_\alpha P ( z ) = ∑ α = 0 Γ h α P α , where each P α P_\alpha P α is a Hermitian Pauli string, the exact evolution is approximated using product-formula simulation.
Let m = t ν m=\frac{t}{\nu} m = ν t , where ν \nu ν denotes the number of time segments. For one segment, let S 2 k ( m ) S_{2k}(m) S 2 k ( m ) denote the standard 2 k 2k 2 k -th order Suzuki product formula. The exact evolution admits the decomposition e − i m P ( z ) = V 2 k ( m ) S 2 k ( m ) e^{-im\mathcal{P}(z)}=V_{2k}(m)S_{2k}(m) e − im P ( z ) = V 2 k ( m ) S 2 k ( m ) , where V 2 k ( m ) V_{2k}(m) V 2 k ( m ) is the remainder operator collecting the product-formula error. Applying the Baker-Campbell-Hausdorff expansion gives
V 2 k ( m ) = exp ( i ∑ s = 2 k + 1 ∞ E 2 k , s m s ) , (58) V_{2k}(m)=\exp\left(i\sum_{s=2k+1}^{\infty}E_{2k,s}m^s\right), \tag*{(58)} V 2 k ( m ) = exp ( i s = 2 k + 1 ∑ ∞ E 2 k , s m s ) , (58)
where the operators E 2 k , s E_{2k,s} E 2 k , s consist of nested commutators of the Pauli components of P ( z ) \mathcal{P}(z) P ( z ) . Consequently V 2 k ( m ) = I + O ( m 2 k + 1 ) V_{2k}(m)=\mathbb{I}+O(m^{2k+1}) V 2 k ( m ) = I + O ( m 2 k + 1 ) , showing that the leading product-formula error begins at order m 2 k + 1 m^{2k+1} m 2 k + 1 .
LCU Approximation of the Remainder Operator
Rather than increasing the Suzuki order directly, the remainder operator is approximated through a truncated LCU representation. Expanding the BCH remainder V 2 k ( m ) = ∑ s = 0 ∞ F 2 k , s ( m ) V_{2k}(m)=\sum_{s=0}^{\infty}F_{2k,s}(m) V 2 k ( m ) = ∑ s = 0 ∞ F 2 k , s ( m ) , where F 2 k , s ( m ) = O ( m s ) F_{2k,s}(m)=O(m^s) F 2 k , s ( m ) = O ( m s ) , the series is truncated after order s c s_c s c as V 2 k ( s c ) ( m ) = I + ∑ s = 2 k + 1 s c F 2 k , s ( m ) V_{2k}^{(s_c)}(m)=\mathbb{I}+\sum_{s=2k+1}^{s_c}F_{2k,s}(m) V 2 k ( s c ) ( m ) = I + ∑ s = 2 k + 1 s c F 2 k , s ( m ) . Each operator F 2 k , s ( m ) F_{2k,s}(m) F 2 k , s ( m ) is subsequently represented as a finite linear combination of efficiently implementable unitary operators, yielding an LCU approximation of the remainder. The compensated short-time propagator is hence U ~ 2 k ( m ) = V 2 k ( s c ) ( m ) S 2 k ( m ) \widetilde{U}_{2k}(m)=V_{2k}^{(s_c)}(m)S_{2k}(m) U 2 k ( m ) = V 2 k ( s c ) ( m ) S 2 k ( m ) . For a total evolution time t = ν m t=\nu m t = ν m , the corresponding approximation is U ~ 2 k ( t ) = [ U ~ 2 k ( m ) ] ν \widetilde{U}_{2k}(t)=\left[\widetilde{U}_{2k}(m)\right]^\nu U 2 k ( t ) = [ U 2 k ( m ) ] ν .
Integration with the Randomised Gaussian Filter
The compensated propagator replaces every exact unitary evolution appearing in the randomised Gaussian filter. Specifically e − i τ x m P ( z ) ⟶ U ~ 2 k ( τ x m ) e^{-i\tau x_m\mathcal{P}(z)}\longrightarrow\widetilde{U}_{2k}(\tau x_m) e − i τ x m P ( z ) ⟶ U 2 k ( τ x m ) for every quadrature node x m x_m x m . Consequently, the implemented randomised filter takes the form
G ~ τ ( P ( z ) ) = ∑ m = − M M w m U ~ 2 k ( τ x m ) , (59) \widetilde{G}_{\tau}(\mathcal{P}(z))=\sum_{m=-M}^{M}w_m\widetilde{U}_{2k}(\tau x_m), \tag*{(59)} G τ ( P ( z )) = m = − M ∑ M w m U 2 k ( τ x m ) , (59)
which differs from the ideal discretised filter only through the Hamiltonian-simulation error of the compensated propagator. This construction preserves the RLCU framework, while replacing each exact real-time evolution by its high-order compensated approximation. The resulting approximation error is analysed in Proposition 2 .
Ancilla-Efficient Measurement of Filtered Observables
Conventional LCU implementations require coherent coefficient encoding using ancillary control registers and multi-controlled select operations. Such constructions introduce substantial qubit overhead and circuit depth, particularly when implementing Gaussian spectral filters composed of many evolution times.
In the present framework this overhead is avoided through the RLCU formulation. Rather than coherently implementing the entire filtered operator, the quantum processor executes only individual real-time evolutions generated by the Hermitian dilation P ( z ) \mathcal{P}(z) P ( z ) , with evolution times sampled from the Gaussian distribution introduced earlier. Each sampled evolution is implemented using the compensated second-order Suzuki product formula described in § 4 D. Consequently, no coherent LCU ancilla registers, block encodings, or multi-controlled select operations are required. The remaining task is to estimate the biorthogonal expectation value ⟨ O ⟩ j = ⟨ L j ∣ O ∣ R j ⟩ ⟨ L j ∣ R j ⟩ \langle O\rangle_j=\frac{\langle L_j|O|R_j\rangle}{\langle L_j|R_j\rangle} ⟨ O ⟩ j = ⟨ L j ∣ R j ⟩ ⟨ L j ∣ O ∣ R j ⟩ .
Embedded Measurement Operators
The corresponding operators acting on the dilated Hilbert space are O = ( 0 O 0 0 ) \mathcal{O}=\begin{pmatrix}0&O\\0&0\end{pmatrix} O = ( 0 0 O 0 ) , I = ( 0 I 0 0 ) \mathcal{I}=\begin{pmatrix}0&I\\0&0\end{pmatrix} I = ( 0 0 I 0 ) . These operators admit a direct representation using the Pauli operators { σ x , σ y } \{\sigma_x,\sigma_y\} { σ x , σ y } acting on an ancillary ”dilation” qubit O = σ x + i σ y 2 ⊗ O \mathcal{O}=\frac{\sigma_x+i\sigma_y}{2}\otimes O O = 2 σ x + i σ y ⊗ O , I = σ x + i σ y 2 ⊗ I \mathcal{I}=\frac{\sigma_x+i\sigma_y}{2}\otimes I I = 2 σ x + i σ y ⊗ I . Since O \mathcal{O} O and I \mathcal{I} I are non-Hermitian, they cannot be measured directly as standard quantum observables. We therefore utilize the Cartesian decomposition to express them in terms of Hermitian operators O = O R + i O I \mathcal{O}=\mathcal{O}_R+i\mathcal{O}_I O = O R + i O I , I = I R + i I I \mathcal{I}=\mathcal{I}_R+i\mathcal{I}_I I = I R + i I I . Using the definition of the Pauli operators, this decomposition maps to
O R = σ x 2 ⊗ O , O I = σ y 2 ⊗ O , \mathcal{O}_R=\frac{\sigma_x}{2}\otimes O,\qquad\mathcal{O}_I=\frac{\sigma_y}{2}\otimes O, O R = 2 σ x ⊗ O , O I = 2 σ y ⊗ O ,
I R = σ x 2 ⊗ I , I I = σ y 2 ⊗ I . \mathcal{I}_R=\frac{\sigma_x}{2}\otimes I,\qquad\mathcal{I}_I=\frac{\sigma_y}{2}\otimes I. I R = 2 σ x ⊗ I , I I = 2 σ y ⊗ I .
These four components { O R , O I , I R , I I } \{\mathcal{O}_R,\mathcal{O}_I,\mathcal{I}_R,\mathcal{I}_I\} { O R , O I , I R , I I } are strictly Hermitian. Consequently, each can be measured independently as a standard quantum observable. The final complex expectation value ⟨ O ⟩ / ⟨ I ⟩ \langle\mathcal{O}\rangle/\langle\mathcal{I}\rangle ⟨ O ⟩ / ⟨ I ⟩ is then reconstructed classically from the measured expectation values of these four components.
Transition-Amplitude Estimation
For each Hermitian component X α ∈ { O R , O I , I R , I I } \mathcal{X}_{\alpha} \in\{\mathcal{O}_{R},\mathcal{O}_{I},\mathcal{I}_{R},\mathcal{I}_{I}\} X α ∈ { O R , O I , I R , I I } , define A X α = ⟨ Φ 0 ∣ U b † X α U a ∣ Φ 0 ⟩ A_{\mathcal{X}_{\alpha}} = \langle\Phi_{0}|U_{b}^{\dagger}\mathcal{X}_{\alpha}U_{a}|\Phi_{0}\rangle A X α = ⟨ Φ 0 ∣ U b † X α U a ∣ Φ 0 ⟩ , where U a U_{a} U a and U b U_{b} U b are independent realisations of the randomised Gaussian filter satisfying E [ U ] = G τ ( P ) \mathbb{E}[U]=G_{\tau}(P) E [ U ] = G τ ( P ) . Hence E [ A X α ] = ⟨ Φ 0 ∣ G τ ( P ) † X α G τ ( P ) ∣ Φ 0 ⟩ \mathbb{E}[A_{\mathcal{X}_{\alpha}}] = \langle\Phi_{0}|G_{\tau}(P)^{\dagger}\mathcal{X}_{\alpha}G_{\tau}(P)|\Phi_{0}\rangle E [ A X α ] = ⟨ Φ 0 ∣ G τ ( P ) † X α G τ ( P ) ∣ Φ 0 ⟩ . If the spectral parameter satisfies z = λ j z=\lambda_{j} z = λ j , then Lemma 1 implies lim τ → ∞ G τ ( P ( λ j ) ) = Π 0 \lim_{\tau\to\infty}G_{\tau}(P(\lambda_{j}))=\Pi_{0} lim τ → ∞ G τ ( P ( λ j )) = Π 0 , where Π 0 \Pi_{0} Π 0 denotes the orthogonal projector onto ker P ( λ j ) \ker P(\lambda_{j}) ker P ( λ j ) . Therefore, lim τ → ∞ E [ A X α ] = ⟨ Φ 0 ∣ Π 0 X α Π 0 ∣ Φ 0 ⟩ \lim_{\tau\to\infty}\mathbb{E}[A_{\mathcal{X}_{\alpha}}]=\langle\Phi_{0}|\Pi_{0}\mathcal{X}_{\alpha}\Pi_{0}|\Phi_{0}\rangle lim τ → ∞ E [ A X α ] = ⟨ Φ 0 ∣ Π 0 X α Π 0 ∣ Φ 0 ⟩ . For a nondegenerate zero mode Π 0 = ∣ χ j ⟩ ⟨ χ j ∣ \Pi_{0}=|\chi_{j}\rangle\langle\chi_{j}| Π 0 = ∣ χ j ⟩ ⟨ χ j ∣ , so that Π 0 ∣ Φ 0 ⟩ = c 0 ∣ χ j ⟩ \Pi_{0}|\Phi_{0}\rangle=c_{0}|\chi_{j}\rangle Π 0 ∣ Φ 0 ⟩ = c 0 ∣ χ j ⟩ , with c 0 = ⟨ χ j ∣ Φ 0 ⟩ c_{0}=\langle\chi_{j}|\Phi_{0}\rangle c 0 = ⟨ χ j ∣ Φ 0 ⟩ . Consequently, lim τ → ∞ E [ A X α ] = ∣ c 0 ∣ 2 ⟨ χ j ∣ X α ∣ χ j ⟩ \lim_{\tau\to\infty}\mathbb{E}[A_{\mathcal{X}_{\alpha}}]=|c_{0}|^{2}\langle\chi_{j}|\mathcal{X}_{\alpha}|\chi_{j}\rangle lim τ → ∞ E [ A X α ] = ∣ c 0 ∣ 2 ⟨ χ j ∣ X α ∣ χ j ⟩ . The factor ∣ c 0 ∣ 2 |c_{0}|^{2} ∣ c 0 ∣ 2 is the overlap weight of the initial state with the zero mode and is independent of the observable component X α \mathcal{X}_{\alpha} X α .
Measurement Protocols
Each transition amplitude may be evaluated either by direct forward-backward evolution or by a standard Hadamard test. For the Hadamard-test implementation, an auxiliary measurement qubit is prepared in ∣ 0 ⟩ + ∣ 1 ⟩ 2 \frac{|0\rangle+|1\rangle}{\sqrt{2}} 2 ∣0 ⟩ + ∣1 ⟩ , followed by the controlled operation ∣ 0 ⟩ ⟨ 0 ∣ ⊗ I + ∣ 1 ⟩ ⟨ 1 ∣ ⊗ ( U b † X α U a ) |0\rangle\langle0|\otimes I+|1\rangle\langle1|\otimes(U_{b}^{\dagger}\mathcal{X}_{\alpha}U_{a}) ∣0 ⟩ ⟨ 0∣ ⊗ I + ∣1 ⟩ ⟨ 1∣ ⊗ ( U b † X α U a ) . Measurement of the ancilla in the X and Y bases yields the real and imaginary parts of the transition amplitude ⟨ X ⟩ = Re ( A X α ) , ⟨ Y ⟩ = Im ( A X α ) \langle X\rangle=\operatorname{Re}(A_{\mathcal{X}_{\alpha}}),\langle Y\rangle=\operatorname{Im}(A_{\mathcal{X}_{\alpha}}) ⟨ X ⟩ = Re ( A X α ) , ⟨ Y ⟩ = Im ( A X α ) . Combining these measurements with the classical RLCU averaging reconstructs the filtered numerator and denominator entering the desired biorthogonal expectation value. All stochastic averaging and coefficient processing are performed classically, while the quantum processor executes only shallow real-time evolutions generated by the Hermitian dilation.
Extension to driven non-Hermitian systems
The Hermitian-dilation construction extends pointwise in time. At each fixed time t t t , define
H z ( t ) = H ( t ) − z I , P ( z , t ) = ( 0 H z ( t ) H z † ( t ) 0 ) . (60) H_{z}(t)=H(t)-zI,\qquad P(z,t)=\begin{pmatrix} 0 & H_{z}(t) \\ H_{z}^{\dagger}(t) & 0 \end{pmatrix}. \tag*{(60)} H z ( t ) = H ( t ) − z I , P ( z , t ) = ( 0 H z † ( t ) H z ( t ) 0 ) . (60)
Its spectrum is spec P ( z , t ) = ± σ k ( H z ( t ) ) k \operatorname{spec}P(z,t)=\pm\sigma_{k}(H_{z}(t))_{k} spec P ( z , t ) = ± σ k ( H z ( t ) ) k . Let the instantaneous right and left eigenvectors satisfy, respectively, H ( t ) ∣ R α ( t ) ⟩ = z α ( t ) ∣ r α ( t ) ⟩ H(t)|R_{\alpha}(t)\rangle=z_{\alpha}(t)|r_{\alpha}(t)\rangle H ( t ) ∣ R α ( t )⟩ = z α ( t ) ∣ r α ( t )⟩ and H † ( t ) ∣ l α ( t ) ⟩ = z α ∗ ( t ) ∣ l α ( t ) ⟩ H^{\dagger}(t)|l_{\alpha}(t)\rangle=z_{\alpha}^{*}(t)|l_{\alpha}(t)\rangle H † ( t ) ∣ l α ( t )⟩ = z α ∗ ( t ) ∣ l α ( t )⟩ . At a simple instantaneous pole z = z α ( t ) z=z_{\alpha}(t) z = z α ( t ) , the dilation has the zero modes
P ( z α ( t ) , t ) ( 0 ∣ l α ( t ) ⟩ ) = 0 , P ( z α ( t ) , t ) ( 0 ∣ r α ( t ) ⟩ ) = 0. P(z_{\alpha}(t),t)\begin{pmatrix}0\\ |l_{\alpha}(t)\rangle\end{pmatrix}=0,\qquad P(z_{\alpha}(t),t)\begin{pmatrix}0\\ |r_{\alpha}(t)\rangle\end{pmatrix}=0. P ( z α ( t ) , t ) ( 0 ∣ l α ( t )⟩ ) = 0 , P ( z α ( t ) , t ) ( 0 ∣ r α ( t )⟩ ) = 0.
Applying the Gaussian filter to the frozen dilation at successive times therefore gives the instantaneous poles and biorthogonal eigenvectors. This snapshot procedure does not, by itself, propagate a state under the time-dependent Hamiltonian. Using the preparation state ∣ s ⟩ |s\rangle ∣ s ⟩ and detection state ⟨ d ∣ \langle d| ⟨ d ∣ defined in the main text, the frozen-time residue and absorption spectrum are
f α ( t ) = ⟨ d ∣ r α ( t ) ⟩ ⟨ l α ( t ) ∣ s ⟩ ⟨ l α ( t ) ∣ r α ( t ) ⟩ , A t ( ω ) ∝ − Im ∑ α f α ( t ) ω − z α ( t ) . (61) f_{\alpha}(t)=\frac{\langle d|r_{\alpha}(t)\rangle\langle l_{\alpha}(t)|s\rangle}{\langle l_{\alpha}(t)|r_{\alpha}(t)\rangle},\qquad A_{t}(\omega)\propto-\operatorname{Im}\sum_{\alpha}\frac{f_{\alpha}(t)}{\omega-z_{\alpha}(t)}. \tag*{(61)} f α ( t ) = ⟨ l α ( t ) ∣ r α ( t )⟩ ⟨ d ∣ r α ( t )⟩ ⟨ l α ( t ) ∣ s ⟩ , A t ( ω ) ∝ − Im α ∑ ω − z α ( t ) f α ( t ) . (61)
Here A t ( ω ) A_{t}(\omega) A t ( ω ) denotes the frozen-time extension of A ( ω ) A(\omega) A ( ω ) . It approximates the response of a slowly varying system; under rapid driving, the response also depends on the time-ordered propagator and cannot be inferred from instantaneous poles and residues alone. On an interval where the poles are simple and the eigenvectors vary smoothly, choose a biorthonormal normalisation ⟨ l α ( t ) ∣ r β ( t ) ⟩ = δ α β \langle l_{\alpha}(t)|r_{\beta}(t)\rangle=\delta_{\alpha\beta} ⟨ l α ( t ) ∣ r β ( t )⟩ = δ α β . For a state expanded as ∣ ψ ( t ) ⟩ = ∑ α c α ( t ) ∣ r α ( t ) ⟩ |\psi(t)\rangle=\sum_{\alpha}c_{\alpha}(t)|r_{\alpha}(t)\rangle ∣ ψ ( t )⟩ = ∑ α c α ( t ) ∣ r α ( t )⟩ , the non-adiabatic connection and its role in the reconstructed state dynamics are
C α β ( t ) = ⟨ l α ( t ) ∣ r ˙ β ( t ) ⟩ , i c ˙ α ( t ) = z α ( t ) c α ( t ) − i ∑ β C α β ( t ) c β ( t ) . (62) C_{\alpha\beta}(t)=\langle l_{\alpha}(t)|\dot{r}_{\beta}(t)\rangle,\qquad i\dot{c}_{\alpha}(t)=z_{\alpha}(t)c_{\alpha}(t)-i\sum_{\beta}C_{\alpha\beta}(t)c_{\beta}(t). \tag*{(62)} C α β ( t ) = ⟨ l α ( t ) ∣ r ˙ β ( t )⟩ , i c ˙ α ( t ) = z α ( t ) c α ( t ) − i β ∑ C α β ( t ) c β ( t ) . (62)
for i ∂ t ∣ ψ ( t ) ⟩ = H ( t ) ∣ ψ ( t ) ⟩ i\partial_{t}|\psi(t)\rangle=H(t)|\psi(t)\rangle i ∂ t ∣ ψ ( t )⟩ = H ( t ) ∣ ψ ( t )⟩ . The connection coefficients depend on the smooth normalisation and phase convention for the modes, while the reconstructed state does not. For a small perturbation H ( t ) ↦ H ( t ) + ϵ V ( t ) H(t)\mapsto H(t)+\epsilon V(t) H ( t ) ↦ H ( t ) + ϵ V ( t ) , first-order perturbation theory at fixed t t t gives for a simple isolated pole
z α ( t ; ϵ ) = z α ( t ) + ϵ ⟨ l α ( t ) ∣ V ( t ) ∣ r α ( t ) ⟩ ⟨ l α ( t ) ∣ r α ( t ) ⟩ + O ( ϵ 2 ) . (63) z_{\alpha}(t;\epsilon)=z_{\alpha}(t)+\epsilon\frac{\langle l_{\alpha}(t)|V(t)|r_{\alpha}(t)\rangle}{\langle l_{\alpha}(t)|r_{\alpha}(t)\rangle}+\mathcal{O}(\epsilon^{2}). \tag*{(63)} z α ( t ; ϵ ) = z α ( t ) + ϵ ⟨ l α ( t ) ∣ r α ( t )⟩ ⟨ l α ( t ) ∣ V ( t ) ∣ r α ( t )⟩ + O ( ϵ 2 ) . (63)
This formula does not apply at an exact EP, where the individual eigenvectors coalesce and a Jordan-chain treatment is required.
General Observable Error Bound
Here we establish a general error bound for the estimation of biorthogonal expectation values. The result is independent of the origin of the approximation error and depends only on the operator-norm distance between the implemented filter and the ideal Gaussian filter. Specific approximation mechanisms are analysed in the following sections. Using the RLCU estimator introduced earlier, we now bound the error of the reconstructed biorthogonal expectation value.
Theorem 1 (General Observable Error Bound). Let G ( P ( z ) ) G(P(z)) G ( P ( z )) denote the ideal Gaussian filter and let G ~ ( P ( z ) ) \widetilde{G}(P(z)) G ( P ( z )) be its randomised implementation introduced in § 4 C, satisfying ∥ G ~ ( P ( z ) ) − G ( P ( z ) ) ∥ ≤ ϵ \lVert\widetilde{G}(P(z))-G(P(z))\rVert\le\epsilon ∥ G ( P ( z )) − G ( P ( z ))∥ ≤ ϵ . For X ∈ { O , I } \mathcal{X} \in\{\mathcal{O},\mathcal{I}\} X ∈ { O , I } , define N G ( X ) = Tr ( G ( P ( z ) ) ρ G ( P ( z ) ) † X ) N_G(\mathcal{X})=\operatorname{Tr}(G(P(z))\rho G(P(z))^\dagger\mathcal{X}) N G ( X ) = Tr ( G ( P ( z )) ρG ( P ( z ) ) † X ) , and N ~ G ( X ) = Tr ( G ~ ( P ( z ) ) ρ G ~ ( P ( z ) ) † X ) \widetilde{N}_G(\mathcal{X})=\operatorname{Tr}(\widetilde{G}(P(z))\rho\widetilde{G}(P(z))^\dagger\mathcal{X}) N G ( X ) = Tr ( G ( P ( z )) ρ G ( P ( z ) ) † X ) . Assume that the transition amplitudes defining N ~ G ( X ) \widetilde{N}_G(\mathcal{X}) N G ( X ) are estimated by independent Hadamard-test measurements with statistical error bounded by ∣ N ^ ( X ) − N ~ G ( X ) ∣ ≤ ∥ X ∥ ϵ s t a t \lvert\widehat{N}(\mathcal{X})-\widetilde{N}_G(\mathcal{X})\rvert\le\lVert\mathcal{X}\rVert\epsilon_{\mathrm{stat}} ∣ N ( X ) − N G ( X )∣ ≤ ∥ X ∥ ϵ stat , with probability at least 1 − ϑ 1-\vartheta 1 − ϑ . Then the estimation errors of the numerator and denominator satisfy
ϵ N : = ∣ N ^ ( O ) − N G ( O ) ∣ ≤ 2 μ 2 ∥ O ∥ ϵ + ∥ O ∥ ϵ s t a t , (64) \epsilon_N:=\left\lvert\widehat{N}(\mathcal{O})-N_G(\mathcal{O})\right\rvert\le2\mu^2\lVert\mathcal{O}\rVert\epsilon+\lVert\mathcal{O}\rVert\epsilon_{\mathrm{stat}}, \tag*{(64)} ϵ N := N ( O ) − N G ( O ) ≤ 2 μ 2 ∥ O ∥ ϵ + ∥ O ∥ ϵ stat , (64) ϵ D : = ∣ N ^ ( I ) − N G ( I ) ∣ ≤ 2 μ 2 ϵ + ϵ s t a t , (65) \epsilon_D:=\left\lvert\widehat{N}(\mathcal{I})-N_G(\mathcal{I})\right\rvert\le2\mu^2\epsilon+\epsilon_{\mathrm{stat}}, \tag*{(65)} ϵ D := N ( I ) − N G ( I ) ≤ 2 μ 2 ϵ + ϵ stat , (65) where μ = ∑ m ∣ w m ∣ \mu=\sum_m\lvert w_m\rvert μ = ∑ m ∣ w m ∣ is the LCU normalisation. Furthermore, if D G = N G ( I ) > ϵ D D_G=N_G(\mathcal{I})>\epsilon_D D G = N G ( I ) > ϵ D , then the reconstructed biorthogonal expectation value obeys
∣ N ^ ( O ) N ^ ( I ) − N G ( O ) D G ∣ ≤ ( ⟨ O ⟩ G + 1 ) ϵ D + ϵ N D G , \left\lvert\frac{\widehat{N}(\mathcal{O})}{\widehat{N}(\mathcal{I})}-\frac{N_G(\mathcal{O})}{D_G}\right\rvert\le\frac{(\langle\mathcal{O}\rangle_G+1)\epsilon_D+\epsilon_N}{D_G}, N ( I ) N ( O ) − D G N G ( O ) ≤ D G (⟨ O ⟩ G + 1 ) ϵ D + ϵ N , where ⟨ O ⟩ G = N G ( O ) D G \langle\mathcal{O}\rangle_G=\frac{N_G(\mathcal{O})}{D_G} ⟨ O ⟩ G = D G N G ( O ) .
Proof. The total estimation error is decomposed into an operator approximation error and a statistical estimation error N ^ ( X ) − N G ( X ) = ( N ^ ( X ) − N ~ G ( X ) ) + ( N ~ G ( X ) − N G ( X ) ) \widehat{N}(\mathcal{X})-N_G(\mathcal{X})=(\widehat{N}(\mathcal{X})-\widetilde{N}_G(\mathcal{X}))+(\widetilde{N}_G(\mathcal{X})-N_G(\mathcal{X})) N ( X ) − N G ( X ) = ( N ( X ) − N G ( X )) + ( N G ( X ) − N G ( X )) . Applying the triangle inequality gives ∣ N ^ ( X ) − N G ( X ) ∣ ≤ ϵ s t a t ( X ) + ϵ o p ( X ) \lvert\widehat{N}(\mathcal{X})-N_G(\mathcal{X})\rvert\le\epsilon_{\mathrm{stat}}^{(\mathcal{X})}+\epsilon_{\mathrm{op}}^{(\mathcal{X})} ∣ N ( X ) − N G ( X )∣ ≤ ϵ stat ( X ) + ϵ op ( X ) . The statistical contribution follows directly from the assumption ϵ s t a t ( X ) ≤ ∥ X ∥ ϵ s t a t \epsilon_{\mathrm{stat}}^{(\mathcal{X})}\le\lVert\mathcal{X}\rVert\epsilon_{\mathrm{stat}} ϵ stat ( X ) ≤ ∥ X ∥ ϵ stat .
For the operator contribution Δ = G ~ ( P ( z ) ) − G ( P ( z ) ) \Delta=\widetilde{G}(P(z))-G(P(z)) Δ = G ( P ( z )) − G ( P ( z )) , where ∥ Δ ∥ ≤ ϵ \lVert\Delta\rVert\le\epsilon ∥ Δ ∥ ≤ ϵ , so that G ~ = G + Δ \widetilde{G}=G+\Delta G = G + Δ . Expanding N ~ G ( X ) = Tr ( ( G + Δ ) ρ ( G + Δ ) † X ) \widetilde{N}_G(\mathcal{X})=\operatorname{Tr}((G+\Delta)\rho(G+\Delta)^\dagger\mathcal{X}) N G ( X ) = Tr (( G + Δ ) ρ ( G + Δ ) † X ) produces N ~ G − N G = Tr ( Δ ρ G † X ) + Tr ( G ρ Δ † X ) + Tr ( Δ ρ Δ † X ) \widetilde{N}_G-N_G=\operatorname{Tr}(\Delta\rho G^\dagger\mathcal{X})+\operatorname{Tr}(G\rho\Delta^\dagger\mathcal{X})+\operatorname{Tr}(\Delta\rho\Delta^\dagger\mathcal{X}) N G − N G = Tr ( Δ ρ G † X ) + Tr ( Gρ Δ † X ) + Tr ( Δ ρ Δ † X ) . Since the RLCU estimator reconstructs filtered expectation values with the normalisation factor μ 2 \mu^2 μ 2 , the induced operator error satisfies ϵ o p ( X ) ≤ 2 μ 2 ∥ X ∥ ϵ + O ( ϵ 2 ) \epsilon_{\mathrm{op}}^{(\mathcal{X})}\le2\mu^2\lVert\mathcal{X}\rVert\epsilon+O(\epsilon^2) ϵ op ( X ) ≤ 2 μ 2 ∥ X ∥ ϵ + O ( ϵ 2 ) , where the quadratic term is neglected for ϵ ≪ 1 \epsilon\ll1 ϵ ≪ 1 . Combining the statistical and operator contributions yields ϵ N ≤ 2 μ 2 ∥ O ∥ ϵ + ∥ O ∥ ϵ s t a t \epsilon_N\le2\mu^2\lVert\mathcal{O}\rVert\epsilon+\lVert\mathcal{O}\rVert\epsilon_{\mathrm{stat}} ϵ N ≤ 2 μ 2 ∥ O ∥ ϵ + ∥ O ∥ ϵ stat , and ϵ D ≤ 2 μ 2 ϵ + ϵ s t a t \epsilon_D\le2\mu^2\epsilon+\epsilon_{\mathrm{stat}} ϵ D ≤ 2 μ 2 ϵ + ϵ stat . Finally
N ^ D ^ − N D = D ( N ^ − N ) + N ( D − D ^ ) D ^ D , \frac{\widehat{N}}{\widehat{D}}-\frac{N}{D}=\frac{D(\widehat{N}-N)+N(D-\widehat{D})}{\widehat{D}D}, D N − D N = D D D ( N − N ) + N ( D − D ) ,
and the triangle inequality gives
∣ N ^ D ^ − N D ∣ ≤ ϵ N + ⟨ O ⟩ G ϵ D D ^ . \left\lvert\frac{\widehat{N}}{\widehat{D}}-\frac{N}{D}\right\rvert\le\frac{\epsilon_N+\langle\mathcal{O}\rangle_G\epsilon_D}{\widehat{D}}. D N − D N ≤ D ϵ N + ⟨ O ⟩ G ϵ D .
Using D G > ϵ D D_G>\epsilon_D D G > ϵ D , so that D ^ ≥ D G − ϵ D \widehat{D}\ge D_G-\epsilon_D D ≥ D G − ϵ D , and replacing the denominator by the weaker bound D G D_G D G yields
∣ N ^ D ^ − N G ( O ) D G ∣ ≤ ( ⟨ O ⟩ G + 1 ) ϵ D + ϵ N D G , \left\lvert\frac{\widehat{N}}{\widehat{D}}-\frac{N_G(\mathcal{O})}{D_G}\right\rvert\le\frac{(\langle\mathcal{O}\rangle_G+1)\epsilon_D+\epsilon_N}{D_G}, D N − D G N G ( O ) ≤ D G (⟨ O ⟩ G + 1 ) ϵ D + ϵ N ,
which proves the theorem.
Decomposition of the Filter Approximation Error
Theorem 1 establishes a general observable-estimation bound in terms of the operator approximation error ∥ G ~ − G ∥ \lVert\widetilde{G}-G\rVert ∥ G − G ∥ . Here we quantify this error by analyzing the successive approximations used to construct the randomised Gaussian filter. Specifically, we consider (i) the finite filter parameter τ \tau τ , (ii) the truncation of the Fourier integral, and (iii) the product-formula approximation of the unitary evolutions. These contributions are combined at the end of the section to obtain the overall operator approximation error.
Finite-τ \tau τ Approximation
The ideal Gaussian filter is obtained in the limit τ → ∞ \tau\to\infty τ → ∞ , for which the filter converges to the projector onto the zero-singular-value subspace of the Hermitian dilation. For finite τ \tau τ , the nonzero singular-value sectors remain exponentially suppressed, resulting in the first contribution to the operator approximation error.
Proposition 5 (Finite-τ \tau τ Approximation). Let P ( z ) \mathcal{P}(z) P ( z ) denote the Hermitian dilation evaluated at z = λ j z = \lambda_j z = λ j , and let G τ ( P ( z ) ) = exp ( − τ 2 2 P ( z ) 2 ) G_\tau(\mathcal{P}(z)) = \exp\left(-\frac{\tau^2}{2}\mathcal{P}(z)^2\right) G τ ( P ( z )) = exp ( − 2 τ 2 P ( z ) 2 ) . If γ j = min ϵ k ≠ 0 ∣ ϵ k ∣ = σ 1 ( H − λ j I ) \gamma_j = \min_{\epsilon_k \ne0} |\epsilon_k| = \sigma_1(H - \lambda_j I) γ j = min ϵ k = 0 ∣ ϵ k ∣ = σ 1 ( H − λ j I ) , then ∥ G τ ( P ( λ j ) ) − Π 0 ∥ ≤ e − τ 2 γ j 2 / 2 \left\|G_\tau(\mathcal{P}(\lambda_j)) - \Pi_0\right\| \le e^{-\tau^2\gamma_j^2/2} ∥ G τ ( P ( λ j )) − Π 0 ∥ ≤ e − τ 2 γ j 2 /2 , where Π 0 \Pi_0 Π 0 denotes the orthogonal projector onto ker ( P ( λ j ) ) \ker(\mathcal{P}(\lambda_j)) ker ( P ( λ j )) .
Proof. By the spectral theorem P ( λ j ) = ∑ k ϵ k ∣ χ k ⟩ ⟨ χ k ∣ \mathcal{P}(\lambda_j) = \sum_k \epsilon_k|\chi_k\rangle\langle\chi_k| P ( λ j ) = ∑ k ϵ k ∣ χ k ⟩ ⟨ χ k ∣ , and therefore G τ ( P ( λ j ) ) = ∑ k e − τ 2 ϵ k 2 / 2 ∣ χ k ⟩ ⟨ χ k ∣ G_\tau(\mathcal{P}(\lambda_j)) = \sum_k e^{-\tau^2\epsilon_k^2/2}|\chi_k\rangle\langle\chi_k| G τ ( P ( λ j )) = ∑ k e − τ 2 ϵ k 2 /2 ∣ χ k ⟩ ⟨ χ k ∣ . Separating the zero-eigenspace gives G τ ( P ( λ j ) ) = Π 0 + ∑ ϵ k ≠ 0 e − τ 2 ϵ k 2 / 2 ∣ χ k ⟩ ⟨ χ k ∣ G_\tau(\mathcal{P}(\lambda_j)) = \Pi_0 + \sum_{\epsilon_k\ne0}e^{-\tau^2\epsilon_k^2/2}|\chi_k\rangle\langle\chi_k| G τ ( P ( λ j )) = Π 0 + ∑ ϵ k = 0 e − τ 2 ϵ k 2 /2 ∣ χ k ⟩ ⟨ χ k ∣ . Hence
G τ ( P ( λ j ) ) − Π 0 = ∑ ϵ k ≠ 0 e − τ 2 ϵ k 2 / 2 ∣ χ k ⟩ ⟨ χ k ∣ . G_\tau(\mathcal{P}(\lambda_j))-\Pi_0 = \sum_{\epsilon_k\ne0}e^{-\tau^2\epsilon_k^2/2}|\chi_k\rangle\langle\chi_k|. G τ ( P ( λ j )) − Π 0 = ϵ k = 0 ∑ e − τ 2 ϵ k 2 /2 ∣ χ k ⟩ ⟨ χ k ∣.
Since ∣ ϵ k ∣ ≥ γ j |\epsilon_k| \ge\gamma_j ∣ ϵ k ∣ ≥ γ j , the spectral theorem implies
∥ G τ ( P ( λ j ) ) − Π 0 ∥ = max ϵ k ≠ 0 e − τ 2 ϵ k 2 / 2 ≤ e − τ 2 γ j 2 / 2 , \left\|G_\tau(\mathcal{P}(\lambda_j))-\Pi_0\right\| = \max_{\epsilon_k\ne0}e^{-\tau^2\epsilon_k^2/2} \le e^{-\tau^2\gamma_j^2/2}, ∥ G τ ( P ( λ j )) − Π 0 ∥ = ϵ k = 0 max e − τ 2 ϵ k 2 /2 ≤ e − τ 2 γ j 2 /2 ,
which completes the proof. □ \square □
Corollary 3. For any normalised state ∣ Φ 0 ⟩ |\Phi_0\rangle ∣ Φ 0 ⟩
∥ G τ ( P ( λ j ) ) ∣ Φ 0 ⟩ − Π 0 ∣ Φ 0 ⟩ ∥ ≤ e − τ 2 γ j 2 / 2 . \left\|G_\tau(\mathcal{P}(\lambda_j))|\Phi_0\rangle-\Pi_0|\Phi_0\rangle\right\| \le e^{-\tau^2\gamma_j^2/2}. ∥ G τ ( P ( λ j )) ∣ Φ 0 ⟩ − Π 0 ∣ Φ 0 ⟩ ∥ ≤ e − τ 2 γ j 2 /2 . In particular G τ ( P ( λ j ) ) ∣ Φ 0 ⟩ = Π 0 ∣ Φ 0 ⟩ + O ( e − τ 2 γ j 2 / 2 ) G_\tau(\mathcal{P}(\lambda_j))|\Phi_0\rangle= \Pi_0|\Phi_0\rangle+ O\left(e^{-\tau^2\gamma_j^2/2}\right) G τ ( P ( λ j )) ∣ Φ 0 ⟩ = Π 0 ∣ Φ 0 ⟩ + O ( e − τ 2 γ j 2 /2 ) .
Proof. The result follows immediately from the operator norm bound
∥ ( G τ ( P ( λ j ) ) − Π 0 ) ∣ Φ 0 ⟩ ∥ ≤ ∥ G τ ( P ( λ j ) ) − Π 0 ∥ ∥ ∣ Φ 0 ⟩ ∥ , \left\|(G_\tau(\mathcal{P}(\lambda_j))-\Pi_0)|\Phi_0\rangle\right\| \le\left\|G_\tau(\mathcal{P}(\lambda_j))-\Pi_0\right\|\left\||\Phi_0\rangle\right\|, ∥ ( G τ ( P ( λ j )) − Π 0 ) ∣ Φ 0 ⟩ ∥ ≤ ∥ G τ ( P ( λ j )) − Π 0 ∥ ∥ ∣ Φ 0 ⟩ ∥ ,
together with ∥ ∣ Φ 0 ⟩ ∥ = 1 \left\||\Phi_0\rangle\right\| = 1 ∥ ∣ Φ 0 ⟩ ∥ = 1 and Proposition 5. □ \square □
Fourier Truncation Error
The Gaussian filter admits the exact Fourier representation
G τ ( P ) = ∫ − ∞ ∞ p ( x ) e − i τ x P d x . G_\tau(\mathcal{P}) = \int_{-\infty}^{\infty} p(x)e^{-i\tau x\mathcal{P}}\,dx. G τ ( P ) = ∫ − ∞ ∞ p ( x ) e − i τ x P d x .
where p ( x ) = 1 2 π e − x 2 / 2 p(x) = \frac{1}{\sqrt{2\pi}}e^{-x^2/2} p ( x ) = 2 π 1 e − x 2 /2 . In practice, the integration domain is truncated to a finite interval [ − X max , X max ] [-X_{\max},X_{\max}] [ − X m a x , X m a x ] , introducing the second contribution to the operator approximation error.
Proposition 6 (Fourier Truncation Error). Define the truncated Gaussian filter
G τ , X ( P ) = ∫ − X max X max p ( x ) e − i τ x P d x . G_{\tau,X}(\mathcal{P}) = \int_{-X_{\max}}^{X_{\max}} p(x)e^{-i\tau x\mathcal{P}}\,dx. G τ , X ( P ) = ∫ − X m a x X m a x p ( x ) e − i τ x P d x . Then ∥ G τ ( P ) − G τ , X ( P ) ∥ ≤ ϵ c \left\|G_\tau(\mathcal{P})-G_{\tau,X}(\mathcal{P})\right\| \le\epsilon_c ∥ G τ ( P ) − G τ , X ( P ) ∥ ≤ ϵ c , where ϵ c = erfc ( X max 2 ) \epsilon_c = \operatorname{erfc}\left(\frac{X_{\max}}{\sqrt{2}}\right) ϵ c = erfc ( 2 X m a x ) . Furthermore ϵ c ≤ 2 π e − X max 2 / 2 X max \epsilon_c \le\sqrt{\frac{2}{\pi}}\frac{e^{-X_{\max}^2/2}}{X_{\max}} ϵ c ≤ π 2 X m a x e − X m a x 2 /2 .
Proof. Since P \mathcal{P} P is Hermitian, ∥ e − i τ x P ∥ = 1 \lVert e^{-i\tau x\mathcal{P}}\rVert= 1 ∥ e − i τ x P ∥ = 1 for every x x x . Therefore
∥ G τ ( P ) − G τ , X ( P ) ∥ = ∥ ∫ ∣ x ∣ > X max p ( x ) e − i τ x P d x ∥ ≤ ∫ ∣ x ∣ > X max p ( x ) ∥ e − i τ x P ∥ d x = ∫ ∣ x ∣ > X max p ( x ) d x = erfc ( X max 2 ) . \begin{aligned}
\lVert G_{\tau}(\mathcal{P})-G_{\tau,X}(\mathcal{P})\rVert
&=\left\lVert\int_{\lvert x\rvert>X_{\max}}p(x)e^{-i\tau x\mathcal{P}}\,dx\right\rVert\\
&\leq\int_{\lvert x\rvert>X_{\max}}p(x)\lVert e^{-i\tau x\mathcal{P}}\rVert\,dx\\
&=\int_{\lvert x\rvert>X_{\max}}p(x)\,dx\\
&=\operatorname{erfc}\left(\frac{X_{\max}}{\sqrt{2}}\right).
\end{aligned} ∥ G τ ( P ) − G τ , X ( P )∥ = ∫ ∣ x ∣ > X m a x p ( x ) e − i τ x P d x ≤ ∫ ∣ x ∣ > X m a x p ( x ) ∥ e − i τ x P ∥ d x = ∫ ∣ x ∣ > X m a x p ( x ) d x = erfc ( 2 X m a x ) .
Using the standard inequality erfc ( x ) ≤ e − x 2 π x \operatorname{erfc}(x)\leq\frac{e^{-x^2}}{\sqrt{\pi}x} erfc ( x ) ≤ π x e − x 2 , with x = X max / 2 x=X_{\max}/\sqrt{2} x = X m a x / 2 yields ϵ c ≤ 2 π e − X max 2 / 2 X max \epsilon_c\leq\sqrt{\frac{2}{\pi}}\frac{e^{-X_{\max}^2/2}}{X_{\max}} ϵ c ≤ π 2 X m a x e − X m a x 2 /2 , which completes the proof.
Quadrature Approximation
The truncated Fourier representation is evaluated numerically by replacing the continuous Gaussian integral with a finite quadrature rule. This introduces the third contribution to the operator approximation error.
Proposition 7 (Quadrature Approximation). Let
G τ , X ( P ( z ) ) = ∫ − X max X max p ( x ) e − i τ x P ( z ) d x , G_{\tau,X}(\mathcal{P}(z))=\int_{-X_{\max}}^{X_{\max}}p(x)e^{-i\tau x\mathcal{P}(z)}\,dx, G τ , X ( P ( z )) = ∫ − X m a x X m a x p ( x ) e − i τ x P ( z ) d x , and let
G τ , Q ( P ( z ) ) = ∑ m = − M M w m e − i τ x m P ( z ) G_{\tau,Q}(\mathcal{P}(z))=\sum_{m=-M}^{M}w_m e^{-i\tau x_m\mathcal{P}(z)} G τ , Q ( P ( z )) = m = − M ∑ M w m e − i τ x m P ( z ) be the corresponding quadrature approximation, where { x m , w m } \{x_m,w_m\} { x m , w m } denote the quadrature nodes and weights, respectively. Assume that the quadrature rule satisfies
∣ ∫ − X max X max p ( x ) f ( x ) d x − ∑ m = − M M w m f ( x m ) ∣ ≤ ϵ q ∥ f ∥ ∞ \left\lvert\int_{-X_{\max}}^{X_{\max}}p(x)f(x)\,dx-\sum_{m=-M}^{M}w_mf(x_m)\right\rvert\leq\epsilon_q\lVert f\rVert_\infty ∫ − X m a x X m a x p ( x ) f ( x ) d x − m = − M ∑ M w m f ( x m ) ≤ ϵ q ∥ f ∥ ∞ for every continuous function f f f on [ − X max , X max ] [-X_{\max},X_{\max}] [ − X m a x , X m a x ] . Then ∥ G τ , X ( P ( z ) ) − G τ , Q ( P ( z ) ) ∥ ≤ ϵ q \lVert G_{\tau,X}(\mathcal{P}(z))-G_{\tau,Q}(\mathcal{P}(z))\rVert\leq\epsilon_q ∥ G τ , X ( P ( z )) − G τ , Q ( P ( z ))∥ ≤ ϵ q .
Proof. Define the operator-valued function F ( x ) = e − i τ x P ( z ) F(x)=e^{-i\tau x\mathcal{P}(z)} F ( x ) = e − i τ x P ( z ) . Since P ( z ) \mathcal{P}(z) P ( z ) is Hermitian, ∥ F ( x ) ∥ = ∥ e − i τ x P ( z ) ∥ = 1 \lVert F(x)\rVert=\lVert e^{-i\tau x\mathcal{P}(z)}\rVert=1 ∥ F ( x )∥ = ∥ e − i τ x P ( z ) ∥ = 1 for every x x x . Applying the quadrature error bound to the operator-valued integrand gives
∥ G τ , X ( P ( z ) ) − G τ , Q ( P ( z ) ) ∥ = ∥ ∫ − X max X max p ( x ) F ( x ) d x − ∑ m = − M M w m F ( x m ) ∥ ≤ ϵ q sup x ∈ [ − X max , X max ] ∥ F ( x ) ∥ = ϵ q . \begin{aligned}
\lVert G_{\tau,X}(\mathcal{P}(z))-G_{\tau,Q}(\mathcal{P}(z))\rVert
&=\left\lVert\int_{-X_{\max}}^{X_{\max}}p(x)F(x)\,dx-\sum_{m=-M}^{M}w_mF(x_m)\right\rVert\\
&\leq\epsilon_q\sup_{x\in[-X_{\max},X_{\max}]}\lVert F(x)\rVert=\epsilon_q.
\end{aligned} ∥ G τ , X ( P ( z )) − G τ , Q ( P ( z ))∥ = ∫ − X m a x X m a x p ( x ) F ( x ) d x − m = − M ∑ M w m F ( x m ) ≤ ϵ q x ∈ [ − X m a x , X m a x ] sup ∥ F ( x )∥ = ϵ q .
Hamiltonian-Simulation Error
The randomised Gaussian filter requires repeated implementations of the unitary evolution U ( t ) = e − i t P ( z ) U(t)=e^{-it\mathcal{P}(z)} U ( t ) = e − i t P ( z ) . Since exact exponentiation is not available on a digital quantum computer, each unitary is replaced by the compensated multiproduct approximation U K ( t ) U_K(t) U K ( t ) from § 4 D. The resulting approximation induces an additional error in the implementation of the Gaussian filter, which we quantify below.
Proposition 8 (Hamiltonian-simulation error). Let U K ( t ) U_K(t) U K ( t ) denote the compensated multiproduct approximation of U ( t ) = e − i t P ( z ) U(t)=e^{-it\mathcal{P}(z)} U ( t ) = e − i t P ( z ) , and suppose that the compensation coefficients are chosen such that all BCH error terms up to order 2 K 2K 2 K are cancelled. Then, for every evolution time satisfying ∣ t ∣ ≤ T max |t|\leq T_{\max} ∣ t ∣ ≤ T m a x the simulation error obeys
∥ U K ( t ) − e − i t P ( z ) ∥ ≤ C K ∣ t ∣ ( δ t ) 2 K Λ 2 K + 1 , (66) \left\|U_K(t)-e^{-it\mathcal{P}(z)}\right\|\leq C_K|t|(\delta t)^{2K}\Lambda^{2K+1}, \tag*{(66)} U K ( t ) − e − i t P ( z ) ≤ C K ∣ t ∣ ( δ t ) 2 K Λ 2 K + 1 , (66) where Λ = max α ∣ h α ∣ \Lambda=\max_\alpha|h_\alpha| Λ = max α ∣ h α ∣ denotes the operator scale of the Pauli decomposition of P ( z ) \mathcal{P}(z) P ( z ) , and C K C_K C K is a constant depending only on the multiproduct construction. Consequently, the approximate Gaussian filter
G ~ τ , Q , K ( P ( z ) ) = ∑ m = − M M w m U K ( τ x m ) \widetilde{G}_{\tau,Q,K}(\mathcal{P}(z))=\sum_{m=-M}^{M}w_mU_K(\tau x_m) G τ , Q , K ( P ( z )) = m = − M ∑ M w m U K ( τ x m ) satisfies ∥ G ~ τ , Q , K − G ~ τ , Q ∥ ≤ ϵ p f \left\|\widetilde{G}_{\tau,Q,K}-\widetilde{G}_{\tau,Q}\right\|\leq\epsilon_{\mathrm{pf}} G τ , Q , K − G τ , Q ≤ ϵ pf , where
ϵ p f = C K ( δ t ) 2 K Λ 2 K + 1 ∑ m = − M M ∣ w m ∣ ∣ τ x m ∣ . (67) \epsilon_{\mathrm{pf}}=C_K(\delta t)^{2K}\Lambda^{2K+1}\sum_{m=-M}^{M}|w_m||\tau x_m|. \tag*{(67)} ϵ pf = C K ( δ t ) 2 K Λ 2 K + 1 m = − M ∑ M ∣ w m ∣∣ τ x m ∣. (67) In particular ϵ p f = O ( τ ( δ t ) 2 K Λ 2 K + 1 ) \epsilon_{\mathrm{pf}}=O\left(\tau(\delta t)^{2K}\Lambda^{2K+1}\right) ϵ pf = O ( τ ( δ t ) 2 K Λ 2 K + 1 ) .
Proof. § 4 D constructs the compensated multiproduct approximation U K ( t ) U_K(t) U K ( t ) by combining second-order Suzuki product formulas in such a way that the BCH contributions through order 2 K 2K 2 K cancel exactly. The BCH remainder hence begins at order 2 K + 1 2K+1 2 K + 1 , implying U K ( t ) = e − i t P ( z ) + O ( ∣ t ∣ ( δ t ) 2 K Λ 2 K + 1 ) U_K(t)=e^{-it\mathcal{P}(z)}+O\left(|t|(\delta t)^{2K}\Lambda^{2K+1}\right) U K ( t ) = e − i t P ( z ) + O ( ∣ t ∣ ( δ t ) 2 K Λ 2 K + 1 ) , which establishes Eq. (66) .
The approximate quadrature filter is obtained by replacing every exact unitary in the quadrature expansion G ~ τ , Q = ∑ m w m e − i τ x m P ( z ) \widetilde{G}_{\tau,Q}=\sum_m w_m e^{-i\tau x_m\mathcal{P}(z)} G τ , Q = ∑ m w m e − i τ x m P ( z ) , with its compensated approximation U K ( τ x m ) U_K(\tau x_m) U K ( τ x m ) . Therefore
G ~ τ , Q , K − G ~ τ , Q = ∑ m w m ( U K ( τ x m ) − e − i τ x m P ( z ) ) . \widetilde{G}_{\tau,Q,K}-\widetilde{G}_{\tau,Q}=\sum_m w_m\left(U_K(\tau x_m)-e^{-i\tau x_m\mathcal{P}(z)}\right). G τ , Q , K − G τ , Q = m ∑ w m ( U K ( τ x m ) − e − i τ x m P ( z ) ) .
Applying the triangle inequality together with Eq. (66) gives
∥ G ~ τ , Q , K − G ~ τ , Q ∥ ≤ ∑ m ∣ w m ∣ ∥ U K ( τ x m ) − e − i τ x m P ( z ) ∥ , \left\|\widetilde{G}_{\tau,Q,K}-\widetilde{G}_{\tau,Q}\right\|\leq\sum_m |w_m|\left\|U_K(\tau x_m)-e^{-i\tau x_m\mathcal{P}(z)}\right\|, G τ , Q , K − G τ , Q ≤ m ∑ ∣ w m ∣ U K ( τ x m ) − e − i τ x m P ( z ) ,
which immediately yields Eq. (67) . The asymptotic scaling follows by observing that the quadrature weights satisfy
∑ m ∣ w m ∣ ∣ x m ∣ = O ( 1 ) , \sum_m |w_m||x_m|=O(1), m ∑ ∣ w m ∣∣ x m ∣ = O ( 1 ) ,
for a convergent Gaussian quadrature rule.
:::
Corollary 4 (Combined Filter Approximation Error). Let G τ ( P ( z ) ) G_\tau(\mathcal{P}(z)) G τ ( P ( z )) denote the ideal Gaussian filter and let G ~ ( P ( z ) ) \widetilde{G}(\mathcal{P}(z)) G ( P ( z )) denote its implemented approximation obtained by (i) finite filter parameter τ \tau τ , (ii) truncation of the Fourier integral, (iii) quadrature discretisation, and (iv) product-formula approximation of each unitary evolution. If the corresponding operator errors satisfy
∥ G ∞ − G τ ∥ ≤ ϵ τ , ∥ G τ − G τ , X ∥ ≤ ϵ c , ∥ G τ , X − G τ , Q ∥ ≤ ϵ q , ∥ G τ , Q − G ~ ∥ ≤ ϵ p f , \left\|G_\infty-G_\tau\right\|\leq\epsilon_\tau,\qquad\left\|G_\tau-G_{\tau,X}\right\|\leq\epsilon_c,\qquad\left\|G_{\tau,X}-G_{\tau,Q}\right\|\leq\epsilon_q,\qquad\left\|G_{\tau,Q}-\widetilde{G}\right\|\leq\epsilon_{\mathrm{pf}}, ∥ G ∞ − G τ ∥ ≤ ϵ τ , ∥ G τ − G τ , X ∥ ≤ ϵ c , ∥ G τ , X − G τ , Q ∥ ≤ ϵ q , G τ , Q − G ≤ ϵ pf , then ∥ G ∞ ( P ( z ) ) − G ~ ( P ( z ) ) ∥ ≤ ϵ τ + ϵ c + ϵ q + ϵ p f \left\|G_\infty(\mathcal{P}(z))-\widetilde{G}(\mathcal{P}(z))\right\|\leq\epsilon_\tau+\epsilon_c+\epsilon_q+\epsilon_{\mathrm{pf}} G ∞ ( P ( z )) − G ( P ( z )) ≤ ϵ τ + ϵ c + ϵ q + ϵ pf .
Proof. By repeated application of the triangle inequality
∥ G ∞ − G ~ ∥ ≤ ∥ G ∞ − G τ ∥ + ∥ G τ − G τ , X ∥ + ∥ G τ , X − G τ , Q ∥ + ∥ G τ , Q − G ~ ∥ , \left\|G_\infty-\widetilde{G}\right\|\leq\left\|G_\infty-G_\tau\right\|+\left\|G_\tau-G_{\tau,X}\right\|+\left\|G_{\tau,X}-G_{\tau,Q}\right\|+\left\|G_{\tau,Q}-\widetilde{G}\right\|, G ∞ − G ≤ ∥ G ∞ − G τ ∥ + ∥ G τ − G τ , X ∥ + ∥ G τ , X − G τ , Q ∥ + G τ , Q − G ,
and substituting the bounds established in Propositions 5-8 yield the result.
Corollary 5 (Final Observable Error). Let G ~ ( P ( z ) ) \widetilde{G}(\mathcal{P}(z)) G ( P ( z )) denote the implemented randomised Gaussian filter obtained from the finite filter parameter, Fourier truncation, quadrature approximation, and product-formula simulation. Then the estimated biorthogonal expectation value satisfies
∣ ⟨ O ~ ⟩ j − ⟨ O ⟩ j ∣ ≤ ( ⟨ O ⟩ j + 1 ) ϵ D + ϵ N D G . \left|\langle\widetilde{O}\rangle_j-\langle O\rangle_j\right|\leq\frac{(\langle O\rangle_j+1)\epsilon_D+\epsilon_N}{D_G}. ⟨ O ⟩ j − ⟨ O ⟩ j ≤ D G (⟨ O ⟩ j + 1 ) ϵ D + ϵ N . where ϵ N = ∥ O ∥ ( 2 μ 2 ϵ + ϵ n ) \epsilon_N=\lVert O\rVert(2\mu^2\epsilon+\epsilon_n) ϵ N = ∥ O ∥ ( 2 μ 2 ϵ + ϵ n ) , ϵ D = 2 μ 2 ϵ + ϵ n \epsilon_D=2\mu^2\epsilon+\epsilon_n ϵ D = 2 μ 2 ϵ + ϵ n , and ϵ = ϵ τ + ϵ c + ϵ q + ϵ p f \epsilon=\epsilon_\tau+\epsilon_c+\epsilon_q+\epsilon_{\mathrm{pf}} ϵ = ϵ τ + ϵ c + ϵ q + ϵ pf . Here, ϵ τ \epsilon_\tau ϵ τ , ϵ c \epsilon_c ϵ c , ϵ q \epsilon_q ϵ q , and ϵ p f \epsilon_{\mathrm{pf}} ϵ pf denote the finite-filter, Fourier truncation, quadrature, and product-formula approximation errors, respectively, while ϵ n \epsilon_n ϵ n is the statistical sampling error.
Proof. By Corollary 4, ∥ G ( P ( z ) ) − G ~ ( P ( z ) ) ∥ ≤ ϵ \lVert G(\mathcal{P}(z))-\widetilde{G}(\mathcal{P}(z))\rVert\leq\epsilon ∥ G ( P ( z )) − G ( P ( z ))∥ ≤ ϵ , where ϵ = ϵ τ + ϵ c + ϵ q + ϵ p f \epsilon=\epsilon_\tau+\epsilon_c+\epsilon_q+\epsilon_{\mathrm{pf}} ϵ = ϵ τ + ϵ c + ϵ q + ϵ pf . Applying Theorem 1 with this operator approximation immediately yields the bounds for ϵ N \epsilon_N ϵ N , ϵ D \epsilon_D ϵ D , and hence the normalised biorthogonal expectation value.
Unbiased Randomised Estimator
The RLCU construction replaces the coherent implementation of the Gaussian filter by stochastic sampling of individual unitary evolutions. Consequently, filtered matrix elements are evaluated as expectation values over a classical probability distribution rather than through direct implementation of the non-unitary operator. For an arbitrary embedded operator M ∈ { O , I } \mathcal{M}\in\{\mathcal{O},\mathcal{I}\} M ∈ { O , I } , the filtered quantity of interest is N G ~ ( M ) = Tr ( G ~ τ ( P ( z ) ) ρ G ~ τ ( P ( z ) ) † M ) N_{\widetilde{G}}(\mathcal{M})=\operatorname{Tr}\left(\widetilde{G}_\tau(\mathcal{P}(z))\rho\widetilde{G}_\tau(\mathcal{P}(z))^\dagger\mathcal{M}\right) N G ( M ) = Tr ( G τ ( P ( z )) ρ G τ ( P ( z ) ) † M ) . Using the RLCU representation G ~ τ ( P ( z ) ) = μ ∑ m p ~ m U m \widetilde{G}_\tau(\mathcal{P}(z))=\mu\sum_m\widetilde{p}_mU_m G τ ( P ( z )) = μ ∑ m p m U m , this quantity becomes
N G ~ ( M ) = μ 2 ∑ m , m ′ p ~ m p ~ m ′ Tr ( U m ρ U m ′ † M ) . (68) N_{\widetilde{G}}(\mathcal{M})=\mu^2\sum_{m,m'}\widetilde{p}_m\widetilde{p}_{m'}\operatorname{Tr}\left(U_m\rho U_{m'}^\dagger\mathcal{M}\right). \tag*{(68)} N G ( M ) = μ 2 m , m ′ ∑ p m p m ′ Tr ( U m ρ U m ′ † M ) . (68)
Since the indices m m m and m ′ m' m ′ are sampled independently from the distribution p ~ \widetilde{p} p , Eq. (68) admits a direct stochastic interpretation.
Proposition 9 (Unbiased Randomised Estimator). Draw two independent random indices m , m ′ ∼ p ~ m,m'\sim\widetilde{p} m , m ′ ∼ p , and define the random variable X ( m , m ′ ; M ) = μ 2 Tr ( U m ρ U m ′ † M ) X(m,m';\mathcal{M})=\mu^2\operatorname{Tr}\left(U_m\rho U_{m'}^\dagger\mathcal{M}\right) X ( m , m ′ ; M ) = μ 2 Tr ( U m ρ U m ′ † M ) . Then E [ X ] = N G ~ ( M ) \mathbb{E}[X]=N_{\widetilde{G}}(\mathcal{M}) E [ X ] = N G ( M ) . Consequently, the empirical estimator N ^ G ~ ( M ) = 1 N s ∑ k = 1 N s X k \widehat{N}_{\widetilde{G}}(\mathcal{M})=\frac{1}{N_s}\sum_{k=1}^{N_s}X_k N G ( M ) = N s 1 ∑ k = 1 N s X k is unbiased, E [ N ^ G ~ ( M ) ] = N G ~ ( M ) \mathbb{E}\left[\widehat{N}_{\widetilde{G}}(\mathcal{M})\right]=N_{\widetilde{G}}(\mathcal{M}) E [ N G ( M ) ] = N G ( M ) , where N s N_s N s denotes the number of independent randomised samples.
Proof. Taking the expectation over the independent sampling of the two indices gives
E [ X ] = ∑ m , m ′ p ~ m p ~ m ′ X ( m , m ′ ; M ) = μ 2 ∑ m , m ′ p ~ m p ~ m ′ Tr ( U m ρ U m ′ † M ) = N G ~ ( M ) , (69) \begin{aligned}
\mathbb{E}[X]=\sum_{m,m'}\widetilde{p}_m\widetilde{p}_{m'}X(m,m';\mathcal{M}) \\
&=\mu^2\sum_{m,m'}\widetilde{p}_m\widetilde{p}_{m'}\operatorname{Tr}\left(U_m\rho U_{m'}^\dagger\mathcal{M}\right) \\
&=N_{\widetilde{G}}(\mathcal{M}), \tag*{(69)}
\end{aligned} E [ X ] = m , m ′ ∑ p m p m ′ X ( m , m ′ ; M ) = μ 2 m , m ′ ∑ p m p m ′ Tr ( U m ρ U m ′ † M ) = N G ( M ) , (69)
which is precisely Eq. (68) . Since the empirical estimator is the arithmetic mean of independent copies of the random variable X X X , N ^ G ~ = 1 N s ∑ k = 1 N s X k \widehat{N}_{\widetilde{G}}=\frac{1}{N_s}\sum_{k=1}^{N_s}X_k N G = N s 1 ∑ k = 1 N s X k . Linearity of expectation immediately yields E [ N ^ G ~ ( M ) ] = N G ~ ( M ) \mathbb{E}\left[\widehat{N}_{\widetilde{G}}(\mathcal{M})\right]=N_{\widetilde{G}}(\mathcal{M}) E [ N G ( M ) ] = N G ( M ) . Therefore the estimator is unbiased.
For the non-Hermitian eigenvalue problem, the transition amplitudes Tr ( U m ρ U m ′ † M ) \operatorname{Tr}\left(U_m\rho U_{m'}^\dagger\mathcal{M}\right) Tr ( U m ρ U m ′ † M ) are evaluated using the Hadamard-test procedure introduced in § 4 E. Applying the estimator to M = O \mathcal{M}=\mathcal{O} M = O and M = I \mathcal{M}=\mathcal{I} M = I yields unbiased estimates of the filtered numerator and denominator, respectively, from which the biorthogonal expectation value is reconstructed.
Necessity of the Hermitian Dilation for Non-Hermitian Gaussian Spectral Filtering
The Gaussian spectral-filtering algorithm relies on the representation of the filter as a weighted superposition of unitary real-time evolutions. For non-Hermitian operators, two distinct obstructions arise. First, the right-eigenvector basis is generally nonorthogonal, so Gaussian filtering does not produce an orthogonal spectral decomposition. Second, the corresponding real-time evolution is non-unitary and therefore cannot be implemented directly within the RLCU framework. We show that both obstructions are removed by the Hermitian dilation introduced earlier.
Direct Gaussian Filtering
Consider the shifted non-Hermitian operator H z = H − z I H_z = H-zI H z = H − z I and the formal Gaussian filter G τ ( H z ) = exp ( − τ 2 2 H z 2 ) G_\tau(H_z)=\exp\left(-\frac{\tau^2}{2}H_z^2\right) G τ ( H z ) = exp ( − 2 τ 2 H z 2 ) . Equivalently, G τ ( H z ) = ∫ − ∞ ∞ d x 2 π e − x 2 / 2 e − i τ x H z G_\tau(H_z)=\int_{-\infty}^{\infty}\frac{dx}{\sqrt{2\pi}}e^{-x^2/2}e^{-i\tau xH_z} G τ ( H z ) = ∫ − ∞ ∞ 2 π d x e − x 2 /2 e − i τ x H z . Suppose that H H H is diagonalisable H = ∑ j λ j ∣ r j ⟩ ⟨ l j ∣ H=\sum_j\lambda_j|r_j\rangle\langle l_j| H = ∑ j λ j ∣ r j ⟩ ⟨ l j ∣ , ⟨ l i ∣ r j ⟩ = δ i j \langle l_i|r_j\rangle=\delta_{ij} ⟨ l i ∣ r j ⟩ = δ ij , where λ j = ω j − i γ j \lambda_j=\omega_j-i\gamma_j λ j = ω j − i γ j . Writing ∣ ψ 0 ⟩ = ∑ j c j ∣ r j ⟩ |\psi_0\rangle=\sum_j c_j|r_j\rangle ∣ ψ 0 ⟩ = ∑ j c j ∣ r j ⟩ , the filtered state becomes
∣ ψ τ ( z ) ⟩ = ∑ j c j e − τ 2 2 ( λ j − z ) 2 ∣ r j ⟩ . (70) |\psi_\tau(z)\rangle=\sum_j c_j e^{-\frac{\tau^2}{2}(\lambda_j-z)^2}|r_j\rangle. \tag*{(70)} ∣ ψ τ ( z )⟩ = j ∑ c j e − 2 τ 2 ( λ j − z ) 2 ∣ r j ⟩ . (70)
Although Eq. (70) resembles the Hermitian case, its norm is
⟨ ψ τ ( z ) ∣ ψ τ ( z ) ⟩ = ∑ j ∣ c j ∣ 2 e − τ 2 2 Re [ ( λ j − z ) 2 ] ⟨ r j ∣ r j ⟩ + ∑ i ≠ j c i ∗ c j e − τ 2 2 ( λ i ∗ − z ∗ ) 2 e − τ 2 2 ( λ j − z ) 2 ⟨ r i ∣ r j ⟩ . (71) \langle\psi_\tau(z)|\psi_\tau(z)\rangle=\sum_j |c_j|^2e^{-\frac{\tau^2}{2}\operatorname{Re}[(\lambda_j-z)^2]}\langle r_j|r_j\rangle
+\sum_{i\ne j}c_i^*c_j e^{-\frac{\tau^2}{2}(\lambda_i^*-z^*)^2}e^{-\frac{\tau^2}{2}(\lambda_j-z)^2}\langle r_i|r_j\rangle. \tag*{(71)} ⟨ ψ τ ( z ) ∣ ψ τ ( z )⟩ = j ∑ ∣ c j ∣ 2 e − 2 τ 2 Re [( λ j − z ) 2 ] ⟨ r j ∣ r j ⟩ + i = j ∑ c i ∗ c j e − 2 τ 2 ( λ i ∗ − z ∗ ) 2 e − 2 τ 2 ( λ j − z ) 2 ⟨ r i ∣ r j ⟩ . (71)
The second summation vanishes only when the right eigenvectors are orthogonal.
Proposition 10. Direct Gaussian filtering of a non-Hermitian operator is not an orthogonal spectral projection.
Proof. For Hermitian operators, ⟨ E i ∣ E j ⟩ = δ i j \langle E_i|E_j\rangle=\delta_{ij} ⟨ E i ∣ E j ⟩ = δ ij and the filtered norm reduces to a sum of independent Gaussian contributions. For a non-Hermitian operator, ⟨ r i ∣ r j ⟩ ≠ 0 \langle r_i|r_j\rangle\ne0 ⟨ r i ∣ r j ⟩ = 0 , ( i ≠ j ) (i\ne j) ( i = j ) , so Eq. (71) contains interference terms coupling different eigenvalues. As a result, the measured response depends not only on the eigenvalues but also on the overlap matrix S i j = ⟨ r i ∣ r j ⟩ S_{ij}=\langle r_i|r_j\rangle S ij = ⟨ r i ∣ r j ⟩ , and thereby no longer represents an orthogonal spectral decomposition. Moreover, the response associated with a single eigenvalue satisfies
∣ e − τ 2 2 ( λ j − z ) 2 ∣ = exp [ − τ 2 2 ( ( ω j − Re z ) 2 − ( γ j + Im z ) 2 ) ] . \left|e^{-\frac{\tau^2}{2}(\lambda_j-z)^2}\right|=\exp\left[-\frac{\tau^2}{2}\left((\omega_j-\operatorname{Re}z)^2-(\gamma_j+\operatorname{Im}z)^2\right)\right]. e − 2 τ 2 ( λ j − z ) 2 = exp [ − 2 τ 2 ( ( ω j − Re z ) 2 − ( γ j + Im z ) 2 ) ] .
Whenever ( γ j + Im z ) 2 > ( ω j − Re z ) 2 (\gamma_j+\operatorname{Im}z)^2>(\omega_j-\operatorname{Re}z)^2 ( γ j + Im z ) 2 > ( ω j − Re z ) 2 , this amplification factor exceeds unity. Hence the direct Gaussian filter is generally not contractive ∥ G τ ( H z ) ∥ > 1 \lVert G_\tau(H_z)\rVert>1 ∥ G τ ( H z )∥ > 1 in contrast to the Hermitian case, where ∥ G τ ( H − ω I ) ∥ ≤ 1 \lVert G_\tau(H-\omega I)\rVert\le1 ∥ G τ ( H − ω I )∥ ≤ 1 . Although this loss of contractivity does not invalidate the operator mathematically, it illustrates that the stability properties of the Hermitian Gaussian filter are lost. Together with the interference terms arising from eigenvector nonorthogonality, this shows that direct Gaussian filtering is neither an orthogonal spectral projector nor a stable filtering procedure for generic non-Hermitian operators.
:::
We now turn to the second limitation.
Proposition 11. The Fourier representation of the direct Gaussian filter is generally not realisable as a RLCU constructed from unitary evolutions.
Proof. Acting on a right eigenvector gives e − i τ x A ( z ) ∣ r j ⟩ = e − i τ x ( ω j − Re z ) e − τ x ( γ j + Im z ) ∣ r j ⟩ e^{-i\tau xA(z)}|r_j\rangle=e^{-i\tau x(\omega_j-\operatorname{Re}z)}e^{-\tau x(\gamma_j+\operatorname{Im}z)}|r_j\rangle e − i τ x A ( z ) ∣ r j ⟩ = e − i τ x ( ω j − Re z ) e − τ x ( γ j + Im z ) ∣ r j ⟩ . Hence ∥ e − i τ x H z ∥ ≠ 1 \lVert e^{-i\tau xH_z}\rVert\ne1 ∥ e − i τ x H z ∥ = 1 , except in the Hermitian limit γ j + Im z = 0 \gamma_j+\operatorname{Im}z=0 γ j + Im z = 0 . The elementary evolution operators entering the Fourier integral are thus non-unitary and cannot serve as the unitary primitives required by the RLCU construction.
:::
Resolution by Hermitian Dilation
The Hermitian dilation replaces the shifted operator H z H_z H z by P ( z ) = ( 0 H z H z † 0 ) \mathcal{P}(z)=\begin{pmatrix}0&H_z\\H_z^\dagger&0\end{pmatrix} P ( z ) = ( 0 H z † H z 0 ) , which is Hermitian for every complex parameter z z z .
Proposition 12. The Hermitian dilation simultaneously removes both obstructions of direct Gaussian filtering.
Proof. Since P ( z ) = P ( z ) † \mathcal{P}(z)=\mathcal{P}(z)^{\dagger} P ( z ) = P ( z ) † , its spectrum is real spec ( P ( z ) ) = { ± σ k ( H Z ) } \operatorname{spec}(\mathcal{P}(z))=\{\pm\sigma_k(H_Z)\} spec ( P ( z )) = { ± σ k ( H Z )} , and its eigenvectors form an orthonormal basis. Consequently, G τ ( P ( z ) ) = exp ( − τ 2 2 P ( z ) 2 ) G_{\tau}(\mathcal{P}(z))=\exp\left(-\frac{\tau^2}{2}\mathcal{P}(z)^2\right) G τ ( P ( z )) = exp ( − 2 τ 2 P ( z ) 2 ) is an orthogonal spectral filter. Moreover, G τ ( P ( z ) ) = ∫ − ∞ ∞ d x 2 π e − x 2 / 2 e − i τ x P ( z ) G_{\tau}(\mathcal{P}(z))=\int_{-\infty}^{\infty}\frac{dx}{\sqrt{2\pi}}e^{-x^2/2}e^{-i\tau x\mathcal{P}(z)} G τ ( P ( z )) = ∫ − ∞ ∞ 2 π d x e − x 2 /2 e − i τ x P ( z ) , where ∥ e − i τ x P ( z ) ∥ = 1 \left\|e^{-i\tau x\mathcal{P}(z)}\right\|=1 e − i τ x P ( z ) = 1 for every x ∈ R x\in\mathbb{R} x ∈ R . Hence, the Gaussian filter admits a representation entirely in terms of unitary evolutions and can be implemented by the RLCU algorithm developed in this work. □
Corollary 6. When z = λ j z=\lambda_j z = λ j , the kernel of P ( z ) \mathcal{P}(z) P ( z ) is ker ( P ( λ j ) ) = span { ( ∣ L j ⟩ 0 ) , ( 0 ∣ R j ⟩ ) } \operatorname{ker}(\mathcal{P}(\lambda_j))=\operatorname{span}\left\{\begin{pmatrix}|L_j\rangle\\0\end{pmatrix},\begin{pmatrix}0\\|R_j\rangle\end{pmatrix}\right\} ker ( P ( λ j )) = span { ( ∣ L j ⟩ 0 ) , ( 0 ∣ R j ⟩ ) } or, equivalently, any nonzero linear combination of these vectors. Hence, the Gaussian filter constructed from P ( z ) \mathcal{P}(z) P ( z ) isolates the left and right eigenspaces associated with the target eigenvalue while maintaining a unitary implementation throughout the algorithm.
Therefore, the filtered zero mode simultaneously encodes the left and right eigenvectors required for evaluating the biorthogonal expectation value ⟨ O ⟩ j = ⟨ L j ∣ O ∣ R j ⟩ ⟨ L j ∣ R j ⟩ \langle O\rangle_j=\frac{\langle L_j|O|R_j\rangle}{\langle L_j|R_j\rangle} ⟨ O ⟩ j = ⟨ L j ∣ R j ⟩ ⟨ L j ∣ O ∣ R j ⟩ . Consequently, the Hermitian dilation resolves both the geometric obstruction caused by eigenvector nonorthogonality and the algorithmic obstruction associated with non-unitary time evolution.