Detecting Hidden Quasiparticle Peaks in Quantum Critical Metals

Some metals, when tuned to a quantum critical point, seem to lose their electrons' particle-like behavior. Simulations of these "strange metals" produce data only in imaginary time, a mathematical stand-in for real time that keeps the simulations stable. Recovering the physically meaningful real-frequency spectrum from that data means inverting a strong smoothing operation, and this inversion is badly unstable. The flagship simulation of an Ising-nematic critical metal (a metal on the edge of spontaneously picking a preferred lattice direction) therefore left its central question open. Near the critical point, the electrons at the "antinodal" parts of the Fermi surface look completely incoherent, but a narrow residual quasiparticle peak "cannot be definitively ruled out" (Superconductivity and non-Fermi liquid behavior near a nematic quantum critical point). A quasiparticle peak is the spectral signature of a long-lived electron-like excitation.

We propose to stop trying to reconstruct the whole spectrum and ask the question that is actually at stake: is there a narrow peak of weight at least $Z$ hidden under a broad continuum, and how small a $Z$ can this data ever reveal? This reframes the problem as event detection. A narrow spectral peak buried in a broad background is the frequency-domain twin of a brief acoustic click buried in a long clip. The audio literature has already learned that averaging destroys such signals and that "is it present?" is far easier to answer than "reconstruct everything." The null hypothesis still needs a physically realistic background. We obtain it from renormalization-group flow equations that have been analytically continued to real frequencies, which compute spectra directly on the real axis with no inversion step.

Core Claim

Whether a quasiparticle peak survives at the antinode of a critical metal is a hypothesis test, and it has a computable detection limit. The noise level and temperature of the imaginary-time data set a minimum detectable peak weight. We will compute that limit, then build detectors that approach it using physically realistic background spectra generated by continued flow equations. Applied to Ising-nematic Monte Carlo data, the result is either a decisive verdict or a proof that the question is undecidable at current statistics, together with the statistics required to decide it. Both outcomes are more useful than a new maximum-entropy picture.

The limit follows from ordinary linear-Gaussian hypothesis testing. Write the imaginary-time Green's function as $G(\tau) = \int K(\tau,\omega)\,[Z\,\delta_\Gamma(\omega) + \rho_{bg}(\omega)]\,d\omega$. Here $K$ is the thermal kernel, $\delta_\Gamma$ is a unit-weight peak of width $\Gamma$ at the Fermi level, and $\rho_{bg}$ is the incoherent background. Let $k_\Gamma(\tau)=\int K\,\delta_\Gamma\,d\omega$ be the peak's imprint in the data. Let $P_\perp$ project out the directions that plausible backgrounds can already produce, and let $C$ be the Monte Carlo covariance matrix. Then the smallest weight detectable at false-alarm rate $\alpha$ with power $1-\beta$ is

$$ Z_{\min}(\Gamma, T) \approx \frac{z_\alpha + z_\beta}{\sqrt{(P_\perp k_\Gamma)^{\top} C^{-1} (P_\perp k_\Gamma)}}, $$

where $z_\alpha$ and $z_\beta$ are standard normal quantiles. The physics enters through $P_\perp$. The kernel is nearly flat for $\omega \ll T$, so a peak narrower than $T$ is almost indistinguishable from any background carrying the same low-frequency weight. That weight can be read off from $G(\beta/2)$, where $\beta = 1/T$. Detectability therefore rests entirely on the curvature of $G(\tau)$ near $\beta/2$, and on how richly the background family can mimic that curvature. A background family that is too generic makes the question undecidable by construction. A family that is physically constrained can make it decidable. This is why the flow-equation spectra matter.

Research Program

The first component is a generator of physically structured spectra. The papers on continued functional renormalization group (FRG) methods take the replacement $i\omega_n \to \omega+i\eta$ inside the flow equations themselves, so real-frequency spectral functions come out without any post-hoc inversion. FRG is a method that follows how a theory's effective interactions change as fluctuations are integrated out from high to low energy. Spectral functions for the quark-meson model phase diagram from the functional renormalization group does this with causal boundary conditions and spatial regulators. Self-consistent spectral functions in the O(N) model from the functional renormalization group solves the flow on a complex momentum-frequency grid and tracks moving poles. These papers supply our method. They also constrain it: both work in relativistic bosonic theories at zero external spatial momentum, and the first names nonzero momentum as the open obstacle.

We will not attempt a controlled solution of the 2D critical metal. Instead we will adapt the continued flow to a Yukawa model of fermions coupled to a damped nematic boson. The fermion self-energy will be evaluated at a small set of fixed Fermi-surface momenta, one nodal and one antinodal. Fixing the external fermion momentum on the Fermi surface sidesteps the general angular branch-cut problem. Sweeping coupling, temperature, and truncation level then produces families of antinodal spectra, some with residual peaks and some without. We use these families as priors and background models, not as the answer. This role is deliberately modest, because FRG truncations are uncontrolled at a critical point.

The second component is the detectors. The principled detector is a generalized likelihood-ratio test: fit the data once with the background family alone and once with background plus a peak, then compare the fits. The learned detector borrows from transient audio detection. It is a multi-scale convolutional network over the imaginary-time profile, with a coarse branch and a fine branch in the style of the macro/micro network of Learning to Recognize Transient Sound Events using Attentional Supervision. Its pooling is max or linear-softmax rather than mean. Towards Weakly Supervised Text-to-Audio Grounding shows that this pooling choice keeps brief events from being averaged away. The learned detector is trained on FRG-generated backgrounds, with and without injected peaks, pushed through the kernel with a realistic noise covariance. The audio work motivates the architecture choice. It does not prove the architecture transfers, so the experiments test transfer directly.

The third component is data. We will run sign-free determinantal quantum Monte Carlo on the Ising-nematic lattice model of Lederer et al., the same model analyzed with Quantum Loop Topography. Unlike that approach, we keep the imaginary-time dependence that it throws away.

Baselines

The naive reference point is the current practice: maximum-entropy continuation (MaxEnt), a regularized inversion that favors smooth spectra, followed by a peak search on the resulting spectrum. This is what Superconductivity and non-Fermi liquid behavior near a nematic quantum critical point did, alongside least-squares fits. That paper reports a qualitative disagreement between the two methods and no numerical detection score, so there is no reported result found on Lacuna for peak-detection accuracy on this model.

The strongest existing reconstruction method is the noise-enhanced neural network of Noise enhanced neural networks for analytic continuation. It is trained with injected Gaussian noise and a Kullback-Leibler divergence loss, and it outperforms MaxEnt on sharp peaks and continua. Its reported advantage is qualitative ("mean absolute error comparable to the exact spectral function"), so again no reported number was found on Lacuna. We will retrain it on the same synthetic spectra and the same noise model as our detectors, and threshold its reconstructed spectrum to make detection calls.

The strongest principled competitor is the likelihood-ratio test with a parametric background, for example a power-law or marginal-Fermi-liquid-like form. It gets the same data and the same fitting budget as every other method.

Two ablations isolate the key ingredients. The first swaps the FRG-generated background family for generic Gaussian mixtures while keeping the detector fixed; this tests whether the flow-equation prior matters. The second swaps max pooling for mean pooling in the learned detector; this tests whether the transient-detection lesson transfers. All methods share the same simulated noise, the same number of time slices, and the same hyperparameter-search budget.

A method beats a baseline if two things hold. It must have higher detection power at a fixed 5% false-alarm rate across a grid of $(Z,\Gamma,T)$. And its empirical $Z_{\min}$ must come within a factor of two of the bound above, while the baseline's does not.

Why not just…?

Why not just run MaxEnt more carefully, with better statistics? Better statistics shrink $C$ and so lower $Z_{\min}$. MaxEnt, however, has no false-alarm control. Its entropy prior actively smooths away narrow features, as the O(N) and neural-continuation papers both document. Our experiments measure MaxEnt's empirical detection power against the bound at matched statistics. If MaxEnt nearly reaches the bound, we will report that plainly.

Why not just use a likelihood-ratio test or Bayes factor and skip the neural network? This is the real threat. When the background family is correctly specified, the likelihood-ratio test is essentially optimal and a learned detector cannot beat it. The learned detector earns its place only under misspecification, for instance when trained at one coupling and tested at another, or when backgrounds come from a different FRG truncation. We test exactly this. If the likelihood-ratio test with the FRG background family matches the learned detector everywhere, we will adopt it as the method. In that case the contribution is the bound plus the physical background family, and the audio architecture is dropped.

Why not just use equal-time features, as in Quantum Loop Topography? Identification of Non-Fermi Liquid Physics in a Quantum Critical Metal via Quantum Loop Topography maps the non-Fermi-liquid region from equal-time Green's functions with a tiny network. Equal-time data integrates over all frequencies, so it cannot separate a narrow peak from a continuum of equal weight. Mapping the phase diagram is a different question from ours.

Why not just solve the continued FRG for the 2D model and read off the answer? Truncated flows at a critical point carry uncontrolled errors, and whether a peak survives is exactly the kind of fine feature a truncation could create or erase. Monte Carlo is exact up to statistical noise. We therefore use FRG only to shape the background hypotheses, and we check sensitivity by varying the truncation.

Proposed Experiments

The first check is controlled and requires no inversion. In the relativistic O(N) model, the continued flow of Self-consistent spectral functions in the O(N) model from the functional renormalization group gives exact real-frequency spectra: a sharp pion pole plus the σ→ππ decay continuum. We forward-map these spectra to imaginary time, add noise, and measure every detector's power against the bound. Because the true spectrum is known exactly, over-claiming detection is caught immediately. A second controlled check uses the 1D transverse-field Ising model setup of Yao et al., where noise-profile mismatch is known to hurt neural continuation. There we deliberately vary the test noise away from the training noise.

The decisive experiment runs on Ising-nematic Monte Carlo data at antinodal and nodal momenta across the superconducting dome's normal state. The nodal points, where quasiparticles are expected to survive, serve as a positive control: every method must detect a peak there. At the antinode, the outcome is one of three things. A detection with calibrated confidence would contradict the complete-decoherence reading. An upper limit on $Z$ below the bound would confirm decoherence down to that weight. Or the analysis shows the question cannot be decided at current statistics, and we state the number of Monte Carlo samples needed to decide it.

As a guard against self-deception, the mean-pooling ablation and the generic-prior ablation are reported alongside the main result. That way readers can see whether the physics prior or the architecture did the work.