Mirror-Paired Chiral Antennas With a Random-Matrix Noise Floor

Chiral gold nanostructures can amplify the faint handedness signal of biomolecules enough to detect them at very low concentrations. They come in left- and right-handed versions, act as "chiral antennas," and are read out by circular dichroism (CD), the difference in how strongly a sample absorbs left- and right-circularly polarized light. The weak point is not raw sensitivity. In real samples such as serum, proteins and other material stick to the particles nonspecifically and shift their optical response. That shift is much larger than the chiral signal and degrades reproducibility. Rare-event physics faced the same problem decades ago and solved it by building observables that the background cannot easily mimic, such as the prompt-light fraction in liquid argon. This note borrows that logic for chiral sensing.

The proposal measures every sample on a mirror-image pair of antennas, one left-handed and one right-handed. By symmetry, any achiral disturbance changes the two CD spectra as exact mirror images, so it cancels in their sum. A chiral analyte breaks the mirror pairing and survives in the sum. The difference channel, which the background dominates, then serves as a built-in background monitor for correcting the leakage that imperfect fabrication lets through. Statistics borrowed from random matrix theory, the mathematics of eigenvalues of large noisy matrices, then decide without labeled data whether any structure remains in the corrected sum channel beyond pure noise. The deliverable is a measured "achiral leakage fraction" for chiral plasmonic sensing in complex media. That number does not appear anywhere in the literature I have access to, and it is what decides whether these sensors work outside clean buffer.

Core Claim

Let $S_L(\lambda)$ and $S_R(\lambda)$ be the CD spectra at wavelength $\lambda$ of a left- and a right-handed antenna exposed to the same sample. Define the sum channel $\Sigma = S_L + S_R$ and the difference channel $\Delta = S_L - S_R$.

For a perfect mirror pair in an achiral environment, $\Sigma = 0$ exactly. A change in that environment, such as nonspecific adsorption changing the local refractive index, therefore moves $\Delta$ strongly and moves $\Sigma$ only through fabrication asymmetry. A chiral analyte forms diastereomeric, non-mirror-related complexes with the two antennas, so it contributes to $\Sigma$ directly. Per antenna pair $k$, this suggests the model

$$ \delta\Sigma_k(\lambda) = K_k(\lambda)\,\delta\Delta_k(\lambda) + c_k\,\chi(\lambda) + \text{noise}. $$

Here $K_k(\lambda)$ is the pair's leakage kernel, estimated once from calibration with purely achiral disturbances. $c_k$ is the analyte amount, and $\chi(\lambda)$ is the analyte's diastereomeric spectral signature.

The claim has two parts. First, subtracting the monitored leakage $K_k\,\delta\Delta_k$ suppresses achiral background in serum by at least two orders of magnitude relative to single-antenna readout, with little loss of chiral signal. Second, a random-matrix certificate on the residuals tells the operator, without labels, when the background model is adequate and how many real signal components remain.

The design also gives a decisive internal check that single-antenna sensing lacks. Swapping the analyte for its mirror-image enantiomer must flip the sign of $c_k\chi$, and a racemic (50/50) mixture must give zero.

Where This Sits in Prior Work

The motivation comes from two papers. The review Gold-Nanoparticle-Based Chiral Plasmonic Nanostructures and Their Biomedical Applications reports detection "down to picomolar or femtomolar levels" and dissymmetry factors up to about 0.2 for discrete particles. The same review names nonspecific adsorption in biological matrices as the main threat to reproducibility. It catalogs fabrication routes for both handedness versions, including chiral-ligand growth and DNA origami, but proposes no measurement design that is immune to background.

Chiral Chaos Enhanced Sensing is the closest conceptual relative. It combines two direction-dependent responses of coupled microcavities and raises the success rate for random-phase signals to "approaching 100%" from below 50% for conventional sensors. Its authors explicitly leave "the impact of background noise" to future work. We share the idea of combining two symmetry-related channels. We differ in using mirror-image antennas with a linear CD readout, and in targeting continuous achiral background rather than phase randomness.

The method draws on three sources. Rare-event discrimination supplies the habit of calibrating leakage with dedicated background sources and quoting it as a single number. First Results from the DEAP-3600 Dark Matter Search with Argon at SNOLAB quotes electronic-recoil leakage below $1.2\times10^{-7}$. α-event characterization and rejection in point-contact HPGe detectors shows that a physically targeted second observable, its delayed-charge-recovery slope, catches what a primary cut misses. That is the role our difference channel plays.

Spectral statistics come from two places. Constructing gene co-expression networks and predicting functions of unknown genes by random matrix theory used a crossover in eigenvalue spacing statistics to separate real structure from noise without a subjective threshold. It also warns that the crossover is a zone, not a point. Anomalous Floquet insulators supplies the unfolding-free spacing-ratio statistic, which avoids fitting the average density of eigenvalues.

Amplifying and Reversing the Chiral Bias in Asymmetric Photo‐Polymerization Reaction is a constraint as much as a resource. It shows that cysteine-coated silver particles impose a strong near-field handedness, and that this bias can dominate other inputs. Antenna handedness should therefore dominate the measured CD, which the mirror-pair argument requires. Matched pairs must also be made under identical conditions, or the "mirror" assumption fails before any analyte is added.

Research Program

The work proceeds in three stages.

Stage 1: simulation with known ground truth. We simulate disordered helical gold-nanoparticle assemblies with a coupled-dipole model. Each mirror pair carries random positional and twist disorder, so it is imperfect in a realistic way. Achiral refractive-index shifts and a chiral molecular layer are added at controlled doses. This produces data where $K_k$ and $c_k$ are known.

The simulation also tests a mechanistic prediction adapted from Optical Spectra of p-Doped PEDOT Nanoaggregates Provide Insight into the Material Disorder. That work used the inverse participation ratio, a number near 1 for a mode confined to one site and near $1/N$ for a mode spread over $N$ sites, to show that disorder localizes some states while optically bright states stay spread out. We predict that leakage $|K_k|$ concentrates at wavelengths where the plasmon modes are localized, because a single defective particle breaks the mirror symmetry of a localized mode far more than that of an extended one. If this holds, the readout should be restricted to delocalized-mode wavelengths. The prediction extends the PEDOT analysis from achiral sliding disorder to twist disorder.

Stage 2: bench experiments. We fabricate left- and right-handed gold structures in matched batches, using chiral-ligand-grown particles and DNA-origami helices as two independent routes. Calibration uses achiral "background sources": sucrose or glycerol refractive-index steps, and adsorption of an achiral polymer. These fit $K_k$ per pair. We then add diluted human serum, and finally enantiopure chiral analytes spiked into serum, using both enantiomers and the racemate of a gold-binding amino acid or short peptide.

Stage 3: random-matrix certificate. After leakage correction, we stack residual sum-channel spectra from $N$ pairs and replicates into a matrix of $N$ samples by $M$ wavelengths and form its $M\times M$ correlation matrix. For pure noise, theory predicts that the eigenvalues fill the Marchenko–Pastur band, a known range set by $N/M$. Spacings inside that band should follow orthogonal-ensemble statistics, with mean adjacent-spacing ratio $\langle r\rangle \approx 0.53$, whereas independent uncorrelated blocks give $\langle r\rangle \approx 0.39$.

Eigenvalues above the band are counted as real components. A bulk $\langle r\rangle$ that deviates from 0.53 flags a broken noise model, for example drifting baselines or heteroscedastic wavelength bands, even when no outlier is visible. Following the gene-network paper's warning about transition zones, we report the certificate as a bootstrap distribution, not a hard cut.

Baselines

Naive reference. The reference point is single-antenna readout: the peak CD shift of one handedness after blank subtraction, thresholded for detection at a fixed false-alarm rate. It runs on exactly the same spectra as the proposed method, using the left-handed half of each pair.

Strongest existing practice. The strongest existing practice is the chiral plasmonic sensing described in Gold-Nanoparticle-Based Chiral Plasmonic Nanostructures and Their Biomedical Applications. Its reported figure is a limit of detection "down to picomolar or femtomolar levels," obtained largely in clean conditions. I found no result on Lacuna for achiral-background leakage or for serum-matrix limits of detection in chiral plasmonic sensing. We therefore re-measure this baseline ourselves in buffer and in serum, on the same antennas, analyte, and acquisition time. Beating it means a serum limit of detection within a factor of three of the buffer value, where the single-antenna baseline degrades by at least an order of magnitude.

Learned baseline. A strong machine-learning competitor is the two-stage design of Deep learning based pulse shape discrimination for germanium detectors: an autoencoder compresses each spectrum, and a small classifier is trained on calibration data. That paper reports 90% background rejection at 90% signal acceptance on germanium waveforms. We give it the same calibration sets used to fit $K_k$, a matched hyperparameter search budget, and single-antenna or concatenated pair spectra as input. We then test it on serum lots it did not see in training. Beating it means higher signal acceptance at matched background rejection on held-out serum lots, where we expect the learned model to suffer the calibration-to-reality mismatch that paper itself flags.

Ablations. Three ablations isolate the ingredients. The first uses the raw sum channel without leakage regression, which tests whether the difference-channel monitor matters. The second applies regression with a fixed number of components chosen by eye, without the random-matrix certificate. The third applies the random-matrix certificate to single-antenna data, which tests whether mirror pairing, not denoising, carries the gain.

Why not just…?

Why not just subtract a serum blank from a single antenna? Nonspecific adsorption varies from sample to sample and drifts during incubation, so a blank taken from one aliquot does not describe another. The review cited above names this variability as the reproducibility problem. The naive baseline is exactly this approach. If it matches the mirror-pair method on held-out serum lots, the pairing is unnecessary, and we would report that result.

Why not just train a classifier on calibration spectra? This alternative is a real threat when train and test conditions match, as the germanium result shows. Both Implementation of the Random Forest method for the Imaging Atmospheric Cherenkov Telescope MAGIC and the germanium study identify mismatch between calibration and reality as the failure mode. Serum varies between donors in ways no calibration set covers. The held-out-lot test settles the question. If the classifier wins there too, the right conclusion is to feed it mirror-paired inputs, and the concatenated-pair variant tests exactly that.

Why not just raise the dissymmetry factor of the antennas? A larger $g$ amplifies the chiral signal, but it also amplifies how strongly the CD spectrum responds to refractive-index changes. The signal-to-background ratio against achiral adsorption need not improve. Stage 1 measures leakage as a function of $g$ across simulated assemblies to check this directly.

Why not just use the Marchenko–Pastur cutoff or the gene-network spacing threshold alone? Both only say how many structured components exist. They cannot say whether a component is chiral or a leftover of background. Mirror pairing and the enantiomer sign flip supply that identification, and the third ablation measures how much each part contributes.

Decisive Checks

The program should be judged on four outcomes. First, whether the measured achiral leakage fraction, defined as corrected sum-channel response per unit difference-channel response under achiral calibration, falls below $10^{-2}$ in serum. Second, whether swapping the enantiomer flips the sign of the recovered $c_k$ in at least 95% of pairs, with the racemate consistent with zero. Third, whether the simulation's prediction that leakage concentrates at localized-mode wavelengths holds on the bench. Fourth, whether the random-matrix certificate flags the injected drift artifacts that fixed-rank denoising misses.

Failure on the sign-flip test would mean the sum channel is not measuring chirality, and the approach should be abandoned in favor of better fabrication matching.