---
title: Light-Cone Feature Selection
url: https://www.emergentmind.com/topics/light-cone-feature-selection
type: topic
---

# Light-Cone Feature Selection

Searching arXiv for recent papers on light-cone feature selection and related formulations.
{"query":"\"Light-cone feature selection\" arXiv", "max_results": 10}
arxiv_search(query="\"Light-cone feature selection\"", max_results=10)
Light-Cone Feature Selection (LCFS) denotes a family of feature-selection procedures in which the organizing structure is a light-cone rather than an unordered list of coordinates. In quantum machine learning, the relevant object is the past causal cone of a measured qubit or parameterized gate in a parameterized quantum circuit, and LCFS selects or ranks those cones through local quantum kernels and kernel-target alignment [2403.18733]. In cosmology, the relevant object is the observational past light cone: feature construction must respect the evolving geometry of galaxy-survey data along the line of sight, and informative angular, radial, tomographic, or summary-statistic features can be ranked directly on the light cone [2603.26533]. A remote-sensing application to methane hyperspectral images adopts the quantum-circuit meaning and explicitly distinguishes it from spatiotemporal past/future light-cones [2504.00793]. The term therefore names a methodological pattern rather than a single domain-specific algorithm.

## 1. Terminology and conceptual scope

The literature uses “light-cone” in at least two distinct senses. In quantum machine learning, a light-cone is the minimal subset of qubits, gates, and layers that can influence a local probe in a circuit. In cosmology, the light cone is the past observational manifold parametrized by redshift and sky direction, where radial evolution breaks translational invariance and changes what constitutes an informative feature [2403.18733].

| Domain | Meaning of “light-cone” | Feature object |
|---|---|---|
| Quantum machine learning | Past causal cone of a measurement or parameterized gate | Local kernel term or cone-defined subspace |
| Methane hyperspectral imaging | Same quantum-circuit causal-cone notion | Spectral band or derived index ranked through cone participation |
| Cosmology | Past light-cone geometry of observations | Angular multipoles, tomographic bins, summary statistics, or learned slice-wise representations |

In the quantum formulation, LCFS is an embedded method: it learns importance weights jointly with the kernel rather than performing a separate wrapper-style subset search. In the cosmological formulation, the same phrase refers to selecting representations that respect line-of-sight evolution and preserve non-Gaussian, multi-scale information. A common misconception is to treat these usages as interchangeable. The methane hyperspectral study states the distinction explicitly: its “light-cone” is “analogous to causal cones in quantum circuits and tensor networks, not to be confused with the past/future light-cones in spatiotemporal dynamical systems” [2504.00793].

## 2. Quantum-machine-learning formulation

The general LCFS framework for quantum machine learning treats each light-cone as a feature and assigns it a nonnegative weight inside a decomposable kernel,
$$
k(x,x')=\sum_l \lambda_l k^{(l)}(x,x'), \qquad \lambda_l \ge 0,\quad \sum_l \lambda_l=1.
$$
Two concrete realizations are developed. For the projected quantum kernel (PQK),
$$
k_{PQ}(x_i,x_j)=\sum_l \lambda_l \mathrm{Tr}\!\left[\rho_i^{(l)}(\theta)\rho_j^{(l)}(\theta)\right],
$$
where $\rho_i^{(l)}(\theta)=\mathrm{Tr}_{\neg l}[\rho_i(\theta)]$ is the single-qubit reduced density matrix on qubit $l$. For the simplified quantum Fisher kernel (QFK),
$$
k_{QF}(x_i,x_j)=\sum_l \lambda_l \mathrm{Tr}\!\left[\rho_0\{\tilde B_{i,\theta_l},\tilde B_{j,\theta_l}\}\right].
$$
In PQK the probe is a one-qubit measurement; in QFK it is the parameterized gate carrying $\theta_l$. The cone for each term is obtained by backward traversal through the circuit graph, including additional qubits whenever a multi-qubit gate is encountered [2403.18733].

Training is organized around kernel-target alignment (KTA),
$$
\mathrm{KTA}(K,Y)=\frac{\langle K,Y\rangle_F}{\|K\|_F\|Y\|_F}
=\frac{\sum_{p,q} y_p y_q k(x_p,x_q)}
{\sqrt{\sum_{p,q} k(x_p,x_q)^2}\sqrt{\sum_{p,q}(y_p y_q)^2}},
$$
with alternating optimization over circuit parameters $\theta$ and cone weights $\lambda$. With $\theta$ fixed, the $\lambda$ update is a convex quadratic program:
$$
v^*=\arg\min_{v\ge 0}\left[v^T M v-2a^T v\right],\qquad
\lambda^{opt}=v^*/\|v^*\|_1,
$$
where
$$
M_{ij}=\sum_{p,q} k^{(i)}(x_p,x_q)k^{(j)}(x_p,x_q),\qquad
a_i=\sum_{p,q} k^{(i)}(x_p,x_q)y_p y_q.
$$
The implementation described in the paper uses Adam with learning rate $0.01$ for about $500$ iterations per outer loop, and solves the quadratic program with CVXOPT [2403.18733].

For classical inputs, cone weights can be mapped back to coordinate-level importance scores. If $w_{\lambda_l}(\mu)$ counts how often classical feature $x_\mu$ appears in cone $l$, then
$$
P(\mu)=\frac{1}{\mathcal N}\sum_l w_{\lambda_l}(\mu)\lambda_l,
\qquad
\mathcal N=\sum_\mu \sum_l w_{\lambda_l}(\mu)\lambda_l.
$$
This compensates for the fact that deeper cones cover more coordinates and therefore would otherwise inflate apparent relevance.

The empirical demonstrations in the foundational LCFS study show four uses of the framework. On the parity$(1,2)$ task with $n=d=8$ and depth $L=3$, QFK reaches $\mathrm{KTA}\approx 0.987$ and cleanly selects $x_1,x_2$, whereas PQK reaches $\mathrm{KTA}\approx 0.605$ and also selects extraneous $x_3,x_8$ because of cone growth. On parity$(1,5)$, LCFS-guided feature-to-qubit reordering changes poor or negative test KTA without reordering—$-0.033$ for QFK and $-0.015$ for PQK—into $0.987$ and $0.560$, respectively. On the Breast Cancer dataset, QFK test KTA improves from $0.719$ to $0.726$ and PQK from $0.366$ to $0.377$ after reordering, while yielding sparser importance profiles. For compression, keeping only the largest-$\lambda$ cone on parity$(1,2)$ reduces QFK KTA from $0.987$ to $0.983$ and PQK from $0.605$ to $0.600$. The same formalism also performs subspace selection for quantum data, where no classical coordinate system is available [2403.18733].

These results establish the central interpretation of LCFS in QML: the “feature” is a causal subcircuit or subspace, not merely an input coordinate. This suggests that LCFS is simultaneously a feature selector, an architecture-search heuristic, and a pruning rule.

## 3. Methane hyperspectral imaging and quantum-kernel LCFS

A concrete application of LCFS appears in methane detection and localization from hyperspectral images collected by AVIRIS-NG over geographically diverse fossil-fuel sites. The data come from the STARCOP dataset. Eight bands are used for analysis—$460$, $550$, $640$, $2004$, $2109$, $2310$, $2350$, and $2360$ nm—and the last three lie in the shortwave infrared region containing strong $\mathrm{CH}_4$ absorption features near $2.31$–$2.36\,\mu\mathrm{m}$ [2504.00793].

The preprocessing pipeline is explicitly superpixel-based. Extended SLIC is computed over all available bands, band values are averaged within each superpixel, a mag1c methane enhancement map is produced and averaged per superpixel, and the binary plume mask assigns labels by majority within each superpixel. The dataset comprises $3{,}208$ hyperspectral images of size $512\times 512\times 8$ for the selected bands. Training uses only $30$ labeled superpixels—$15$ methane and $15$ background—because the kernels are computed in simulation; the test set $T'$ contains all methane superpixels plus randomly sampled background superpixels, with $|T'|=1{,}476$ [2504.00793].

The quantum model uses $n=8$ qubits and $L=4$ alternating layers of one-qubit data-reupload gates $V(x_j)$ and two-qubit entanglers $U(\Theta_{k,l})$. Measuring each qubit yields a local kernel
$$
\kappa_i(x,y)=T[\rho_x^i\rho_y^i],
$$
and the global kernel is
$$
\kappa(x,y)=\sum_{i=1}^n \lambda_i \kappa_i(x,y),
$$
where $\lambda_i$ is set by centered alignment and normalized so that $\sum_i \lambda_i=1$. Feature importance is defined through light-cone participation counts $w_i(x_j)$ of feature $j$ in the past cone of measured qubit $i$. The paper displays
$$
P_j=\frac{1}{\mathcal N \sum_{i=1}^n w_i(x_j)\lambda_i},
$$
but also states that features with high values of $P$ are those “re-uploaded many times in the most influential local kernels.” The text therefore identifies a mismatch between the displayed inverse formula and the intended monotonic behavior [2504.00793].

Three SVM baselines are compared: linear, RBF, and quantum-kernel SVM. Without mag1c, the reported metrics on $T'$ are: for SVML, accuracy $=0.599$, sensitivity $=0.617$, specificity $=0.578$, F-score $=0.622$, MCC $=0.195$; for SVMRBF, accuracy $=0.805$, sensitivity $=0.301$, specificity $=0.571$, F-score $=0.668$, MCC $=0.123$; for SVMQ, accuracy $=0.517$, sensitivity $=0.566$, specificity $=0.540$, F-score $=0.546$, MCC $=0.083$ [2504.00793].

When mag1c is inserted under a fixed $n=8$ qubit budget by dropping one band, two quantum configurations stand out. Dropping band $6$ ($2310$ nm) yields accuracy $=0.623$, sensitivity $=0.615$, specificity $=0.619$, F-score $=0.637$, MCC $=0.237$, and $p=0.210$, which the paper interprets as no significant difference versus ground truth. Dropping band $8$ ($2360$ nm) yields accuracy $=0.692$, sensitivity $=0.486$, specificity $=0.596$, F-score $=0.647$, MCC $=0.182$, and $p=0.112$ [2504.00793].

The ablation results show that feature ranking is not fully stable. The top three mean importance scores are bands $j=6$ ($0.130\pm 0.031$), $j=1$ ($0.129\pm 0.030$), and $j=3$ ($0.128\pm 0.010$), whereas the top three mean sums of metrics are $j=1$ ($2.40\pm 0.32$), $j=5$ ($2.44\pm 0.28$), and $j=4$ ($2.43\pm 0.28$). The paper states that both measures have relatively large standard deviations and agree only on the importance of feature $1$ [2504.00793]. A plausible implication is that LCFS in this setting is sensitive to circuit topology and small-sample variability, even when it improves downstream metrics by enabling mag1c inclusion.

## 4. Cosmological light-cone geometry and feature construction

In cosmology, the light cone is the natural data domain of galaxy surveys. Observations are indexed by $x\equiv (z,\hat n)$, and an observable is modeled as
$$
D(x)=\bar D(z)[1+\delta D(x)].
$$
Because the radial coordinate mixes space and time, translational invariance is broken along the line of sight, while angular directions retain statistical isotropy. This is why spherical harmonics and spherical Fourier–Bessel (SFB) decompositions are natural, and why feature-selection rules derived from cubic-box intuition can mis-rank informative modes in wide or deep surveys [1905.08262].

The SFB decomposition writes
$$
\delta\mathcal D(x)=\sum_{\ell m}\int_0^\infty dk\,\sqrt{\frac{2}{\pi}}\,k\, j_\ell(k\chi)Y_{\ell m}(\hat n)\, s_{\ell m}(k),
$$
with covariance
$$
\langle s_{\ell m}(k)s^*_{\ell' m'}(k')\rangle=\delta_{\ell\ell'}\delta_{mm'}S_\ell(k,k').
$$
The radial spectrum $S_\ell(k,k')$ is generally non-diagonal because light-cone evolution mixes radial modes. For angular or tomographic data, the corresponding Fisher matrix is
$$
F_{\alpha\beta}=f_{\rm sky}\sum_\ell \frac{2\ell+1}{2}
\mathrm{Tr}\!\left[C_\ell^{-1}\frac{\partial C_\ell}{\partial \theta_\alpha}
C_\ell^{-1}\frac{\partial C_\ell}{\partial \theta_\beta}\right].
$$
The same framework yields per-$\ell$ and per-bin information measures for selecting observables, multipole ranges, cross-spectra, or tomographic bins directly on the past light cone [1905.08262].

The more recent machine-learning study translates this geometric constraint into representation design. It uses AbacusSummit halo lightcone mocks from a corner of the main box plus two periodic copies, selecting $\mathrm{RA}/\mathrm{DEC}\in (0^\circ,40^\circ)$ and $0.3<z<0.8$, with halo selection $M_{\min}=2\times 10^{11}\,h^{-1}M_\odot$ and number densities $\sim 8$–$11\times 10^{-3}\,(h/\mathrm{Mpc})^3$. Redshift-space distortions are included through
$$
s_\parallel=r_\parallel+\frac{(v\cdot \hat z)}{aH(a)}\hat z,
$$
and an Alcock–Paczynski mapping to a fiducial cosmology is applied:
$$
s_\perp=s_\perp^0 \left[\frac{d_A^f(z)}{d_A(z)}\right],\qquad
s_\parallel=s_\parallel^0 \left[\frac{H(z)}{H^f(z)}\right].
$$
The key methodological claim is that a 3D CNN on a single gridded lightcone implicitly imposes translational invariance along the line of sight and can therefore mix features from different lookback times or learn spurious invariances [2603.26533].

This motivates a geometry-aware “slice–project–analyze” strategy: divide the light cone into thin redshift slices, project each slice to a HEALPix sphere, and then analyze the stack with a 2D CNN. In the reported implementation, nine slices cover $0.3<z<0.8$, HEALPix uses $N_{\rm side}=512$, each slice is mapped to a $512\times 512$ Cartesian patch over $40^\circ\times 40^\circ$, and the resulting $9\times 512\times 512$ tensor is processed jointly [2603.26533].

## 5. Summary statistics, learned features, and comparative performance in cosmology

The cosmological comparison includes four feature families: slice-wise image features processed by a 2D CNN, spherical harmonic coefficients $a_{\ell m}$, wavelet scattering transform (WST) coefficients, and the angular two-point correlation function $w(\theta)$, all evaluated on AbacusSummit halo lightcones [2603.26533].

For the harmonic representation, each field is expanded as
$$
X(\hat n)=\sum_\ell \sum_m a_{\ell m}Y_{\ell m}(\hat n),\qquad
a_{\ell m}=\int d\Omega\, X(\hat n)Y^*_{\ell m}(\hat n),
$$
with angular power spectrum
$$
C_\ell=\frac{1}{2\ell+1}\sum_{m=-\ell}^{\ell}|a_{\ell m}|^2,
$$
and
$$
w(\theta)=\sum_{\ell=0}^{\ell_{\max}} \frac{2\ell+1}{4\pi} C_\ell P_\ell(\cos\theta).
$$
In practice, with $N_{\rm side}=512$ and $\ell_{\max}=256$, the $a_{\ell m}$ are compressed by PCA to $185$ components per slice explaining $99\%$ of the variance. For WST, the full lightcone density is gridded into $512\times 512\times 64$, tiled into $64$ sub-regions of size $64\times 64\times 64$, and each sub-volume yields $784$ coefficients for $J=6$ scales and $L=6$ orientations; PCA reduces each to $60$ components explaining $99\%$ of the variance. For the 2PCF, the estimator is Landy–Szalay,
$$
w(\theta)=\frac{DD-2DR+RR}{RR},
$$
with $26$ linear bins in $0.17^\circ<\theta<8.5^\circ$ per slice and jackknife covariance over $16$ subregions per slice [2603.26533].

The CNN+2D architecture has four convolutional blocks with filters $(32,64,128,256)$, $3\times 3$ convolutions, batch normalization, ReLU, and $2\times 2$ max-pooling, followed by adaptive average pooling to $256\times 16^2$, flattening to $65{,}536$ features, and a dense regressor with Dense$(512)$, ReLU, Dropout$(0.3)$, and Output$(6)$ for $\{\Omega_b,\Omega_m,h,A_s,n_s,\sigma_8\}$. FC models for $a_{\ell m}$, WST, and 2PCF use the same Dense$(512)$ + ReLU + Dropout$(0.3)$ + Output$(6)$ head. Training uses Adam with learning rate $10^{-4}$, batch size $4$, and mean-squared error loss [2603.26533].

The reported test losses are:
- FC+WST: $0.205$
- FC+$a_{\ell m}$: $0.294$
- CNN+2D: $0.353$
- FC+2PCF: $0.430$

Sample parameter-level metrics illustrate the ranking. For $\Omega_m$, CNN+2D gives $R^2=0.868$, RMSE $=0.014$; FC+WST gives $0.922$, $0.011$; FC+$a_{\ell m}$ gives $0.905$, $0.012$; FC+2PCF gives $0.833$, $0.016$. For $\sigma_8$, CNN+2D gives $R^2=0.626$, RMSE $=0.024$, while FC+WST gives $0.826$, $0.021$ and FC+2PCF gives $0.453$. For $n_s$, CNN+2D gives $R^2=0.405$, RMSE $=0.023$, FC+WST gives $0.703$, $0.016$, and FC+2PCF gives $0.360$, $0.024$ [2603.26533].

The fiducial multiple-realization experiment modifies the ranking. For cosmology c000 with $21$ realizations, CNN+2D yields $\Omega_b=0.0491\pm 0.0014$, $\Omega_m=0.3244\pm 0.0085$, $h=0.6783\pm 0.0083$, $10^9A_s=2.0144\pm 0.0539$, $n_s=0.9711\pm 0.0094$, and $\sigma_8=0.8071\pm 0.0132$. FC+WST remains competitive and is tighter for some parameters, for example $n_s=0.9722\pm 0.0077$, but the paper states that CNN+2D attains the smallest statistical uncertainties overall across the fiducial ensemble [2603.26533].

The qualitative interpretation is explicit. WST outperforms 2PCF-based methods because it captures higher-order, multi-scale statistics through cascaded wavelet moduli and low-pass averages, including filamentary anisotropy and halo/void morphology across scales. Appendix-level comparisons show larger inter-cosmology differences at high multipoles in $C_\ell$ and at small angles in $w(\theta)$, while WST coefficients vary with scale $j$ and orientation $l$, reflecting multi-scale anisotropy. These regimes contribute strongly to constraints on $\Omega_m$ and $\sigma_8$, and $n_s$ benefits from broad-scale sensitivity in both WST and high-$\ell$ $C_\ell$ [2603.26533].

## 6. Limitations, ambiguities, and future directions

Across the literature, LCFS is limited by the geometry it assumes and by the objects it treats as features. In PQK, cone growth with depth can create false positives by sweeping many classical coordinates into a measurement cone; this is why the QFK formulation is described as sharper for precise localization on the parity tasks [2403.18733]. In the methane study, feature-importance scores have relatively large standard deviations, centered-alignment optimization of $\Theta$ is described as problematic, and the training set is restricted to $30$ samples because kernel computation is simulator-limited [2504.00793]. These facts indicate that topology dependence, training-sample scarcity, and simulator cost remain core constraints for quantum LCFS.

In cosmology, limitations arise from both modeling and representation. The CNN+2D pipeline assumes approximate stationarity within each redshift slice and does not explicitly model temporal dynamics along the line of sight. The $a_{\ell m}$ and 2PCF summaries are fundamentally two-point descriptions; under complex survey masks, pseudo-$C_\ell$ with mode-coupling matrices is required to debias $C_\ell$ estimates. WST sensitivity depends on configuration choices such as wavelet family, $J$, $L$, and $\sigma$, and the reported implementation uses a 3D Euclidean gridding rather than a spherical scattering construction. More generally, supervised learning depends on the realism of mocks and halo-occupation prescriptions, which motivates domain adaptation, nuisance marginalization, and cross-survey validation [2603.26533].

Future directions are correspondingly domain-specific. For quantum LCFS, the literature points to more robust alignment optimization, regularization for $\lambda_i$ and $\Theta$, systematic cone design and topology search, larger training sets, and extension to other Earth-observation tasks or other quantum-data settings [2504.00793]. For cosmology, the reported directions include spatiotemporal CNNs or transformers across slices, spherical CNNs on HEALPix, learned scattering networks, hybrid CNN+WST models, emulator-based WST likelihoods, and hybrid machine-learning-plus-likelihood pipelines [2603.26533]. The broader Fisher-based light-cone program also emphasizes multi-tracer combinations, cross-correlations, and explicit treatment of angular versus radial information as the route to overcoming cosmic-variance ceilings on the past light cone [1905.08262].

Taken together, the literature supports a narrow but technically consistent characterization: LCFS is effective when relevance is localized by a causal or observational geometry, and when the feature map preserves that geometry rather than erasing it through inappropriate invariances. In QML, the relevant objects are cone-defined subspaces and local kernels. In cosmology, they are light-cone-consistent angular, radial, tomographic, or learned summaries. The unifying principle is that feature relevance is determined by what can physically or causally influence the target observable on the chosen light-cone.

Source: https://www.emergentmind.com/topics/light-cone-feature-selection