Foreground Wedge in 21-cm Cosmology

• Foreground Wedge is the region in 21-cm Fourier space where smooth astrophysical foregrounds leak into higher modes due to instrumental chromaticity.
• It originates from the chromatic nature of interferometers, which mix sky signals and produce a characteristic wedge shape in (k⊥, k∥) or delay-space analyses.
• Mitigation strategies such as foreground avoidance, improved calibration, advanced array design, and machine-learning recovery are crucial for isolating the cosmological signal.

The foreground wedge is the region of 21-cm Fourier space in which spectrally smooth astrophysical foregrounds become contaminated by instrumental chromaticity and leak from low line-of-sight modes into higher ones, producing a characteristic wedge-shaped contamination pattern in $(k_\perp,k_\parallel)$ or delay-space representations. In the 21-cm literature, it is the central geometric expression of mode mixing in radio interferometers and the standard rationale for “foreground avoidance,” in which wedge modes are discarded and only the cleaner complement—commonly called the EoR window—is used for cosmological inference (Pober et al., 30 Jul 2025, Thyagarajan et al., 2015).

1. Physical origin and conceptual definition

The wedge arises from the combination of two facts that recur across 21-cm cosmology. First, the dominant foregrounds—Galactic synchrotron emission, diffuse Galactic emission, and extragalactic radio sources—are smooth in frequency. Second, a radio interferometer is chromatic: baselines are measured in wavelengths, the synthesized beam changes with frequency, and the same sky structure is mixed differently across frequency channels. When visibilities are Fourier transformed along the frequency axis, this chromatic mode mixing pushes power that would otherwise remain near $k_\parallel \approx 0$ into larger $k_\parallel$, especially on longer baselines, thereby generating a triangular contamination region in cylindrical Fourier space (Pober et al., 30 Jul 2025).

A visibility-based formulation makes the same point in geometric-delay language. For a baseline $\boldsymbol{b}$, the visibility contains the phase factor associated with geometric delay across the sky, and the delay transform maps smooth-spectrum sources to delay modes bounded by the horizon delay limit, $|\tau| \le |\boldsymbol{b}|/c$. This baseline-dependent delay ceiling becomes a baseline-dependent contamination boundary in $(k_\perp,k_\parallel)$, which is why the wedge broadens toward larger $k_\perp$ (Thyagarajan et al., 2015).

The wedge is therefore not merely an artifact of imperfect subtraction. It is an instrumental mapping of smooth foreground structure into line-of-sight Fourier modes. This distinction is central to the empirical interpretation of early measurements with PAPER, which showed that wedge-like contamination is present directly in measured 2D power spectra even without aggressive foreground subtraction (Pober et al., 2013).

2. Geometric and mathematical descriptions

In the standard cosmological mapping reviewed for dark-ages and EoR analyses, interferometric coordinates $(u,v,\eta)$ map to cylindrical Fourier variables through

$$\mathbf{k}_\perp = \frac{2\pi \mathbf{u}}{D_C(z)}, \qquad k_\parallel = \frac{2\pi \nu_{21} H_0 E(z)}{c(1+z)^2}\,\eta,$$

with $E(z)=\sqrt{\Omega_m(1+z)^3+\Omega_\Lambda}$ in the form used in that analysis (Pober et al., 30 Jul 2025). In this language, the horizon-limit wedge boundary is

$k_\parallel \approx 0$0

Equivalently, the wedge slope is the redshift-dependent quantity

$k_\parallel \approx 0$1

which steepens toward higher redshift, shrinking the uncontaminated window (Pober et al., 30 Jul 2025).

A source-angle formulation used in image-domain wedge-cut studies writes the contamination track from a source at angle $k_\parallel \approx 0$2 from the field center as

$k_\parallel \approx 0$3

with the horizon limit at $k_\parallel \approx 0$4 defining the largest geometric wedge extent. That same work extends the boundary in 3D Fourier space to account for spectral leakage from finite Fourier windows: $k_\parallel \approx 0$5 where $k_\parallel \approx 0$6 is an extra buffer in $k_\parallel \approx 0$7-pixels introduced to absorb window-function leakage beyond the geometric boundary (Kittiwisit et al., 2022).

For power-spectrum forecasting and anisotropic analyses, the wedge is often rewritten as a cut in the cosine to the line of sight,

$k_\parallel \approx 0$8

with all modes below a redshift-dependent $k_\parallel \approx 0$9 treated as contaminated. For 21-cm BAO forecasts, this yields a conical exclusion in 3D Fourier space, operationally equivalent to removing the low-$k_\parallel$0 region and retaining the cleaner window at higher $k_\parallel$1 (Seo et al., 2015).

The conventional horizon line is not universal. For wide-field, phase-tracking instruments, the flat-sky wedge can underestimate the true contamination region. A full-sky formulation derives LST-dependent snapshot and full-synthesis wedge boundaries; for a snapshot at phase-center altitude $k_\parallel$2,

$k_\parallel$3

so the effective horizon line depends on telescope latitude, phase-center declination, and LST rather than remaining a single time-independent line (Munshi et al., 2024).

3. Wide-field structure, horizon emission, and observed wedge morphology

Wide-field interferometry makes the wedge unavoidable in a stronger sense than the small-field picture suggests. The solid-angle element

$k_\parallel$4

grows rapidly toward the horizon, so equal-sized delay bins subtend increasingly large sky areas near the horizon. At the same time, baselines are foreshortened toward the horizon, making even long baselines sensitive to larger angular scales in those directions. The result is that diffuse emission near the horizon can dominate the wedge, including on antenna spacings that would naively be associated only with small angular scales (Thyagarajan et al., 2015).

This wide-field behavior produces the “pitchfork” signature described in delay-space analyses: a strong central feature from compact emission in the primary field of view, plus two enhanced edge features near the positive and negative horizon-delay limits from diffuse near-horizon emission. In that description, the central tine is compact in-beam emission near $k_\parallel$5, while the outer tines are produced by diffuse wide-field emission near $k_\parallel$6 (Thyagarajan et al., 2015).

Instrument design strongly modulates the amplitude of this morphology. Simulations comparing antenna types found EoR-window contamination levels of roughly $k_\parallel$7 for a dipole, $k_\parallel$8 for a phased array, and $k_\parallel$9 for a dish, reflecting the greater horizon sensitivity of broader beams (Thyagarajan et al., 2015).

Observationally, the wedge is real but not perfectly sharp. PAPER measurements at 150 MHz showed wedge-like foreground confinement below a baseline-dependent boundary, with power extending past the nominal wedge edge due to spectral structure in the foregrounds and instrumental response. In that analysis, the practical CLEAN boundary was placed 50 ns beyond the physical maximum horizon delay, and the strongest supra-horizon extension was observed on the shortest baselines. The filtered images suggested that problematic contamination beyond the nominal wedge boundary is not easily localized to a few bright sources, but is instead associated with aggregate faint emission and diffuse structure (Pober et al., 2013).

4. Foreground avoidance and the EoR window

Foreground avoidance is the most direct operational use of the wedge: contaminated modes are excised, and only the EoR window is analyzed. Its attraction is methodological. Avoidance relies on wedge geometry rather than a detailed foreground model, and for that reason it became a standard conservative strategy in EoR power-spectrum work and later in tomographic and intensity-mapping analyses (Kittiwisit et al., 2022).

The cost is loss of cosmological information. In wedge-filtered tomographic images, ionized bubbles are heavily distorted, the image becomes anisotropic, and structures can be lost in both transverse and line-of-sight directions. For imaging, this is not a mild degradation but a direct consequence of removing large-scale Fourier modes that are signal-bearing as well as foreground-dominated (Gagnon-Hartman et al., 2021).

A concrete image-domain implementation was developed for HERA-like mock data in the study of one-point statistics. The workflow selects a frequency subband, multiplies by a Blackman-Nuttall window in frequency, Fourier transforms to $\boldsymbol{b}$0-space, applies a binary wedge filter, inverse transforms back to image space, divides by the same Blackman-Nuttall window, and then measures the PDF and moments. That analysis varied subband bandwidths from 1 to 8 MHz, wedge opening angles $\boldsymbol{b}$1 from $\boldsymbol{b}$2 to $\boldsymbol{b}$3, and leakage buffers $\boldsymbol{b}$4 from 0 to 4 $\boldsymbol{b}$5-pixels. It found that the center frequency channel of a band is the least susceptible to bias from wedge-cutting, which motivated a rolling filter method that stitches together the least biased central channels from overlapping subbands to reconstruct a wedge-cut cube across the full bandwidth (Kittiwisit et al., 2022).

The wedge boundary itself can be defined more or less conservatively. Some analyses adopt the horizon limit, $\boldsymbol{b}$6, as a pessimistic choice because sidelobes can pick up bright emission far from the main beam, while others treat the field-of-view wedge and horizon wedge separately and identify an intermediate region between them. Full-sky tracking calculations further show that the physically relevant horizon line can depend strongly on observing geometry, so “the wedge” is not a single immutable line across all instruments and observing modes (Gagnon-Hartman et al., 2021, Munshi et al., 2024).

5. Biases and impact on inferred observables

The wedge affects more than signal-to-noise. Because the redshift-space 21-cm signal is anisotropic, restricting measurements to the EoR window changes the quantity being estimated. In simulations of the EoR 21-cm power spectrum, the resulting “wedge bias” is strongest at high redshift, where foreground-avoidance measurements over-estimate the power spectrum by around 100 per cent. Later in reionization the bias becomes negative and smaller in magnitude, typically $\boldsymbol{b}$7 per cent, with only weak dependence on spatial scale and reionization topology (Jensen et al., 2015).

A correction based on clustering wedges recasts the window-restricted monopole as a controlled combination of the monopole, quadrupole, and hexadecapole through a Legendre expansion. In MCMC forecasts with semi-numerical reionization simulations, this procedure yielded unbiased recovery of $\boldsymbol{b}$8 and $\boldsymbol{b}$9, with statistical uncertainties comparable to ideal full-$|\tau| \le |\boldsymbol{b}|/c$0-space analyses and, in some cases, improved degeneracy breaking (Raut et al., 2018).

The impact is also regime-dependent. For 21-cm BAO surveys at $|\tau| \le |\boldsymbol{b}|/c$1–2, treating the wedge as inaccessible in Fisher forecasts increases the errors on angular diameter distances by 3–4.4 times and the errors on $|\tau| \le |\boldsymbol{b}|/c$2 by a factor of 1.5–1.6, showing that the wedge is especially damaging for transverse information (Seo et al., 2015). In low-redshift interferometric H I intensity mapping with MeerKAT-like specifications, visibility-space component separation was found to estimate H I clustering more accurately than pure foreground avoidance, with uncertainties 30 per cent smaller (Chen et al., 2022).

For higher-order EoR observables, wedge-cutting can erase the very non-Gaussian structures that motivate the measurement. In HERA-like forward modeling of one-point statistics, more aggressive wedge-cutting washed out ionized-bubble structure, drove the variance toward nearly zero, made skewness and kurtosis noisy and trend toward zero, and removed the distinctive non-Gaussian PDF shape. Yet the same study found that the rise in skewness and kurtosis near the end of reionization survives in the least aggressive and most realistic wedge-cut cases; with the rolling wedge-cut method, $|\tau| \le |\boldsymbol{b}|/c$3, $|\tau| \le |\boldsymbol{b}|/c$4, 277 hours of integration, and 1 MHz binning, skewness and kurtosis near the upper end of the band reached maximum signal-to-noise ratios around 5 (Kittiwisit et al., 2022).

At dark-ages redshifts, the wedge becomes substantially more severe. A recent analysis found that the wedge typically occupies $|\tau| \le |\boldsymbol{b}|/c$5 of the 2D $|\tau| \le |\boldsymbol{b}|/c$6-space region for $|\tau| \le |\boldsymbol{b}|/c$7, so foreground avoidance leaves only a thin slice of usable modes and usually costs about an order of magnitude in detection significance. For a fiducial $|\tau| \le |\boldsymbol{b}|/c$8 array, avoidance alone is therefore insufficient even for the $|\tau| \le |\boldsymbol{b}|/c$9 signal, leading to the conclusion that some level of foreground subtraction is necessary for realistic dark-ages 21-cm cosmology (Pober et al., 30 Jul 2025). A wedge-aware Fisher treatment for primordial non-Gaussianity at dark-ages redshifts reached the same qualitative conclusion: wedge avoidance removes a substantial fraction of Fourier modes in both the power spectrum and reduced bispectrum, with roughly two orders of magnitude total SNR reduction over the considered redshift range and especially strong degradation of bispectrum constraints (Zhang et al., 22 Jun 2026).

6. Mitigation, suppression, and recovery beyond simple avoidance

The wedge has motivated a large methodological spectrum that extends from improved estimators to calibration, subtraction, array design, and direct mode recovery. In covariance-aware power-spectrum estimation, the optimal quadratic estimator

$(k_\perp,k_\parallel)$0

was shown to suppress foregrounds by an extra factor of $(k_\perp,k_\parallel)$1 in power at the peripheries of the EoR window relative to a basic estimator, increasing the detection significance in a fiducial midpoint-of-reionization model from $(k_\perp,k_\parallel)$2 to $(k_\perp,k_\parallel)$3. That same framework introduced decorrelation and foreground-isolation procedures intended to enlarge the EoR window and, in principle, enable measurements deep within the wedge without formal cosmological signal loss in the final projection step (Liu et al., 2014).

Component separation in visibility space offers a different mitigation path. For MeerKAT/MIGHTEE-like low-redshift data, PCA performed directly in visibility space suppressed foreground contamination at large line-of-sight scales strongly enough that measurements could be pushed closer to the wedge boundary than with pure avoidance. In that case, two foreground modes were removed in the low-noise tests, and the converged avoidance criterion could be relaxed to $(k_\perp,k_\parallel)$4 for $(k_\perp,k_\parallel)$5 after PCA, rather than $(k_\perp,k_\parallel)$6 for pure avoidance (Chen et al., 2022).

Calibration can also be formulated to act directly on wedge leakage. The nucal method extends redundant-baseline calibration by exploiting spectral redundancy: visibilities measuring the same angular Fourier mode at different frequencies are jointly fit with a DPSS model of the beam-weighted sky. In simulations with realistic source and beam chromaticity, this framework solved for unsmooth bandpass features, exposed narrowband interference systematics, and suppressed smooth-spectrum foregrounds below the level of 21cm reionization models, even within much of the wedge region. For well-sampled angular Fourier modes, the associated signal loss was about $(k_\perp,k_\parallel)$7 (Cox et al., 2023).

Array layout enters at an even more structural level. A semi-analytical study of baseline layouts argued that the wedge is fundamentally driven by interferometric chromaticity plus incomplete $(k_\perp,k_\parallel)$8-sampling, and found that radially dense, logarithmically regular layouts can reduce foreground power in the wedge and window by $(k_\perp,k_\parallel)$9–$k_\perp$0, although complete suppression was judged practically unachievable for realistic arrays (Murray et al., 2018). By contrast, the RULES algorithm approached the problem as a $k_\perp$1-completeness design criterion and, in simulations over 130–150 MHz, produced an array that suppressed wedge power by sixteen orders of magnitude relative to a compact hexagonal reference and by about thirteen orders of magnitude more than a random layout of comparable size. That result was explicitly sensitive to antenna position errors and missing baselines, and the paper proposed minimum redundancy requirements and tighter $k_\perp$2 packing density as practical mitigations (MacKay et al., 18 Sep 2025). This juxtaposition suggests that the achievable degree of wedge suppression is strongly dependent on the exact array-completeness criterion, estimator, and implementation assumptions.

Machine-learning and field-level methods pursue recovery rather than suppression. A 3D U-Net trained on wedge-filtered EoR simulations was shown to recover the morphology of the largest ionized regions, including their shape, size, and location, while failing on Gaussianized null tests, indicating reliance on genuinely non-Gaussian mode correlations (Gagnon-Hartman et al., 2021). A light-cone extension using SKA1-Low-like noise reached similar conclusions: the largest ionized regions were reconstructed reliably, recovery was strongest at low $k_\perp$3, and the maps could be used to guide high-redshift galaxy searches and characterize ionized or neutral environments around existing galaxy catalogues (Kennedy et al., 2023). SegU-Net v2, when tested across PCA, GPR, polynomial fitting, and wedge removal, found wedge removal to be the fastest but least effective preprocessing method, whereas the overall framework still achieved about 71% average accuracy and AUC centered around 95% over $k_\perp$4 under foreground-contaminated conditions (Bianco et al., 2023).

The most direct attempts to reconstruct the missing wedge modes now operate at the field level. One recent study introduced both an EFT-based field-level inference pipeline and a diffusion-based generative model for reconstructing full 21-cm fields from wedge-filtered observations. On EFT-generated mock data, the field-level method recovered over 50% of the Fourier modes inside the wedge for $k_\perp$5, with a transfer function within about 5% of unity, while the diffusion model achieved nearly identical large-scale performance. Applied to 21cmFAST mocks, both methods remained competitive and substantially outperformed naive interpolation (Chen et al., 18 Aug 2025).

The foreground wedge thus remains both a geometric contaminant region and a design principle for 21-cm analysis. It defines the conservative boundary for foreground avoidance, sets the dominant systematic geometry for wide-field interferometry, and organizes a substantial literature on estimator theory, calibration, subtraction, tomographic reconstruction, and array architecture. Across that literature, the central issue is unchanged: the wedge is where bright smooth foregrounds, instrument chromaticity, and finite measurement operators overlap most destructively with cosmological information.