Cosmology on Ultralarge Scales with Intensity Mapping of the Neutral Hydrogen 21 cm Emission: Limits on Primordial Non-Gaussianity

Quick read on Camera, Santos, Ferreira & Ferramacho. It's a forecast, not a measurement, and it should be read that way. The Fisher machinery is standard and correctly assembled, the noise models for dishes and interferometers are written out explicitly, and the closing survey discussion — MeerKAT, SKA1, APERTIF, ASKAP, GMRT, SKA2, and which of them can actually reach 400 MHz — is the most useful thing in the paper. That's real value for instrument planning. The new bit relative to Joudaki et al. 2011 is the post-reionization window, z ≈ 1–5, with an explicit dish-versus-interferometer optimization. That's a genuine addition, not a rehash. The classifier record says no limitations and no methods were found; that's wrong on both counts — the paper states its limitations plainly and describes its method in detail. Ignore that part of the record. The soft spots are the ones the authors themselves flag, and they are load-bearing for the headline number. The signal goes as (b_HI − 1), so essentially all the gain comes from z ≈ 2–4 where b_HI is unmeasured; they build it from a Sheth–Tormen mass function with M_HI ∝ M_halo and a circular-velocity cutoff, normalized to a single Ω_HI b_HI measurement. They call the bias evolution a matter of 'fundamental ignorance' — honest, but then the Fisher analysis uses one fiducial b_HI(z) and treats f_NL as the only parameter, asserting that's 'a valid approximation on ultralarge scales' without showing it. Marginalizing over a bias amplitude per frequency bin, partly degenerate with the k⁻² shape over ℓ ≲ 60, would inflate σ_f_NL. The 75×75 covariance is approximated block-diagonally at 20×20 and never validated. Foreground cleaning is called an open question, which it is. None of this breaks the qualitative conclusion — a deep, low-frequency, wide-bandwidth survey is the right route to ultralarge-scale f_NL — but σ_f_NL ≈ 1 is an optimistic target, not a projected error bar. The stress-test note lands; I'd only add that the single-parameter Fisher is a defensible first pass for a forecast, so it's a caveat rather than a flaw. Who should read it: anyone costing low-frequency HI arrays, and anyone working on scale-dependent bias who wants a realistic sense of what intensity mapping can add. Worth a serious referee — the fixes are quantification, not demolition.

The signal is the scale-dependent bias Δb ∝ (b_HI − 1), so the paper's own simplified expression gives σ_f_NL ∝ b/(b−1). At z ≈ 1, b_HI ≈ 1 and the signal vanishes; essentially all of the gain comes from z ≈ 2–4, where b_HI(z) is not measured. The authors build it by integrating a Sheth–Tormen mass function with M_HI ∝ M_halo and a circular-velocity cutoff, then fixing the normalization with Ω_HI b_HI from Ref. [8] (Switzer et al., arXiv:1304.3712, a preprint at the time). They call the redshift evolution of the bias a matter of 'fundamental ignorance,' which is honest, but the Fisher analysis then uses a single fiducial b_HI(z) and treats f_NL as the only parameter, asserting that this is 'a valid approximation on ultralarge scales' without showing it. Two unquantified consequences follow. First, if the true b_HI(z) stays closer to unity or evolves differently, σ_f_NL degrades roughly as b/(b−1). Second, even at the fiducial b_HI, marginalizing over the bias amplitude per frequency bin—which is partly degenerate with the k^−2 f_NL shape over the limited ℓ range—would inflate σ_f_NL. The block-diagonal 20×20 approximation to the 75×75 covariance is also stated but not validated. None of this breaks the qualitative conclusion that a deep low-frequency survey is the right route, but it means the quoted σ_f_NL ≈ 1 is a best case rather than a projected error bar.

This is a well-posed forecast, and the reader's CONDITIONAL verdict is the right frame. The Fisher machinery is standard, the dish and interferometer noise models are explicit, and the instrument survey (MeerKAT, SKA1, APERTIF, ASKAP, GMRT, SKA2) is the most useful part of the paper. The authors are also candid about foreground cleaning and about the bias evolution. My concern is not that they hide the bias problem but that they do not propagate it: the headline number is computed at one fiducial b_HI(z) and with f_NL as the sole parameter, so the reader has no way to see how much of the σ_f_NL ≈ 1 is signal and how much is model choice. The same applies to the block-diagonal covariance approximation, which is stated but not validated. These are exactly the kind of checks a forecast paper needs before its number is used for instrument costing, and they are cheap to run. I would keep the CONDITIONAL verdict and ask for a sensitivity table or figure showing σ_f_NL under variations of the HI–halo model and under marginalization over the bias amplitude. The qualitative conclusion—that a deep, low-frequency, wide-bandwidth HI survey is the right route to ultralarge-scale f_NL—is not in doubt.

The paper's genuine strength is that it identifies a concrete, currently under-exploited observable — the largest angular scales of large-scale HI clustering at z ≈ 1–5 — where the scale-dependent bias signature of primordial non-Gaussianity is maximally accessible, and it pairs that with explicit, instrument-level noise models for both dish and interferometer modes. The forecast is falsifiable in a useful sense: it makes specific, checkable predictions for the survey area, integration time, and angular resolution needed to reach σ_f_NL ≈ 1, which can be tested as MeerKAT and SKA come online. The scale-dependent bias formula and the Fisher machinery are standard and correctly assembled, and the paper is candid about two of its limitations (foregrounds and bias evolution). If the headline number survives the systematic checks requested below, the result would be a well-motivated science case for a dedicated low-frequency intensity-mapping experiment.

The paper asks whether a survey that maps the collective 21 cm emission of neutral hydrogen, without resolving individual galaxies, can measure the scale-dependent signature that primordial non-Gaussianity imprints on the clustering of matter. The authors argue that the answer is yes, in principle: a purpose-built low-frequency intensity mapping experiment covering $z$ between roughly 0.5 and 4.5 could reach an uncertainty on the non-Gaussianity amplitude $f_{NL}$ of order 1. The prize is a constraint on a parameter that separates canonical single-field inflation, which predicts $|f_{NL}|$ of order $10^{-2}$, from models that generate values of order 1 through post-inflationary evolution. The argument rests on the fact that intensity mapping reaches very large spatial scales at high redshift, where the non-Gaussian signal is strongest and where conventional galaxy surveys are slow and cosmic-variance limited.

What carries the argument. The load-bearing mechanism is the scale-dependent bias induced by primordial non-Gaussianity: a tracer's bias acquires a correction $Delta b_X(z,k) = 3 (b^G_X(z) - 1) Omega_m H_0^2 delta_c / [c^2 k^2 T(k) D_+(z)] f_{NL}$, which grows toward small $k$ and vanishes for tracers with bias equal to one. The authors model the neutral hydrogen bias by integrating a Sheth-Tormen halo mass function with hydrogen mass proportional to halo mass, bounded by a circular-velocity cutoff, and normalising to an external measurement of $Omega_{HI} b_{HI}$. The forecast itself is a Fisher matrix built from tomographic 21 cm angular power spectra computed with the CAMB-sources code, including redshift-space distortions and a top-hat frequency window, with noise models for both dish arrays and interferometers.

Core claim. The central claim is a forecast, not a detection. The authors compute the angular power spectrum of 21 cm intensity fluctuations in redshift slices, add the standard scale-dependent correction to the bias that a non-zero $f_{NL}$ induces, and run a Fisher matrix over the tomographic spectrum to estimate the marginal error on $f_{NL}$. For a cosmic-variance-limited survey, the error scales as $[b^G_X T(k_alpha) D_+(z) c^2]^2 / ([6(b^G_X - 1) Omega_m H_0^2 delta_c]^2 sum_alpha N_alpha k_alpha^{-4})$, so two things drive the result: the number of modes, which grows with survey volume and depth, and how far the tracer's bias sits from unity. Neutral hydrogen is underbiased at low redshift and strongly biased at early times, and the authors show that a deep, wide survey reaching frequencies below about 400 MHz is what makes $sigma_{f_{NL}} sim 1$ achievable. They then translate that requirement into concrete instrument specifications, noting that a dish array with roughly 10,000 small elements near 400 MHz, or an interferometer with a field of view near 1000 square degrees, would be needed; existing and planned instruments such as MeerKAT, APERTIF, ASKAP and GMRT fall short on frequency coverage, sensitivity or field of view.

The premise it rests on. The forecast assumes a specific, unmeasured redshift evolution of the neutral hydrogen bias, derived from a halo mass function with hydrogen mass proportional to halo mass; if the real bias stays closer to one at high redshift, the projected error on $f_{NL}$ degrades well beyond the quoted value of about 1.

A measurement of $sigma_{f_{NL}} sim 1$ would discriminate between canonical single-field inflation and models producing order-unity non-Gaussianity through post-inflationary evolution, a distinction the cosmic microwave background alone cannot make at that precision.

The same survey would deliver a three-dimensional map of the matter distribution on ultralarge scales out to $z sim 4$, useful for testing general-relativistic and gauge effects that become non-negligible on those scales.

The forecast sets a concrete design target for a dedicated instrument: a compact array working near 400 MHz with a wide field of view and a large bandwidth, rather than the high angular resolution that current interferometer designs prioritise.

Because the constraint improves as the tracer bias moves away from unity, the result argues for spanning a broad redshift range rather than optimising a single deep slice.

The forecast's sensitivity to the bias evolution suggests a cheap intermediate test: cross-correlating 21 cm intensity maps with an overlapping galaxy survey at $z sim 1$-$2$ would measure the effective hydrogen bias directly, and that measurement could be fed back into the Fisher analysis before any dedicated instrument is built.

The authors treat foreground removal as tractable because Galactic synchrotron emission varies slowly in frequency, but the largest angular scales are exactly where the $f_{NL}$ signal lives; a dedicated study of how much of the $ell < 60$ range survives realistic cleaning would be the natural next step and could shift the quoted error substantially.

If the hydrogen bias at $z sim 3$-$4$ turned out to be closer to one than the Sheth-Tormen model predicts, the same survey would still constrain $f_{NL}$, but the required observing time would grow faster than the survey area, which changes which instrument concept is cheapest rather than whether the measurement is possible.

The scale-dependent bias signature is shared by any tracer of the dark matter, so the intensity mapping case is best read as an argument that the very largest scales are worth instrumenting at all, independent of which tracer is eventually used.

This review was written by a machine. No person has reviewed this paper on Pith yet. A scientist listed on Pith can review it for a fee, and their review is published here whatever it concludes.

Request a human review