WorksPhysics and Astronomy
A high-resolution search for axion dark-matter flows with ORGAN-Q
Abstract
We search for narrow axion dark-matter flows using the high-resolution data acquired during the ORGAN-Q experiment. The analysis covers nine potential axion linewidths from 0.466Hz to 119.209Hz, equivalent to axion effective quality factors to , over the range 6.1550GHz to 6.3496GHz, corresponding to axion masses of approximately 25.455 to 26.260. Once contributions from native-bin correlations, data-quality and interference vetoes, frequency-dependent system noise, the cavity response, and filter attenuation are accounted for, no candidate satisfies the required persistence ratio in any search bandwidth. Across the nine assumed linewidths, the strongest KSVZ axion-flow density limits range from 22.8 for a signal to for a signal, with both attained near .
Introduction
The axion was originally introduced as a consequence of the Peccei–Quinn mechanism, which provides a dynamical solution to the strong CP problem of quantum chromodynamics [1, 2, 3]. In addition to this particle-physics motivation, axions produced non-thermally in the early universe can behave as cold dark matter and may constitute some or all of the Galactic dark-matter density [4, 5, 6]. Their small mass and weak interactions make direct detection challenging, but the axion-photon coupling provides a measurable electromagnetic signature.
In a microwave-cavity haloscope, Galactic axions may convert into photons in the presence of a strong static magnetic field [7, 8]. When a cavity mode is tuned near the axion conversion frequency, the deposited power is resonantly enhanced and can be detected through the receiver chain. Conservation of energy relates the frequency of the resulting photon to the energy of a nonrelativistic axion according to
where is Planck’s constant, is the axion mass, is the speed of light, and is the axion velocity relative to the detector. For an axion flow with velocity dispersion , this relation gives a fractional signal linewidth
where is the signal linewidth, is the axion rest-mass frequency, and is the velocity dispersion of the flow. The central signal frequency therefore
primarily identifies the mass, while the signal linewidth and temporal frequency variation contain information about the local axion velocity distribution.
The Oscillating Resonant Group Axion (ORGAN) experiment is a microwave-cavity haloscope designed primarily to search for axions at comparatively high masses [9]. The ORGAN-Q experiment used a tunable conducting-rod resonator, a strong solenoidal magnetic field, and operated at milli-kelvin with a Josephson parametric amplifier to search the frequency interval from approximately 6.16 GHz to 6.35 GHz. It provided constraints on the virialised axion mass range over 25.45 eV to 26.27 eV [10].
Conventional haloscope analyses generally adopt a smooth, approximately virialised Galactic halo, for which characteristic velocities of order produce a fractional axion linewidth of order [10, 11, 12]. At the ORGAN-Q operating frequency, this corresponds to a signal distributed over several kilohertz. However, galactic dark matter may consist of more coherent populations. Fine-grained substructure may produce flows with much smaller velocity dispersions and correspondingly narrower spectral features. Such models include cold infall and caustic ring flows, together with streams produced by the tidal disruption of axion miniclusters [13, 14, 15]. The assumptions of a caustic ring model for example, produce an upper limit to the fractional energy dispersion of the “Big Flow” dark matter stream of [14], with the root-mean-square energy dispersion of the flow. This corresponds to a lower limit of effective quality factor or an upper limit on linewidth of less than 1.5 Hz at ORGAN-Q frequencies. This model also provides an axion density central estimate of 11.2 GeV/cm [16], providing a model-dependent benchmark for interpreting limits on narrow flows. High-resolution cavity searches have previously been used to probe such non-virialised axion populations [17, 18, 19, 20].
The ORGAN-Q acquisition system generated two simultaneous data streams: a medium resolution stream, previously used to report on a virialised halo search, and a high-resolution stream, used here to search for narrower signals (see § 2 B for details). Motivated by these possibilities, we use the ORGAN-Q high-resolution data to search for persistent narrow spectral features and constrain the density of non-virialised KSVZ axion flows.
Experiment and Data Set
ORGAN apparatus and operating configuration
The ORGAN-Q experiment [10] employed a 0.708 L oxygen-free high-conductivity copper cavity inside a 7 T solenoidal magnet, for which the average magnetic field over the cavity volume was T. An off-axis aluminium tuning rod varied the frequency of the mode between approximately 6.15 GHz to 6.35 GHz without encountering a mode crossing, yielding typical mode form factors of . Over the rod-tuning range of the experiment, the loaded quality factor was (), whilst the antenna coupling had an average value of ().
The experimental configuration is shown schematically in Figure 1. The cavity and first-stage receiver components operated at millikelvin temperatures. The mixing chamber and cavity were maintained near 80 mK and 100 mK, respectively, although motion of the tuning rod produced a higher characteristic rod temperature of approximately 500 mK [10]. Signals from the strongly coupled cavity port were routed through a cryogenic circulator to a flux-driven Josephson parametric amplifier (JPA), followed by a cryogenic high-electron-mobility-transistor (HEMT) amplifier. The JPA was operated in three-wave-mixing-mode, with a pump frequency near twice the signal frequency, and provided more than 18 dB of gain over most of the searched range. Following amplification, the signal was mixed to an intermediate frequency using an in-phase and quadrature mixer, whose outputs were combined using a 90° hybrid to provide image rejection before digitisation.
Figure 1. Schematic of the ORGAN-Q apparatus and receiver chain used to acquire the high-resolution data. A potential axion signal generated in the tunable cavity is routed through the cryogenic receiver to the flux-driven JPA and subsequent HEMT amplification stage. At room temperature, the amplified signal is down-converted using an image-rejecting mixer before being processed by the FPGA digitiser. The vector network analyser, cryogenic switch and heated load provide the measurements required for a cavity and receiver calibration. Piezoelectric actuators control the tuning-rod position and antenna coupling. The schematic is reproduced from Ref. [10].
At each tuning position, vector-network-analyser measurements tracked the cavity resonance and JPA gain response. The tuning increment was approximately one fifth of the cavity linewidth, and acquisition at an individual tuning position typically lasted between 30 min to 60 min. The medium- and high-resolution data streams were acquired simultaneously by the field-programmable gate array (FPGA) digitiser.
Data acquisition
The medium-resolution stream performed real-time averaging and produced 6554-bin spectra over a 6.25 MHz interval centred near an intermediate frequency of 45 MHz, corresponding to a bin spacing of Hz. These data also supplied the tuning and receiver calibration metadata used in the present analysis. In parallel, the high-resolution stream operated at one quarter of the primary sampling rate and without real-time spectral averaging. For the high-resolution stream, time-series data were transferred continuously, without gaps, from the FPGA to the host computer, where FFTs were computed to produce the stored power spectra. Each stored high-resolution spectrum contains 3,355,444 bins spanning 1.5625 MHz with frequency spacing Hz. However, after analysing the data, frequency-bin correlations were detected. These are consistent with rectangular records zero-padded by a representative oversampling factor , as discussed and accounted for in Appendix A. This corresponds to a representative record duration s, and a rectangular-window equivalent noise bandwidth Hz.
Two passes over the frequency range were acquired from 13–26 February 2024 and 6–10 March 2024. The second pass over selected frequency regions were to provide sufficient coverage after accounting for interference or poor receiver performance.
System-noise calibration
The frequency-dependent system-noise temperature referred to the cavity output , was determined using the signal-to-noise ratio improvement (SNRI) procedure discussed in Ref. [10] for each tuning position. This combined the measured HEMT noise temperature and receiver-line attenuation with JPA-off transfer functions measured using a vector network analyser and noise-power spectra recorded using a spectrum analyser. The same cuts and calibration employed in the medium-resolution were used for this search, satisfying .
Analysis pipeline
Baseline Filtering
Each raw high-resolution spectrum may contain smooth frequency-dependent structure arising from the receiving gain and noise response. This baseline must be estimated and divided out before the statistical properties of the residual power can be tested. Savitzky–Golay (SG) filters are commonly used for this purpose in halo-scope analyses. In the present analysis, the baseline was instead estimated using the modified-sinc (MS) smoother introduced in Ref. [21]. At a matched 3-dB cutoff, the MS filter provides substantially stronger stop-band suppression than a conventional SG filter [21]. The transfer functions of the two filters at a demonstrative window are shown in Figure 2. Despite stronger stop-band suppression, this alone does not establish improved recovery of narrow spectral signals. Such signals have broad Fourier spectra, and their attenuation after baseline normalisation depends on the full filter response and the signal profile. The MS smoother was therefore selected for its strong suppression of fine-scale structure in the estimated baseline, with its signal-recovery evaluated using injection tests described in § 6 B.
Figure 2. Representative magnitude responses of degree-four Savitzky–Golay (SG) and modified-sinc (MS) smoothers matched at their −3-dB cutoff. The responses are normalised at zero spatial frequency, expressed as cycles per native spectral bin. The MS kernel exhibits substantially lower stop-band sidelobes.
Denoting the linear power in native frequency bin by , the estimated baseline was
where is the estimated baseline in bin , is the normalised symmetric MS kernel at bin offset , and is the kernel half-width. The analysis used the degree parameter and , corresponding to a bin kernel. At the boundaries of each spectrum, the data required by the convolution were extrapolated using cosine-squared-weighted linear extrapolation. The convolution was evaluated using zero-padded FFT convolution to assist with edge-handling.
Bins belonging to manually identified strong-features masks discussed in § 3 E, had their power replaced by the mean power measured ten masked-interval widths either side of the mask. The masked bins were omitted from the Erlang test in § 3 B and excluded from the subsequent trigger search.
Noise Distribution Cuts
Following baseline normalisation, the power distribution of each raw spectrum was tested for consistency with an Erlang distribution;
where is the baseline-normalised power, is the shape parameter and is the scale parameter. For an individual, unaveraged spectrum, and is its mean power, reducing Eq. 4 to the exponential distribution expected for stationary Gaussian receiver noise. The evolution of the erlang distribution is shown in Figure 3 and compared to a sample dataset.

Figure 3. Erlang evolution expectation with comparison to the data.
The agreement between the measured distribution of a spectrum and the expected Erlang distribution is quantified using the reduced Pearson’s chi-square statistic [22],
where and are the observed and expected counts in histogram bin , respectively, and is the degrees of freedom. Every tenth native frequency bin was sampled, known strong signals masked, constructed using 1000 equal-width bins with those excluded. The expected counts were normalised to the corresponding observed total.
To establish the expected variation of the goodness-of-fit statistic for noise-only spectra, we generated 100,000 synthetic spectra with independently exponentially distributed powers. Each was analysed using the same procedure as the experimental data. The resulting distribution of reduced , shown in Figure 4, had a mean and standard deviation of 1.00 and 0.0629, respectively. These values were used to set the data-quality acceptance criteria: , corresponding to approximately . The cut rejected two spectra from the first pass and two from the second-pass out of the eligible tuning folders. The first failures exhibited anomalous background behaviour, while the second failures had substantially elevated values and resembled corrupted start-up records.

Figure 4. Reduced chi-square distribution obtained from 100,000 synthetic Erlang-distributed spectra. The adopted quality cuts are shown by the vertical dotted lines.
Averaging
Within each dataset and cavity tuning, the accepted time-ordered spectra were partitioned into consecutive, non-overlapping groups containing at most 100 spectra, with group sizes made as uniform as possible. This provides an enhancement of signal-to-noise ratio of each group, whilst the conservative probability of a doppler drift of one bin remains low ( bin width drift per 1000 spectra). Each group was represented by the bin-wise arithmetic mean of its baseline-normalised spectra, maintaining the actual group size , which was used in subsequent statistical calculations. This procedure produced 17, 421 and 4361 first and second pass averaged groups, respectively, prior to the subsequent calibration quality and interference vetoes.
Native-bin correlations
The raw FFT powers exhibit correlations between neighbouring native frequency bins. For each dataset, the Pearson correlation coefficient between bins was evaluated before MS filtering using 4420 first-pass and 1370 second-pass spectra, and the median was taken across spectra. The correlations are dominated by the first three lags, with for both datasets. The representative FFT oversampling factor was obtained by jointly fitting the first twelve correlation lags of the first and second-pass datasets to the rectangular-record model derived in Appendix A.
Let denote the number of stored frequency bins in the sliding search window. For the mean , assuming locally stationary bin statistics with a variance , summing the covariance terms gives [23]
where
Here is the integer separation between bins. Averaging independent spectra therefore gives variance . Equating this to the variance of a unit-mean gamma distribution yields an effective shape parameter . For the widest search window, , the measured values are for both datasets. These correlations were used in subsequent trigger-threshold and sensitivity calculations.
Interference Masks and Vetoes
Four classes of contaminated intermediate-frequency (IF) intervals were catalogued: tuning-specific strong instrumental features, dataset-wide IF lines, tuning-specific radio-frequency interference (RFI), and spectral combs, summarised in Table I.
Each tuning-specific RFI signal was carefully analysed over the duration at which they appeared. The key indication that these were pickup of an instrumental source elsewhere was inferred from its stability on the order of minutes, followed by discontinuous frequency jumps up to 100 Hz.
The strong feature masks were applied before baseline estimation to prevent prominent pump and mixing products from distorting the MS baseline.
For a -bin search, each interval was expanded on either side by , where is the native bin spacing, so that any sliding window intersecting a masked bin was vetoed. All manual, IF, RFI, and comb vetoes were applied before persistence matching, and consequently, vetoed triggers could neither seed nor support a candidate. Additionally, affected groups provide no coverage of that trial frequency.
Multibin trigger catalogues
Each averaged spectrum was searched using bin-aligned top-hat templates formed from sliding means of adjacent native bins, evaluated with a one-bin stride. The tested window sizes were , corresponding to bandwidths starting from the native bin spacing , up to 119.209 Hz.
For the averaged group , a window was recorded as a trigger when its mean power exceeded
where is the inverse survival function of the gamma distribution, and is the effective shape parameter defined in Sec. III D. The value provides a manageable number of triggers for processing ( per spectra), whose IF frequency and sliding mean power were stored.
Candidate Selection
An axion flow would produce a narrow excess of microwave power above the receiver-noise background, where the intrinsic linewidth is set by the flow’s velocity dispersion. For the persistent flows we consider here, a dark matter signal should recur in observations covering the same radio-frequency range within allowed relative Doppler drifts of the detector. Candidate signals are therefore identified by testing whether an excess is persistent across temporally-separated measurements within the allowed frequency drift. Candidate selection was performed separately for each search width using the combined first and second-pass trigger catalogues. Only the groups and triggers passing the calibration and interference selections discussed earlier enter this stage.
Doppler-Drift Model
The axion signal present at the detector may drift in frequency due to model-dependent Doppler effects. The axion energy depends on the relative velocity , where is the axion flow velocity, and is the detector velocity, using the galactic coordinate system. The detector velocity is broken up into several components: . is the Sun’s total galactocentric velocity, including the contributions of local standard of rest and the solar peculiar orbit, is the velocity of the Earth around the sun, and is the velocity of the experiment due to rotation of the Earth in Perth, Australia. The axion frequency is given by Eq. (1). One can treat and as constant over the duration of the experiment, and define the axion rest frequency . For two spectra taken at and , the predicted frequency change is
| Veto class | Count | Reason/Description |
| Strong instrumental features | 11 | Linear combinations of the pump, synth, and/or sampling frequencies result in measured signals. |
| Dataset-wide IF lines | 10 | Approximately fixed IF across cavity tunings, indicating receiver-chain interference. |
| Tuning-specific RFI | 2 | Discontinuous frequency jumps over time, consistent with instrumental pickup. |
| Spectral combs | 7 | Periodic instrumental interference combs with 99 999.5 Hz spacing. |
Table 1. Interference vetoes across the data.
highlighting the time dependence of the relative velocity .
The detector velocity was evaluated by constructing the laboratory barycentric state and transforming it into the Astropy Galactocentric frame. The Earth’s barycentric position and velocity at each spectrum time were obtained from the JPL DE440 ephemeris [24], while the laboratory topocentric velocity due to the Earth’s rotation was calculated at the UWA laboratory position using Astropy’s EarthLocation.get_gcrs_posvel routine [25]. The topocentric GCRS vector is geocentric but oriented consistently with the barycentric celestial axes, so it was combined with the Earth’s barycentric state before transformation to the Galactocentric frame. The Galactocentric transformation used the Astropy default solar parameters, corresponding to [25]. The resulting detector velocity was used with an assumed axion-flow velocity to compute the relative velocity , which determines the expected frequency drift between spectra through Eq. (9).
The axion flow speed was fixed at the galactic escape velocity [26, 27] to allow maximal Doppler drift. Its direction was chosen from the two directions parallel and antiparallel to , retaining the direction that maximised the absolute frequency change between the first and last data timestamps of the experimental run. For a seed trigger at RF frequency in group and a comparison group , Eq. (9) was evaluated between the start of the earlier group and the end of the later group, using in place of . A
trigger at frequency was considered persistent when
where is the corresponding Doppler displacement and is the native bin spacing.
Persistence Requirement
For each seed trigger , was the number of eligible averaged groups whose unvetoed RF span contained . The corresponding counted groups containing at least one same-width trigger satisfying Eq. (10). Each group, including the seed group, contributed at most once. The persistence ratio was therefore
This is a conservative estimation under the assumption that we expect an incoming axion signal to be persistent at the same frequency within an allowable Doppler drift.
Sensitivity and Density-Limit Construction
Signal and cavity response
For this high-resolution analysis, we constrain the axion-flow density with the axion–photon coupling fixed to the KSVZ benchmark. Using the normalisation adopted in Ref. [12], the coupling magnitude is
where is the fine-structure constant and is the adopted quantum chromodynamics parameter. The axion rest energy is approximated by , giving the coupling in GeV.
For group , the expected axion-conversion power at frequency and trial flow density is [12, 28]
where is the benchmark coupling evaluated at in units of , and is the flow density expressed in . The Lorentzian cavity response is
The remaining parameters are the fitted cavity frequency , cavity full width at half max , magnetic field , cavity volume , interpolated form factor , loaded quality factor , antenna coupling , , and vacuum permeability .
The signal is modelled by an intrinsic top-hat power spectrum of width , centred on the -bin search window. For the assumed top-hat signals, we define the effective axion quality factor as , where is the full intrinsic linewidth. The finite record duration broadens this spectrum according to the rectangular-window FFT response. Its expected contribution to the baseline-normalised statistic is
Here is Boltzmann’s constant, is the system-noise temperature, is the MS filter efficiency, and with argument of oversampling factor , is the dimensionless signal-response correction due to oversampling that is derived in Appendix A. The quoted signal widths refer to intrinsic linewidths before instrumental broadening. The limits assume that the signal is centred on the search window.
Modified-sinc filter attenuation
For each , a weak, unit-area, aligned top-hat was added to a flat baseline, the MS normalisation was applied, and the maximum recovered -bin sliding sum was compared with the injected sum. The resulting efficiencies are shown in Table II.
| (Hz) | ||
| 1 | 0.466 | 0.999 960 |
| 2 | 0.931 | 0.999 920 |
| 4 | 1.863 | 0.999 840 |
| 8 | 3.725 | 0.999 680 |
| 16 | 7.451 | 0.999 360 |
| 32 | 14.901 | 0.998 721 |
| 64 | 29.802 | 0.997 441 |
| 128 | 59.605 | 0.994 883 |
| 256 | 119.209 | 0.989 766 |
Table II. Modified-sinc filter signal-recovery efficiency for each aligned top-hat search width. The efficiency is the ratio of recovered to injected -bin signal power after baseline normalisation.
Detection probability and 95% limit
Using the correlation-corrected shape parameter and threshold defined in § 3, the pipeline represents the signal as an additive shift of the normalised statistic. The probability that a group crosses its threshold is
where is survival function of a gamma distribution. At each trial frequency , and search width , let denote the number of eligible averaged groups covering that frequency. As in [29], groups are treated as independent Bernoulli trials with a trigger probability , and thus, the total number of triggering groups, , follows a Poisson-binomial distribution, calculated by successively convolving pairs .
For each and , the 95% KSVZ flow-density limit is obtained numerically by solving
where , rounded upwards, accounts for the 30% persistence requirement.
Result and discussion
Data-quality and Candidate Outcomes
The averaged-group data quality cuts are summarised in Table III. Missing-metadata (No meta.) groups correspond to tuning folders that were not included in the eligible medium-resolution analysis and are therefore not used in the high-resolution search. Averaged-groups with suspicious values, and missing spectrum analyser files (No cal.) are also excluded (see § 2.C).
| Data set | Accepted | Bad | No meta. | No cal. | total |
| First | 14,256 | 505 | 2,631 | 29 | 17,421 |
| Second | 3,227 | 850 | 268 | 16 | 4,361 |
| Combined | 17,483 | 1,355 | 2,899 | 45 | 21,782 |
Table III. Averaged-group calibration status before frequency-dependent interference cuts.
No triggers survive the 30% persistence requirement after calibration and interference vetoes.
KSVZ flow-density limits
Figure 5 presents the nominal 95% KSVZ axion-flow density limits sampled every 1 kHz. Each coloured curve corresponds to a distinct assumed top-hat linewidth. At evaluated points, densities above a curve are excluded for that linewidth; missing points denote no valid post-cut coverage.
Figure 5. 95% KSVZ axion-flow density limits sampled on a 1 kHz grid for the nine aligned top-hat signal widths. At evaluated frequencies, densities above each curve are excluded under the intrinsic top-hat model. The calculation includes MS filter attenuation and the measured native-bin correlation correction.
The strongest 95% limits placed on KSVZ axion-flow density range from 22.8 GeV cm for a 0.466 Hz signal to 153.7 GeV cm for a 119.209 Hz signal.
The effect of strong nearby bin correlations on the order of bins manifests as reduced sensitivity particularly to the narrowest few-bin-width searches.
Uncertainty Analysis
Following the uncertainty prescription of Ref. [10], we adopt fractional uncertainties of 11.6% in and 5% in . Assuming independence, and holding other experimental parameters as fixed, the signal statistic scales as . Consequently, the density limit scales as . First-order propagation gives
Interpretation and Comparison
The present search complements the previous ORGAN-Q virialised-halo analysis [10] by testing substantially narrower spectral features within the same mass region. The earlier analysis constrained the axion–photon coupling for an assumed local density. Here, we fix the coupling to the mass-dependent KSVZ benchmark and constrain the density of a persistent axion flow under the signal model described in § 6.
Previous ADMX searches placed limits on nonvirialised axion populations in the 1.98 eV to 2.17 eV and 3.3 eV to 3.69 eV mass regions, considering different flow velocity dispersions and treatments of Doppler modulation [17, 18, 19]. More recently, Bartram et al. reported frequency-dependent KSVZ flow-density limits of approximately 0.003 GeV/cm to 0.01 GeV/cm at 95% confidence across the searched portions of 2.81 eV to 3.31 eV [20]. Hipp et al. extended the coverage to 3.3 eV to 4.2 eV, reporting KSVZ flow-density limits of approximately 0.004 GeV/cm in their finest-resolution analysis at 95% confidence. CAPP-12 TB also searched for axions associated with the Sagittarius tidal stream, excluding KSVZ axion densities above approximately 0.025 GeV/cm over 4.51 eV to 4.59 eV at 90% confidence under the adopted stream model [30].
The present work provides complementary constraints at substantially higher axion masses, 25.455 eV to 26.260 eV, for nine assumed linewidths between 0.466 Hz and 119.209 Hz. Comparisons between these searches must account for their different signal shapes, velocity dispersions, Doppler treatments, and confidence levels. The benchmark of the “Big Flow” discussed in § 1 is not excluded, with our strongest
density limits presented in this study approximately twice this value. This limit is also roughly 51 times larger than the conventional local-halo density benchmark of 0.45 GeV/cm [31].
Conclusion
The ORGAN-Q high-resolution data were analysed for persistent narrow axion signals over the mass range 25.455 eV to 26.260 eV, using nine assumed linewidths between 0.466 Hz and 119.209 Hz. After calibration and interference vetoes, no candidate satisfies the 30% persistence requirement. Under the assumed signal model and KSVZ axion–photon coupling, we obtain 95% upper limits on the local axion-flow density. The strongest limits are 22.8 GeV/cm for a 0.466 Hz linewidth, and 153.7 GeV/cm for a 119.209 Hz linewidth, both obtained near an axion mass of 25.857 eV.
Future ORGAN experiments will have a better understanding of the acquisition and analysis pipeline, maturing the pipeline to scout for signs of correlations earlier and improving the efficiency of data processing. Additionally, acquisition procedures that minimise the total data stored will be investigated, aiming to implement real-time filtering and threshold triggers.
By probing spectral features substantially narrower than those expected from a virialised halo, this analysis extends the ORGAN-Q experiment to non-virialised axion populations. These density constraints complement the previously reported virialised-halo coupling limits within the same mass range.
Acknowledgments
This work is supported by the Australian Research Council grant CE200100008. Michael T. Hatzon acknowledges support from the Defence Science Centre, an initiative of the State Government of Western Australia, and from The University of Western Australia through the Hackett Postgraduate Research Scholarship. Ben T. McAllister acknowledges support from grant DE250100933.
FFT Response due to Oversampling
For white Gaussian noise sampled by a rectangular window, the correlation between nearby values due to oversampling follows the squared-sinc form derived in Appendix A of Ref. [32]. This is given by
where is the separation in stored bins, , and .
For a rectangular window of duration , the unit-area spectral response to a monochromatic signal, evaluated at a frequency offset from the signal, is
We note here that , therefore a monochromatic signal of total power has a measured signal PSD of .
Next, we consider the top-hat signal, described by an intrinsic power spectrum that is uniform over the full width , with intrinsic PSD
Here is the central signal frequency, is the frequency offset from , and for , and zero otherwise. A frequency interval within this band contains mean signal power,
At the evaluation frequency , the contribution from a component at is . For the assumed stationary signal with an intrinsic top-hat PSD, the expected measured signal PSD is obtained by convolving the intrinsic PSD with the sampled response :
where is the top-hat shape after the instrumental response, with contributions added in mean power.
For a signal centred on a -bin window, the frequency bin offsets are
Denoting the local noise PSD as for Boltzmann’s constant and system temperature , before any filter attenuation the expected excess above the unit baseline of the normalised -bin mean is
which defines the dimensionless response factor
For the adopted intrinsic width , define
Substituting Eqs. (20) and (23) into Eq. (26), and changing the integration variables gives
Evaluating the integral where
is an anti-derivative of , and
yields the final form of the implemented response factor:
For the measured value of , representative factors are , , and .
At fixed axion–photon coupling, the conversion power is proportional to the flow density. Consequently, a smaller signal response requires a larger density to produce the same expected excess.
The combined signal response is modelled by the factorised approximation , using MS filter efficiencies obtained for unbroadened top-hat profiles. The correlated-noise variance is treated specifically using the measured frequency-bin correlations.
References
- [1]R. D. Peccei and H. R. Quinn, Physical Review Letters 38, 1440 (1977).
- [2]S. Weinberg, Physical Review Letters 40, 223 (1978).
- [3]F. Wilczek, Physical Review Letters 40, 279 (1978).
- [4]J. Preskill, M. B. Wise, and F. Wilczek, Physics Letters B 120, 127 (1983).
- [5]L. F. Abbott and P. Sikivie, Physics Letters B 120, 133 (1983).
- [6]M. Dine and W. Fischler, Physics Letters B 120, 137 (1983).
- [7]P. Sikivie, Phys. Rev. Lett. 51, 1415 (1983).
- [8]P. Sikivie, Physical Review D 32, 2988 (1985).
- [9]B. T. McAllister, G. Flower, E. N. Ivanov, M. Goryachev, J. Bourhill, and M. E. Tobar, Physics of the dark universe 18, 67 (2017).
- [10]A. P. Quiskamp, G. R. Flower, S. Samuels, B. T. McAllister, P. Altin, E. N. Ivanov, M. Goryachev, and M. E. Tobar, Physical Review D 111, 095007 (2025).
- [11]M. S. Turner, Phys. Rev. D 42, 375 (1990).
- [12]B. M. Brubaker, L. Zhong, S. K. Lamoreaux, K. W. Lehnert, and K. A. van Bibber, Phys. Rev. D 96, 123008 (2017).
- [13]P. Sikivie, Physics Letters B 432, 139 (1998).
- [14]N. Banik and P. Sikivie, Phys. Rev. D 93, 103509 (2016).
- [15]P. Tinyakov, I. Tkachev, and K. Zioutas, Journal of Cosmology and Astroparticle Physics 2016 (01), 035.
- [16]S. S. Chakrabarty, Y. Han, A. H. Gonzalez, and P. Sikivie, Physics of the Dark Universe 33, 100838 (2021).
- [17]L. D. Duffy, P. Sikivie, D. B. Tanner, S. J. Asztalos, C. Hagmann, D. Kinion, L. J. Rosenberg, K. van Bibber, D. B. Yu, and R. F. Bradley, Phys. Rev. D 74, 012006 (2006).
- [18]J. Hoskins, J. Hwang, C. Martin, P. Sikivie, N. S. Sullivan, D. B. Tanner, M. Hotz, L. J. Rosenberg, G. Rybka, A. Wagner, S. J. Asztalos, G. Carosi, C. Hagmann, D. Kinion, K. van Bibber, R. Bradley, and J. Clarke, Phys. Rev. D 84, 121302(R) (2011).
- [19]J. Hoskins, N. Crisosto, J. Gleason, P. Sikivie, I. Stern, N. S. Sullivan, D. B. Tanner, C. Boutan, M. Hotz, R. Khatriwada, D. Lyapustin, A. Malagon, R. Ottens, L. J. Rosenberg, G. Rybka, J. Sloan, A. Wagner, D. Will, G. Carosi, D. Carter, L. D. Duffy, R. Bradley, J. Clarke, S. O’Kelley, K. van Bibber, and E. J. Daw, Phys. Rev. D 94, 082001 (2016).
- [20]C. Bartram, T. Braine, R. Cervantes, N. Crisosto, N. Du, C. Goodman, M. G. Guzzetti, C. Harnett, S. Lee, G. Leum, L. J. Rosenberg, G. Rybka, J. S. W. Dishaw, M. H. Awida, D. Bowring, A. S. Chou, M. Hollister, S. Knirck, A. Sonnenschein, W. Wester, R. Khatiwada, J. Brodsky, G. Carosi, L. D. Duffy, M. Goryachev, B. McAllister, A. Quiskamp, C. Thomson, M. E. Tobar, C. Boutan, M. Jones, B. H. LaRocque, E. Lentz, N. E. Man, N. S. Oblath, M. S. Taubman, J. Yang, J. Clarke, I. Siddiqi, A. Agrawal, A. V. Dixit, J. R. Gleason, Y. Han, A. T. Hipp, S. Jois, P. Sikivie, N. S. Sullivan, D. B. Tanner, E. J. Daw, M. G. Perry, J. H. Buckley, C. Gaikwad, J. Hoffman, K. W. Murch, and J. Russell (ADMX Collaboration), Phys. Rev. D 109, 083014 (2024).
- [21]M. Schmid, D. Rath, and U. Diebold, ACS Measurement Science Au 2, 185 (2022).
- [22]K. Pearson, The London, Edinburgh, and Dublin Philosophical Magazine and Journal of Science 50, 157 (1900).
- [23]H. Flyvbjerg and H. G. Petersen, The Journal of Chem- R. S. Park, W. M. Folkner, J. G. Williams, and D. H. Boggs, The Astronomical Journal 161, 105 (2021).
- [25]Astropy Collaboration, A. M. Price-Whelan, P. L. Lim, N. Earl, N. Starkman, L. Brad ley, D. L. Shupe, A. A. Patil, L. Corrales, C. E. Brasseur, M. Nöthe, A. Donath, E. Tollerud, B. M. Morris, A. Ginsburg, E. Vaher, B. A. Weaver, J. Tocknell, W. Jamieson, M. H. van Kerkwijk, T. P. Robitaille, B. Merry, M. Bachetti, H. M. Günther, T. L. Aldcroft, J. A. Alvarado-Montes, A. M. Archibald, A. Bödi, S. Bapat, G. Barentsen, J. Bazán, M. Biswas, M. Boquien, D. J. Burke, D. Cara, M. Cara, K. E. Conroy, S. Conseil, M. W. Craig, R. M. Cross, K. L. Cruz, F. D’Eugenio, N. Dencheva, H. A. R. Devillepoix, J. P. Dietrich, A. D. Eigenbrot, T. Erben, L. Ferreira, D. Foreman-Mackey, R. Fox, N. Freij, S. Garg, R. Geda, L. Glattly, Y. Gondhalekar, K. D. Gordon, D. Grant, P. Greenfield, A. M. Groener, S. Guest, S. Gurovich, R. Handberg, A. Hart, Z. Hatfield-Dodds, D. Homeier, G. Hossainzadeh, T. Jenness, C. K. Jones, P. Joseph, J. B. Kallmbach, E. Karamehmetoglu, M. Kaluszyński, M. S. P. Kelley, N. Kern, W. E. Kerzendorf, E. W. Koch, S. Kulumani, A. Lee, C. Ly, Z. Ma, C. MacBride, J. M. Maljaars, D. Muna, N. A. Murphy, H. Norman, R. O’Steen, K. A. Oman, C. Pacifici, S. Pascual, J. Pascual-Granado, R. R. Patil, G. I. Perren, T. E. Pickering, T. Rastogi, B. R. Roulston, D. F. Ryan, E. S. Rykoff, J. Sabater, P. Sakurikar, J. Salgado, A. Sanghi, N. Saunders, V. Savchenko, L. Schwardt, M. Seifert-Eckert, A. Y. Shih, A. S. Jain, G. Shukla, J. Sick, C. Simpson, S. Singanamalla, L. P. Singer, J. Singhal, M. Sinha, B. M. Sipőcz, L. R. Spitler, D. Stansby, O. Streicher, J. Šumak, J. D. Swinbank, D. S. Taranu, N. Tewary, G. R. Tremblay, M. de Val-Borro, S. J. Van Kooten, Z. Vassović, S. Verma, J. V. de Miranda Cardoso, P. K. G. Williams, T. J. Wilson, B. Winkel, W. M. Wood-Vasey, R. Xue, P. Yoachim, C. Zhang, A. Zonca, and Astropy Project Contributors, Astrophysics. J. 935, 167 (2022), arXiv:2206.14220 [astro-ph.IM].
- [26]H. H. Koppelman and A. Helmi, Astronomy & Astrophysics 649, A136 (2021).
- [27]L. Necib and T. Lin, The Astrophysical Journal 926, 189 (2022).
- [28]S. Al Kenany, M. Anil, K. Backes, B. Brubaker, S. Cahn, G. Carosi, Y. Gurevich, W. Kindel, S. Lamoreaux, K. Lehnert, S. Lewis, M. Malnou, D. Palken, N. Rapidis, J. Root, M. Simanovskaia, T. Shokair, I. Urdinaran, K. van Bibber, and L. Zhong, Nuclear Instruments and Methods in Physics Research Section A: Accelerators, Spectrometers, Detectors and Associated Equipment 854, 11 (2017).
- [29]A. Hipp, A. Quiskamp, T. Caligiure, J. Gleason, Y. Han, S. Jois, P. Sikivie, M. Solano, N. Sullivan, D. Tanner, et al., Physical Review D 112, L101101 (2025).
- [30]A. K. Yi, S. Ahn, Ç. Kutlu, J. Kim, B. R. Ko, B. I. Ivanov, H. Byun, A. F. van Loo, S. Park, J. Jeong, O. Kwon, Y. Nakamura, S. V. Uchikhin, J. Choi, S. Lee, M. Lee, Y. C. Shin, J. Kim, D. Lee, D. Ahn, S. Bae, J. Lee, Y. Kim, V. Gkika, K. W. Lee, S. Oh, T. Seong, D. Kim, W. Chung, A. Matlashov, S. Youn, and Y. K. Semertzidis, Phys. Rev. D 108, L021304 (2023).
- [31]J. I. Read, Journal of Physics G: Nuclear and Particle Physics 41, 063101 (2014).
- [32]C. Koen, Monthly Notices of the Royal Astronomical Society 449, 1098 (2015), https://academic.oup.com/mnras/article-pdf/449/1/1098/4138927/stv288.pdf.


