Datasets:

SouthernCrossAI
/

Arxiv

Dataset card Data Studio Files Files and versions Community

Split (1)

train · 132k rows

Filename stringlengths 22 64	Paragraph stringlengths 8 5.57k
Processed_Optimising_MWA_EoR_data_processing_for_improved_21.txt	Radio interferometers measure data referred to as visibili- ties, computed using the visibility measurement equation (Thompson et al., 2017).
Processed_Optimising_MWA_EoR_data_processing_for_improved_21.txt	direction of the incoming signal respectively. Assuming coplanar antenna locations (w = 0) or a small ﬁeld 1 − l2 − m2 ≈ 1), the Van Cittert Zernike of view ( theorem relates the measurement equation to the 2D Fourier transform of the sky. However, multiple system- atics corrupt the signal in its propagation path, resulting into vmeas discrepant from the true sky visibilities vtrue. Calibration is the process that aims to correct for this discrepancy using a myriad of methods that revolve around minimising the diﬀerence between an estimate model of vtrue and vmeas to obtain correction factors, typically referred to as calibration gain solutions or sim- ply gains. See Smirnov 2011 and their associated works for a full description on calibration. The calibration step is crucial as it dictates not only the credibility of the subsequent PS estimation step but also how close the output PS upper limits are to the 21 cm signal level. Since the RTS calibration software is a key component of the analysis presented in this work, a distinction be- tween the direction-dependent (DD) and -independent (DI) calibration steps, as implemented by the RTS, is of importance to this work.
Processed_Optimising_MWA_EoR_data_processing_for_improved_21.txt	The total beam-attenuated visibilities accumulated from a pool of the brightest calibrator sources to be corrected for the ionosphere are ﬁrst subtracted from the data in a step referred to as prepeeling. Prepeel- ing allows for calculation of ionospheric gains in the direction of each source individually, without sidelobes confusion from other sky directions. In descending order of brightness, the sources are one by one re-added back into the residual visibilities and the catalogued position of the source is rotated to be at the phase centre of the observation. This rotation implies that for a source in its correct sky position in the data, the (l, m) values from equation 1 should be zero. Any oﬀsets for the source per frequency channel are then ﬁtted for the λ2 spectral signature of the ionosphere, and a model of ionospheric refraction on the source is obtained; see Equation 5 in Mitchell et al. 2008.The gains obtained for the source direction are primarily used in subtracting it from the data, and the process is repeatedly carried out for all the remaining calibrators.
Processed_Optimising_MWA_EoR_data_processing_for_improved_21.txt	The power spectrum, P (k), is the Fourier transform of the 2-point correlation function (Equation 2, e.g., Zaldarriaga et al. 2004) and it probes the ﬂuctuations of the 21 cm brightness temperature along the line of sight (frequency modes) as well as spatially (angular modes).
Processed_Optimising_MWA_EoR_data_processing_for_improved_21.txt	The PS obtained from the steps above is 3 dimen- sional, and with (u, v, η) dimensions where where η is the Fourier dual to frequency. It is common practice to compute its weighted averages in either cylindrical or spherical shells to obtain the 2D or 1D PS versions, respectively. The 2D PS has the distinct EoR window and foregrounds wedge morphology, with axis units of k⊥ = p(k2 v), kk ∝ η obtained from the two (u, v) baseline directions and the line-of-sight (frequency) com- ponents respectively. The occurrence of these features is due to the mode mixing phenomenon (Datta et al., 2010; Liu & Tegmark, 2011; Parsons et al., 2012; Vedantham et al., 2012; Morales et al., 2012). Mode mixing results from spectrally smooth foreground components that lie at small kk modes interacting with the instrument chro- maticity. As a result, mode mixing leaks power to higher kk as a function of k⊥ resulting in the foregrounds wedge feature. The EoR window remains as a region with lower foregrounds power and the most promising for detection of the 21 cm signal. The 2D PS is a crucial diagnostic in examining how diﬀerent aspects of the 21 cm experiment lead to varying power in both the wedge and the window modes. The 1D PS on the other hand is mostly used to draw the power limits and is usually presented as a dimensionless quantity by integrating the total power on a given spatial scale over the volume. A full description together with the conversions from k to cosmological dimensions can be found in Morales & Hewitt 2004; McQuinn et al. 2006.
Processed_Optimising_MWA_EoR_data_processing_for_improved_21.txt	In this work, we use both the 1D and 2D PS computed using the Cosmological HI PS Estimator (c h i p s, (Trott et al., 2016)) to describe our results. CHIPS is optimised for estimation of MWA data power spectra and has been used in previous literature (e.g. Trott et al. 2020; Rahimi et al. 2021; Yoshiura et al. 2021).
Processed_Optimising_MWA_EoR_data_processing_for_improved_21.txt	When including positional oﬀsets induced by the iono- sphere, an ionospheric diﬀerential phase oﬀset, φ0, term evaluated between two lines of sight through the iono- sphere from each antenna pair in the array is added to the φ term in equation 1. To model this eﬀect in mock MWA visibilities, we use the s i v i o package (Chege et al., 2021), which makes thin phase screen models of ionospheric refraction as well as MWA pa- rameters to compute the visibilities in Equation 1. We modelled phase screens with Kolmogorov (K) and sine- like (S) spatial structures. The K screen can be de- scribed by an isotropic 2-dimensional (2D) Gaussian random ﬁeld with a power law power spectrum of the form ∼ \|k\|11/3 where k is the 2D Fourier wavenumber (e.g., Vedantham & Koopmans, 2016). The S screen is described by a smooth-varying 2D sine function of the (cid:17) form γ × sin with (x, y) being the two axis of the screen, γ modifying the number of ‘ridges’ and (α, β) modulating their shape.
Processed_Optimising_MWA_EoR_data_processing_for_improved_21.txt	channels of 40 kHz width each between ∼ 140 - 170 MHz, a replica of real MWA low band data. The simulated data provides ionospheric eﬀects that are dominant over any other systematics therefore testable.
Processed_Optimising_MWA_EoR_data_processing_for_improved_21.txt	Figure 1. 2D power spectra for 5000 sources simulations with Kolmogorov-like ionospheric models. The rows indicate increasing ionospheric turbulence levels from top to bottom. For each panel, the same amount of sources is used for both DI and ionospheric calibration; indicated at the top of each column and increasing from left to right. A constant 1000 sources were subtracted from the calibrated visibilities in all panels. The black solid contour shows the 1010.5 mK2h−3Mpc3 power level. The EoR window is noise-dominated even for the best ionospheric conditions and calibration. For this reason, we draw conclusions of ionospheric impact using lower noise simulations shown in Figure 2. The dashed and dotted lines represent the ‘horizon’ and the ‘primary ﬁeld of view’ foreground power limits, based on source positions in the sky and the MWA primary beam, respectively.
Processed_Optimising_MWA_EoR_data_processing_for_improved_21.txt	bilities. Typical values for the MWA are Tsys = 240 K, Aeﬀ = 21 m2, ∆ν = 40 kHz and ∆t = 8 s.
Processed_Optimising_MWA_EoR_data_processing_for_improved_21.txt	As ionospheric eﬀects are best studied using com- pact sources, we do not add any large-scale sky model components such as the galactic diﬀuse emission to the visibilities. Since the EoR signal is relatively weak when compared to the above foregrounds, we leave it out of the visibility simulations.
Processed_Optimising_MWA_EoR_data_processing_for_improved_21.txt	Figure 2. 2D power spectra for 5000 sources simulations with Kolmogorov-like ionospheric models and scaled down thermal noise. The rows indicate increasing ionospheric turbulence levels from top to bottom. For each panel, the same amount of sources is used for both DI and ionospheric calibration; indicated at the top of each column and increasing from left to right. A constant 1000 sources were subtracted from the calibrated visibilities in all panels. The black solid contour shows the 1010.5 mK2h−3Mpc3 power level. Increasing the number of sky model sources in all calibration steps reduces power in all power spectrum modes. The most improvement is obtained when the ionosphere is most inactive.
Processed_Optimising_MWA_EoR_data_processing_for_improved_21.txt	Figure 1 shows a suite of 2-minute MWA simulations, calibrated and used to obtain the 2D PS as fully de- scribed in Section 2. The number of DI and ionospheric calibration sources used during calibration are equal for each run but increasing across the columns from 250 to 4000, while the ionospheric turbulence increases for each row from top to bottom. For all panels, a constant 1000 sources was subtracted after DD calibration. The third column is synonymous with the parameters (1000 sources for both DI and DD) used for calibration in the recent best limits results by (Trott et al., 2020). The ﬁrst two rows show ionospheric oﬀsets within 0.15 ar- cmin and therefore, in real data, they would be used for EoR power estimation as done in Yoshiura et al. 2021. To date, we have not established a maximum cutoﬀ of acceptable ionospheric contamination when aggregat- ing data for PS estimation. We expect that, in general, less ionospheric activity is better. However, we wish to better understand how poor ionospheric conditions are allowed to be before we discard data, to ﬁnd the optimum compromise.
Processed_Optimising_MWA_EoR_data_processing_for_improved_21.txt	eﬀects to manifest without many hours of averaged data, we decide to reduce the thermal noise in the simulations.
Processed_Optimising_MWA_EoR_data_processing_for_improved_21.txt	Figure 3. The log of the ratio between the 2D PS with and with- out RTS ionospheric correction for a k100 simulation and scaled down noise. The dominant blue colour indicates that ionospheric correction does indeed reduce contamination due to ionospheric activity.
Processed_Optimising_MWA_EoR_data_processing_for_improved_21.txt	We ﬁrst demonstrate the beneﬁts of using more sources in calibration by using simulations with a minimal ther- mal noise level. Figures 2 is similar to Figure 1 but with more individual runs included and with the noise in each observation scaled down by a factor of 104. The eﬀect of both the ionospheric activity as a function of the cali- brator sources is more evident. More DI and DD sources result in lower power levels for all ionospheric condi- tions. As we approach the top right panel, where the ionosphere is most inactive, and with maximum amount of calibration sources, less and less power bleeds to the EoR window. As expected, the bottom left represents the extreme opposite. For the most extreme ionosphere, the foreground subtraction results in structured residu- als around sources due to a highly distorted point-spread function (see Koopmans 2010). The structured artefacts, especially around the brightest sources, appear as the vertical high-power stripes visible in the two bottom rows of Figure 2.
Processed_Optimising_MWA_EoR_data_processing_for_improved_21.txt	Figure 4. EoR window 1D PS comparison of varying the number of sources in DI and DD with k-screen ionospherically contaminated visibilities. The simulation is composed of 5000 point sources with scaled down noise levels. Here, the DD value represents the number of sources that were both ionospherically corrected and subtracted. This ﬁgure shows how a combination of inactive ionospheric conditions, more complete DI calibration sky models and subtraction of more compact foregrounds results in reduced contamination in the EoR window.
Processed_Optimising_MWA_EoR_data_processing_for_improved_21.txt	from the best run (solid blue). The dotted lines further show that without DD calibration, a better DI model does not provide any signiﬁcant improvements in the power spectrum. Conversely, with the DI calibrators kept constant, more DD calibrators results in signiﬁcantly lower power.
Processed_Optimising_MWA_EoR_data_processing_for_improved_21.txt	The red solid line shows the run with 1000 DI sources and 4000 DD sources, while the blue dashed line repre- sents the reverse (4000 DI sources and 1000 DD sources). In almost all ionospheric cases, these two RTS conﬁgu- rations show similar power levels, but the former shows slightly less power in some modes. This indicates how increased calibrator sources are important in not just one step of calibration, but in both the DI and the DD for maximum improvements to be realised.
Processed_Optimising_MWA_EoR_data_processing_for_improved_21.txt	Higher ionospheric levels result in higher 1D power. For the most extreme ionospheres (last two plots, k200 and k300, note the diﬀerent y-axis range. Similarly ob- served for s200 and s300.), the calibration performance has become extremely poor and all the RTS DD conﬁgu- rations make no diﬀerence. This is also accompanied by an increased power spectrum slope, resulting from fail- ure by the RTS to accurately subtract extremely oﬀset sources.
Processed_Optimising_MWA_EoR_data_processing_for_improved_21.txt	Ideally, source subtraction should only reduce power in the wedge. However, a key observation from the simu- lations is that the RTS calibration process is not perfect; regardless of conﬁguration, power leaks into the EoR window from the wedge. By subtracting power from the wedge, we are by extension reducing the power that can leak to the window and helping to keep the window as contamination free as possible for 21 cm detection.
Processed_Optimising_MWA_EoR_data_processing_for_improved_21.txt	For the simulated data, no signiﬁcant diﬀerence in power is observed between the K and the S screens (see Figure B.2 for the S-screen analogue of Figure 4). The reason for this observation has not yet been established and will warrant more investigation in future. Similarly, in the other simulation experiments carried out in this work, no signiﬁcant result diﬀerences were observed from using either a K or S-screen.
Processed_Optimising_MWA_EoR_data_processing_for_improved_21.txt	Figure 5. Summary of 452 real data observations used in the analysis. The dashed line shows a typical data cut, where only observations with lower median oﬀsets are analysed further. A distribution of diﬀerent ionospheric conditions observed within a 1 hour LST interval (LST 0hrs) was targeted. The marker size shows increasing ionospheric QA metric; a linear combination of the two axes values. The PCA value signiﬁes spatial anisotropy in the ionosphere. The observations were carried out in 2014 and 2015.
Processed_Optimising_MWA_EoR_data_processing_for_improved_21.txt	Figure 6. A 2D PS for the best 306 MWA 2-minute observations. The dash-dotted line encloses the modes used for to obtain the 1D spherically averaged PS limits.
Processed_Optimising_MWA_EoR_data_processing_for_improved_21.txt	Table 2 A summary of diﬀerent RTS runs investigated. Each cell represents a calibration procedure performed over the whole dataset, the numbers are the amount of sources included in the sky model in the respective calibration step, e.g. ‘1000 DI’ means a direction independent run with a source catalogue comprising 1000 sources. DD here is used to imply the source subtraction step only, separate from the ionospheric correction step denoted by ‘ionocal’. In row A, only direction independent calibration was done. Row B has both DI and DD, but no ionocal step applied. Row C combines all DI, ionocal and DD, while row D performs an additional calibration iteration with the sky model updated using ionospheric oﬀsets obtained from the ﬁrst ionocal run. See sections 6 and 7 for the discussion motivating the bright sources selection criteria applied in column 3 as well as the results. The analysis strategy in this table was applied to the real data only, but not the simulations .
Processed_Optimising_MWA_EoR_data_processing_for_improved_21.txt	the simulated data. The data is observed in a shift and drift method where the MWA analogue beamformer is electronically pointed to a speciﬁc direction (‘pointing‘) and the sky is allowed to drift overhead for a given obser- vation duration. We use data from the zenith pointing and 2 non-zenith ones, labelled as pointings 2 and 4 (see Beardsley et al., 2016). We select data with a range of ionospheric activity levels, varying from calm to turbu- lent. However, variations in diﬀuse emission in the real data would dominate over ionospheric eﬀects, rendering them unquantiﬁable. We therefore choose observations that were observed within the same LST hour, assuming the variation in the EoR0 ﬁeld sky temperature over an hour to be negligible (∼ 315 K at 154 MHz). In total, we use a set of 452 individual datasets, altogether amounting to ∼ 14 hours of data.
Processed_Optimising_MWA_EoR_data_processing_for_improved_21.txt	in the ionosphere. For a 2-minute observation processed using the 14 calibration timestamps, the median oﬀset per source was used as the representative oﬀset value for each source. The median oﬀset over all calibrator sources is then used as the model oﬀset value for the observation. In each observation and taking only the brightest ∼ 800 sources, we used C t h u l h u software (Jordan et al., 2017), to extract the m and p parame- ters from the calibration outputs. We then categorised our dataset into 111, 102, 140 and 99 observations for types 1, 2, 3 and 4 respectively. Figure 5 summarises the dataset as a function of the ionospheric quality metrics.
Processed_Optimising_MWA_EoR_data_processing_for_improved_21.txt	Table 2 details the diﬀerent RTS calibration runs iter- ated over the dataset. The power spectrum is computed for each RTS conﬁguration integrated over the whole dataset, and where indicated, over each ionosphere type.
Processed_Optimising_MWA_EoR_data_processing_for_improved_21.txt	not to include the aﬀected data in the PS estimation, and only show results for the 306 observations with the best calibration solutions. Figure 6 shows the 2D PS for these 306 observations processed according to procedure A1. The ﬁgure shows the expected 2D PS morphology with the foregrounds power concentrated in the wedge while the window has signiﬁcantly less power. When comparing such 2D spectra from diﬀerent cells, say A1 and A2, we plot the log of the ratio of the 2 analogous to Figure 3 and with a similar interpretation.
Processed_Optimising_MWA_EoR_data_processing_for_improved_21.txt	The dash-dotted line encloses the k-modes within 3.5kk > k⊥, 0.01 < k⊥ < 0.04 and kk > 0.1 which are typically used to obtain the 1D spherically averaged PS limits. This region excludes the entire wedge, as well as avoiding the supra-horizon emission. Additionally, this is a region well sampled by MWA uv-coverage. In this work, we apply these exact power cuts as well to compute our real data 1D power spectra.
Processed_Optimising_MWA_EoR_data_processing_for_improved_21.txt	MWA PS results have been found to vary with dif- ferent sky pointings due to beam modelling errors as well as enhanced ionospheric eﬀects from non-zenith sky directions (e.g. Barry et al., 2019). Figure 8 shows runs C1 and C3 1D power spectra for the best 37 zenith point- ing observations that showed low ionospheric activity (Types 1 and 3 from Jordan et al. 2017). There is an overall factor of ∼ 2 improvement at the 0.1 hMpc−1 scales. This improvement was consistent at this level for the 3 sky pointings that comprised the analysed observations. We now describe the main factors that contributed to this improvement.
Processed_Optimising_MWA_EoR_data_processing_for_improved_21.txt	Figure 7. 2D PS for the best 37 zenith observations with low ionospheric activity, left: 4000 sky model sources in DI and DD calibration and sources with high SNR corrected for the ionosphere (C3), middle: 1000 sky model sources in DI, ionospheric, and DD calibration (C1), Right: Log of the ratio between ﬁrst two (log(C3/C1)). The bluer colour in the log-ratio plot shows less power in the C3, implying less calibration errors, which in turn reduces contamination in both the foregrounds wedge and the EoR window.
Processed_Optimising_MWA_EoR_data_processing_for_improved_21.txt	Figure 8. C1 and C3 1D PS for the window modes and zenith pointing data with minimal ionospheric activity. The dashed line corresponds to the thermal noise level.
Processed_Optimising_MWA_EoR_data_processing_for_improved_21.txt	advantageous as it reduces the power level that can leak into the EoR window as a result of any other systematic.
Processed_Optimising_MWA_EoR_data_processing_for_improved_21.txt	As found in the simulation results, the performance of RTS ionospheric correction deteriorates at extreme ionospheric activity. This can lead to inaccuracies during source subtraction and, potentially, signal loss. Such over- subtraction was found in residual images for observations with active ionosphere (types 2 and 4). No signiﬁcant over-subtraction was observed in low ionospheric activity (types 1 and 3, Figure A.1) images, except for sources with incorrect sky models.
Processed_Optimising_MWA_EoR_data_processing_for_improved_21.txt	Figure 9. The log of the ratio between the 2D PS for the best observations for a single sky pointing, processed according to cells A2 and A1. In both runs, DD calibration has not been applied. The bluer region in the lower window modes shows improvements from improved DI sky modelling.
Processed_Optimising_MWA_EoR_data_processing_for_improved_21.txt	Figure 10. The log of the ratio between the 2D PS for the best 306 MWA 2-minute observations from cells B3 and B2, log(B3/B2). The B3 wedge shows excess power in the small scales as compared to B2. This is because in B3, only ∼ 800 sources with > 1 Jy ﬂux density were subtracted, and they are less than the 4000 subtracted in B2. The window ratio is noise-like.
Processed_Optimising_MWA_EoR_data_processing_for_improved_21.txt	which is propagated to the RTS-derived ionospheric oﬀsets. Their inclusion in ionospheric calibration could be more detrimental than not using them at all. How faint we can allow a source to be in order for it to be included in the ionospheric calibration is a pertinent question. We investigate this eﬀect by comparing pairs of simulated 2-minute observations, identical in all aspects including the ionosphere except that one has no thermal noise included. The ionospheric S-screens were used with magnitude hyperparameters of 10, 25, 50 and 100 (maximum median oﬀset of 0.2 arcmin). Figure 12 shows how position errors evolve as sources get fainter for the s50 simulation, with the accurate oﬀsets being the RTS oﬀsets from the simulation with lower noise. Noise confusion results in higher RTS position errors for the faint sources. The RMS level from the central 10deg by 10deg region of a 2-minute snapshot image made from the visibilities with noise using robust 0 Briggs weighting was found to be ∼ 5 mJy/beam. This noise level is consistent with the predicted MWA snapshot thermal noise (e.g., Wayth et al. 2015). The observed trend in Figure 12 was consistent in the four turbulence levels used, and we found a maximum percentage error of between ∼ 5 − 20%.
Processed_Optimising_MWA_EoR_data_processing_for_improved_21.txt	Figure 11. The log of the ratio between the 2D PS for the best 306 MWA 2-minute observations from cells C2 and B2, log(C2/B2). The only diﬀerence in the two runs is that C2 has ionospheric correction while B2 has not. The C2 wedge shows less power as compared to B2, but despite this, the EoR window remains noise-like. We can conclude that the ionospheric correction is still indirectly advantageous, as it reduces the power level that can leak into the EoR window as a result of any other systematic, as shown in Figure 3.
Processed_Optimising_MWA_EoR_data_processing_for_improved_21.txt	threshold to categorise between bright and faint sources in the ionospheric correction of the real data runs in the ﬁnal column of table 2. This cutoﬀ is qualitative, chosen at the ‘elbow’ of the trend observed in the simulated oﬀset errors. We note however that the RTS does average over multiple channels during ionospheric correction, and the ∼ 1 Jy cutoﬀ would result in bright sources with suﬃcient signal-to-noise ratio regardless of the dominant noise cause.
Processed_Optimising_MWA_EoR_data_processing_for_improved_21.txt	We obtained the system temperature as the sum of the recorded sky temperature for the respective observation with ∼ 50K receiver temperature for the MWA (Tingay et al., 2013; Wayth et al., 2015). Similar to the simula- tion, the chosen ∼ 1 Jy ﬂux threshold corresponded to a ∼ 2σ SNR for a single channel at the centre of the fre- quency band (154 MHz) over 8s duration. This threshold resulted in ∼ 800 sources that we deemed bright enough to be corrected for the ionosphere without introducing errors.
Processed_Optimising_MWA_EoR_data_processing_for_improved_21.txt	Figure 12. Position oﬀset errors from the RTS ionospheric oﬀsets estimation as a function of source brightness. The larger dots represent sources with higher SNR. The position errors increase with lower SNR. The dashed line represents a qualitative ﬂux threshold chosen at the ‘elbow’ of the trend, for categorising ’faint’ and ’bright’ sources during calibration.
Processed_Optimising_MWA_EoR_data_processing_for_improved_21.txt	From the diﬀerent categories of ionospheric activity de- scribed, the real data did not show conclusive diﬀerences in the power spectrum. We attribute this to the presence of other dominant systematics e.g., Trott et al., 2020. The major sources of these systematics are errors in the models of the instrument and the sky (the diﬀuse component of the sky not being included in the calibra- tion model), as well as errors inherent to the calibration and power spectrum estimation algorithms. The eﬀect of active ionosphere is, however, already clear from the simulations. Trott et al. 2018 uses analytical models to show how ionospheric spatial structures would introduce biases in the power spectrum, for various reasons that make any direct comparison of real data runs from dif- ferent ionospheric categories unfair, a result similar to theirs is not observed in this work.
Processed_Optimising_MWA_EoR_data_processing_for_improved_21.txt	Figure 13. log(C3/C2). The noise-like window implies that we do not see signiﬁcant diﬀerences in the window based on the number of sources that ionospheric correction is applied to. The ionospheric correction to fewer sources in C3 is however apparent, signiﬁed by the apparent redness in the wedge.
Processed_Optimising_MWA_EoR_data_processing_for_improved_21.txt	more accurate description of the measured visibilities for that observation and should result in more accurate cal- ibration gain solutions. Yoshiura et al. 2021 performed this procedure for MWA ultra-low data, where the iono- spheric impact is much higher than at the MWA EoR low band. Here, we investigate whether this method is viable in the low band.
Processed_Optimising_MWA_EoR_data_processing_for_improved_21.txt	In order to perform such ionospheric correction proce- dures accurately, we need to ascertain that the observed source positional oﬀsets are predominantly ionospheric and not caused by other systematics. We test this by ob- taining a list of common sources in all the 452 datasets and investigating the distribution of the oﬀsets from each individual source. The oﬀset magnitude per source is assumed to be static over the 2 minutes dataset dura- tion, and is approximated by the median of the oﬀset magnitudes obtained for the 14 calibration timestamps. Ionospherically induced oﬀsets on a source caused by multiple random ionospheric conditions at diﬀerent times should be normally distributed around zero. Any other distribution centred at a diﬀerent value implies a sys- tematic error in the catalogued position of the sources. Figure 14 shows the distribution of oﬀsets obtained for the total 452 observations. Except for the outliers, there is no distinctive disparity of the source oﬀsets from the expected distribution, signifying lack of signiﬁcant systematics.
Processed_Optimising_MWA_EoR_data_processing_for_improved_21.txt	Figure 14. Oﬀsets distribution per source over all observations. Such a Gaussian-like distribution is expected for purely ionospheric oﬀsets.
Processed_Optimising_MWA_EoR_data_processing_for_improved_21.txt	during prior calibration. We obtained the scaling factors applied to the amplitude gains for each source over all the datasets. A scaling factor distant from unity would signify a systematic in the properties of the calibrator source. The amplitude scales showed a maximum spread of ∼ 20% around unity, implying no major systemat- ics in the source ﬂux densities are introduced by the ionospheric calibration.
Processed_Optimising_MWA_EoR_data_processing_for_improved_21.txt	After this correction, there were minimal improve- ments observed across all modes, but varying across the 3 sky pointings. The most improvement was seen in pointing 2 with the possible reason for this being that pointing 2 had more datasets as compared to the other 2 pointings and also has less galactic contami- nation (Beardsley et al., 2016). Figure 15 shows the log of the ratio between the 2D PS from cells D2 and C2; log(D2/C2) from the pointing 2 data. The ﬁgure shows that the sky model correction applied does indeed reduce ionospheric contamination, albeit at a minimal level, which at the current EoR calibration precision levels, can easily get dominated by other systematics.
Processed_Optimising_MWA_EoR_data_processing_for_improved_21.txt	To further investigate the minimal improvement ob- served, we present Figure 16 which shows the ratio of the source position oﬀsets and the gain amplitudes before and after applying sky model updates. The ﬁgure shows a clear positive correlation between ionospheric activity with both the gains amplitude and position oﬀsets. The updated sky model results in a factor of ∼ 1.2 to 3 (up to ∼ 67%) reduction in position errors. However, the maximum change in the amplitudes is less than 1%. This variation in the magnitude of the two eﬀects is direct evidence to the well known dominance of ionospheric ﬁrst order eﬀects (refraction phase errors) over visibili- ties amplitude scintillation; a higher order ionospheric eﬀect.
Processed_Optimising_MWA_EoR_data_processing_for_improved_21.txt	Figure 15. The log of the ratio between the 2D PS for cells D2 and C2, log(D2/C2) obtained for pointing 2 data. There is marginal improvement observed in the wedge and the window.
Processed_Optimising_MWA_EoR_data_processing_for_improved_21.txt	treatment would require an oﬀset-corrected sky model for every individual frequency channel, a requirement that is not feasible with the current tools.
Processed_Optimising_MWA_EoR_data_processing_for_improved_21.txt	Figure 16. Ratio of the source position oﬀsets and the gains amplitude before and after the sky model updates. The markers represent the 4 ionospheric categories, while the black dashed line shows the correlation trend with a 99% conﬁdence interval (grey shaded region). The trend shows a positive correlation between ionospheric activity and both gains amplitudes and position oﬀsets. Updating the sky model has a much higher impact on the oﬀsets (up to a factor of 3 reduction) as compared to the visibility amplitudes (< 1% reduction).
Processed_Optimising_MWA_EoR_data_processing_for_improved_21.txt	in green. The ﬁducial simulated ionospheric structure is well recovered, resulting in minimal residuals, and we conclude that this suﬃciently corrects for the spurious noise-confused oﬀsets obtained for the fainter sources by the calibration algorithm. However, no signiﬁcant improvement was observed on the PS as a result of this smoothing procedure when compared to the D2 and D3 runs.
Processed_Optimising_MWA_EoR_data_processing_for_improved_21.txt	To avoid errors associated with sources confused by the thermal noise level, we apply the optimum ﬂux den- sity threshold for the RTS ionospheric correction based on the recorded per observation parameters. However, the errors introduced by applying ionospheric corrections to low SNR sources are found not to be signiﬁcant. These errors might not observable as a result of the limited amount of data integrated, and future more sensitive integrations could unearth the errors. Nevertheless, this would be an especially important factor to consider for future deeper analyses.
Processed_Optimising_MWA_EoR_data_processing_for_improved_21.txt	The DI step alone provides the biggest improvement. This is in line with other literature in the ﬁeld, which have shown that all kinds of calibration (e.g., sky-based and redundant calibration) suﬀer from sky model in- completeness errors (Byrne et al., 2019; Barry et al., 2016; Patil et al., 2016). Foreground subtraction of more sky sources combined with increased ionospheric correc- tion in the DD step accounts for more contamination reduction.
Processed_Optimising_MWA_EoR_data_processing_for_improved_21.txt	Figure 17. Top: Ionospheric spatial structure interpolated using the diﬀerential RA (left) and Dec (right) oﬀsets for an EoR0 s i v i o simulation over the main lobe of the MWA. Bottom: The recovered RA and Dec oﬀset values for the interpolants with their residuals. The ﬁducial oﬀsets are the ones measured during DD calibration. Interpolation corrects spurious ionospheric gains obtained for low SNR sources during DD calibration.
Processed_Optimising_MWA_EoR_data_processing_for_improved_21.txt	the power spectrum of the 21 cm signal. They ﬁnd that correction of the ionosphere eﬀects both before source subtraction and afterwards in the residuals is key to getting rid of this bias.
Processed_Optimising_MWA_EoR_data_processing_for_improved_21.txt	Our additional ionospheric correction process does not account for the variation of the oﬀsets over the fre- quency band pass. This would require use of diﬀerent sky models per frequency channel, which is not sup- ported by the RTS. Actual development of the RTS capabilities is beyond the scope of this work. However, future calibration algorithms that aim to fully correct for the ionosphere while minimising spectral errors should take this process into account. Despite this, based on the marginal improvements observed by applying the additional ionospheric corrections, we conclude that a stringent calibration such as the one done in the C3 run is suﬃcient, and the ionosphere is not a showstopper for EoR science at frequencies above 100MHz. This is in agreement with previous literature (e.g. Vedantham & Koopmans, 2016).
Processed_Optimising_MWA_EoR_data_processing_for_improved_21.txt	The approach of calibration using select sources from composite catalogues of the target EoR ﬁeld of view is not unique to the MWA. Neither is the need for suﬃcient SNR from the number of sources used as well as the cor- rection of direction dependent errors while attempting to minimise computation costs, see e.g. Mertens et al. 2020; Patil et al. 2017; Yatawatta 2016 for similar en- deavours as applied to the LoFAR EoR experiment. This work adds to such eﬀorts targeting to achieve improved calibration for MWA EoR analysis with the currently available resources.
Processed_Optimising_MWA_EoR_data_processing_for_improved_21.txt	the recommended calibration strategy from this work for EoR limits in future will incorporate an end-to-end signal loss analysis.
Processed_Optimising_MWA_EoR_data_processing_for_improved_21.txt	Obtaining the best performing calibration routine us- ing the available RTS/CHIPS MWA pipeline was one of the main aims of this paper. After ﬁnding that the iterative calibration for ionospheric correction provides only marginal improvement to the PS in the frequency range of this work, we recommend the C3 and C2 cal- ibration runs as the current most optimum strategies. The enhancements in those runs result in a ∼ 2 factor improvement in the EoR window PS. Future work will involve applying this strategy to a larger MWA dataset for improved PS limits. This work not only provides an improved EoR calibration strategy, but also contributes to the need for end to end pipeline veriﬁcation, which is getting stronger as the EoR science community gets closer to detecting the signal.
Processed_Optimising_MWA_EoR_data_processing_for_improved_21.txt	Kariuki Chege thanks Siyanda Matika for valuable discus- sions during this work. This work was supported by the Cen- tre for All Sky Astrophysics in 3 Dimensions (ASTRO3D), an Australian Research Council Centre of Excellence, funded by grant CE170100013. CMT is supported by an ARC Fu- ture Fellowship through project number FT180100321. SY is supported by JSPS KAKENHI Grant No 21J00416 and JSPS Research Fellowships for Young Scientists. The Inter- national Centre for Radio Astronomy Research (ICRAR) is a Joint Venture of Curtin University and The University of Western Australia, funded by the Western Australian State government. This scientiﬁc work makes use of the Murchi- son Radio-astronomy Observatory, operated by CSIRO. We acknowledge the Wajarri Yamatji people as the traditional owners of the Observatory site. Support for the operation of the MWA is provided by the Australian Government (NCRIS), under a contract to Curtin University adminis- tered by Astronomy Australia Limited. We acknowledge the Pawsey Supercomputing Centre, which is supported by the Western Australian and Australian Government.
Processed_Optimising_MWA_EoR_data_processing_for_improved_21.txt	Deboer D. R., Parsons A. R., Aguirre J. E., Alexander P., Ali Z. S., 2017, Publ. Astron. Soc. Paciﬁc, 129 Fan X., Carilli C. L., Keating B., 2006, Annu. Rev.
Processed_Optimising_MWA_EoR_data_processing_for_improved_21.txt	Over-subtraction of sources due to inaccurate sky models or inaccurate ionospheric source oﬀset corrections can lead to power loss. Figure A.1 shows the central region of residual images made from visibilities calibrated using procedures C1 (A.1a) and C3 (A.1b). Both images are made from the same 2-minutes observation carried out on 14th November 2014 and using the same imaging parameters. No over-subtraction is observed in both images.
Processed_Optimising_MWA_EoR_data_processing_for_improved_21.txt	quadrant of both images, and the new improved source catalogue by (Lynch et al., 2021) will enable more accu- rate subtraction of sources.
Processed_Optimising_MWA_EoR_data_processing_for_improved_21.txt	Figure A.1. C1 (top) and C3 (bottom) residual images after subtraction of 1000 and 4000 sources respectively.
Processed_A_formula_for_Euler_characteristic_of_line_singula.txt	Abstract. We prove an algebraic formula for the Euler characteristic of the Milnor ﬁbres of functions with critical locus a smooth curve on a space which is a weighted homogeneous complete intersection with isolated singularity.
Processed_A_formula_for_Euler_characteristic_of_line_singula.txt	The question we are interested in is that if there is a way to express χ(F ) by some “computable” invariants determined only by (f, Σ, X). When X is Cm, the singular locus Σ of f is a one dimensional complete intersection with isolated singularity, and the transversal singularity type of f along Σ is Morse, Pellikaan [9] answered this question positively. Pellikaan’s formula expresses the Euler characteristic in terms of the Jacobian number j(f ), δ and the Milnor number µ(Σ) of Σ. These numbers can be computed directly by counting the dimensions of certain ﬁnite dimensional vector spaces. The development of computer algebra makes this kind of algebraic In this article we answer the formulae more and more important and popular. question by proving a similar formula for function germs with line singularities on a weighted homogeneous space X with isolated complete intersection singularity (see Proposition 7). Remark that for a function germ f with isolated singularity on a weighted homogeneous complete intersection with isolated singularity, Bruce and Robert [1] have proved an algebraic formula for the Milnor number of f .
Processed_A_formula_for_Euler_characteristic_of_line_singula.txt	Let (X, 0) ⊂ (Cm, 0) be the germ of a reduced analytic subspace X of Cm deﬁned by an ideal h of O, generated by h1, . . . , hp ∈ O. Let g be the ideal generated by g1, . . . , gn ∈ O. The germ of the analytic space deﬁned by g at 0 is denoted by (Σ, 0). Write OX := O/h and OΣ := O/g.
Processed_A_formula_for_Euler_characteristic_of_line_singula.txt	Let Der (O) denote the O-module of germs of analytic vector ﬁelds on Cm at 0. Then Der (O) is a free O-module with ∂ as basis, where z1, . . . , zm are the local coordinates of (Cm, 0). Der (O) is a Lie algebra with the bracket deﬁned by [ξ, η] := ξη − ηξ for all ξ, η ∈ Der.
Processed_A_formula_for_Euler_characteristic_of_line_singula.txt	Deﬁne Der h(O) := {ξ ∈ Der (O) \| ξ(h) ⊂ h}, which is the O-module of loga- rithmic vector ﬁelds along (X, 0) and a Lie subalgebra of Der (O) [10, 1]. When h is a radical ideal deﬁning the analytic space X, Der h(O) is often denoted by DX. Geometrically, DX consists of all the germs of vector ﬁelds that are tangent to the smooth part of X. When X is a weighted homogeneous complete intersection with isolated singularity, one can write down precisely all the generators of DX (cf. [13]). For f ∈ O, the ideal JX(f ) := {ξ(f ) \| ξ ∈ DX} is called the (relative) Jacobian ideal of f . Obviously, when X is the whole space Cm, namely, h = {0}, then JX(f ) = J(f ), the Jacobian ideal of f .
Processed_A_formula_for_Euler_characteristic_of_line_singula.txt	Let S = {Sα} be an analytic stratiﬁcation of X, f : (X, 0) −→ (C, 0) an analytic function germ. The critical locus ΣS f of f relative to the stratiﬁcation S is the union of the critical loci of f restricted to each of the strata Sα, namely, ΣS Sα∈S Σf \|Sα. We denote ΣS f is not positive, we say that f deﬁnes (or has) isolated singularities on X. If the dimension of ΣS f is positive, we say that f deﬁnes (or has) non-isolated singularities on X. If ΣS f is one dimensional smooth complex manifold, we say that f deﬁnes (or has) a line singularity on X.
Processed_A_formula_for_Euler_characteristic_of_line_singula.txt	logarithmic stratiﬁcation Slog := {Sα} of U (see [10, 1]). locally ﬁnite. If Slog is locally ﬁnite, then X is said to be holonomic.
Processed_A_formula_for_Euler_characteristic_of_line_singula.txt	log = {X ∩ Sα \| Sα ∈ Slog} is a stratiﬁcation of X which will be called the logarithmic stratiﬁcation of X in this article. Especially, when X has isolated singularity in 0, then {0} and the connected components of X \ {0} form a holonomic logarithmic stratiﬁcation of X. So 0 is always a critical point of any germ f : (X, 0) −→ (C, 0) relative to this stratiﬁcation. Hence for f ∈ m, Σf = {p ∈ X \| ξ(f )(p) = 0 for all ξ ∈ DX} is the critical locus of f relative to the logarithmic stratiﬁcation. Obviously Σf = X ∩ V(JX (f )).
Processed_A_formula_for_Euler_characteristic_of_line_singula.txt	g := {f ∈ g \| ξ(f ) ⊂ g for all ξ ∈ Derh(O)}.
Processed_A_formula_for_Euler_characteristic_of_line_singula.txt	X . Denote X(f ) = {ξ(f ) \| ξ ∈ D0 If f is clear from the context we just write J 0 and J 1.
Processed_A_formula_for_Euler_characteristic_of_line_singula.txt	Proof. The inequality means that Σ ∩ V((J 1 + h) : J 0) = {0}. For ǫ > 0 small enough, let P ∈ Σ ∩ {z0 = t}, 0 < \|t\| < ǫ. P /∈ V((J 1 + h) : J 0) if and only if (J 0)P ⊂ (J 1)P , where (J i)P is the localization of the ideal at P . Since z0 = 0 intersects both Σ and X transversally and X is smooth at P , we can choose the local coordinates such that locally the branch of Σ containing P is the ﬁrst axis. Furthermore, we can arrange the coordinate transformation such that under this transformation D0 X are preserved. By this we mean that any derivation of X with the ﬁrst component non-zero at P will remain non-zero at P and any derivation of X with ﬁrst component zero at P will remain zero at P . We let x = z0, y1, . . . , yn−p be the new local coordinates in a neighborhood of P in X, then ∂f at P , (J 0)P = OX,P . By [9], the inequality ∂y1 , . . . , (cid:3) is equivalent to the constancy of the transversal singularity type of f .
Processed_A_formula_for_Euler_characteristic_of_line_singula.txt	On the other hand, there is an integer n1 >> 0 such that when k ≥ n1 (see (3.2)) xkg ⊂ (h + JX(fk)) ∩ g ⊂ ˜J ⊂ g.
Processed_A_formula_for_Euler_characteristic_of_line_singula.txt	Remark 8. Formula (4.2) allows us to use a computer program to compute the Euler characteristic eﬀectively. In fact, we have a small Singular [2] program to calculate χ(F ) for a function with critical locus a line on a hypersurface X. We use it to check the examples in Example 10.
Processed_A_formula_for_Euler_characteristic_of_line_singula.txt	ay2 + byz + cz2, for example, take g = y2 − yz + 1 µ(g) = 6l − 3.
Processed_A_formula_for_Euler_characteristic_of_line_singula.txt	if s = 0, if 1 ≤ s ≤ l − 1, if s ≥ l.
Processed_Bayesian_calibration_of_simulation_models:_A_tutor.txt	Simulation models of epidemiological, biological, ecological, and environmental processes are increasingly being calibrated using Bayesian statistics. The Bayesian approach provides simple rules to synthesise multiple data sources and to calculate uncertainty in model output due to uncertainty in the calibration data. As the number of tutorials and studies published grow, the solutions to common diﬃculties in Bayesian calibration across these ﬁelds have become more apparent, and a step-by- step process for successful calibration across all these ﬁelds is emerging. We provide a statement of the key steps in a Bayesian calibration, and we outline analyses and approaches to each step that have emerged from one or more of these applied sciences. Thus we present a synthesis of Bayesian calibration methodologies that cut across a number of scientiﬁc disciplines. To demonstrate these steps and to provide further detail on the computations involved in Bayesian calibration, we calibrated a compartmental model of tobacco smoking behaviour in Australia. We found that the proportion of a birth cohort estimated to take up smoking before they reach age 20 years in 2016 was at its lowest value since the early 20th century, and that quit rates were at their highest. As a novel outcome, we quantiﬁed the rate that ex-smokers switched to reporting as a ‘never smoker’ when surveyed later in life; a phenomenon that, to our knowledge, has never been quantiﬁed using cross-sectional survey data.
Processed_Bayesian_calibration_of_simulation_models:_A_tutor.txt	There are many approaches to the calibration of epidemiological, biological, environmental, or ecological simulation models. Where data are abundant and model-parameters are identiﬁable with such data, or where uncertainty is not of interest, almost any calibration framework will provide an acceptable point estimate. What, then, to make of situations where there are less data, and signiﬁcant uncertainty exists that is important either as a research question itself or in decision-making? Consider a government health department using a model of a screening program to inform a decision on screening frequency; if the estimate of harm at a given screening frequency is highly uncertain, a policy-maker may prefer a more conservative frequency depending on the (modelled) marginal eﬀect on disease-burden. Statement and minimisation of uncertainty is vital in this and similar contexts, and a Bayesian approach to calibration provides simple rules to quantify uncertainty [1] and reduce uncertainty by incorporating all available data [2, 3, 4].
Processed_Bayesian_calibration_of_simulation_models:_A_tutor.txt	We are not the ﬁrst to write a tutorial on Bayesian model calibration and its beneﬁts over conventional ﬁtting procedure-like methods. Past tutorials reveal its widespread use, and have collectively described both generic-model uncertainty analysis [1], and model calibration in health policy [2, 3] and systems biology [5]. These ﬁelds beneﬁt from the ability to synthesise disparate data sources, using graphical models to both map out the dependence structure of the data [2, 6] and to identify a procedure to sample (posterior) values of parameters via languages such as BUGS [7]. Analysis of identiﬁability of parameters has emerged as a theme in successful calibration [8, 9], and Bayesian- speciﬁc approaches to such analyses have emerged [10]. Furthermore, Bayesian calibration may succeed where other approaches fail through greater use of prior information.
Processed_Bayesian_calibration_of_simulation_models:_A_tutor.txt	The aim of this tutorial was to strengthen connections between existing knowledge on Bayesian calibration from systems biology, epidemiology, ecology, infectious disease modelling, and health economic evaluation. In this tutorial we outline the common steps and tools used across these ﬁelds in successful Bayesian calibration of models and we discuss how to use the results of a calibration to make statements of uncertainty.
Processed_Bayesian_calibration_of_simulation_models:_A_tutor.txt	factor from epidemiology, governed by a class of equations more familiar to systems biologists (where they have been used to model breast cancer cell growth [12]) can be calibrated using Bayesian statistics. Using the calibrated model, we described the value and uncertainty of the model’s; proportion that initiated smoking in birth cohorts born between 1910 and 1996; rate of quitting smoking between 1930 and 2016, and; rate that ex-smokers switched to reporting as a never smoker when surveyed later in life. This ﬁnal quantity has never been quantiﬁed using cross-sectional smoking survey data.
Processed_Bayesian_calibration_of_simulation_models:_A_tutor.txt	A mathematical simulation model, which we refer to simply as a ‘model’, is an abstract representation of a process (or system) in which a parameterised set of equations transform inputs into outputs or predictions. Model calibration is a procedure by which values of the parameters are selected. Calibration involves taking measurements of inputs and outputs from studies of the process and then computationally searching for parameter values for which the model predictions closely match the observations.
Processed_Bayesian_calibration_of_simulation_models:_A_tutor.txt	In a Bayesian calibration, both the parameters and model predictions are deﬁned as random variables. Calibration is reframed as the calculation of the posterior distributions of these random variables given the observed data. The posterior distribution, or simply ‘the posterior’, describes the uncertainty owing to the randomness of the collected data. The knowledge of the parameters prior to data collection is described by the prior distribution, or simply, ‘the prior’.
Processed_Bayesian_calibration_of_simulation_models:_A_tutor.txt	Capital letters, such as Θ and X, represent real-valued random variables, speciﬁc values are denoted with lower case letters such as θ and x.
Processed_Bayesian_calibration_of_simulation_models:_A_tutor.txt	The probability density (or mass) function of a random variable is denoted as fΘ(θ). • The random variable Θ conditional on X is denoted Θ\|X, with density (or mass) denoted as fΘ\|X (θ\|x).
Processed_Bayesian_calibration_of_simulation_models:_A_tutor.txt	A function’s argument may be suppressed when it is obvious, e.g., fΘ instead of fΘ(θ).
Processed_Bayesian_calibration_of_simulation_models:_A_tutor.txt	These steps are not prescriptive; however, this structure may reduce a modeller’s workload by providing both early warning of failure and potential remedies. Next, we will discuss each of these steps in more detail.
Processed_Bayesian_calibration_of_simulation_models:_A_tutor.txt	Our ﬁrst step in any calibration, Bayesian or otherwise [13], is to determine for which parameters will values be estimated. For this we should investigate structural identiﬁability, which is a property of the model, some ideal input and output, and the parameters. If a set of parameters is structurally identiﬁable, then one and only one distribution of the parameter-values is allowed for any input and output of the given ideal form. We loosely paraphrase this by saying structural identiﬁability means the set of parameters can be uniquely determined by the ideal input and output (more formal discussion of identiﬁability and the related concept of parameter redundancy can be found in Cole 2020 [14]). Structural identiﬁability is a necessary (but insuﬃcient) condition for a more important characteristic called practical identiﬁability, i.e. that the set of parameters can be uniquely determined by the observed calibration data.
Processed_Bayesian_calibration_of_simulation_models:_A_tutor.txt	Structurally non-identiﬁable parameters cannot be calibrated with data only from observed output of the modelled process; other data sources are necessary. For example, if the cell division rate is a parameter of the model, but it is not identiﬁable, then some other experimental study, direct measurement, published estimate, or expert-elicited estimate is required to calibrate the model.
Processed_Bayesian_calibration_of_simulation_models:_A_tutor.txt	Formal methods to determine structural identiﬁability have been developed in ecology [14] and systems biology [15, 16], although uptake is limited in the latter ﬁeld [17] and in ﬁelds that use similar mathematical models such as infectious disease modelling and epidemiology [8]. Nonetheless, any investigation of structural identiﬁability can guide us towards which additional data will be required other than observations of the process. A priori practical identiﬁability is diﬃcult to establish [8] and, as such, practical identiﬁability is assessed in a later step with observed calibration data and sometimes using simulated data.
Processed_Bayesian_calibration_of_simulation_models:_A_tutor.txt	We assign a prior, fΘ, to the parameters which reﬂects the available prior knowledge about their value. Priors can be either ‘informative’, in which case they quantify established certainty that the parameter takes one value over another, or ‘non-informative’, where the intention is that the corresponding posterior reﬂects only the data used in calibration. Informative priors can be estimated from literature, taken directly from a previous analysis, or the result of an elicitation procedure, and must be obtained independently of the calibration data. Non-informative priors are designed to fulﬁl a chosen set of conditions or optimise some abstract criteria, for example the Jeﬀreys prior always satisﬁes a criterion of invariance under any change of variables.
Processed_Bayesian_calibration_of_simulation_models:_A_tutor.txt	may be diﬃcult to the point of impracticality. While a proper posterior may guarantee convergence, seemingly-converged output from the sampling algorithms we discuss below is no guarantee that the posterior is proper [20]. Any evidence of lack of convergence is, at the least, discouraging, and we discuss this further in the sampling step.
Processed_Bayesian_calibration_of_simulation_models:_A_tutor.txt	In conventional calibrations the observation processes by which the calibration data were collected may not elicit much modelling interest. These processes can include experiments, surveys, a Census, scraping from websites or apps, or some other type of study. In Bayesian calibration we assume each observation process is a random process, and the collected data are viewed as one outcome. With few exceptions, such as ‘Approximate Bayesian Computation’ [21], we must construct a statistical model of the observation process. This requirement is not exclusive to Bayesian calibration; the likelihood of the observed data, given a statistical model thereof, is one option for the objective function in ﬁtting procedure-based calibrations.
Processed_Bayesian_calibration_of_simulation_models:_A_tutor.txt	where the subscript i refers to the ith component of a vector, and n is a vector of the number of trials from the observation process. The right-hand side is the product of the likelihood of independent Binomial trials (given our model M ).

Subsets and Splits

No community queries yet

The top public SQL queries from the community will appear here once available.