LA-UR-25-31241 Approved for public release; distribution is unlimited. Title: A New Probability Table Methodology with Covariances on Probability Tables Author(s): Fritsch, William Nicholas Neudecker, Denise Walton, Noah Anthony Wy Intended for: Report Issued: 2025-11-14 Los Alamos National Laboratory, an affirmative action/equal opportunity employer, is operated by Triad National Security, LLC for the National Nuclear Security Administration of U.S. Department of Energy under contract 89233218CNA000001. By approving this article, the publisher recognizes that the U.S. Government retains nonexclusive, royalty-free license to publish or reproduce the published form of this contribution, or to allow others to do so, for U.S. Government purposes. Los Alamos National Laboratory requests that the publisher identify this article as work performed under the auspices of the U.S. Department of Energy. Los Alamos National Laboratory strongly supports academic freedom and a researcher's right to publish; as an institution, however, the Laboratory does not endorse the viewpoint of a publication or guarantee its technical correctness. A New Probability Table Methodology with Covariances on Probability Tables William Fritsch, Denise Neudecker, Noah Walton October 2025 Abstract Through this work, we test a new methodology of evaluating proba- bility tables informed by a combination of physics models and experimen- tal data. This methodology uses generalized least squares to produce a joint mean and covariance for average cross sections and probability ta- bles. This methodology has been tested and validated using total Monte Carlo. Attempts were made to parameterize the probability table us- ing a formulation provided in [1]. Although the normal inverse Gaussian fit was successful for the total cross section probability distribution, the rational-function fit to the capture, inelastic, and fission cross sections was unsuccessful due to time constraints. 1 Introduction Traditionally, probability tables are generated using NJOY [2] using only aver- age cross sections are informed by experiments. However, we wish to expand this methodology to include high-fidelity physics models that provide average cross sections and probability tables (PTs) into the evaluation. Additionally, we wish to produce evaluated probability tables with associated covariances. The process to generate evaluated average cross sections, PTs, and associ- ated covariances is shown in the flow diagram in Fig. 1: first, the basic method- ology will take model average cross sections from CoH3 [3], and generate proba- bility tables from the CoH3 average cross sections, using a high-fidelity sampler with resonance statistics. CoH3 is a Hauser-Feshbach physics model that pre- dicts the magnitude and trend of average cross sections. Next, the average cross sections and probability tables will be updated with cross section measurement via generalized least squares (GLS). The output will be a set of probability tables with associated covariance matrices. At the time of this report, there is no ENDF-6 [4] format for probability tables nor covariance matrices because they are usually handled after evaluation within the processing codes such as NJOY [2]. This work is a part of a larger NEUP project which will use this high-fidelity probability table generator to generate probability tables for OpenMC transport [5]. The probability table uncertainty will be propagated to system observables in a benchmark example. 1 Figure 1: The methodology for probability table evaluation, expanding upon the normal methodology. 2 Methodology 2.1 Generalized Least Squares Methodology The prior model data is provided in 500 random samples of average cross sections and PT parameters across energy. First, we will combine all of the data within a sample into a vector, ysamp . This includes the average cross sections for each reaction, and probability table parameters, such as bin edges, bin probabilities, and cross sections for each reaction in those bins.  tot el fis bin 1 T ⃗ysamp = ⟨σsamp ⟩(E1 ) ⟨σsamp ⟩(E1 ) ... ⟨σsamp ⟩(E1 ) Psamp (E1 ) ... The sample mean and covariance matrix can be calculated from ⃗ysamp as follows: 1 X ŷprior = ⃗ysamp (1) Nsamp samp 1 X T Vy,prior = (⃗ysamp − ŷprior ) (⃗ysamp − ŷprior ) (2) Nsamp − 1 samp Let ẑdata and Vz,data represent the mean and covariance matrix of the experi- mental data that will be used to update the model prior. Let S be the sensi- tivity matrix that maps z to y, where ẑprior = Sŷprior and ẑpost = Sŷpost . The remapping, S, is needed to 1) select the quantities in ŷ that appear in ẑ and 2) interpolate the quantities of interest at the ŷ energies to the ẑ energies. The generalized least squares equations are the following:  −1 −1 Vy,post = I + Vy,prior ST Vz,data S Vy,prior (3)  −1   −1 −1 ŷpost = I + Vy,prior ST Vz,data S ŷprior + Vy,prior ST Vz,data ẑdata (4) The GLS equations are written specifically such that Vprior is never directly inverted, since it is not full rank. The non-full rank status is due to equalities such as: 2 • The bin probabilities must sum to 1. • The average reaction cross sections must sum to the total average cross section. • The reaction cross section bins must sum up to the total cross section bins. Numerical instabilities occur if Vdata is not strictly positive definite, so a diag- onal “regularization uncertainty” may need to be added for stability. Typically, this is about 1% of the total uncertainty so that no correlation exceeds 99%. If there are multiple statistically independent datasets, the mean and covari- ance for each experiment, given by j, can be separated. The GLS equations can be reduced to the following: X  −1 Q = I + Vy,prior STj Vz,data,j Sj (5) j Vy,post = Q−1 Vy,prior (6)   X  −1 ŷpost = Q−1 ŷprior + Vy,prior STj Vz,data,j ẑdata,j  (7) j In practice, the sensitivity matrix, S, can be separated into two parts: a selection matrix that selects only the quantity or quantities measured in the experimental data, and a design matrix for interpolating from the prior energy grid to the data energy grid. For the design matrix, linear interpolation is used to put the model prior on the same grid as the experimental data. Experimental data at energies outside the bounds of the model (2.5 keV to 30 keV) are ignored. 2.1.1 Peelle’s Pertinent Puzzle and IRLS The uncertainties provided in the field of nuclear data are typically in the form of relative uncertainties, i.e., relative to the true quantity (average cross sections in this case). Of course, the true quantities are unknown and GLS is used to estimate such quantities. A naive approach where the uncertainties are set rela- tive to the data results in biased results, commonly known as Peelle’s Pertinent Puzzle [6]. The best approach is to iteratively update the experimental covari- ance matrix with the best estimate of the mean until convergence is reached; this process is called iteratively reweighted least squares (IRLS) [7]. The data covariance matrices are updated as follows: (i+1) (i+1) ! !  (i+1)   (i)  ẑpost ẑpost Vz,data = Vz,data (i) (i) (8) jk jk ẑpost j ẑpost k (i+1) (i+1) ! !  (i)  Sŷpost Sŷpost = Vz,data (i) (i) (9) jk Sŷpost j Sŷpost k 3 (0) (0) We start with Vz,data evaluated relative to the experimental data with ẑpost = ẑdata . ŷpost is recalculated iteratively along with equation (8) until ŷpost con- verges. 2.2 Total Monte Carlo Methodology Total Monte Carlo (TMC) will be used to validate the GLS approach. TMC is not limited by the linearity and normality constraints that are assumed in the GLS approach, making it the more accurate approach. In TMC, the model sam- ples are each weighted by likelihoods calculated from the data. The likelihoods can be calculated using the following:   1 T −1 L(ẑdata , Vz,data |⃗ysamp ) ∝ exp − (S⃗ysamp − ẑdata ) Vdata (S⃗ysamp − ẑdata ) 2 (10) Weights are calculated from the likelihoods for each sample, normalized to unity. L(ẑdata , Vz,data |⃗ysamp ) wsamp = P (11) samp′ L(ẑdata , Vz,data |⃗ysamp′ ) The sample mean and covariance matrix can be calculated statistically using the weights: X ŷpost = wsamp · ⃗ysamp (12) samp ESS X T Vy,post = wsamp (⃗ysamp − ŷpost ) (⃗ysamp − ŷpost ) (13) ESS − 1 samp The left term on the right side of equation (13) inflates the uncertainty to account for bias from an insufficient number of samples. ESS is the effective sample size, calculated with equation (14). The effective sample size is a useful metric for convergence of TMC results. P 2 samp wsamp ESS = P 2 (14) samp wsamp Like GLS, IRLS is applied to the TMC algorithm to combat the effects of Peelle’s Pertinent Puzzle. It should be noted that TMC is limited by both the discrepancy between the prior and data, and the dimensionality of the data. If either the discrepancy magnitude or data size is large, the effective sample size decreases, and the uncertainty inflates. This is why TMC is considered a secondary validation approach to GLS. 4 2.3 Fitting to a Functional Parameterization Gavin Ridley introduced a parameterized, functional form for probability ta- bles [1] as a means of accelerating probability table sampling which is slow and memory-intensive. We will fit our evaluated probability tables with the func- tional parameterization 2.3.1 Normal Inverse Gaussian Fit The total cross section probability distribution was fit with a Normal Inverse Gaussian (NIG) distribution (equation (15)), which has four parameters, α, β, δ, and µ. K1 represents the first order Bessel function.  p  αδK1 α δ 2 + (x − µ)2 NIG(x) = p exp (δγ + β(x − µ)) (15) π δ 2 + (x − µ)2 The probability density function for the NIG distribution is given by the function, PNIG . The loss function for NIG fitting is given by equation (16). The loss function is similar to the one provided in the thesis, except an additional term on the right of the equation to ensure the mean of the NIG distribution fit is the same as the reported average total cross section. This term has a large (103 ) coefficient, c, to force the consistency in the average total cross section. Phist is the histogram PDF from the tabular form of the PT. Phist is defined in equation (17), where {pj }nj=1bins are the probabilities for each bin, {σt,j }nj=1 bins are edge nbins +1 the binned total cross sections, and {σj }j=1 are the bin edges. Z  2 1/2 1 p p LOSS = √ Phist (σt ) − PNIG (σt ) dσt + c(E[σt |PNIG ] − σ̄t )2 2 (16)  edge 0  σt < σt,1 p /(σ edge − σ edge ) σ edge ≤ σ < σ edge  1 t,2 t,1 t,1 t t,2 Phist (σt ) = (17) ...   edge 0 σt ≥ σt,n  bins Fitting to the probability density histogram function as we are doing here is ap- propriate when the PTs are generated with many resonances. However, when there is a lack of resonances and the PT histogram is not fully converged. There- fore, perhaps a fit to the empirical CDF would be more appropriate when the PT distribution is not fully converged. The empirical CDF averages much of the statistical variation seen in the PT histogram. See Figure 6 in the results for further explanation. We found the NIG distribution difficult to fit consistently due to the limita- tions of the canned scipy optimizers. Hence, a few constraints had to be made to stabilize the fit using SciPy [8] minimize. The chosen optimization method was Nelder-Mead. First, since |β| < |α|, we define β̃ = β/α as the replacement optimization parameter for β and define limits [10−3 , 0.999]. Strictly speaking, 5 β does not have to be greater than zero, but as discussed in Gavin Ridley’s thesis [1], this condition covers the shape of the probability tables. α and δ were given bounds of [10−3 , 104 ]. µ was given bounds between the minimum and maximum bin edges for the total cross section. The initial parameters were derived to minimize the squared error in the difference between the mean, vari- ance, skewness, and kurtosis of the probability table distribution and the NIG distribution. 2.3.2 Barycentric Rational Fit The independent reaction cross sections, excluding elastic (capture, fission, and inelastic) are defined by a rational function fit with respect to the total cross section. The elastic cross section is inferred from sum rule with the other reac- tions. Unfortunately, due to time constraints at the end of the internship, a stable optimization routine for the rational function fit was not discovered. However, we will discuss the methodology that was taken as an approach. Let σx signify the reaction cross section (capture, inelastic, or fission in this case). We define the functional form of the fit below as the same as in the thesis [1]: P  m wj fj,x j=1 σ −z σxBRF (σt ) =  Pm tw j  σt j (18) j=1 σt −zj Consistent with the work of Gavin Ridley, the node positions, {zj }m j=1 , and weights, {wj }mj=1 are the same for all reactions. Gavin Ridley took m = 5, but we found it to result in underfitting of the low-cross section tail, so instead we take m = 6 which is more consistent with the distributions present. For initial parameters, {wj }m j j=1 have been set to alternate such that wj = (−1) , {zj }j=1 m has been defined such that the values are equal to total cross sections that are linearly inter-spaced from 0 to 1 in cumulative probability, and {fj,x }m j=1 has been taken as the interpolated value of the tabular cross section ratio. {wj }m j=1 are bounded from -1 to 1, {zj }m j=1 are bounded between the 0 and the maximum bin edge, and {fj,x }m j=1 are bound between 0 and 1. The loss function was defined to be similar as in the thesis, but with some changes: 1. The bin probabilities, pj , are not squared in the following loss function, unlike as written in the thesis. The thesis states that squaring the proba- bility provides a better shape, but we find it neglects the low-cross section behavior if pj is squared. 2. An additional term to force the node positions, {zj }m j=1 , to space out instead of clumping up. Clumping results in sharp fluctuations that are not represented in the tabular fit. The coefficient, c1 , is given a nominal value of 10−4 . 6 3. An additional term with a large (103 ) coefficient, c2 , ensures that the aver- age reaction cross section from the rational fit equals the reported average cross section, σ̄x . This part is the same as with the NIG distribution loss function. We define nbins as the number of bins in the probability table, pj as the proba- bility for those bins, and σxj and σtj as the reaction and total cross sections for those bins. nX bins  2 LOSS = pj σxj − σx (σtj ) (19) j=1 m−1 X 1 +c1 R zj+1 j=1 | zj P hist (z)dz| +c2 (E[σxBRF (σt )|Phist ] − σ̄x )2 As mentioned before, the quality we could reach for the barycentric rational fit was very much limited by time constraints at the end of the internship. These constraints are due to 1) unpredictable vertical asymptotes in the fitting function and 2) inability to fit the lower cross section parts of the reaction/total cross section curves. Perhaps future work will make the barycentric rational fit more reliable. 2.3.3 Monte Carlo Sampling for Mean and Covariance To fit these new distributions with provided uncertainty, a Monte Carlo ap- proach is taken. 20 Gaussian samples are taken from the mean and covariance for the tabular PT parameters. For each sample, the tabular PT was fit with the NIG and barycentric rational functions. Then, the mean and covariance for the new parameters is calculated. 3 Preparing the Data 3.1 Model Prior Data The prior model data was provided by Kazuki Fujio. The average cross sections were sampled using the CoH3 model [3]. Paired with the average cross sections, probability tables (PTs) were sampled using Gaussian Orthogonal Ensemble resonance sampling. The probability tables were generated by binning the sampled cross section realizations. If the total cross section lie within a bin, its frequency is tallied as a probability. The average total, elastic, capture, inelastic, and fission cross sections are also recorded for the bin. In the case of the samples provided, Kazuki Fujio used 20 bins with consistent probabilities. At the time of this report, 500 samples were provided. The number and quality of these samples is subject to change. 7 Figure 2: The full prior model correlation matrix across energy and quantity is shown on the left. The right plot shows the correlation matrix at 30 keV, where each quantity is labeled appropriately. The average cross sections bins are total, elastic, capture, inelastic, and fission respectively. The white streaks on the full correlation matrix result from the inelastic channel not being open ≤7.5 keV. Due to the way that the PTs were generated, the average of the cross sections from the provided PTs did not match the provided average cross sections. The PTs were scaled linearly to consistency with the corresponding average cross sections. This subsequently broke the sum rule for the probability tables: the reaction cross section in each bin of the probability table must sum to the total in each bin. This sum rule was restored by recalculating the elastic cross section as the total cross section minus each reaction cross section excluding elastic for each bin. At the time of this report, Kazuki Fujio is working to resolve this issue as well. The mean and covariance matrix was calculated from the samples using equations (1) and (2). The covariance matrix for the prior model data is shown in figure 2. The mean and covariance matrix includes all of the PT parameters, including bin edges, probabilities, and the average cross sections within the bins. 3.2 Experimental Data Three experimental datasets were used. Harvey 1988 (EXFOR No. 13632018) [9] for total cross section, Mosby 2018 (EXFOR No. 14794002) [10] for capture cross section, and Weston 1992 (EXFOR No. 13488003) [11] for fission cross section. Uncertainty quantification had been previously performed by Denise Neudecker and Noah Walton for all three datasets using ARIADNE [12]. The uncertainty-quantified data was used. The Harvey data was averaged, but the energy bins in EXFOR were not large enough to account for the cross section fluctuation. Therefore, a larger bin structure was used, provided by Noah Walton. Harvey reported negligible statistical uncertainty, manifesting as unrealistically high correlation, and in- stability in the GLS algorithm. A regularization uncertainty was applied. This 8 Figure 3: The data correlation matrices for Harvey (left), Mosby (middle), and Weston (right). regularization uncertainty was a diagonal uncertainty component equal to 1% of the variance. The Mosby data was well averaged and had no problems. The Mosby data was used as provided by ARIADNE. The Weston data was given on a fine grid. The fine-gridded Weston data were averaged into large bins. No self-shielding correction was required because of the small target thicknesses used in the fission experiments. The statistical uncertainty of the data was quite low like the Harvey data so a regularization uncertainty was applied. This regularization uncertainty was a diagonal uncer- tainty component equal to 1% of the variance. The data correlation matrices for the three datasets is shown in figure 3. 4 Results Figure 4 shows the prior model and posterior model alongside the experimental EXFOR data. The posterior results move towards the experimental data while maintaining the shape of the prior model. The posterior correlation matrix is shown in figure 5. Figure 6 shows the probability table as a probability density function and as a cumulative probability function. Consistent with the average cross sections in figure 4, the curve scaled to match the lower posterior average cross section. The variance clearly reduces in CDF-space, but remains consistent in PDF-space. The PDF does not reduce in variance because of the large correlation between cross-sections. The capture and fission cross sections at 15 keV are shown in Figure 7. Analogous to the update in the average capture and fission cross sections in figure 4, the mean values for the capture cross section bins in figure 7 do not move and the fission cross section bins move to lower fission cross sections. Based on the statistical variation in the cross section bins for capture and fission, more prior samples are suggested. 4.1 Validation Various validation cases were undertaken to ensure GLS was performing as intended. First, there are various sum rules that should be preserved from the 9 Figure 4: The prior model average cross sections (red), posterior model aver- age cross sections (blue), and experimental data (black) for each experiment. Shaded error bands are shown for the prior and posterior average cross sections. Figure 5: The posterior correlation matrix, analogous to figure 2. Figure 6: The probability density function (left) and cumulative probability function (right) for the prior and posterior total cross section probabilities. 10 Figure 7: The PT distributions for the capture cross section (left) and the fission cross section (right) at 15 keV. prior. For these tests, we take a random Gaussian sample from the posterior mean and covariance matrix. These rules include the following: 1. For each energy, the sum of the bin probabilities should be 1. 2. For each energy, the average cross section as calculated from the proba- bility table should equal the reported average cross section. 3. For each energy, the average total cross section should be the sum of the average cross sections for each reaction. 4. For each energy, and each PT bin, the bin total cross section should be the sum of the bin reaction cross sections. 5. The total cross sections for each bin should lie between the bin edges. The first four rules were forced to be true in the prior and they still hold true for the posterior results. The last rule is true except occasionally for the first bin. This is because the first bin is heavily weighted towards the upper bin boundary. Ocasionally, the first bin total cross section exceeds the upper bin boundary by a negligible amount. 4.1.1 Total Monte Carlo (TMC) Validation TMC was used to validate whether the linearity assumptions inherent to the GLS algorithm were appropriate to make. GLS is limited by assumptions of linearity and normality which certainly do not hold for the PT parameters (see figure 8). TMC is not limited by linearity and normality. By comparing the results of GLS to TMC, we can see if the assumption violation produces significant error in the final results. TMC produces inaccurate results when all three datasets are used. This is because the effective sample size (equation (14)) is nearly 1 when all three datasets are used, inflating the uncertainty to infinity. This is a result of three factors: 11 Figure 8: This plot displays the non-Gaussian nature of the prior data. Each line is a prior model sample for the total cross section bins. Although the average cross sections are indeed normally-distributed, the PT generator is non-linear resulting in non-Gaussian PT parameters. The non-Gaussian nature is clearly shown by the bin total cross sections in this figure. There is a clear skew towards lower cross sections for each PT bin for bin 5 onwards. 12 Figure 9: Updated GLS and TMC average cross sections (left) and fission prob- ability table distribution (right). 1. Small data uncertainty, reducing the amount of prior samples that are considered “good” by the data. 2. The large quantity of data points, increasing the number of conditions needed to have a “good” sample. 3. The discrepancy between the prior average cross sections and the measured average cross sections, reducing the likelihood of sampling average cross sections that are “good” according to the experimental data. To allow for a larger effective sample size, only a subset of Weston data is used for a cross-validation case between GLS and TMC. With only a small portion of Weston data, the effective sample is a small, but workable effective sample size of 45. The results are shown in figure 9. GLS and TMC appear to be in agreement on both the average cross section and probability table distribution. 4.2 Conversion to Parameterized Probability Tables The NIG and barycentric rational functions were fit to 20 samples drawn from the posterior mean and covariance matrix for the PT parameters. A test was performed to ensure the sampler is unbiased and not subject to numerical errors. One must be careful with sampling from large mean and covariance matrices. We found that eigen-decomposition of the covariance matrix was the most stable approach for generating samples (see equations (20) and (21)). Vy,post = OΛOT (20) ⃗r ∼ N (0, 1) samp ⃗ypost = ŷpost + OΛ1/2⃗r (21) It is uncommon, but there is occasionally a negative sample from this sampling method. In this case, we take the absolute value of the parameters. An attempt 13 Figure 10: The normal inverse Gaussian fit to the tabular fit with uncertainty. The PDF in log-y space shown in the top left, the PDF in linear space is shown in the top right and the CDF is shown at the bottom. was made to sample from a multivariate log-normal distribution to avoid neg- ative samples, but the covariance matrix does not meet all of the constraints needed for the multivariate log-normal distribution. The fit to the NIG distribution is shown in figure 10. The NIG fit slightly underfits the low cross section tail and overfits the bulk of the cross section. The covariance matrix for the new parameterization is shown in figure 11. 5 Acknowledgments We would like to thank the Nuclear Energy University Program (NEUP) for funding this work. References [1] Gavin Ridley. “GPU-Oriented Algorithms for Continuous Energy Monte Carlo Neutron Transport”. PhD thesis. Massachusetts Institute of Tech- nology, 2024. [2] R. E. MacFarlane. The NJOY Nuclear Data Processing System, Version 2016. 2016. 14 Figure 11: The correlation matrix for the Normal Inverse Gaussian fit of the posterior data, analogous to figure 5. [3] Toshihiko Kawano. “CoH3: The Coupled-Channels and Hauser-Feshbach Code”. In: Springer International Publishing; Cham, 2020. Chap. Part I. [4] D. A. Brown. ENDF-6 Formats Manual. Sept. 2023. [5] Paul K. Romano et al. “OpenMC: A State-of-the-Art Monte Carlo Code for Research and Development”. In: Ann. Nucl. Energy 82 (2015), pp. 90– 97. [6] R.W. Peelle. Peelle’s Pertinent Puzzle. (Informal ORNL memorandum, Oak Ridge Oct. 13, 1987). [7] Donald B. Rubin. Iteratively Reweighted Least Squares. University of Wis- consin, 1981. url: https://apps.dtic.mil/sti/tr/pdf/ADA114534. pdf. [8] SciPy Documentation. SciPy. url: https : / / docs . scipy . org / doc / scipy/index.html. [9] J.A.Harvey et al. “High-Resolution Neutron Transmission Measurements on 235U, 239Pu, 238U”. In: Nuc. Data for Sci. and Tech. (1988). [10] S. Mosby et al. “Unifying measurement of 239Pu(n,g) in the keV to MeV energy regime”. In: Phys. Rev. (2018). [11] L. W. Weston and J. H. Todd. “High-Resolution Fission Cross-Section Measurements of 235U and 239Pu”. In: Nuc. Sci. and Eng. (1992). [12] Cyrille de Saint Jean and Denise Neudecker. “ARIADNE – a program estimating covariances in detail for neutron experiments”. In: EPJ - Nu- clear Sciences Technologies 4 (2018). issn: 2491-9292. doi: https : / / doi.org/10.1051/epjn/2018012. url: https://www.sciencedirect. com/science/article/pii/S2491929218000390. 15 [13] Noah Walton et al. “Methodology for physics-informed generation of syn- thetic neutron time-of-flight measurement data”. In: Computer Physics Communications 294 (2024), p. 108927. issn: 0010-4655. doi: https : / / doi . org / 10 . 1016 / j . cpc . 2023 . 108927. url: https : / / www . sciencedirect.com/science/article/pii/S0010465523002722. [14] Jesse Brown. “Measurements, Evaluation, and Validation, and Validation of Ta-181 Resolved Resonance and Unresolved Resonance Regions”. PhD thesis. Rensselaer Polytechnic Institute, 2019. A Averaging Cross Sections For this work, we had to average cross sections for Harvey1988 [9] and We- ston1992 [11]. A discussion arose on how exactly to average cross sections. Three methods were discussed. (1) One could do simple averaging on the data, {dj }N j=1 : N 1 X dj (22) N j=1 Or (2) variance-weighted averaging could be done on the data: PN j=1 wj dj PN (23) j=1 wj wj = 1/Var(dj ) Or (3) the experimentalist could take a coarser binning structure for the mea- surement. Equivalently, the counts could be summed over many bins. The conclusion was that (3) was the only universally safe solution. Simple averaging fails when there is variable uncertainty, resulting in bias. Weighted averaging fails because it is equivalent to regressing URR cross section to a flat line. However, there is still cross section fluctuation in the URR domain, so a flat cross section is a poor model. Poor models often result in biased results. A.1 Testing with a Synthetic Data Generator Using a synthetic data generator, we can see the bias of these methods. The synthetic data generator uses synthetic 6 mm Ta-181 transmission realizations provided by SynDat [13], replicating data by Jesse Brown [14] in the 3-keV to 4-keV range. The bin spacing also matches the real experiment: 345 bins, uni- form in time of flight, starting at around 2.3-eV and ending at around 3.6-eV. The synthetic data generator is rather simple: no corrections or systematic un- certainties were involved. Background and differential efficiencies are neglected for this toy problem. 16 First, we define the expected transmission, Texp , and the expected cross section, σexp , calculated from equation 24. broad 1 σexp =− ln(Texp ) (24) n By averaging, we get the average transmission, ⟨Texp ⟩, and average cross section, broad ⟨σexp ⟩. Note, the average cross section is broadened and therefore not self- shielding corrected. First, we assume a constant open count rate, Copen = 20 counts per bin. The expected sample counts, Cexp , is calculated from the expected transmission, and the open counts. Random Poisson samples are generated from the expected bin counts with propagated uncertainty. Cexp = Texp · Copen (25) Cdat ∼ Poisson(Cexp ) (26) p δCdat = Cdat Next, the measured transmission, Tdat , is calculated from the open counts with propagated uncertainty. Tdat = Cdat /Copen (27) δTdat = δCdat /Copen broad The measured cross sections, σdat , are calculated from the measured trans- mission using a sample thickness of n = 0.03358 at/b. Note: the cross sections are experimental and not self-shielding corrected, like the true average cross section. broad 1 σdat = − ln(Tdat ) (28) n  broad 1 δTdat δσdat = n Tdat As seen in figure 12, weighted averaging has the expected biased estimate. Simple and bin averaging in transmission space produces an unbiased estimator of the average transmission. Simple averaging works for this case since transmis- sion is a linear transformation of the raw counts which are Poisson distributed. A linear transformation of Poisson data is still Poisson. Simple averaging produces an unbiased estimate of Poisson distributed data. However, there is a non-linear transformation between transmission and cross section, shown by equation (28). This non-linear transformation makes the data no longer Poisson distributed, resulting in the simple averaging bias seen in figure 12. Any thick-sample yield measurement is subject to the same problems. Fission measurements are mostly unaffected because they use thin samples. Simple averaging in transmission/yield space is not a cure-all fix. Non- Poisson data still appears if the open counts are not smoothed or for fission clean ratio measurements. A new methodology should be created to better estimate the average cross sections for these spaces. To experimentalists, we suggest providing bin-averaged cross sections along with fine-grid data for truly unbiased average cross sections. 17 Figure 12: Performance of various averaging methods in transmission space (left) and cross section space (right). 18