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_The_Spatial_Structure_of_Young_Stellar_Clusters._I.txt	Fig. 9.— Left: Histogram of all 142 subclusters’ ellipticities is shown. The ellipticities range from ∼0, or nearly circular in projection, to 0.9, a 10 to 1 aspect ratio, with low ellipticity subclusters favored. Right: The scatter plot shows subcluster core size vs. core radius.
Processed_The_Spatial_Structure_of_Young_Stellar_Clusters._I.txt	The MPCM samples have been designed to improve the statistical characterization of stellar populations in MSFRs (Feigelson et al. 2013). Using a rich statistical sample of 16,608 MPCMs, “ﬂattened” to remove X-ray spatial sensitivity variations, and the ﬁnite-mixture- model technique to identify subclusters among these young stars, new details about the spatial structure of young stellar clusters emerge. In the diverse MSFRs surveyed by the MYStIX project, we see a picture of star-forming regions with multiple non-coeval subclusters that probably formed with similar small sizes, but have expanded to a wide range of sizes. A similar ﬁnding of cluster expansion in a diﬀerent sample is studied by Pfalzner (2009). The catalog of 142 MYStIX subclusters derived here, combined with inferred intrinsic stellar populations (Paper II) and age estimates for the subclusters (Getman et al. 2014a), will help provide powerful links between observational and theoretical analysis star-cluster formation.
Processed_The_Spatial_Structure_of_Young_Stellar_Clusters._I.txt	The ﬁnite mixture model, using isothermal ellipsoids, is a novel cluster analysis method for MSFRs that can reveal information that is hard to obtain from other cluster-ﬁnding methods. Pfalzner (2011) ﬁnds that multimodal properties in regions with varying cluster surface density and richness are diﬃcult to determine from surface density distributions alone. But our method is an objective and mathematically founded procedure to separate multiply-clustered patterns into distinct isothermal ellipsoidal subclusters. The procedure is ﬂexible, ﬁnding statistically signiﬁcant structures on all scales simultaneously, even treating cases such as small clusters reside inside large ones. This is important because, even in relatively simple MSFRs like the Orion Nebula, layered subcluster structures are present and can aﬀect inferred subcluster properties. Furthermore, our algorithm does not require a single threshold in surface-density or separation to divide subclusters, which can help avoid bias when subclusters with very diﬀerent properties exist in the same star-forming region. However, the ﬁnite-mixture-model method requires additional assumptions about region structure and subcluster shape.
Processed_The_Spatial_Structure_of_Young_Stellar_Clusters._I.txt	We validate the subcluster results using residual maps, χ2 tests of source counts, and comparison with results from previous studies of these regions. The models reproduce the ob- served surface densities reasonably well, although residual maps reveal some extra structure. The improved stellar census in dense regions improves the accuracy of subcluster radius and peak surface density maps. On larger scales, we caution that the MPCM sample coverage of widely distributed young stars is limited, and isothermal ellipsoid ﬁtting is not sensitive to any truncation of cluster sizes that may be present.
Processed_The_Spatial_Structure_of_Young_Stellar_Clusters._I.txt	The numbers of subclusters and their arrangements vary signiﬁcantly from region to region. Heuristically, the clustering morphologies are divided into four classes: isolated clusters, core-halo clusters, clumpy clusters, and linear chains of clusters (Section 5). The dominant subcluster in a region (i.e. the subcluster with most stars, N4,obs) often sits at the center of a group of subclusters (e.g. Subcluster L in M 17), but this is not always true (e.g. subcluster B in the Eagle Nebula). In regions with linear chain structure there may be no subcluster that is particularly dominant. While, in contrast, regions with simple or core-halo structure may be much more centrally concentrated.
Processed_The_Spatial_Structure_of_Young_Stellar_Clusters._I.txt	In DR 21, NGC 2264, and NGC 6334, the linear chain cluster structure maps onto molecular ﬁlaments in the regions, and it is clear that the star cluster structure was at least partially inherited from the ﬁlamentary chain of molecular cloud clumps. More detail about correspondence between subcluster locations and molecular cloud clump locations will be given in a later MYStIX study. Thus, regions with linear chain structure have not evolved much from their initial state at the time of the birth of stars. The more centrally concentrated clumpy, core-halo, and simple cluster structures are more likely to have dynamically evolved and, in cases where the subclusters are well ﬁt by single isothermal ellipsoids, may have achieved some dynamical relaxation.
Processed_The_Spatial_Structure_of_Young_Stellar_Clusters._I.txt	The morphological classiﬁcation suggests an evolutionary progression from linear chain structure to clumpy structure to core-halo structure to simple cluster structure. Hierarchical cluster formation, often in molecular ﬁlaments, would evolve into dynamical relaxed uni- modal structures. However, the MYStIX regions indicate that this progression is not always followed. For example, the linear chain region NGC 1893 is almost entirely unembedded, and the chain of subclusters in NGC 6334 exhibit a wide range of absorptions.
Processed_The_Spatial_Structure_of_Young_Stellar_Clusters._I.txt	The distribution of subcluster properties show trends common to all the regions. A lognormal distribution of subcluster size peaks at 0.17 pc, and many subclusters have el- lipticities of (cid:15) ≤ 0.3, but few have ellipticities greater than 0.5. No relationships between ellipticity and size are seen.
Processed_The_Spatial_Structure_of_Young_Stellar_Clusters._I.txt	size of a subcluster is found, most easily explained if both absorption and size are related to age. This issue is pursued in the accompanying study by (Getman et al. 2014a) where spatial-age gradients in MYStIX regions are found based on a new age indicator for individual pre-main sequence stars. A similar relationship was found for subclusters within the Rosette Molecular Cloud by (Ybarra et al. 2013) who infer a gas removal e-folding timescale of 0.4 Myr.
Processed_The_Spatial_Structure_of_Young_Stellar_Clusters._I.txt	Cluster expansion is expected to produce a wide range of stellar surface densities seen in many star-forming regions (Gieles et al. 2012). Early cluster expansion arises principally from the loss of the gravitational potential of the molecular material, but stellar dynamics can also produce stellar halos around dense cores through three-body interactions with hard binaries and gravothermal core collapse associated with mass segregation (Giersz & Heggie 1996). This issue will be discussed further in Paper II when intrinsic populations, rather than observed population subject to diﬀerent sensitivity eﬀects, are derived.
Processed_The_Spatial_Structure_of_Young_Stellar_Clusters._I.txt	The anti-correlation between radius and absorption provides evidence that subcluster expansion is important part of subcluster dynamical evolution. From numerical simulations, Moeckel & Bate (2010) ﬁnd that even the very dense simulated cluster produced by Bate (2009a), with a half-mass radius of ∼0.05 pc, grew to >2 pc over a period of ∼2 Myr. The typical small size that we ﬁnd for the embedded clusters indicates that the high initial densities from the simulation in Bate (2009a) are realistic, which can make competitive accretion and dynamical interaction between protostars much more important than if clusters form ∼1 pc in size.
Processed_The_Spatial_Structure_of_Young_Stellar_Clusters._I.txt	Simulation of cluster expansion after gas expulsion by Goodwin & Bastian (2006) show a relationship between star-formation eﬃciency and the age-radius relationship. They ﬁnd that, after 2.5 Myr, core radii will grow by a factor of ∼1.5 for a star-forming eﬃciency of 60% and a factor of ∼3 for a star-forming eﬃciencies of 10–30%. Larger expansion factors are found in models of very rich clusters by Banerjee & Kroupa (2013). For comparison, MYStIX results indicate that subcluster core radii start out at ∼0.08 pc and grow to the typical ∼0.2 pc for the typical unabsorbed MYStIX subclusters (a factor of more than 2 expansion). The stars in MYStIX regions are generally <5 Myr old, but more precise age determination is necessary to determine which age-radius relation provides the better ﬁt. Pfalzner (2011) ﬁnd a trend consistent with a star-formation eﬃciency of ∼30% using the clusters in Lada & Lada (2003).
Processed_The_Spatial_Structure_of_Young_Stellar_Clusters._I.txt	The ellipticities of MYStIX subclusters could be inherited from the structure of the parental molecular cloud or could arise from subcluster mergers. The subcluster ellip- ticity distribution resembles those found by numerical simulations of merging subclusters (Maschberger et al. 2010, their Figure 10b) in which merger products have a distribution of ellipticities, including a tail at high ellipticities.
Processed_The_Spatial_Structure_of_Young_Stellar_Clusters._I.txt	The cluster morphological classes may correspond to structures seen in star formation simulations, such as those performed by Bate et al. (2003); Bate (2009a) or Parker & Meyer (2012). Simulations often show multiple pockets of star formation occurring along gas ﬁl- aments as the cloud collapses, which can end up merging and dynamically relaxing as the simulation progresses. Clumpy clusters may indicate early stages of partial merging, core- halo structures may be an intermediate stages of mergers, and isolated clusters with relaxed structure a ﬁnal stage. Getman et al. (2014b) add an important constraint to such a sce- nario: in at least two MYStIX regions, the dense core of the cluster has younger stars than the outer regions. This observational result requires some special behavior, such as continual feeding of gas towards the cluster center or dispersal of older core stars into the halo.
Processed_The_Spatial_Structure_of_Young_Stellar_Clusters._I.txt	The simple clusters may represent those that have completed subcluster merging and dynamical relaxation. Subcluster morphology can also inﬂuence these rates. Smith et al. (2011) ﬁnd that the cluster formation rate is related to a ﬁlling factor measuring the fraction of volume containing subclusters. When this ﬁlling factor becomes high, tidal interactions accelerate subcluster mergers and formation of a simple unimodal cluster is promoted. This phenomenon may be occurring today in crowded, clumpy MYStIX regions like M 17.
Processed_The_Spatial_Structure_of_Young_Stellar_Clusters._I.txt	The core-halo structures in 5 out of 17 MYStIX regions may also be a result of subcluster mergers. Bate (2009a) report a hydrodynamical and N-body simulation of a MSFR resulting in a cluster with a half-mass radius of 0.05 pc and a halo of stars extending out ∼0.4 pc. Smith et al. (2011) found that gravitational interactions between subclusters of young stars in simulated star-forming environments can scatter stars out of the clumps, which can lead to a dense core surrounded by a less dense halo of scattered stars. Maschberger et al. (2010) also identify a halo of scattered low-mass stars around the dominant cluster, albeit with a surface-density–radius relation opposite of the eﬀect we ﬁnd with concentric ellipsoids. How- ever, Pfalzner & Kaczmarek (2013) report that 2-body encounters can produce a population of cluster stars with high ellipticities which will spend most of the time in a halo. The accompanying paper by Getman et al. (2014b) describe star formation scenarios that may account for core-halo structures. All of these issues relating to cluster formation and MSFR star formation histories will be pursued in MYStIX papers by Getman et al. (2014a) and Paper II where quantitative measures of subcluster intrinsic populations and cluster ages will be combined with the structural results obtained here.
Processed_The_Spatial_Structure_of_Young_Stellar_Clusters._I.txt	We thank the anonymous referee for taking the time and eﬀort to provide useful feed- back for this paper. The MYStIX project is supported at Penn State by NASA grant NNX09AC74G, NSF grant AST-0908038, and the Chandra ACIS Team contract SV4-74018 (G. Garmire & L. Townsley, Principal Investigators), issued by the Chandra X-ray Center, which is operated by the Smithsonian Astrophysical Observatory for and on behalf of NASA under contract NAS8-03060. M.A.K. also received support from NSF SI2-SSE grant AST- 1047586 (G. J. Babu, PI). We thank Leisa Townsley for providing advice on this paper and for the development of Chandra tools. This research made use of data products from the Chandra Data Archive. This work is based on observations made with the Spitzer Space Telescope, obtained from the NASA/IPAC Infrared Science Archive, both of which are oper- ated by the Jet Propulsion Laboratory, California Institute of Technology under a contract with the National Aeronautics and Space Administration. This research has also made use of SAOImage DS9 software developed by Smithsonian Astrophysical Observatory and NASA’s Astrophysics Data System Bibliographic Services.
Processed_The_Spatial_Structure_of_Young_Stellar_Clusters._I.txt	The procedure for ﬁtting ﬁnite mixture models is illustrated here for the MYStIX region NGC 6357 with the code from the comprehensive R public domain statistical software system (R Core Team 2013). The R code below makes use of online source code and machine readable tables. Two packages from the Comprehensive R Analysis Network (CRAN) are used: plotrix (Lemon 2006) and spatstat (Baddeley et al. 2005).
Processed_The_Spatial_Structure_of_Young_Stellar_Clusters._I.txt	The structure of the example is as follows: Step 1 loads the required R functions and data; Step 2 creates a ﬁrst guess of the ellipsoid parameters; Step 3 conducts the ﬁrst iteration of model ﬁtting using the Nelder-Mead simplex algorithm; Step 4 displays the results of the ﬁrst iteration; Step 5 reﬁnes the model with further ﬁtting iterations; and Step 6 presents the results of the ﬁt, producing ﬁgures analogous to Figures 2 and 4a in the paper. These commands may be pasted into an R session following the instructions below.
Processed_The_Spatial_Structure_of_Young_Stellar_Clusters._I.txt	function star.ppp creates a point pattern object (an internal format of the spatstat package) for the region NGC 6357 assuming a distance of 1.7 kpc. Finally, the function unique removes any duplicated positions in the catalog.
Processed_The_Spatial_Structure_of_Young_Stellar_Clusters._I.txt	To illustrate the analysis, we ﬁt a model with k = 6 subclusters and the uniform den- sity component. The data structure “param2” speciﬁes: the number of model components (6 + 1 = 7), the model that will be used for each component (ellipsoid or constant), and the number of additional parameters for each component (5 or 0 for ellipsoid or constant, respectively). The function ell.model is an implementation of Equation 4 for the isother- mal ellipsoid in Section 3.1. The array “param.init” contains the initial guess parameters for all components of the ﬁnite mixture model — we can inspect these parameters more easily by running the param2ellipse command, which returns a data-frame object with k rows (subclusters) and 6 columns: “x” – ellipsoid x coordinate; “y” – ellipsoid y coordinate; “core” – core radius in parsecs; “theta” – ellipse orientation in radians; “b” – axis ratio; and “mix” – the log of the mixture coeﬃcient corresponding to the relative peak surface densities for each subcluster (the log mixture coeﬃcient for the ﬁrst component is assumed to be 0.0 and not included in the “param” array.) The ﬁnal entry in “param” is the log mixture coeﬃcient for the unclustered component.
Processed_The_Spatial_Structure_of_Young_Stellar_Clusters._I.txt	The function multi.model is the ﬁnite mixture model, which is speciﬁed by the variables “k,” “param,” and “param2.” The function model.lik calculates log-likelihood of this model using Equation 6. Model ﬁtting is done using the Nelder-Mead simplex algorithm via optim. This R function takes the star cluster data, the model form and parameters, and the log- likelihood function. It is unlikely that the absolute maximum of the likelihood will be found on the ﬁrst attempt, so this step must be performed iteratively — with certain parameters held ﬁxed in turn while other are free to vary — until the likelihood converges on the maximum.
Processed_The_Spatial_Structure_of_Young_Stellar_Clusters._I.txt	The commands below plot the spatial point pattern object (plot) and draws ellipses (draw.model) indicating the initial guess parameters (red) and new parameters (green). The log-likelihood “L” and the AIC are calculated — these values will improve with each iteration.
Processed_The_Spatial_Structure_of_Young_Stellar_Clusters._I.txt	the entire (6k+1)-dimensional parameter space. The commands below improve, sequentially, subcluster positions, subcluster sizes, subcluster orientations, subcluster shapes, subcluster positions again, and, ﬁnally, all parameters again. After each iteration, the choice of the next ﬁtting function to run depends on inspection of the results of the previous iteration. Nevertheless, each iteration will always result in an AIC value that is greater or equal to the previous iteration’s AIC values, so the order of these commands is not particularly important as long as there are enough iterations for the AIC to converge. (AIC is calculated again at the end of this ﬁtting sequence.) The time for these steps to run depends on the number of points and the dimensionality of the parameter space being searched, so ﬁts with frozen parameters will run more quickly than ﬁts where all parameters are free to vary.
Processed_The_Spatial_Structure_of_Young_Stellar_Clusters._I.txt	Now that the model has been ﬁt, we can visualize the results. The adaptive.density command produces the adaptively smoothed surface density maps described in Section 4.1. (Here, we only run 10 iterations rather than 100 used to produce Figure 2.) The function make.ﬁg2 produces an image of subcluster ellipses overlaid on the non-parametric surface density map similar to Figure 2. The function make.ﬁg2 produces an image of the spatial ﬁt residuals similar to Figure 4.
Processed_The_Spatial_Structure_of_Young_Stellar_Clusters._I.txt	A similar procedure is applied for diﬀerent numbers of ellipsoids (with the initial guess parameters “n”, “param”, and “param2” adjusted accordingly) to ﬁnd the number of sub- clusters that optimizes the AIC value. It is useful to try the ﬁtting procedure starting with several diﬀerent initial guesses in order to ensure that the results are independent of the initial parameters used.
Processed_The_Spatial_Structure_of_Young_Stellar_Clusters._I.txt	Baddeley, A. 2007, in Statistical Challenges in Modern Astronomy IV (eds. G. J. Babu & E.
Processed_The_Spatial_Structure_of_Young_Stellar_Clusters._I.txt	Clarke, C. J., Bonnell, I. A., & Hillenbrand, L. A. 2000, in Protostars and Planets IV (V.
Processed_The_Spatial_Structure_of_Young_Stellar_Clusters._I.txt	Kuhn, M. A., Povich, M. S., Luhman, K. L., Getman, K. V., Busk, H. A. & Feigelson, E. D.
Processed_Fourier-Based_Augmentations_for_Improved_Robustnes.txt	Diverse data augmentation strategies are a natural approach to improving robustness in computer vision models against unforeseen shifts in data distribution. However, the ability to tailor such strategies to inoculate a model against speciﬁc classes of corruptions or attacks—without incurring substantial losses in robustness against other classes of corruptions—remains elusive. In this work, we successfully harden a model against Fourier-based attacks, while producing superior-to-AugMix accu- racy and calibration results on both the CIFAR-10-C and CIFAR-100-C datasets; classiﬁcation error is reduced by over ten percentage points for some high-severity noise and digital-type corruptions. We achieve this by incorporating Fourier-basis perturbations in the AugMix image-augmentation framework. Thus we demon- strate that the AugMix framework can be tailored to effectively target particular distribution shifts, while boosting overall model robustness.
Processed_Fourier-Based_Augmentations_for_Improved_Robustnes.txt	Despite the chart-topping performances of CNN-based models across both standard and domain- specialized computer vision benchmarks, the tendency for these models to behave unreliably in the face of subtle changes to data distribution make them likely to fail in deployment [3, 4, 13]. This lack of robustness has been broadly attributed to the proclivity for CNNs to ﬁt on unintuitive and superﬁcial patterns that exist in the training data [3, 9, 10]. Thus there is a need for more comprehensive benchmarks—including those that test models against common corruptions and perturbations [5, 6, 8, 12, 14, 16, 17]—as well as for more varied data augmentation strategies.
Processed_Fourier-Based_Augmentations_for_Improved_Robustnes.txt	diverse set of augmentation primitives— including low, mid, and high-frequency characteristics— have proven to be effective at boosting overall model robustness [1, 7]; the AugMix augmentation method is particularly outstanding in this regard [7].
Processed_Fourier-Based_Augmentations_for_Improved_Robustnes.txt	Still outstanding is the need to build a model with robustness to particular distribution shifts—in a targeted manner—while maintaining robustness to common corruptions. There are well-known universal adversarial perturbations (UAP) that can be applied uniformly across images in order to induce high error rates across convolution-based computer vision models [13]. Indeed, it has been shown that even simple sinusoidal noise, designed to target particular frequency-and-orientation regimes, can serve as an effective UAP; these so-called Fourier-basis attacks can reduce image- classiﬁcation models to near-guessing performance [7, 19, 21]. It is of high importance that models can be made robust both to common, naturally occurring corruptions, as well as known black-box adversarial perturbations such as UAPs.
Processed_Fourier-Based_Augmentations_for_Improved_Robustnes.txt	In this work we investigate the capacity for AugMix to be modiﬁed to target particular data distribution shifts, while maintaining its outstanding ability to instill robustness against common corruptions. We speciﬁcally seek protection against Fourier-basis attacks, which are noteworthy for their simplicity, their efﬁcacy at diminishing model performance, as well as their ability to be tailored to target any frequency regime in input space.
Processed_Fourier-Based_Augmentations_for_Improved_Robustnes.txt	Figure 1: Classiﬁcation error heatmaps for a model trained using three data augmentation techniques, in association with Fourier-basis perturbations applied to all CIFAR-10 test images. AugSV (i.e, AugMix [7]) improves robustness to some high and mid-frequency regions. AugSVF, which explicitly incorporates distinct Fourier-basis augmentations, greatly reduces the model’s susceptibility across all frequency regimes.
Processed_Fourier-Based_Augmentations_for_Improved_Robustnes.txt	A real-valued 2D Fourier-basis matrix of shape dx × dy, U ((cid:126)k) ∈ Rdx×dy , has its elements deﬁned by a 2D sinusoidal plane wave that is evaluated on the discrete grid P = I{0,...,dx−1}×{0,...,dy−1}, i.e., Ux,y((cid:126)k) = A cos(2πf(cid:126)r · ˆk − φ) (1) where (cid:126)r = [x, y] ∈ P . A is selected such that (cid:107)U (cid:107)2 is ﬁxed.2 Each basis matrix is characterized by a so-called wave vector, (cid:126)k = [kx, ky] ∈ K ⊆ P , whose magnitude determines the frequency of the sinusoid, f = (cid:107)(kx/dx, ky/dy)(cid:107)2, and whose direction determines the orientation of the wave. K (⊆ P ) corresponds to the set of non-degenerate plane waves with a common phase shift of φ = π 4 .3 Three example basis matrices are depicted in Figure 1. A Fourier-basis perturbation consists of adding U ((cid:126)k), elementwise, to each color channel of a shape- dx × dy image, with strength controlled by (cid:107)U (cid:107)2. The perturbation can also be applied to each channel with a randomly-drawn sign applied to it; this is referred to as a “random-ﬂip” perturbation.
Processed_Fourier-Based_Augmentations_for_Improved_Robustnes.txt	Susceptibility heatmaps for channel-aligned perturbations with (cid:107)U (cid:107)2 = 2 are shown in Figure 1. It reports dx × dy metric values that measure a model’s performance on dx × dy perturbed test sets.
Processed_Fourier-Based_Augmentations_for_Improved_Robustnes.txt	Table 1: A comparison of augmentation strategies across datasets and performance metrics. CIFAR-C results are averaged across all corruptions and severities. AugSVF produces superior calibration metrics across the board, and yields outstanding robustness on corruption datasets.
Processed_Fourier-Based_Augmentations_for_Improved_Robustnes.txt	Entry [kx, ky] (∈ K) corresponds to a perturbation using U ((cid:126)k = [kx, ky]) applied identically to every image in the test set; the color of the pixel at [kx, ky] thus conveys the model’s classiﬁcation error rate on that perturbed test set. The heatmap is arranged so the lowest-frequency resides at its center; it is symmetrized about its center, due to the noted degeneracies among (cid:126)k ∈ P .
Processed_Fourier-Based_Augmentations_for_Improved_Robustnes.txt	The standard AugMix implementation involves a set of ﬁve spatial primitives (S) protate, pshear{x,y} , ptrans{x,y} , and a set of four “vision” primitives (V ) pcontrast, pEQ, pposter, psolar. AugS, AugV, and AugSV will indicate the inclusion of the primitives in S, V , and S ∪ V in the AugMix framework, respectively. Thus AugSV corresponds to the original implementation in [7] (see Appendix for more details). We can naturally include a Fourier-basis perturbation primitive (denoted by F ) among the primitives used by AugMix. The elements of stochasticity for this augmentation result from sampling: (cid:126)k, φ, and (cid:107)U (cid:107)2. We sample over frequencies ((cid:107)(cid:126)k(cid:107)2) uniformly.4 For example, AugSVF includes spatial, vision, and Fourier primitives. Lastly, there are four primitives that overlap with the CIFAR-C corruptions, pbrightness, pcolor, pcontrast, and psharp, which we will denote as C.
Processed_Fourier-Based_Augmentations_for_Improved_Robustnes.txt	We train a Wide ResNet architecture [22] on both CIFAR-10 and CIFAR-100, respectively, using a variety of data-augmentation strategies. We evaluate the following augmentation strategies: Base, AugF, AugS, AugSF, AugSV, AugSVC, and AugSVF. For each augmentation technique, we train the model using four independent random seeds and report the average performance. The baseline augmentation entails randomly ﬂipping and cropping an image, and then normalizing it. We evaluate the model’s classiﬁcation error as well as the RMS calibration error [11]. To assess the robustness of these models to unforeseen, common image corruptions, we include evaluations against all categories and severities of corruptions in the CIFAR-10-C and CIFAR-100-C datasets [6]. Our architecture, hyperparameters, and training methods exactly match those used in [7]; all models are trained using Jensen-Shannon divergence as a consistency loss [20].
Processed_Fourier-Based_Augmentations_for_Improved_Robustnes.txt	The performance metrics reported in Table 1 reveal the efﬁcacy of incorporating Fourier-basis perturbations in the AugMix framework. AugSVF produces the best classiﬁcation performance on the CIFAR-C datasets and superior calibration results on all benchmarks. The inclusion of AugSVC helps to demonstrate that simply adding any additional primitives to AugMix does not necessarily net a gain in performance. Indeed, AugSVC harms calibration across the board and is inconsistent in its impact on classiﬁcation error. This is despite the fact that AugSVC explicitly incorporates corruptions that overlap with CIFAR-C.
Processed_Fourier-Based_Augmentations_for_Improved_Robustnes.txt	Figure 2: (Left) Enhanced robustness in classiﬁcation accuracy over the baseline, across severity-5 corruption categories for CIFAR-100-C. AugSVF yields a signiﬁcant boost in robustness to “noise” and “digital” corruptions, without any signiﬁcant trade-offs elsewhere. (Right) CIFAR-100-C classiﬁcation accuracies averaged over corruptions, for each corruption severity.
Processed_Fourier-Based_Augmentations_for_Improved_Robustnes.txt	Table 2: Comparing vulnerabilities to Fourier-based attacks. AugSVF greatly reduces the impact that Fourier- basis corruptions have on the model’s performance. The phase shift used for all perturbations was held out at train time for AugSVF, nor did AugSVF train on (cid:107)U (cid:107)2 exceeding 2.
Processed_Fourier-Based_Augmentations_for_Improved_Robustnes.txt	Examining these results prior to aggregation over severity and corruption-type reveals that the gains in robustness achieved by AugSVF grows relative to AugMix with increasing corruption severity. Figure 2 shows, for each severity-5 corruption in CIFAR-100-C, the improvement that AugMix and AugSVF achieve over the baseline’s classiﬁcation accuracy for that category. AugSVF offers substantial boosts in robustness for several noise-type and digital-type corruptions, with improvements exceeding ten percentage points for multiple categories. Despite these marked improvements against distinctly “high- frequency” corruptions [21], AugSVF does not suffer any substantial trade-offs in other categories, other than against contrast and impulse noise; it otherwise sustains the gains in robustness achieved by AugMix. The error bars denote standard deviations across independent seeds.
Processed_Fourier-Based_Augmentations_for_Improved_Robustnes.txt	did AugSVF include perturbations with (cid:107)U (cid:107)2 > 2 during training. Table 2 reports results for both random and targeted Fourier attacks of varying strengths, as measured via mean and max error rates, respectively. The maximum error metrics correspond to a targeted attack where a single, worst-case Fourier-basis is selected to modify the entire test set.
Processed_Fourier-Based_Augmentations_for_Improved_Robustnes.txt	Figure 1 and Table 2 report that AugSVF is highly robust to corruptions of all frequencies, whereas Base and AugSV are reduced to near-guessing for targeted attacks with (cid:107)U (cid:107)2 ≥ 2, and a random (cid:107)U (cid:107)2 = 1 corruption roughly doubles their clean error rates. Thus incorporating Fourier perturbations in AugMix is effective at targeting this distribution shift.
Processed_Fourier-Based_Augmentations_for_Improved_Robustnes.txt	In this work we modify AugMix to protect against Fourier-basis attacks—a particular form of data distribution shift that can be used as an effective universal adversarial perturbation—while improving robustness against common corruptions and uncertainty calibration. We ﬁnd that inoculating a model against such a fundamental class of perturbations does not degrade performance, despite the noted propensity for models to “attach” to spurious statistics. Furthermore, although these perturbations explicitly span all frequencies, there is only an enhancement—and no loss—in robustness to a wider variety of corruptions. Indeed, we obtain outstanding results on CIFAR-10-C and CIFAR-100-C, while greatly ameliorating the model’s susceptibility to Fourier-based attacks.
Processed_Fourier-Based_Augmentations_for_Improved_Robustnes.txt	Future work will investigate whether achieving robustness to Fourier-basis attacks provides robustness against other universal adversarial perturbations that have distinctive frequency characteristics. For instance, the UAP developed by Moosavi-Dezfooli et al. [13] has been shown to exploit narrow regimes of high-frequency patterns to fool models [19]; however, we have shown here that our AugSVF augmentation method dramatically reduces a model’s susceptibility in these regimes. This gets at a broader question to be investigated: how reliable is the Fourier perspective on robustness? In particular, does improved robustness to plane wave perturbations in a particular subdomain of frequencies and orientations reliably indicate that a model will be more robust to corruptions that are prominently comprised of these plane wave components? Lastly, the simplicity and composability of Fourier bases may make them an especially effective augmentation strategy in applied domains, such as medical imaging, where other common photography-inspired augmentations are not appropriate; thus Fourier-based augmentations should be tested on a broad range of domain-speciﬁc datasets.
Processed_Fourier-Based_Augmentations_for_Improved_Robustnes.txt	Conference on Computer Vision and Pattern Recognition. 2021, pp. 15262–15271. Joern-Henrik Jacobsen et al. “Excessive Invariance Causes Adversarial Vulnerability”. In: International Conference on Learning Representations. 2018. Jason Jo and Yoshua Bengio. “Measuring the tendency of cnns to learn surface statistical regularities”. In: arXiv preprint arXiv:1711.11561 (2017).
Processed_Fourier-Based_Augmentations_for_Improved_Robustnes.txt	Although our formulation of U ((cid:126)k) presented in Section 2, which explicitly deﬁnes a real-valued plane wave sinusoid, does not immediately resemble the Fourier-based deﬁnition introduced in [21], the two are exactly equivalent, as are their implementations.
Processed_Fourier-Based_Augmentations_for_Improved_Robustnes.txt	A channel-aligned Fourier-basis perturbation is applied to an image by adding U , which is scaled to have a speciﬁed (cid:96)2-norm in order to control the strength of the perturbation, to each of the image’s color channels. The image’s pixel values are assumed to reside in [0, 1] and thus they are “clipped” to this domain following the channel-wise addition.
Processed_Fourier-Based_Augmentations_for_Improved_Robustnes.txt	A random-ﬂip perturbation follows this same process, but includes a randomly-drawn factor from {−1, 1} that multiplies U in association with each color channel. We include both channel-aligned and random-ﬂip perturbations in our analyses, whereas the random-ﬂip type is exclusively considered in some prior works [7, 21].
Processed_Fourier-Based_Augmentations_for_Improved_Robustnes.txt	The AugMix data processing technique processes an image through parallel chains of randomly- selected and composed augmentation primitives; it randomly weights and sums these parallel- processed images, and then randomly mixes the result with the original image [7].
Processed_Fourier-Based_Augmentations_for_Improved_Robustnes.txt	An example involving a particular AugMix conﬁguration, for AugSVF, for processing a single image is detailed in equation 2.
Processed_Fourier-Based_Augmentations_for_Improved_Robustnes.txt	Here the “depth”—the number of augmentation primitives composed for each chain—of each of the three chains is drawn uniformly from {1, 2, 3} with replacement. Once a chain’s depth has been determined, that number of augmentations is drawn uniformly from S ∪ V ∪ F . The example in 2 thus shows depths of 2, 1, and 2, respectively.
Processed_Fourier-Based_Augmentations_for_Improved_Robustnes.txt	Prior to this AugMix processing, a random ﬂip-and-crop is applied to the image—as in the “baseline” augmentation strategy. Following the AugMix process, the per-channel data normalization—enforcing a per-channel mean and standard deviation of 0.5—is applied to the “mixed” image.
Processed_Fourier-Based_Augmentations_for_Improved_Robustnes.txt	Our Wide ResNet architecture conﬁguration, hyperparameters, and training methods match exactly those used in [7] except for one detail: for each seed, we select the best model checkpoint based on performance on a holdout validation set, whereas the cited work evaluates directly on the test set. We leveraged hydra-zen [18] to produce all of our results in a self-documenting and reproducible manner.
Processed_Fourier-Based_Augmentations_for_Improved_Robustnes.txt	Figure 3 is the complement to Figure 2: it reports the gains in robustness—due to AugSV and AugSVF, respectively—for the various corruption categories in CIFAR-10-C.
Processed_Fourier-Based_Augmentations_for_Improved_Robustnes.txt	Here we include heatmaps corresponding to Fourier attacks of various channel alignments and strengths (i.e. (cid:107)U (cid:107)2) against the Base, AugSV, and AugSVF models, which were each trained separately on CIFAR-10 and CIFAR-100. Note the dramatic reduction in susceptibility that is achieved by AugSVF for all channel alignments and strengths. It should be noted that perturbations with (cid:107)U (cid:107)2 = 4 are roughly comparable to severity-5 corruptions from the CIFAR-C datasets, in terms of how visually disruptive they are.
Processed_Fourier-Based_Augmentations_for_Improved_Robustnes.txt	It is interesting to see the different susceptibility patterns that manifest, between the channel-aligned and random-ﬂip attacks; the former tend to have stronger effects in the high-frequency regimes, while the latter impact mid-to-low frequency regimes more severely. Prior papers [7, 21] only considered random-ﬂip channels, which overlooked an important class of Fourier perturbations and their impact on model performance. For instance AugSV appears to be somewhat robust to Fourier perturbations, in comparison to Base, when only considering random-ﬂip attacks. However, e.g, for CIFAR-10, we see that high error-rates (80%) manifest when AugSV encounters high-frequency, channel-aligned perturbations that are only of strength (cid:107)U (cid:107)2 = 2.
Processed_Fourier-Based_Augmentations_for_Improved_Robustnes.txt	Figure 3: (Left) Enhanced robustness in classiﬁcation accuracy over the baseline, across severity-5 corruption categories for CIFAR-10-C. AugSVF yields a signiﬁcant boost in robustness to “noise” and “digital” corruptions, without any signiﬁcant trade-offs elsewhere. (Right) CIFAR-10-C classiﬁcation accuracies averaged over corruptions, for each corruption severity.
Processed_Optimized_Bit_Mappings_for_Spatially_Coupled_LDPC_.txt	Abstract—In many practical communication systems, one bi- nary encoder/decoder pair is used to communicate over a set of parallel channels. Examples of this setup include multi-carrier transmission, rate-compatible puncturing of turbo-like codes, and bit-interleaved coded modulation (BICM). A bit mapper is commonly employed to determine how the coded bits are allocated to the channels. In this paper, we study spatially coupled low-density parity check codes over parallel channels and optimize the bit mapper using BICM as the driving example. For simplicity, the parallel bit channels that arise in BICM are replaced by independent binary erasure channels (BECs). For two parallel BECs modeled according to a 4-PAM constellation labeled by the binary reﬂected Gray code, the optimization results show that the decoding threshold can be improved over a uniform random bit mapper, or, alternatively, the spatial chain length of the code can be reduced for a given gap to capacity. It is also shown that for rate-loss free, circular (tail-biting) ensembles, a decoding wave effect can be initiated using only an optimized bit mapper.
Processed_Optimized_Bit_Mappings_for_Spatially_Coupled_LDPC_.txt	This work was partially funded by the Swedish Research Council under grant #2011-5961 and by the European Community’s Seventh’s Framework Programme (FP7/2007-2013) under grant agreement No. 271986. The calcu- lations were performed in part on resources provided by the Swedish National Infrastructure for Computing (SNIC) at C3SE.
Processed_Optimized_Bit_Mappings_for_Spatially_Coupled_LDPC_.txt	a region where reliable decoding is possible and where it is not. Under the assumption of inﬁnite codeword length, density evolution (DE) can be applied in order to ﬁnd the threshold for LDPC codes and BP decoding [5].
Processed_Optimized_Bit_Mappings_for_Spatially_Coupled_LDPC_.txt	SC-LDPC codes have been studied in the context of BICM systems in [6] and [7] assuming a uniform random bit mapping. Many authors have studied the optimization of bit mappers for irregular LDPC code ensembles. For example, in [8], [9], the authors use extrinsic information transfer charts to ﬁnd approximate decoding thresholds and subsequently optimize bit mappers, where in [9] special attention is payed to the short block length regime. In [10], a downhill algorithm is used to ﬁnd optimized bit mappings for two different LDPC codes. In [11], the authors substitute the parallel BICM bit channels by binary-input additive white Gaussian noise (AWGN) channels and optimized bit mappers are found for the codes and modulations speciﬁed in the DVB-T2 standard. Bit mappers with a simple implementation structure are designed in [12]. Some authors have also devised heuristic bit mapping strategies [13]. Furthermore, there exists a substantial amount of literature dealing with both code optimization for a ﬁxed bit mapping (e.g., [14]) as well as the joint optimization of the code ensemble and the bit mapper (e.g., [15]).
Processed_Optimized_Bit_Mappings_for_Spatially_Coupled_LDPC_.txt	In this paper, we take a similar approach as in [12] and substitute the parallel bit channels that arise in BICM with independent binary erasure channels (BECs). This is justiﬁed by the observation that the decoding threshold is mainly deter- mined by the mutual information of the channel rather than the channel details, see the discussion in [12, Sec. I]. Compared to [11], where the parallel channels are approximated as binary- input AWGN channels, the numerical complexity of the DE equations is greatly simpliﬁed when studying BECs.
Processed_Optimized_Bit_Mappings_for_Spatially_Coupled_LDPC_.txt	uncoupled ensembles due to the absence of a termination boundary. We show that by using an optimized bit mapper, the different qualities of the parallel channels can be exploited to obtain a decoding wave effect as for two-sided ensembles, i.e., the channels are effectively used to induce a termination boundary.
Processed_Optimized_Bit_Mappings_for_Spatially_Coupled_LDPC_.txt	We consider a communication system where one binary encoder/decoder pair is used to communicate over a set of parallel channels and a bit mapper determines the allocation of coded bits to the channels. A block diagram of the system model is shown in Fig. 1. In the following, the individual blocks are described in more detail.
Processed_Optimized_Bit_Mappings_for_Spatially_Coupled_LDPC_.txt	Many practical transmission scenarios can be modeled as a set of parallel channels and we take BICM as an example throughout this paper. To this end, consider the real AWGN channel Y = X + N , where X ∈ X is the channel input taking on values from a discrete signal constellation X and N ∼ N (0, 1). If we label each element in the constellation with a unique binary string of length m = log2 \|X \|, then, conceptually, we may view this setup as having m parallel bit channels from Bi to Y , where Bi, 1 ≤ i ≤ m, denotes the ith bit in the binary strings (counting from left to right) [4, Sec. 2-C]. Each of these bit channels can be characterized by an individual channel quality parameter αi (cid:44) 1 − I(Bi; Y ) ranging from 0 (perfect channel) to 1 (useless channel). The mutual information I(Bi; Y ) is commonly parameterized by the signal-to-noise ratio (SNR) and depends on the signal constellation as well as the binary labeling [4].
Processed_Optimized_Bit_Mappings_for_Spatially_Coupled_LDPC_.txt	For simplicity, we replace these bit channels by parallel, independent BECs with erasure probabilities εi = αi. In Fig. 2, two PAM constellations labeled by the BRGC are shown, and in Fig. 3 we plot the corresponding εi as a function of the average erasure probability ¯ε = ((cid:80)m i=1 εi)/m. Note that ¯ε (ranging from 0 to 1) is implicitly parameterized by the SNR (ranging from +∞ dB to −∞ dB), indicated by the top scale in Fig. 3, and this parameterization is different for the two constellations. Henceforth, ¯ε is used as the parameter to char- acterize the overall quality of the set of parallel channels (cf. Fig. 1) and the correspondence between ¯ε and the individual channel qualities is according to Fig. 3.
Processed_Optimized_Bit_Mappings_for_Spatially_Coupled_LDPC_.txt	variable node (VN) and check node (CN) degrees, L the spatial chain length, and w is a “smoothing” parameter. The construction of the two-sided ensemble is explained in detail in [2, Sec. 2-B] and extended to circular ensembles in [16, Sec. V]. For completeness, we review the construction with the help of the example depicted in Fig. 4, where dv = 3, dc = 6, L = 5, and w = 2, starting with the two-sided case. M VNs are placed at each spatial position 1 to L and dvM/dc CNs are placed at each position 1 to L + w − 1. For the asymptotic case, i.e., inﬁnite codeword length, it is assumed that M → ∞. The red circles in Fig. 4 correspond to known VNs which are initialized to zero erasure probability and placed at positions −w + 2 to 0 and L + 1 to L + w − 1. The connections between VNs and CNs are as follows. It the dvM edges originating from VNs at is assumed that position j are uniformly and independently distributed to CNs at positions j to j + w − 1, whereas the dvM edges from CNs at position j are assumed to be uniformly and independently distributed to VNs at positions j − w + 1 to j. In the ﬁgure, this is represented by the interleaver blocks that uniformly spread out the edges from the VNs and CNs. For circular ensembles, one can apply the same construction as above, but all position indices are now interpreted modulo L and no known VNs are present [16, Sec. V]. For the example in Fig. 4, this would correspond to removing the red nodes and edges and appropriately connecting the VNs at position 5 to the CNs at position 1. Also, no CNs are placed at position 6 due to the modulo indexing.
Processed_Optimized_Bit_Mappings_for_Spatially_Coupled_LDPC_.txt	Fig. 4. (3, 6, 5, 2) spatially coupled ensemble. Known VNs are shown in red.
Processed_Optimized_Bit_Mappings_for_Spatially_Coupled_LDPC_.txt	certain fraction of CNs is only connected to known VNs. For the circular ensemble, the rate loss is zero since all VNs and CNs are fully connected.
Processed_Optimized_Bit_Mappings_for_Spatially_Coupled_LDPC_.txt	for −w + 2 ≤ j ≤ L + w − 1, where l denotes the iteration number and εj the input erasure probability for the VNs at position j. The initial conditions for the two-sided ensemble are p(0) j = εj, where εj = 0 for j < 1 and j > L due to the known VNs. For circular ensembles, j ∈ {1, · · · , L} and the index arithmetic in (2) is performed modulo L [16, Sec. V].
Processed_Optimized_Bit_Mappings_for_Spatially_Coupled_LDPC_.txt	For a ﬁxed bit mapper, i.e., for a ﬁxed assignment matrix A, the decoding threshold ¯ε∗(A) is deﬁned as the largest p(l) j = 0, 1 ≤ j ≤ L, cf. ¯ε ∈ [0, 1] such that liml (2). This condition corresponds to successful decoding, i.e., all erased VNs can be recovered using BP decoding. In practice, to obtain the threshold with a certain precision δ, one ﬁxes a target erasure probability ptar and a maximum number of iterations lmax. Then, starting from ¯ε = δ, one iteratively computes (2) until the average erasure probability ((cid:80)L j )/L is either smaller than ptar (successful decoding) or the number of iterations exceeds lmax (decoding failure). In the ﬁrst case, ¯ε is increased by δ until the decoding fails. For a given channel quality parameter ¯ε up to the decoding threshold, we denote the number of iterations until successful decoding by ls(A, ¯ε).
Processed_Optimized_Bit_Mappings_for_Spatially_Coupled_LDPC_.txt	We are interested in optimizing A in terms of the decoding threshold for a given code ensemble. The baseline bit mapper realizes a uniform random mapping of coded bits to channels. For this case, we have ai,j = 1/m ∀i, j, and the corresponding assignment matrix is denoted by Auni. To establish the amount of threshold gain we can hope for by ﬁnding a better A, consider the following. Each BEC has capacity Ci = 1 − εi and the average capacity is ¯C = 1 i=1 Ci = 1 − ¯ε, where m 1 − ¯ε would correspond to the capacity of a BEC with erasure probability ¯ε. By employing the baseline mapper, however, the channel is effectively a BEC with erasure probability ¯ε and the two-sided (dv, dc, L, w) ensemble can approach the capacity of this channel for appropriately chosen (dv, dc) and L → ∞, w → ∞ [2, Th. 10]. Hence, for very long chain length L and smoothing parameter w, one would expect the potential gains in terms of threshold improvement to be rather small.3 However, for ﬁnite L and w, which is our main region of interest in this paper, signiﬁcant gains may still be possible. This is also an important region for practical systems since increasing L and w leads to large block lengths (assuming a ﬁxed and ﬁnite M ) and high decoding complexity. Moreover, for the same average channel quality ¯ε, an optimized bit mapper may signiﬁcantly reduce the number of decoding iterations until successful decoding compared to the baseline bit mapper. Finally, these potential gains come only at a small cost, i.e., by replacing the baseline bit mapper.
Processed_Optimized_Bit_Mappings_for_Spatially_Coupled_LDPC_.txt	the fact that ﬁnding the threshold is computationally expen- sive. This is especially pronounced for SC-LDPC codes and even for the moderate chain lengths considered in this paper (L ≤ 40), the computational cost attached to one threshold computation is signiﬁcant in the context of an optimization routine. An alternative, yet practical, approach, is to start with a certain channel quality parameter ¯ε and then optimize the convergence behavior of the ensemble in terms of decoding iterations. Then, one can calculate the new threshold for the obtained assignment matrix and repeat the whole procedure. Such an iterative approach was proposed in [17, Sec. IV] to ﬁnd optimized degree distributions for irregular LDPC codes. Based on this idea, we use the following iterative optimization routine in order to ﬁnd bit mappers with good decoding thresholds.
Processed_Optimized_Bit_Mappings_for_Spatially_Coupled_LDPC_.txt	With this procedure, the computational complexity can be signiﬁcantly reduced. However, it is not guaranteed to be (cid:54)= A∗ equivalent to a true threshold optimization, i.e., Aopt in general.
Processed_Optimized_Bit_Mappings_for_Spatially_Coupled_LDPC_.txt	In the following, we present optimization results assuming two parallel BECs according to Fig. 3(a). We focus on the two- sided and circular versions of the (4, 8, L, w) ensemble, where L ∈ {10, 15, . . . , 40} and w ∈ {2, 4}. Threshold values are 6, and lmax = 5000. computed assuming δ = 10− For a code ensemble with design rate R and an assignment matrix A, the gap to capacity is deﬁned as ∆ (cid:44) 1− ¯ε∗(A)−R.
Processed_Optimized_Bit_Mappings_for_Spatially_Coupled_LDPC_.txt	the results for the two-sided ensembles are shown. Fig. 5(a) shows both the decoding threshold (solid lines) and the design rate (dashed lines) and Fig. 5(b) shows the corresponding gap to capacity. The red and blue lines correspond to w = 2 and w = 4, respectively.
Processed_Optimized_Bit_Mappings_for_Spatially_Coupled_LDPC_.txt	Fig. 5. Optimization results for the two-sided (4, 8, L, w) ensemble. Color is helpful.
Processed_Optimized_Bit_Mappings_for_Spatially_Coupled_LDPC_.txt	Fig. 6. Values in the ﬁrst row of the optimized assignment matrices A∗ for the two-sided (4, 8, L, 2) ensemble with different chain lengths.
Processed_Optimized_Bit_Mappings_for_Spatially_Coupled_LDPC_.txt	w = 4, respectively.4 Furthermore, the ensemble with stronger coupling (w = 4) exhibits a signiﬁcantly larger rate loss.
Processed_Optimized_Bit_Mappings_for_Spatially_Coupled_LDPC_.txt	The thresholds that can be achieved with the optimized bit mappers are shown by the star markers. It can be observed that it is possible to improve over the “uniform” thresholds and, as expected, the absolute threshold improvement becomes smaller with increasing chain length. For L → ∞, it was already mentioned that the expected threshold gains will tend to zero. The results shown in Fig. 5(b) incorporate both the design rate and the decoding threshold and allow for a comparison between the two coupling parameters. We can conclude that a small coupling parameter is beneﬁcial in terms of ∆, due to the large rate loss for w = 4 for both the uniform and optimized bit mappers. Moreover, for a ﬁxed gap to capacity, the optimized bit mappers allow for a signiﬁcant chain length reduction. As an example, for ∆ = 0.04 and w = 4, the chain length can be reduced from approximately L = 40 to L = 25 and for ∆ = 0.02 and w = 2 a chain length reduction from L = 35 to L = 25 is possible.
Processed_Optimized_Bit_Mappings_for_Spatially_Coupled_LDPC_.txt	Fig. 7. Visualization of the iterative decoding behavior for (a) the two-sided and (b) the circular version of the (4, 8, 20, 2) ensemble with optimized bit mappers. The optimized bit mapper induces a one-sided wave propagation for the two-sided ensemble in (a).
Processed_Optimized_Bit_Mappings_for_Spatially_Coupled_LDPC_.txt	behavior. In Fig. 6, the values in the ﬁrst row of the optimized assignment matrices A∗ are plotted for w = 2 and spatial lengths L ∈ {10, 20, 30, 40}. The values in the ﬁrst row (i.e., i = 1) determine the fraction of VNs at a particular position to be sent over the good channel, cf. Fig. 3(a). Certain regions of the spatial dimension are shaded in different colors and, when ¯ε = ¯ε∗(A∗), these regions correspond to the part where a so-called decoding wave will start (green), end (red), and propagate at a roughly constant speed (blue). First, let us focus on Fig. 6(a) to gain some insight into the general structure of the optimized bit mappers. It is visible that the VNs at the last position are allocated only to the bad channel and there are proportionally more VNs allocated to the bad channel at the ﬁrst position (i.e., a1,1 is slightly less than 0.5). For the positions shaded in blue, the values of the assignment matrix are roughly constant. Similar observations can be made for L = 20. To illustrate the effect of the optimized bit mappers, in Fig. 7(a) we provide a visualization of the iterative decoding behavior for the two-sided (4, 8, 20, 2) ensemble at the threshold value of ¯ε∗(A∗) = 0.5024. As seen from the ﬁgure, the optimized bit mapper induces a one-sided wave propagation even though the ensemble is two-sided, and the wave propagates at a roughly constant speed. The direction of the wave is arbitrary due to the symmetry of the Tanner graph describing the ensemble, i.e., ﬂipping each row in the assignment matrix leads to a wave propagating from right to left with otherwise unchanged behavior.
Processed_Optimized_Bit_Mappings_for_Spatially_Coupled_LDPC_.txt	Fig. 8. Optimization results for the circular (4, 8, L, w) ensemble. In (b) and (c), we plot the values in the ﬁrst row of the optimized assignment matrices for w = 2. Color is helpful.
Processed_Optimized_Bit_Mappings_for_Spatially_Coupled_LDPC_.txt	propagation leads to a faster convergence compared to one- sided propagation for a given chain length L, even though a one-sided propagation may be better in terms of threshold. However, due to the ﬁxed maximum number of decoding iterations, the iterative optimization routine converges to these solutions for larger L.
Processed_Optimized_Bit_Mappings_for_Spatially_Coupled_LDPC_.txt	From the structure of the optimized bit mappers, an intuitive explanation for the decoding threshold improvement can be given as follows. In some sense, certain VN classes are “overprotected” and proportionally more of the VNs from these classes can be allocated to the bad channel without harming the overall iterative decoding performance. In turn, this allows for the remaining VN classes to be allocated more to the good channel, i.e., channel uses corresponding to the good channel become available and are spread out evenly among the VNs in the regions indicated in blue. In fact, all optimized bit mappers for w = 2 are such that in the blue regions, the values for a1,j are all slightly greater than 0.5. From this one can also explain why asymptotically the threshold gain will tend to zero, since asymptotically as L → ∞ this effect is not noticeable any more and a1,j → 0.5 in the blue regions.
Processed_Optimized_Bit_Mappings_for_Spatially_Coupled_LDPC_.txt	The optimized bit mappers for w = 4 in principle show a similar structure, but tend to be more unstable and wiggle around a certain average value in the part of the spatial chain that is shaded in blue.
Processed_Optimized_Bit_Mappings_for_Spatially_Coupled_LDPC_.txt	decoding threshold of the regular, uncoupled (4, 8) ensemble due to the absence of a termination boundary. By employing the optimized bit mappers, a signiﬁcant threshold gain is possible, which directly translates into a signiﬁcant reduction in terms of the gap to capacity. Contrary to the two-sided case, the threshold gain with respect to a uniform bit mapper increases for longer chain lengths L. Compared to the results for the two-sided ensembles, one can achieve a slightly better performance for w = 4, while for w = 2, the corresponding curves in Fig. 8(a) virtually overlap. The improvement for w = 4 can be explained by the large rate loss for the two- sided ensemble, and it shows that a more careful design of the termination boundary for small L may be beneﬁcial to achieve a better trade-off between rate loss and decoding threshold performance.
Processed_Optimized_Bit_Mappings_for_Spatially_Coupled_LDPC_.txt	Similarly as before, in Fig. 8(b) and (c), we show some of the optimized bit mappers in the form of the values in the ﬁrst row of A∗ for w = 2 and L ∈ {20, 40}. The actual results from the optimization routine have been adjusted (i.e., appropriately ﬂipped and shifted), so as to make the ﬁgures look similar. It can be observed that for circular ensembles, the optimized bit mappers are such that the VNs over a small spatial range in the green region are exclusively allocated to the good channel. In the red regions, the optimized allocation resembles that of the two-sided ensembles for L ∈ {30, 40}, cf. Fig. 6(c) and (d). The resulting iterative decoding behavior is illustrated in Fig. 7(b) for the circular version of the (4, 8, 20, 2) ensemble the decoding threshold value of ¯ε∗(A∗) = 0.4772. In at essence, the different channel qualities are exploited and a virtual termination boundary is created by the optimized bit mapping schemes. This allows for local convergence of the BP decoder at these positions within the ﬁrst few iterations and consequently two waves propagate outwards and eventually end in the region shaded in red.
Processed_Optimized_Bit_Mappings_for_Spatially_Coupled_LDPC_.txt	An extension to scenarios where more than two channels are present, e.g., for three BECs as shown in Fig. 3(b), is straightforward but comes at the price of increased optimiza- tion complexity due to the dimensionality increase of the problem. However, once a good bit mapper for two BECs is found, one can easily try the following. Let A∗ ∈ A2 L be an optimized assignment matrix for a given code ensemble of length L and ¯ε∗(A∗) the corresponding threshold. The input erasure probabilities for the VNs (ε1, . . . , εL) are thus ﬁxed. One can then try to ﬁnd a feasible A ∈ A3 L that satisﬁes εA = (ε1, . . . , εL), where ε = (ε1, ε2, ε3) is a vector with the individual erasure probabilities for ¯ε = ¯ε∗(A∗) in Fig. 3(b). This can be accomplished using standard numerical optimization routines.
Processed_Optimized_Bit_Mappings_for_Spatially_Coupled_LDPC_.txt	determines the allocation of coded bits to the channels. Com- pared to a uniform random bit mapper, the decoding threshold can be improved or, alternatively, the spatial chain length can be reduced. For circular ensembles, the different qualities of the channels can be exploited to obtain a wave-like decoding behavior similar to terminated, e.g., two-sided, ensembles. Future work includes the study of protograph-based ensembles for ﬁnite length code design and the application to different channel types. Further, it would be of much practical value to ﬁnd an analytical characterization of the optimal mappers.
Processed_Constraints_on_coronal_turbulence_models_from_sour.txt	The ﬁgure shows that the highest frequency predicted for R > 1 by the Leblanc et al.
Processed_Constraints_on_coronal_turbulence_models_from_sour.txt	in the integrand in Eqs (10) and (11), we start the integration at R0 = 1.056 R⊙.
Processed_Constraints_on_coronal_turbulence_models_from_sour.txt	the hybrid model from now on, for it is by far the most realistic one.
Processed_Constraints_on_coronal_turbulence_models_from_sour.txt	it follows that the rest of the baselines in the problem automatically satisfy this criterion.
Processed_Constraints_on_coronal_turbulence_models_from_sour.txt	this is immaterial, since most of the contribution to θc takes place well within 30 R⊙.

Subsets and Splits

No community queries yet

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