Stress testing the dark energy equation of state imprint on supernova data
Ben Moews, Rafael S. de Souza, Emille E. O. Ishida, Alex I. Malz, Caroline Heneka, Ricardo Vilalta, Joe Zuntz
aa r X i v : . [ a s t r o - ph . C O ] J u l Stress testing the dark energy equation of state imprint on supernova data
Ben Moews, ∗ Rafael S. de Souza, Emille E. O. Ishida, AlexI. Malz, Caroline Heneka, Ricardo Vilalta, and Joe Zuntz (COIN collaboration) Institute for Astronomy, University of Edinburgh, Royal Observatory, Edinburgh EH9 3HJ, UK Department of Physics and Astronomy, University of North Carolina at Chapel Hill, NC 27599-3255, USA Universit´e Clermont Auvergne, CNRS/IN2P3, LPC, F-63000 Clermont-Ferrand, France Center for Cosmology and Particle Physics, New York University, 726 Broadway, NY 10004, USA Scuola Normale Superiore, Piazza dei Cavalieri 7, 56126 Pisa, Italy Department of Computer Science, University of Houston, 3551 Cullen Blvd., TX 77204-3010, USA (Dated: July 9, 2019)This work determines the degree to which a standard ΛCDM analysis based on type Ia super-novae can identify deviations from a cosmological constant in the form of a redshift-dependent darkenergy equation of state w ( z ). We introduce and apply a novel random curve generator to simulateinstances of w ( z ) from constraint families with increasing distinction from a cosmological constant.After producing a series of mock catalogs of binned type Ia supernovae corresponding to each w ( z )curve, we perform a standard ΛCDM analysis to estimate the corresponding posterior densities ofthe absolute magnitude of type Ia supernovae, the present-day matter density, and the equation ofstate parameter. Using the Kullback-Leibler divergence between posterior densities as a differencemeasure, we demonstrate that a standard type Ia supernova cosmology analysis has limited sensi-tivity to extensive redshift dependencies of the dark energy equation of state. In addition, we reportthat larger redshift-dependent departures from a cosmological constant do not necessarily manifesteasier-detectable incompatibilities with the ΛCDM model. Our results suggest that physics beyondthe standard model may simply be hidden in plain sight. I. INTRODUCTION
The standard model of cosmology, in which the Uni-verse is composed primarily of cold dark matter (CDM)and a cosmological constant (Λ), is mainly supportedby three observational pillars: Big Bang nucleosynthesis(BBN) [1], the cosmic microwave background radiation(CMB) [2–7], and the discovery of late-time acceleratingcosmic expansion [8–10].BBN occurred within the first 20 minutes after the BigBang and is responsible for the production of the lightestnuclides, providing sensitive constraints on the ΛCDMmodel (e.g., [11–13]). Similarly, the estimated CMB tem-perature evolution with redshift is corroborated by rota-tional excitation of molecules and the Sunyaev-Zel’dovicheffect [14, 15]. The discovery of accelerated cosmic ex-pansion relies on the observational evidence that type Iasupernovae (SN Ia) appear fainter than it would be ex-pected in a decelerating universe [8, 9].The condition for late-time acceleration requires theequation of state parameter of dark energy to be w < − /
3, where w ≡ p/ρ is the ratio of its pressure p and en-ergy density ρ . The postulate of a cosmological constantcorresponds to w = − ∗ [email protected] has garnered the interest of cosmologists for the last twodecades [8, 23, 24]. The general notion of a cosmolog-ical constant predates the discovery of the acceleratingexpansion of the Universe (e.g., [25–27]). The conceptof dark energy, however, is much broader and has longserved as a generic placeholder for the physical cause ofan accelerating expansion, which is not necessarily re-stricted to a constant w (see, e.g., [23] for a review).Typical attempts to probe deviations from the ΛCDMmodel assume modifications at the background level,which can be described as a relativistic fluid with aneffective time-dependent equation of state. The formof the variable equation of state depends on the the-ory involved, subject to underlying kinetic and potentialterms, which can result in considerable variations of w as a function of redshift z . This also leads to proposalslike the Chevallier-Polarski-Linder (CPL) parametriza-tion [28, 29]. Examples of other non-constant models ofdark energy include quintessence [30] and, more gener-ally, scalar-tensor theories [31], with many of them fallingunder the umbrella of w CDM models [32].Theories relying on non-constant parametrizations of w have been tested on real datasets, with no evidenceof statistically significant deviations from ΛCDM beingreported [33–36]. The same inability to rule out compet-ing theories of dark energy is reported when using SNIa data under a specialized hypothesis test for ranges of w , though future survey data could provide stronger con-straints [37]. This competition between a constant anda variable, often redshift-dependent, equation of state isa matter of continuing debate [38]. A recent exampleof efforts in testing the CPL parametrization is carriedout using the Pan-STARRS Medium Deep Survey SN Iadata in combination with CMB measurements [39, 40].Apart from common parametrizations of w ( z ) [41, 42],non-parametric approaches make use of linear or cubicspline interpolation as well as Gaussian processes (GPs)[43–46]. The latter replace the need for placing a lim-ited number of nodes for an interpolation with the choiceof a suitable covariance function K ( z, z ′ ) [47, 48]. Re-lated research also makes use of non-parametric Bayesianmethods based on correlated priors [49].Regardless of the preferred representation for the equa-tion of state, the standard analysis consists of includingthe chosen w ( z ) model in the supernova likelihood andevaluating the results with the ΛCDM model as the nullhypothesis. In this scenario, the goal is to determinewhich type of behavior is allowed by the data in the con-text of a given dark energy model, with the prevailingconclusion that currently allowed behaviors are indistin-guishable from the ΛCDM model [50].In light of these results, we aim to address the contra-positive question: How robust is a standard SN Ia anal-ysis pipeline to deviations from ΛCDM in the data? Wethus investigate whether the traditional ΛCDM analysisframework is, in this context, a meaningful process to be-gin with. By creating arbitrary realizations of w ( z ), westress-test the viability of currently wide-spread methodsto measure w via SN Ia data for the assessment of darkenergy models. To accomplish this goal, we explore cur-rent capabilities to discriminate between different mod-els beyond a cosmological constant by running a standardcosmological inference pipeline on random fluctuations ofthe dark energy parameter w that adhere to physicallymotivated constraints.This work is organized as follows: The SN Ia mocksamples generated for subsequent experiments are de-scribed in Section II, along with our procedure for gen-erating data perturbations and the theoretical considera-tions that have to be taken into account when constrain-ing w ( z ). The analysis is performed according to the pro-cedure outlined in Section III, which provides an overviewof the cosmological inference pipeline, the choice of pri-ors, and the measure of posterior differences. We presentand discuss the results of both the primary investiga-tion and additional experiments for relaxed constraintsin Section IV and provide our conclusions in Section V. II. DATA
In order to test the limits of a standard SN Ia cosmo-logical pipeline, we generate a series of mock catalogs,each one corresponding to a universe with a different un-derlying behavior for the dark energy equation of state https://panstarrs.stsci.edu/ FIG. 1. Schematic flowchart of the generation for
Pan-theon -based SN Ia simulations. Dotted rectangles denotecalculated values, whereas rounded rectangles and circles in-dicate known values and random variables, respectively. Dot-ted lines mark operations performed at a given point duringthe process. parameter. The individual w ( z ) curves are obtained us-ing a smooth random curve generator described in Sec-tion II A, coupled with physically motivated constraintsexplained in Section II B. The generated curves are sub-sequently fed into into a SN Ia simulation pipeline, basedon the statistical properties and redshift distribution ofthe Pantheon
SN Ia sample [51]. Details on our sim-ulation, the process for which is shown in Figure 1, aregiven in Section II C.
A. Generating perturbations of Λ CDM
The construction of mock type SN Ia datasets that canmimic universes with varying dark energy equations ofstate requires the ability to create w ( z ) realizations un-der arbitrarily flexible sets of constraints, for example todefine vertical intervals and regulate the maximum num-ber of gradient sign changes. To this end, we introducea general-purpose smooth random curve generator thatsatisfies the need for extensive constraints, together withan easy-to-handle implementation for the wider researchcommunity. While we use this generator to create real-izations of w ( z ), our method is applicable to a wide arrayof problems in which generic curves are needed. In thiscontext, curve realizations can also be used for functionperturbations of arbitrary measurement detail, treatingthe value at each measurement point as a multiplier forthe respective value in a function that is to be smoothlyperturbed.Both node-dependent interpolation approaches andGPs present some significant drawbacks. Linear splineslead to sharp changes in the generated functions, whilecubic splines are prone to introducing spurious features.Similarly, GPs require setting a covariance function and,depending on the kernel, may lack smoothness [52]. Inaddition, the aforementioned methods hamper the abil-ity to easily subject the generated curves to customizedsets of constraints.To overcome such limitations, we introduce and em-ploy Smurves , a random smooth curve generator thatallows for highly customizable and physically motivatedconstraints to be placed on the curve-generating process.The source code of the curve generator, as well as a tuto-rial and examples, can be found in a public code reposi-tory . Based on the concept of changes in gravitationaldirection and magnitude along projectile paths, the gen-erator employs Newtonian projectile motion, adapted toallow for negative values, as the basis for generatingcurves.Given a set of user-specified constraints, Smurves gen-erates smooth curves through uniform-random samplingof the number of changes in gravitational direction andthe locations of such changes, while adhering to the spec-ified constraints. The path is segmented at the sampledchange points, and uniform-random samples of the gravi-tational acceleration are drawn within the bounds of pos-sible curve paths, while respecting the set of interval con-straints. The method used for curve segment calculationsis further summarized, including a pseudocode represen-tation, in Appendix A.While we primarily make use of the ability to set inter-vals and the number of maximum gradient sign changesfor this paper,
Smurves features a variety of additionaloptions that make it applicable to a wider array of prob-lems. Examples of other capabilities include the use oflogarithmic scales and the capacity for perfect conver-gence in a specified point along the generated curves’paths.The next section describes the use of
Smurves to cre-ate 50 w ( z ) curves per constraint family, which imposesboundaries in both dimensions, z and w , on each curvesampled at 500 equally-spaced redshift bins on a linearscale. For brevity, we call each such constraint familygenerated with Smurves a “SmurF”.
B. Constraints on w(z)
We explore families of w ( z ) curves that evolve withinthe redshift range covered by the binned Pantheon data, 0 . < z < . w models, with a broadestenvelope of − / < w < − / w = − / https://github.com/moews/smurves FIG. 2. Smooth random w ( z ) curves generated with Smurves to create SN Ia mock observations. The figure shows curvesfrom four different constraint families (“SmurFs”), with 50curves per family, while adhering to a maximum of one gra-dient sign change for a given curve. The varying parametersare the upper and lower boundaries of w ( z ) for each family. time driven by dark energy. For each component i of theUniverse, this limit corresponds to P i ( ρ i + 3 p i ) < w i ≡ p i /ρ i , pressure p i , and energy density ρ i ofenergy component i [53]. The limit of w < − / w results fromthe requirement that a so-called Big Rip scenario cannothave occurred within the age of the Universe of roughlyone Hubble time H − . The previous term implies thatphantom energy, with w < −
1, becomes infinite in finitetime and overcomes all other forms of energy, rippingapart everything, from cosmic structure to atoms, withthe Universe ending in a “Big Rip” [54]. We also notethat phantom dark energy violates the null energy con-dition [55].While the lowest redshift for the
Pantheon data is z = 0 . z = 0 so that w (0) = −
1. This is to agreewith near- z cosmological probes bearing small scatter atthe lowest redshift bin. The resulting set of constrained w ( z ) curves, shown in Figure 2, exhibits behaviors thatcan be found, among others, in effective fluid descriptionsof f ( R ) models [56], scaling, or interacting, dark matter[28], and bimetric theories of gravity [57].In practice, this approach means that we evolve theFriedmann equation while including both matter anddark energy as energy components. For a flat Universe,this implies H ( z ) = H h Ω m (1 + z ) + Ω Λ (1 + z ) w ) i / , (1)where Ω m and Ω Λ represent the dark matter and darkenergy density parameters, respectively. For a flat Uni-verse, we note that Ω Λ = 1 − Ω m . The current age t > H − of the Universe sets a lower limit on w fora given Ω m . The more negative a phantom component( w < −
1) is, the faster we reach a Big Rip scenario. Alower boundary of w & − m = 0 . m = 0 . w & − .
2, and Ω m = 0 .
01 yields w & − /
3. Therefore,we constrain our broadest envelope of w ( z ) curves to alower limit of w = − /
3, conservatively corresponding toa very low matter density and yielding symmetric inter-vals for the curve limits.For the three remaining SmurFs, we halve the pre-ceding symmetric interval around w = − w ( z ) = − w ( z ) curves in linewith shapes found in research discussed in Section I,but explore an increased maximum number of gradientsign changes, as well as the effect of an omission of the w ( z ) = 0 constraint, later in Section IV B. C. SN Ia data simulation
Observations sensitive to the background expansionsuch as SN Ia data can be employed to measure the lu-minosity distance, d L ( z ) = (1 + z ) d H Z z d z ′ E ( z ′ ) , (2)where the Hubble distance is d H = c/H and the Hub-ble parameter is E ( z ) = H ( z ) /H , with H ( z ) given byEquation 1. This is related to the peak B-band magni-tude, m B i = 5 log d L ( z i ) + M, (3)of a given supernova i at redshift z i , with absolute mag-nitude M . We generate SN Ia peak B-band magnitudecatalogs by inserting each w ( z ) curve seen in Figure 2into Equation 1 and following the process shown in Fig-ure 1.Our mock data are constructed to mimic the statisti-cal properties and redshift distribution of the Pantheon
SN Ia sample , which consists of a total of 1048 SN Iaat redshifts 0 . < z < .
3, representing the largest com-bined sample of SN Ia observations to date [51]. We use https://archive.stsci.edu/prepds/ps1cosmo/index.html FIG. 3. Peak B-band magnitudes m B as a function of red-shift z for different dark energy equation of state ( w ( z )) real-izations. The figure shows the diagrams for the ΛCDM model(dashed line), as well as 50 random w ( z ) curves for each of thefour constraint families, which represent increasing deviationsfrom ΛCDM. Black points depict the Pantheon dataset andrespective uncertainties, and the insets highlight w ( z ) mod-els regarding ΛCDM as mostly falling within the data uncer-tainty, even at redshifts as high as z & . the publicly available catalog, which is summarized by40 redshift bins from z = 0 . z = 1 . w between the binned and un-binned versions are smaller than (1 / σ for statisticalmeasurements [51], which makes this an adequate andeasy-to-handle data representation for a large number ofanalysis pipeline runs.We propagate the curves through a simulation pipelineusing CosmoSIS , as described in Section III A. The sim-ulation pipeline also takes into account the full covari-ance matrix, which includes effects due to photometricerror, the uncertainty in the mass step correction, uncer-tainty from peculiar velocity and redshift measurement,distance bias correction, and uncertainty from stochasticlensing and intrinsic scatter. Peak B-band magnitudesfor w ( z ) curves are shown in Figure 3 to demonstratethe similarity of results even at high redshifts. III. METHODS
We run a full analysis pipeline that assumes a constant- w dark energy model, hereafter called Ψ w const , to inferthe posterior probability distribution of w , Ω m , and M as described in Section III A. In Section III B, we listand justify our choice of priors for parameters. Finally,in Section III C, we introduce the metric by which wecompare simulation-based posteriors and those from realSN Ia Pantheon data.
TABLE I. Priors for the estimation of cosmological andnuisance parameters. U( · ) denotes a uniform distribution,whereas we use “fixed” to indicate a Dirac delta function with δ ( x ) = ∞ for an x from the column of initial values.Parameter Prior Initial valueΩ m U(0.01, 0.6) 0.307 M U(-20.0, -18.0) -19.255 w U(-2.0, -0.3333) -1.026Ω k fixed 0Ω b fixed 0.04 h fixed 0.7324 A. Pipeline with
CosmoSIS
CosmoSIS is a cosmological parameter estimation code[58], which models cosmological likelihoods and calcula-tions as a sequence of independent modules that read andwrite their inputs and outputs to a central data storageblock. The package has been used extensively for param-eter estimation by the Dark Energy Survey (
DES ) (e.g.,[59–63]), among others [64–67].We utilize two
CosmoSIS pipelines; the first simulatesdata using the w ( z ) realizations described above, and thesecond analyzes the simulated data using the emcee sam-pler [68, 69] under a standard cosmological model. Theprocess of emcee is described in detail in Appendix B.We connect these two pipelines in a script to iter-ate the process over the curves from each SmurF usingfour standard library modules: consistency , which com-putes the complete set of cosmological parameters, camb [70], which, in our case, calculates cosmological back-ground functions, and pantheon , which computes the Pantheon likelihood. A custom module is used to readin tabulated w ( z ) functions and cast them to the formused in camb .For Gaussian likelihoods, CosmoSIS automatically gen-erates simulated outputs incorporating both the signalbased on the used model and noise, as described in Ap-pendix C. Employing the reported uncertainties on m B and the full covariance matrix, we use this process tosimulate peak B-band magnitudes at the same redshiftvalues as reported for the real data in the binned Pan-theon sample. The distributions of these mock peakB-band magnitudes are provided in Figure 4.
B. Choice of priors
We vary our cosmology via the present-day matter den-sity Ω m and the dark energy equation of state w . We as-sume a flat Universe with Ω k = 0 and, therefore, a darkenergy density of Ω Λ = 1 − Ω m . We keep the present-day Hubble parameter fixed to h = 0 . b = 0 .
04 [72]. An additionalnuisance parameter is the absolute magnitude of SN Ia M , which is degenerate with the Hubble parameter. Our set of estimated parameters from the emcee sam-pler is { Ω m , w, M } . We choose uniform priors for allparameters, with bounds given in Table I. The range forthe absolute magnitude M encompasses previous con-straints given, for example, by the SDSS-II/SNLS3 JointLight-Curve Analysis ( JLA ) [73]. The central startingvalue of M = − .
255 is chosen from a preliminarymaximum likelihood run with
Pantheon data. Theprior over Ω m covers allowed parameter ranges as esti-mated by present-day SN Ia samples like JLA and
Pan-theon . The starting point for the dark matter parame-ter is Ω m = 0 . Pantheon w CDM best-fit value. Analogously, the central value for w is set to w = − .
026 [51].The prior range on w coincides with the allowed val-ues for the families of w ( z ) curves considering the priorupper bound of Ω m = 0 . w -interval). For our param-eter estimation, we loosen the symmetric lower-boundrequirement, with w = − m from SN Ia at 3 σ . C. Comparison criteria
Conventional error contours, used ubiquitously in cos-mology, are estimated from samples from posterior prob-ability distributions p ( θ | D, Ψ) of parameters of interest,in our case θ = { Ω m , w, M } , conditioned on the cosmo-logical model Ψ and data D = { d i } N , where i runs overthe number N of observations. For Pantheon , the datais presented as D Pantheon = { z i , m B i , σ m B ,i } for bins i .We consider each individual w ( z ) curve separately,but group them by constraint family S k , as depictedin Figure 1, for interpretability (see Appendix D fora detailed justification of this procedure). For j ∈{ , , . . . , } , each of 50 simulated data sets D j is gener-ated with the curve w j ( z ), and our experimental designyields samples from the posteriors p j ≡ p ( θ | D j , Ψ w const ).Each posterior corresponds to the probability of parame-ters from a cosmological model Ψ w const conditioned onthe data generated from w j ( z ). We also apply thesame pipeline to 50 realizations of the data under theΛCDM model, producing p Λ j ≡ p ( θ | D ΛCDM j , Ψ w const ),and to the real Pantheon data, producing p Pantheon ≡ p ( θ | D Pantheon , Ψ w const ).To compare the samples from each mock universe totheir ΛCDM counterparts, we adopt a measure suited toquantifying the difference between probability distribu-tions. The KullbackLeibler divergence ( D KL ) [74], D KL = Z ∞−∞ p ( x ) ln (cid:20) p ( x )ˆ p ( x ) (cid:21) d x, (4)is the directional difference between a reference proba-bility distribution p ( x ) and a proposed approximatingprobability distribution ˆ p ( x ). The D KL has been appliedwithin astronomy only to a limited extent, but is gainingpopularity [46, 75–79]. FIG. 4. Visualization of peak B-band magnitude ( m B ) residuals between our simulated data and ΛCDM, as well as betweenobserved Pantheon data and the ΛCDM model. In both cases, ΛCDM corresponds to Ω m = 0 .
307 and M = − . z ) bins show a rotated kernel density plot of the distributions of values for eachof 50 different realizations for one SmurF per panel. Black dots indicate binned Pantheon data, with vertical black linesrepresenting the error bars of one standard deviation. The comparison is plotted as the difference between the respective peakB-band magnitudes and expected ΛCDM values, m B − m BΛCDM , to show both the deviation from theoretical values and thedistributions of simulated SN Ia data around observed values.
Unlike symmetric measures of the distance betweentwo probability distributions, such as the familiar root-mean-square-error, the D KL is defined as the directionalloss of information due to using an approximation inplace of the truth; we must designate one distributionas a reference from which the proposal distribution di-verges. A generic example of a pair of reference andproposal distributions can be defined by posterior sam-ples derived from a large set of observations, as opposedto posterior samples derived from a small subset thereof.There is, therefore, an implicit assumption that the for-mer is closer to the truth than the latter, which may bean approximation when the rest of the observations areunavailable.In our case, the samples from p Pantheon always serve asthe reference distribution, and the samples from p j and p Λ j always act as the proposal distribution. IV. RESULTS AND DISCUSSION
In the previous sections, we describe both the data andour methodology. In Section IV A, we present the resultsof primary experiments, together with a discussion of theunderlying causes and implications for SN Ia investiga- tions. In addition, we relax the different constraints fortwo of the constraints families in Section IV B to explorethe impact such changes have on the resulting D KL dis-tributions. In the first of these two additional experi-ments, we generate w ( z ) curves with an increased maxi-mum number of gradient sign shifts, whereas the secondexperiment eliminates the requirement that w ( z ) = − A. Primary experiments
For each SmurF, as described in Section II A, we gen-erate 50 w ( z ) curves that are fed into the CosmoSIS sim-ulation and analysis pipeline described in Section III A.This results in four sets of 50 posterior distributions forparameters { Ω m , w, M } , or p S k ,j , where k ∈ { , , , } identifies the SmurF and j ∈ { , , . . . , } denotes itsrealizations (see Section III C for details on notation). Inaddition, 50 datasets from a ΛCDM model are generatedto illustrate the impact allowed by current statistics andsystematic uncertainties. We feed these simulations, aswell as the original binned Pantheon dataset, into thesame analysis pipeline. Posteriors derived from all simu-lated data are then compared to the
Pantheon resultsusing the Kullback-Leibler divergence D KL , described in FIG. 5. Histograms of the Kullback-Leibler divergence( D KL ) for different sets of constraints. The shown histogramsdepict the distribution of D KL values for the ΛCDM case andeach SmurF used to generate simulated SN Ia peak B-bandmagnitudes. D KL values are calculated for the posterior dis-tributions of parameters obtained through a standard ΛCDManalysis pipeline that considers only constant w models. Section III C.Figure 5 shows histograms of D KL values for eachSmurF along with those from ΛCDM simulations. In ac-cordance with our expectations, the distributions of D KL values for constraint families with increasingly wider w -intervals, from SmurF 1 through 4, show a systematicshift towards higher means, larger variances, and multi-modality. These differences are, however, small enoughthat the bulk of D KL values for each SmurF coincideswith the D KL range covered by the ΛCDM case, present-ing a serious obstacle for the detection of deviations froma cosmological constant.This effect is better visualized by a representative w ( z )function for each SmurF and the respective posteriors,shown in each column of Figure 6. The top row shows w ( z ) curve associated with the median D KL value foreach SmurF, as well as the constant w = − w ( z ) = − w inter-vals of its respective constraint family, thus confirmingthe applicability of a median- D KL approach for choos-ing a representative SmurF instance. The bottom threerows show two-dimensional posterior distributions, forparameters { Ω m , w, M } , for each SmurF and the ΛCDM case (colored contours) superimposed on the posteriorsfrom Pantheon data (black contours). Similarly, poste-rior distributions from the ΛCDM model, together withSmurFs 1, 3, and 4, go from agreement to disagreementwith
Pantheon . Posteriors from SmurF 2, on the otherhand, show an unexpected visual match with both real
Pantheon data results and the ΛCDM case, despite itsassociated w ( z ) exhibiting larger deviations from w = − z and high- z regimens, meaningthat larger deviations from the λ CDM case do not neces-sarily result in posteriors considerably different from theones produces by w ( z ) = − w ( z ) and compliant posterior estimates derivesfrom the fact that, while w ( z ) can change widely, theobservable signature of w ( z ) relies on the peak B-bandmagnitude m B . The dependence of m B on the integralof the Hubble parameter leads to a statistical degeneracythat makes such posteriors indistinguishable from ΛCDMwithin the current magnitude precision level and probedredshift range. Coupled with the large D KL overlap be-tween SmurF instances and ΛCDM results seen in Fig-ure 5, this directly extends to a considerable chance ofmistaking an equation of state varying significantly withredshift for one in reasonable agreement with a cosmo-logical constant.A more detailed view of all posteriors over w is shownin the ridgeline plots of Figure 7, in which the means,as well as the bulk of the probability, fall within the95% credible intervals of the Pantheon results undera constant- w hypothesis. SmurF 2, in particular, showsmore constrained posteriors, which offers an explanationfor the agreement of the median- D KL representative’sposterior with the ΛCDM case. It does, however, alsofeature four obvious outliers reaching far beyond the leftboundary of the credible interval, which demonstratesthe variability in the agreement of w -posteriors withinthe same constraint family.Naturally, all of the the aforementioned results arebounded by the Pantheon -like quality of our simula-tions. Current surveys such as
DES continue to con-tribute to the number of SN Ia observations [59]. Thoughthe
DES
SN Ia samples used in combination with addi-tional external samples amount to less than a third of
Pantheon ’s sample size,
DES results indicate smallerintrinsic scatter in the Hubble diagram, taking one stepfurther in the attainment of higher-quality SN Ia samples[80]. These new and future datasets will certainly in-crease our ability to discriminate between different mod-els for the dark energy equation of state parameter.It is, however, important to highlight the non-intuitiveand unavoidable behavior derived from the nature of dis-tance measurements as an integral over the Hubble pa-rameter. Given a dataset with sufficiently low measure-ment and systematic uncertainties, especially at high red-
FIG. 6. First row: Representative redshift-dependent dark energy equation of state ( w ( z )) curves associated with the median D KL per constraint family (full lines) and the ΛCDM case (dashed line). Second row: Posteriors for w and dark matter densityΩ m per constraint family. The four plots depict the posterior distributions for the above-mentioned curves (colored contours),as well as the posteriors for the Pantheon analysis case (black contours). Third and fourth row: With M as the absolutemagnitude, the plots show two-dimensional posteriors for M × Ω m and M × w , respectively. shifts, discrimination between phenomenologically closemodels is possible, but we cannot rely on the assumptionthat substantial redshift-dependent changes in w ( z ) willnecessarily result in detectable biases under a constant- w analysis. This is especially the case for SN Ia-onlyanalyses [50, 81–83].Caution should be exercised in using other cosmologi-cal observables to break the degeneracy via constrainingadditional parameters. This strategy is wide-spread inthe literature, to the point that recent research questionsthe use of SN Ia data without such additional observ-ables [84]. It is, however, important to keep in mind thatsupernovae are the primary dynamical observable thatprobe the line of sight directly, and consequently imposeboundaries in the behavior of w . The use of additional probes such as weak lensing can, with insufficient infor-mation on the baryonic physics involved, introduce newbiases, for example in the CPL parametrization [85].In summary, we recognize the need to combine com-plementary observables, for example baryon acoustic os-cillations and CMB data, while making use of carefulstatistical analyses capable of probing more subtle be-haviors of the dynamical evolution of dark energy. Al-though paramount for a more general discussion of thistopic, the addition of extra observables exceeds the scopeof this paper. FIG. 7. Ridgeline plots for the dark energy equation of stateparameter w . Each row depicts the posterior densities of w forall 50 curves, for each of the four constraint families as well asthe simulations for the ΛCDM case. The transparent bandscovering the middle section of each column show the 95%credible interval for the Pantheon sample, analyzed under aconstant- w model. B. Relaxed constraints on w(z)
In a bid to push our analysis a bit further, we relax theconstraints put on the curve generator for SmurFs 2 and4 for illustrative purposes. For SmurF 2, we increase themaximum number of gradient sign changes from one to10, allowing for more complicated functions to be real-ized. In contrast, for SmurF 4, we omit the requirementthat w (0) = − w ( z ) interval. The respectivecurves used in these additional experiments are depictedin Figure 8.To assess the impact of these further constraint re-laxations, their D KL distributions are shown in Figure 9,along with those from SmurF 2, SmurF 4, and the ΛCDMcase. The D KL distribution of SmurF 2.1 still holds thesame overall shape of SmurF 2 and occupies a range of D KL values between those covered by SmurF 2 and 4.This demonstrates that the use of more complicated func-tions, for example the larger maximum number of gradi-ent sign changes in SmurF 2.1, has a lesser impact thansimpler functions allowed to vary in a larger interval, asis the case for SmurF 4.1, when constrained to the same w ( z ) intervals and initial conditions. The complexity of w ( z ) curves does, as a result, seem to have less of an ef-fect on distinguishability than the intervals in which they FIG. 8. Smooth random dark energy equation of state ( w ( z ))curves generated with Smurves to create mock SN Ia obser-vations for additional experiments. The figure shows curvesfrom two different constraint families, SmurF 2.1 and SmurF4.1, with 50 curve realizations per family. live. This is, again, a consequence of the dependence of m B on the integral over the Hubble parameter, meaningthat faster variations in w ( z ) tend to be smoothed out ob-servationally. Residual additional variations, which arestill present, lead to the slightly higher spread in the cor-responding D KL distribution.When we omit the w (0) = − w ( z ) curves to exhibit stark variationsfrom the ΛCDM case at very low redshifts, we find our-selves confronted with a very different result. Relative toSmurF 4, SmurF 4.1 exhibits larger D KL values with aconsiderably wider spread. We also note that the distri-bution of D KL values is much flatter than for distribu-tions constrained to w (0) = −
1, without a peak at low D KL values. This wider spread and flattened distributioncan be attributed to introducing an offset in our observ-able m B , since m B averages over w ( z ) via the Hubbleparameter. Curves like those in SmurF 4.1 can, for ex-ample, always lie above or below -1, with an additionaloffset of varying magnitude depending on its w (0) value,leading to a posterior very different from the ΛCDM case.Intuitively, choosing random w (0) anchoring points leadsto a roughly flat distribution of D KL values until reachinga maximal possible deviation from ΛCDM that dependson our allowed w (0) prior range. V. CONCLUSION
Searching for new physics beyond the standard ΛCDMmodel inherently requires the capability to discriminatebetween competing models for the dark energy equationof state. This work scrutinizes the pitfalls of standardcosmological analysis pipelines in their ability to detect0
FIG. 9. Histograms of the Kullback-Leibler divergence( D KL ) for different constraint families. The histograms showthe distributions of D KL values, with a total of 50 redshift-dependent dark energy of state curves w ( z ) per family. Indoing so, this figure facilitates the comparison of two previ-ous constraint families, SmurF 2 and SmurF 4, with furtherrelaxed constraint families, namely SmurF 2.1 and SmurF 4.1,as well as with the ΛCDM case. signals of ΛCDM deviations.For this task, we introduce a novel smooth randomcurve generator, Smurves , which uses random samplingand modified Newtonian projectile motion as the meansfor its generative process. This method is highly cus-tomizable and facilitates the use of physically motivatedconstraints into the curve-generating process. Whileapplied to a specific cosmological case in this paper,
Smurves represents a general multi-purpose methodologyfor constrained curve generation and function perturba-tion. We also provide a user-friendly implementation ofthe code for the sake of reproducible science.We employ
Smurves to generate mock SN Ia observa-tions representing four constraint families, or SmurFs,each one representing increasing degrees of deviationfrom the ΛCDM model. Making use of 50 random w ( z )curves per SmurF, we run a Bayesian cosmological in-ference pipeline for each curve to subsequently produce200 joint posteriors of Ω m , w , and M . We then com-pare these posteriors to those from an analysis of the Pantheon sample derived under the assumption of aconstant- w model.We show that SN Ia cosmology observables under ex-tensive redshift dependencies of the dark energy equa-tion of state are virtually indistinguishable from thoseof ΛCDM models using current state-of-the-art analy-sis pipelines. Notably, w ( z ) realizations that exhibit a stronger deviation from w = − m , w , and M exhibiting a slightly betteragreement with ΛCDM than realizations with lesser lev-els of deviation. This result highlights a fundamentaland generally unstated caveat underpinning the currentmethodology used to estimate w from SN Ia observations:If ΛCDM is assumed as the null hypothesis in a test forcompatibility with observational SN Ia data, the inabilityto rule out the standard model could, in a given case, bebased on such similarities in posteriors with potentiallylarge underlying deviations due to statistical degenera-cies.In addition, we test the effect of both an increasednumber of gradient sign changes, leading to more com-plex curves, and of larger deviations from w ( z ) = − w (0) = − z = 0reduces ΛCDM compliance considerably. We recommendfurther research on the topic, specifically in terms of aninvestigation focused on different curve characteristics toreduce the set of viable candidate hypotheses. In doingso, further insights into the specific features of redshift-dependent dark energy equations of state can be gainedby identifying regions of w ( z ) parametrizations that favorcertain cosmologies.The upcoming arrival of larger and higher-qualitydata sets, especially at high redshifts, will certainly im-prove our capability to distinguish between dark energymodels. There are, however, intrinsic characteristics ofdistance-based observables that can render the identifi-cation of strong deviations unattainable. The applica-tion of redshift-dependent analyses, parametric or non-parametric, alongside the constant- w scenario and thecareful use of additional cosmological observables, arecrucial steps in providing a realistic picture of our currentknowledge regarding properties of dark energy. Due tothese caveats, and given the significant loss in precisionwhen redshift-dependence is taken into account, physicsbeyond the standard model may be hidden in plain sight. ACKNOWLEDGEMENTS
We would like to express our gratitude to Eric D.Feigelson, Arya Farahi, and Alberto Krone-Martins forhelpful discussions and suggestions, as well as to Bran-don Sanderson for the initial inspiration for the curvegenerator used in this work. This work was created dur-ing the 5 th COIN Residence Program (CRP (COIN) is a non-profit orga-nization whose aim is to nourish the synergy between astrophysics, cosmology, statistics, and machine learningcommunities. [1] G. Gamow, Nature , 680 (1948).[2] R. H. Dicke, P. J. E. Peebles, P. G. Roll, and D. T.Wilkinson, Astrophys. J. , 414 (1965).[3] A. A. Penzias and R. W. Wilson,Astrophys. J. , 419 (1965).[4] J. C. Mather, E. S. Cheng, R. E. Eplee, Jr., R. B. Isaac-man, S. S. Meyer, R. A. Shafer, R. Weiss, E. L. Wright, et al. , Astrophys. J. Lett. , L37 (1990).[5] G. F. Smoot, C. L. Bennett, A. Kogut, E. L. Wright,J. Aymon, N. W. Boggess, E. S. Cheng, G. de Amici, et al. , Astrophys. J. Lett. , L1 (1992).[6] D. N. Spergel, R. Bean, O. Dor´e, M. R. Nolta, C. L.Bennett, J. Dunkley, G. Hinshaw, N. Jarosik, et al. ,Astrophys. J. Suppl. Ser. , 377 (2007).[7] Planck Collaboration, R. Adam, P. A. R. Ade,N. Aghanim, Y. Akrami, M. I. R. Alves, F. Arg¨ueso,M. Arnaud, et al. , Astron. Astrophys. , A1 (2016).[8] A. G. Riess, A. V. Filippenko, P. Challis, A. Clocchiatti,A. Diercks, P. M. Garnavich, R. L. Gilliland, C. J. Hogan, et al. , Astron. J. , 1009 (1998).[9] S. Perlmutter, G. Aldering, G. Goldhaber, R. A. Knop,P. Nugent, P. G. Castro, S. Deustua, S. Fabbro, et al. ,Astrophys. J. , 565 (1999).[10] P. J. Peebles and B. Ratra,Rev. Mod. Phys. , 559 (2003).[11] R. H. Cyburt, B. D. Fields, K. A. Olive, and T. H. Yeh,Rev. Mod. Phys. , 015004 (2016).[12] R. S. de Souza, C. Iliadis, and A. Coc,Astrophys. J. (2019), 10.3847/1538-4357/aafda9.[13] R. S. de Souza, S. R. Boston, A. Coc, and C. Iliadis,Phys. Rev. C (2019), 10.1103/PhysRevC.99.014619.[14] P. Noterdaeme, P. Petitjean, R. Srianand, C. Ledoux,and S. L´opez, Astron. Astrophys. , L7 (2011).[15] G. Luzzi, R. T. G´enova-Santos, C. J. A. P.Martins, M. De Petris, and L. Lamagna,J. Cosmol. Astropart. Phys. , 011 (2015).[16] A. G. Riess, L.-G. Strolger, S. Casertano, H. C. Ferguson,B. Mobasher, B. Gold, P. J. Challis, A. V. Filippenko, et al. , Astrophys. J. , 98 (2007).[17] W. M. Wood-Vasey, G. Miknaitis, C. W. Stubbs, S. Jha,A. G. Riess, P. M. Garnavich, R. P. Kirshner, C. Aguil-era, et al. , Astrophys. J. , 694 (2007).[18] R. Amanullah, C. Lidman, D. Rubin, G. Aldering,P. Astier, K. Barbary, M. S. Burns, A. Conley, et al. ,Astrophys. J. , 712 (2010).[19] E. Komatsu, K. M. Smith, J. Dunkley, C. L. Bennett,B. Gold, G. Hinshaw, N. Jarosik, D. Larson, et al. ,Astrophys. J. Suppl. Ser. , 18 (2011).[20] M. Sullivan, J. Guy, A. Conley, N. Regnault,P. Astier, C. Balland, S. Basa, R. G. Carlberg, et al. ,Astrophys. J. , 102 (2011).[21] N. Suzuki, D. Rubin, C. Lidman, G. Aldering, R. Aman-ullah, K. Barbary, L. F. Barrientos, J. Botyanszki, et al. ,Astrophys. J. , 85 (2012). https://cosmostatistics-initiative.org [22] L. Anderson, E. Aubourg, S. Bailey, D. Bizyaev, M. Blan-ton, A. S. Bolton, J. Brinkmann, J. R. Brownstein, et al. ,Mon. Not. R. Astron. Soc. , 3435 (2012).[23] J. A. Frieman, M. S. Turner, and D. Huterer,Annu. Rev. Astron. Astrophys. , 385 (2008).[24] C. O’Raifeartaigh, M. O’Keeffe, W. Nahm, and S. Mit-ton, Eur. Phys. J. H , 73 (2018).[25] A. Einstein, Sitzungsberichte der K¨oniglich PreußischenAkademie der Wissenschaften (Berlin) , 142 (1917).[26] A. Friedman, Z. Astrophys. , 377 (1922), [Gen. Rel.Grav. 31, 1991(1999)].[27] G. H. J. E. Lemaˆıtre, The gravitational field in a fluidsphere of uniform invariant density according to the the-ory of relativity , Ph.D. thesis, MIT (1927).[28] M. Chevallier and D. Polarski,Int. J. Mod. Phys. D , 213 (2001).[29] E. V. Linder, Phys. Rev. Lett. , 091301 (2003).[30] P. J. E. Peebles and B. Ratra,Astrophys. J. Lett. , L17 (1988).[31] R. Gannouji, D. Polarski, A. Ranquet, and A. A.Starobinsky, J. Cosmol. Astropart. Phys. , 016 (2006).[32] E. J. Copeland, M. Sami, and S. Tsujikawa,Int. J. Mod. Phys. D , 1753 (2006).[33] P. M. Garnavich, S. Jha, P. Challis, A. Clocchiatti,A. Diercks, A. V. Filippenko, R. L. Gilliland, C. J.Hogan, et al. , Astrophys. J. , 74 (1998).[34] S. Hannestad and E. Mortsell,J. Cosmol. Astropart. Phys. , 001 (2004).[35] R. Ch´avez, M. Plionis, S. Basilakos, R. Terlevich, E. Ter-levich, J. Melnick, F. Bresolin, and A. L. Gonz´alez-Mor´an, Mon. Not. R. Astron. Soc. , 2431 (2016).[36] A. Tripathi, A. Sangwan, and H. K. Jassal,J. Cosmol. Astropart. Phys. , 012 (2017).[37] C. Genovese, P. Freeman, L. Wasserman, R. Nichol, andC. Miller, Ann. Appl. Stat. , 144 (2009).[38] D. Huterer and D. L. Shafer,Rep. Prog. Phys. , 016901 (2018).[39] D. O. Jones, D. M. Scolnic, A. G. Riess, A. Rest,R. P. Kirshner, E. Berger, R. Kessler, Y.-C. Pan, et al. ,Astrophys. J. , 51 (2018).[40] D. O. Jones, D. M. Scolnic, R. J. Foley, A. Rest,R. Kessler, P. J. Challis, K. C. Chambers, D. A. Coulter, et al. , arXiv:1811.09286.[41] H. K. Jassal, J. S. Bagla, and T. Padmanabhan,Mon. Not. R. Astron. Soc. , L11 (2005).[42] A. De Felice, S. Nesseris, and S. Tsujikawa,J. Cosmol. Astropart. Phys. , 029 (2012).[43] G.-B. Zhao, D. Huterer, and X. Zhang,Phys. Rev. D , 121302 (2008).[44] P. Serra, A. Cooray, D. E. Holz, A. Melchiorri, S. Pan-dolfi, and D. Sarkar, Phys. Rev. D , 121302 (2009).[45] J. A. V´azquez, M. Bridges, M. P. Hobson, and A. N.Lasenby, J. Cosmol. Astropart. Phys. , 020 (2012).[46] S. Hee, J. A. V´azquez, W. J. Hand-ley, M. P. Hobson, and A. N. Lasenby,Mon. Not. R. Astron. Soc. , 369 (2017). [47] T. Holsclaw, U. Alam, B. Sans´o, H. Lee,K. Heitmann, S. Habib, and D. Higdon,Phys. Rev. D , 103502 (2010).[48] T. Holsclaw, U. Alam, B. Sans´o, H. Lee,K. Heitmann, S. Habib, and D. Higdon,Phys. Rev. Lett. , 241302 (2010).[49] R. G. Crittenden, G.-B. Zhao,L. Pogosian, L. Samushia, and X. Zhang,J. Cosmol. Astropart. Phys. , 048 (2012).[50] T. M. C. Abbott, S. Allam, P. Andersen, C. Angus,J. Asorey, A. Avelino, S. Avila, B. A. Bassett, et al. ,(), arXiv:1811.02374.[51] D. M. Scolnic, D. O. Jones, A. Rest, Y. C. Pan,R. Chornock, R. J. Foley, M. E. Huber, R. Kessler, et al. ,Astrophys. J. , 101 (2018).[52] C. E. Rasmussen and C. K. I. Williams, Gaussian pro-cesses for machine learning (adaptive computation andmachine learning) (MIT Press, Cambridge, 2005).[53] S. Dodelson,
Modern cosmology (Academic Press, SanDiego, 2003).[54] R. R. Caldwell, M. Kamionkowski, and N. N. Weinberg,Phys. Rev. Lett. , 071301 (2003).[55] S. M. Carroll, M. Hoffman, and M. Trodden,Phys. Rev. D , 023509 (2003).[56] R. Arjona, W. Cardona, and S. Nesseris,arXiv:1811.02469.[57] F. Koennig, Y. Akrami, L. Amendola, M. Motta, andA. R. Solomon, Phys. Rev. D , 124014 (2014).[58] J. Zuntz, M. Paterno, E. Jennings, D. Rudd, A. Man-zotti, S. Dodelson, S. Bridle, S. Sehrish, andJ. Kowalkowski, Astron. Comput. , 45 (2015).[59] T. M. C. Abbott, F. B. Abdalla, A. Alarcon, J. Aleksi´c,S. Allam, S. Allen, A. Amara, J. Annis, et al. ,Phys. Rev. D , 043526 (2018).[60] T. M. C. Abbott, A. Alarcon, S. Allam, P. Andersen,F. Andrade-Oliveira, J. Annis, J. Asorey, A. Avelino, et al. , (), arXiv:1811.02375.[61] T. M. C. Abbott, F. B. Abdalla, J. An-nis, K. Bechtol, J. Blazek, B. A. Benson,R. A. Bernstein, G. M. Bernstein, et al. ,Mon. Not. R. Astron. Soc. , 3879 (2018).[62] J. Elvin-Poole, M. Crocce, A. J. Ross, T. Giannanto-nio, E. Rozo, E. S. Rykoff, S. Avila, N. Banik, et al. ,Phys. Rev. D , 042006 (2018).[63] M. A. Troxel, N. MacCrann, J. Zuntz, T. F. Eifler,E. Krause, S. Dodelson, D. Gruen, J. Blazek, et al. ,Phys. Rev. D , 043528 (2018).[64] A. Barreira, B. Li, E. Jennings, J. Merten,L. King, C. M. Baugh, and S. Pascoli,Mon. Not. R. Astron. Soc. , 4085 (2015).[65] I. Harrison, S. Camera, J. Zuntz, and M. L. Brown,Mon. Not. R. Astron. Soc. , 3674 (2016).[66] E. Krause and T. Eifler,Mon. Not. R. Astron. Soc. , 2100 (2017).[67] W. Lin and M. Ishak, Phys. Rev. D , 083532 (2017).[68] J. Goodman and J. Weare,Commun. Appl. Math. Comput. Sci. , 65 (2010).[69] D. Foreman-Mackey, D. W. Hogg, D. Lang, and J. Good-man, Publ. Astron. Soc. Pac. , 306 (2013).[70] A. Lewis, A. Challinor, and A. Lasenby,Astrophys. J. , 473 (2000).[71] A. G. Riess, L. M. Macri, S. L. Hoffmann, D. Scolnic,S. Casertano, A. V. Filippenko, B. E. Tucker, M. J. Reid, et al. , Astrophys. J. , 56 (2016). [72] R. J. Cooke, M. Pettini, R. A. Jorgenson, M. T. Murphy,and C. C. Steidel, Astrophys. J. , 31 (2014).[73] M. Betoule, R. Kessler, J. Guy, J. Mosher,D. Hardin, R. Biswas, P. Astier, P. El-Hage, et al. ,Astron. Astrophys. , A22 (2014).[74] S. Kullback and R. A. Leibler,Ann. Math. Statist. , 79 (1951).[75] M. Kilbinger, D. Wraith, C. P. Robert, K. Benabed,O. Capp´e, J.-F. Cardoso, G. Fort, S. Prunet, et al. ,Mon. Not. R. Astron. Soc. , 2381 (2010).[76] A. Ben-David, H. Liu, and A. D. Jackson,J. Cosmol. Astropart. Phys. , 051 (2015).[77] R. S. De Souza, M. L. L. Dantas, M. V. Costa-Duarte, E. D. Feigelson, M. Killedar, P.-Y.Lablanche, R. Vilalta, A. Krone-Martins, et al. ,Mon. Not. R. Astron. Soc. , 2808 (2017).[78] A. I. Malz, P. J. Marshall, M. L. DeRose, J. Graham,S. J. Schmidt, R. Wechsler, and (LSST Dark EnergyScience Collaboration), Astron. J. , 35 (2018).[79] A. Nicola, A. Amara, and A. Refregier,arXiv:1809.07333.[80] D. Brout, M. Sako, D. Scolnic, R. Kessler, C. B.D’Andrea, T. M. Davis, S. R. Hinton, A. G. Kim, et al. ,arXiv:1811.02378.[81] V. Miranda and C. Dvorkin,Phys. Rev. D D98 , 043537 (2018).[82] G.-B. Zhao et al. , Nat. Astron. , 627 (2017).[83] B. L’Huillier, A. Shafieloo, E. V. Linder, and A. G. Kim,Mon. Not. R. Astron. Soc. , 2783 (2019).[84] J. Sol`a Peracaula, A. G´omez-Valent, and J. de Cruz P´erez,Phys. Dark Universe (2019), 10.1016/j.dark.2019.100311.[85] D. Copeland, A. Taylor, and A. Hall,Mon. Not. R. Astron. Soc. , 2247 (2018). Appendix A: Constrained curve generation
The segmented path calculation of
Smurves follows, inits broadest terms, the classical Newtonian calculation ofa projectile path: Given a velocity, an acceleration mag-nitude as a force acting on the projectile, and a launchangle, a flight path can be easily computed as verticalaxis values along a set of measurement points on the hor-izontal axis. At the end of the partial path computation,the function returns the path measurements, the impactangle, and the final velocity of the projectile. Dependingon the number of sampled change points, and on whetherparts of the full path are not yet calculated, a new forceacting in the opposite direction of the previous one issampled, and previously returned values are re-used asinputs to the same function. This lets the projectile con-tinue its flight with the same characteristics, but withchanged gravitational magnitude and direction, to en-sure a smooth curve evolution that easily lends itself tosubsequent splining.The corresponding method for curve segment calcula-tions is specified, as pseudocode, in Algorithm 1 to allowfor an easier replication and easier understanding bothof our approach and the accompanying open-source codeimplementation for smooth random curve generation.3
Algorithm 1:
Partial trajectory calculation
Data: v := velocity α := step size β := direction s := partial steps p := start point f := vertical force θ := launch angle Result:
Path p , impact angle θ imp , velocity v Set the initial horizontal displacement to zero ∆ x ←− Calculate the horizontal and vertical velocities v x ←− v cos( θ ) v y ←− v sin( θ ) Initialize start velocity and path measurements v ←− vp ←− p Loop over the given x-axis measurement points for i ← to length( s ) do Horizontal distance, displacement and time d ←− s [ i ]∆ x ←− ∆ x + αt ←− ∆ xv x Calculate vertical velocity and displacement v y ←− v sin( θ ) − ft ∆ y ←− − (cid:0) v sin( θ ) t − ft (cid:1) Total velocity and directional displacement v ←− p v x + v y D ←− β ∆ x Append the projectile location at that point p ←− append( p, ( d, p + D )) end Calculate the impact angle for the partial path θ imp ←− arctan( − v y v x ) return p, θ imp , v Appendix B: Parameter estimation with emcee
For our parameter estimation, we employ emcee ,a popular pure-Python implementation of the affine-invariant MCMC ensemble sampler. This approach ex-tends the classic Metropolis-Hastings algorithm with aparallel “stretch move”.A number K of walkers explore the parameter space,with their respective steps drawn from a proposal dis-tribution that depends on other walkers’ positions. Awalker at position Y is drawn by chance to propose anew position X ′ for the walker that is to be updated andcurrently at position X , meaning that X → X ′ = Y + Z [ X − Y ] . (B1)Here, Z acts as a random variable with S := [0 . ,
2] and Z ∼ g ( z ) ∝ S ( z ) · √ z − , with the indicator function S ( z ) taking a value of one for all z ∈ S and a value ofzero for all z / ∈ S . Alternatively, this can be written as g ( z ) ∝ ( √ z if z ∈ (cid:2) , (cid:3) . (B2) The “parallel stretch” mentioned above splits the K walkers into two equal-sized subsets and updates all walk-ers of one subset using the other, followed by the corre-sponding opposite procedure, which allows for the paral-lelization of this computationally expensive update step.An affine-invariant MCMC algorithm satisfies X a ( t ) = AX b ( t ) + b for different starting points X a and X b , andtwo probability densities π and π A,b , for any affine trans-formation Ax + b . The independence of the aspect ratio inhighly anisotropic distributions offers a speed advantagein highly skewed distributions. Appendix C:
CosmoSIS noise addition and bug fix
CosmoSIS generates simulations of peak B-band mag-nitudes as m B ( z i ) double arrays based on binned Pan-theon
SN Ia data. From the
Pantheon noise covari-ance C ≡ h nn T i , we can generate this simulation usingits (unique) Cholesky decomposition C = LL T and arandom vector r , where each element is a random nor-mal value with r i ∼ N (0 , n = L · r as our noise simulation, as the noise covariance is then h nn T i = h Lrr T L T i = h LL T i = C. (C1)As a consequence, the total simulated values m B sim ob-tained through CosmoSIS are m B sim = m B truth ( z i ) + L · r, (C2)for true values m B truth . Initial experiments to comparethe original Pantheon data with SN Ia data generatedusing flat w ( z ) curves as a null test uncovered a bugin CosmoSIS . After this was reported and subsequentlyfixed, the flat-curve simulations of SN Ia peak B-bandmagnitudes returned to expected values of m B . Appendix D: Interpretation of posterior samples
Given the way in which posteriors from w ( z ) curverealizations from the same constraint family are used inthis paper, one might ask why posterior samples obtainedfrom instances of the same SmurF are not simply com-bined to arrive at a posterior for the constraint family.Considering error contours as being comprised of samplesfrom p ( θ | D, Ψ), as introduced in Section III C, neglectsthe role of the initial conditions C that have been im-plicitly marginalized out as p ( θ | D, Ψ) = Z p ( C , θ | D, Ψ) dC . (D1)Since we generally cannot constrain the initial conditionsas such, an obvious question to ask is why they matter.When combining constraints on cosmological parame-ters from different probes D and D ′ , we are really askingfor p ( θ | D, D ′ , Ψ) when we have p ( θ | D, Ψ) and p ( θ | D ′ , Ψ).4To make use of the independence of the datasets, wewould expand this in terms of Bayes’ Rule as p ( θ | D, D ′ , Ψ) = Z p ( C , θ | D, D ′ , Ψ) dC (D2)= Z p ( D, D ′ | C , θ, Ψ) p ( C , θ | Ψ) p ( D, D ′ | Ψ) dC . (D3)If D and D ′ are our standard independent probes, everyterm in Equation D2 is well-defined. This means thatthe integral is separable and we can recover the intuitiveway to combine the posteriors.The situation investigated in this paper, however, isdifferent. In our case, D and D ′ correspond to differentSmurF instances j and j ′ . These two datasets are inher- ently contradictory; they could never be observed in thesame instantiation of the universe, even under the samephysical model and values of the cosmological parameters θ . In other words, p ( D, D ′ | C ) = 0 for any pair of mock- Pantheon data we consider. What distinguishes oneSmurF from another is rolled into the initial conditions C , leading to well-defined p ( θ | D, Ψ) and p ( θ | D ′ , Ψ), butto an internally inconsistent p ( θ | D, D ′ , Ψ). Thus, itwould be inappropriate to combine samples of the cos-mological parameters obtained through a Markov chainMonte Carlo (MCMC) method from any collection ofSmurF instances with different w ( zz