Literature review: error structure in interatomic
\npotential predictions
\nClassical force fields for metals exhibit structured, low-dimensional prediction errors that
\ncan be understood through the convergence of three distinct intellectual traditions: sloppy
\nmodel theory from statistical physics, Simpson’s paradox from meta-analytic methodology,
\nand systematic interatomic potential benchmarking. This review assembles the key
\nreferences across these areas to support a paper showing that FCC elastic constant errors
\n(C11, C12, C44) occupy a manifold with effective dimensionality ~1.66/3, while pooled BCC
\ndata produce a spurious correlation inversion attributable to Simpson’s paradox.
\nArea 1: Sloppy model theory and the geometry of poorly
\nconstrained models
\nFoundational framework
\nThe sloppy model framework emerged from computational biology and statistical physics,
\nestablishing that multiparameter models generically have parameter sensitivities spanning
\nmany orders of magnitude. Brown and Sethna (2003) introduced the concept, showing
\nthat biochemical signaling models with ~48 parameters have Hessian eigenvalue spectra
\nspanning many decades, with only a few “stiff” parameter combinations controlling
\npredictions and many “sloppy” directions allowing parameters to vary freely (Brown KS,
\nSethna JP, “Statistical mechanical approaches to models with many poorly known
\nparameters,” Physical Review E 68, 021904, 2003). PubMed
\nWaterfall et al. (2006) demonstrated that sloppiness constitutes a universality class by
\nconnecting the characteristic eigenvalue spectrum to Vandermonde matrix ensembles. The
\nlogarithms of eigenvalues are roughly uniformly spaced — a signature now recognized
\nacross physics, biology, and engineering. PubMed American Physical Society Crucially, this
\npaper explicitly noted that sloppiness had also been demonstrated “in three multiparameter
\ninteratomic potentials fit to electronic structure” Cornell (Waterfall JJ, Casey FP,
\nGutenkunst RN, Brown KS, Myers CR, Brouwer PW, Elser V, Sethna JP, “Sloppy-model
\nuniversality class and the Vandermonde matrix,” Physical Review Letters 97, 150601, 2006).
\nGutenkunst et al. (2007) tested 17 systems biology models and found that every one
\nexhibited sloppy parameter sensitivities, cementing the universality claim. They argued that
\ncollective fits yield well-constrained predictions even when individual parameters remain
\npoorly determined, and that direct parameter measurements must be “formidably precise
\nand complete” to be more useful than collective fitting PubMed (Gutenkunst RN, Waterfall
\nJJ, Casey FP, Brown KS, Myers CR, Sethna JP, “Universally sloppy parameter sensitivities in
\nsystems biology models,” PLoS Computational Biology 3(10), e189, 2007).
\nGeometric and information-theoretic development
\nThe geometric interpretation of sloppiness matured through a series of papers by
\nTranstrum, Machta, and Sethna. Transtrum et al. (2010) introduced the model manifold
\nperspective: the set of all possible predictions as parameters vary forms a “hyper-ribbon” in
\ndata space with a geometric hierarchy of widths. This narrowness explains both why fitting
\nis computationally difficult and why predictions are low-dimensional arXiv (Transtrum MK,
\nMachta BB, Sethna JP, “Why are nonlinear fits to data so challenging?” Physical Review
\nLetters 104, 060201, 2010). Transtrum et al. (2011) provided the comprehensive
\ngeometric treatment, showing that model manifolds universally exhibit a geometric series of
\nwidths, extrinsic curvatures, and parameter-effect curvatures American Physical Society
\n(Transtrum MK, Machta BB, Sethna JP, “Geometry of nonlinear least squares with
\napplications to sloppy models and optimization,” Physical Review E 83, 036701, 2011).
\nMachta et al. (2013) connected sloppiness to fundamental physics by showing that
\nparameter space compression underlies emergent theories. Using the Fisher Information
\nMatrix (FIM) as a Riemannian metric on parameter space, they demonstrated that the
\nemergence of effective theories — continuum limits, renormalization group fixed points —
\ncorresponds to compression from many microscopic parameters to few macroscopic ones.
\nAIP Publishing PubMed Stiff FIM directions become emergent parameters; sloppy
\ndirections correspond to irrelevant microscopic detail arXiv (Machta BB, Chachra R,
\nTranstrum MK, Sethna JP, “Parameter space compression underlies emergent theories and
\npredictive models,” Science 342(6158), 604–607, 2013). This framework directly implies
\nthat the ~1.66 effective dimensions for elastic constant errors correspond to approximately
\ntwo emergent combinations of potential parameters dominating elastic response.
\nTranstrum et al. (2014) introduced the Manifold Boundary Approximation Method (MBAM)
\nfor systematic model reduction by following geodesics to the manifold boundary, removing
\nsloppy parameter combinations while preserving stiff predictions PubMed (Transtrum MK,
\nQiu P, “Model reduction by manifold boundaries,” Physical Review Letters 113, 098701,
\n2014). The comprehensive review by Transtrum et al. (2015) synthesized the full program,
\nincluding FIM-based analysis, hyperribbon structure, MBAM, and connections to emergent
\ntheories across physics and biology AIP Publishing ADS (Transtrum MK, Machta BB, Brown
\nKS, Daniels BC, Myers CR, Sethna JP, “Perspective: Sloppiness and emergent theories in
\nphysics, biology, and beyond,” Journal of Chemical Physics 143(1), 010901, 2015).
\nQuinn et al. (2019) provided the first rigorous mathematical explanation for sloppiness,
\nusing Chebyshev approximation theory to derive universal bounds on model manifold widths
\nas a consequence of model smoothness alone American Physical Society (Quinn KN, Wilber H,
\nTownsend A, Sethna JP, “Chebyshev approximation and the global geometry of model
\npredictions,” Physical Review Letters 122, 158302, 2019). The most recent comprehensive
\nreview is Quinn et al. (2023), covering model manifold structure, MBAM, information
\ntopology, optimal Bayesian priors, and visualization via intensive PCA PubMed Central
\n(Quinn KN, Abbott MC, Transtrum MK, Machta BB, Sethna JP, “Information geometry for
\nmultiparameter models: New perspectives on the origin of simplicity,” Reports on Progress
\nin Physics 86, 035901, 2023).
\nApplications to interatomic potentials
\nThe seminal application of sloppy model theory to force fields is Frederiksen et al. (2004),
\nwhich developed a Bayesian ensemble approach for estimating prediction errors from
\ninteratomic potentials. Working with EAM-type potentials for molybdenum fitted to DFT
\nforce databases, they showed that the potentials exhibit sloppy eigenvalue spectra identical
\nin character to biological models, and that Bayesian error bars on elastic constants,
\ngamma-surface energies, structural energies, and dislocation properties provided realistic
\nuncertainty estimates. PubMed Parameters varied wildly across the ensemble while
\npredictions remained constrained Cornell — the hallmark of low-dimensional prediction
\nstructure PLOS (Frederiksen SL, Jacobsen KW, Brown KS, Sethna JP, “Bayesian ensemble
\napproach to error estimation of interatomic potentials,” Physical Review Letters 93, 165501,
\n2004).
\nWen et al. (2017) applied Fisher information theory to a Stillinger-Weber potential for MoS₂,
\ncomputing the FIM eigenvalue spectrum and verifying parameter identifiability. They used
\nthe geodesic Levenberg-Marquardt algorithm for fitting and demonstrated that the FIM
\nanalysis provides uncertainty bounds for elastic constants and other predicted properties
\n(Wen M, Li J, Brommer P, Elliott RS, Sethna JP, Tadmor EB, “A force-matching Stillinger-
\nWeber potential for MoS₂: Parameterization and Fisher information theory based sensitivity
\nanalysis,” Journal of Applied Physics 122, 244301, 2017).
\nKurniawan et al. (2022) provided the most thorough study of parametric uncertainty in
\nclassical potentials using the OpenKIM framework. Working with Lennard-Jones, Morse, and
\nStillinger-Weber potentials, they confirmed that interatomic potentials are sloppy models
\narXiv with bounded manifolds exhibiting a hierarchy of widths DNTB and low effective
\ndimensionality. arXiv Many parameter combinations are unidentifiable, ResearchGate yet
\npredictions remain constrained (Kurniawan Y, Petrie CL, Williams KJ, Transtrum MK, Tadmor
\nEB, Elliott RS, Karls DS, Wen M, “Bayesian, frequentist, and information geometric
\napproaches to parametric uncertainty quantification of classical empirical interatomic
\npotentials,” Journal of Chemical Physics 156(21), 214103, 2022).
\nMortensen et al. (2005) extended the Bayesian sloppy-model framework to density
\nfunctional theory itself, showing that DFT exchange-correlation functionals are also sloppy
\n— sloppiness pervades the entire hierarchy of computational materials science (Mortensen
\nJJ, Kaasbjerg K, Frederiksen SL, Nørskov JK, Sethna JP, Jacobsen KW, “Bayesian error
\nestimation in density-functional theory,” Physical Review Letters 95, 216401, 2005).
\nNeural networks and modern potentials
\nMao et al. (2024) showed that deep neural networks with widely different architectures,
\nsizes, optimizers, and regularization all traverse the same low-dimensional manifold
\nduring training, with hyper-ribbon structure characteristic of sloppy models. ADS This
\nsuggests that neural network potentials (NequIP, MACE, etc.) should also produce low-
\ndimensional elastic constant error patterns regardless of architecture (Mao J, Griniasty I,
\nTeoh HK, Ramesh R, Yang R, Transtrum MK, Sethna JP, Chaudhari P, “The training process
\nof many deep networks explores the same low-dimensional manifold,” Proceedings of the
\nNational Academy of Sciences 121(12), e2310002121, 2024). An analytical treatment
\nfollowed in Mao et al. (2026), deriving conditions for hyper-ribbon emergence in linear
\nmodels and characterizing phase boundaries (Mao J, Griniasty I, Sun Y, Transtrum MK,
\nSethna JP, Chaudhari P, “Analytical characterization of sloppiness in neural networks:
\nInsights from linear models,” Physical Review E 113, 015306, 2026).
\nArea 2: Simpson’s paradox and meta-analytic methodology for
\ngrouped correlations
\nFoundational references
\nSimpson (1951) demonstrated that collapsing two-way contingency tables can reverse
\nassociations, establishing the foundational paradox Wiley Online Library (Simpson EH, “The
\ninterpretation of interaction in contingency tables,” Journal of the Royal Statistical Society:
\nSeries B 13(2), 238–241, 1951). Blyth (1972) formalized the paradox and connected it to
\nSavage’s sure-thing principle, Wikipedia providing the probabilistic framework (Blyth CR,
\n“On Simpson’s paradox and the sure-thing principle,” Journal of the American Statistical
\nAssociation 67(338), 364–366, 1972).
\nThe most celebrated real-world demonstration is the Berkeley admissions case (Bickel et
\nal., 1975), where aggregate data showed apparent bias against women that reversed at the
\ndepartment level Science ResearchGate — women applied disproportionately to
\ncompetitive departments, a confounding structure directly analogous to pooling elastic
\nconstant errors across elements with different physical properties (Bickel PJ, Hammel EA,
\nO’Connell JW, “Sex bias in graduate admissions: Data from Berkeley,” Science 187(4175),
\n398–404, 1975).
\nPearl (2014) argued that Simpson’s paradox can only be resolved through causal
\nreasoning, introducing the do-calculus and back-door criterion for determining whether
\npooled or stratified analysis gives the correct inference. In the interatomic potential context,
\nelement identity acts as a causal confounder through distinct physical properties — atomic
\nradius, electron configuration, bonding character — that independently influence both the
\npotential’s error structure and the target elastic constants (Pearl J, “Comment:
\nUnderstanding Simpson’s paradox,” The American Statistician 68(1), 8–13, 2014; also Pearl
\nJ, Causality: Models, Reasoning, and Inference, 2nd ed., Cambridge University Press,
\n2009).
\nRobinson (1950) established the closely related ecological fallacy, showing that state-level
\ncorrelations between demographic variables can completely reverse from individual-level
\ncorrelations. Wikipedia The pooled interatomic potential error correlation across elements
\nis precisely this type of “ecological correlation” (Robinson WS, “Ecological correlations and
\nthe behavior of individuals,” American Sociological Review 15(3), 351–357, 1950).
\nCorrelation testing methodology
\nJackson and Somers (1991) is the key methodological reference for handling spurious
\ncorrelations in grouped data. They demonstrated that ratios and indices sharing common
\ncomponents generate correlations even from random data, and advocated randomization
\n(permutation) tests that generate a null distribution accounting for mathematical coupling.
\nTheir central recommendation is that the appropriate null correlation is frequently non-zero,
\nand standard tests against r = 0 are inappropriate when variables share structural
\ncomponents PubMed Springer (Jackson DA, Somers KM, “The spectre of ‘spurious’
\ncorrelations,” Oecologia 86(1), 147–151, 1991. DOI: 10.1007/BF00317404).
\nArchie (1981) established the mathematical coupling problem: when derived variables
\nshare components (e.g., X/Z versus Y/Z), significant correlations arise even from
\nuncorrelated raw data. This is the foundational reference for the non-zero null hypothesis
\nissue — physical constraints among elastic constants (e.g., the Cauchy relation, stability
\nrequirements) create inherent baseline correlations that must not be attributed to
\nsystematic potential errors (Archie JP, “Mathematic coupling of data: A common source of
\nerror,” Annals of Surgery 193(3), 296–303, 1981).
\nKievit et al. (2013) provided a practical detection guide for Simpson’s paradox in
\ncontinuous bivariate data with categorical grouping variables, including an R toolbox and
\nstatistical markers for the paradox. Frontiers Their methods are directly applicable to
\nelastic constant error data grouped by element (Kievit RA, Frankenhuis WE, Waldorp LJ,
\nBorsboom D, “Simpson’s paradox in psychological science: A practical guide,” Frontiers in
\nPsychology 4, Article 513, 2013).
\nRandom-effects meta-analysis for correlations
\nThe standard methodology for properly combining correlations across heterogeneous
\ngroups uses Fisher’s z-transformation within a random-effects framework. Hedges and
\nOlkin (1985) established the foundational approach: transform each group-specific
\ncorrelation using z = arctanh(r), combine via inverse-variance weighting, and back-
\ntransform (Hedges LV, Olkin I, Statistical Methods for Meta-Analysis, Academic Press,
\n1985). Borenstein et al. (2009) is the standard textbook reference, explaining fixed versus
\nrandom-effects models, heterogeneity assessment (Q-statistic, I²
\n, τ²), and interpretation
\n(Borenstein M, Hedges LV, Higgins JPT, Rothstein HR, Introduction to Meta-Analysis, Wiley,
\n2009). Borenstein et al. (2010) clarified when random-effects models are appropriate —
\nspecifically when group-specific effects (element-specific correlations) are expected to
\ndiffer, as they do for different BCC metals (Borenstein M, Hedges LV, Higgins JPT, Rothstein
\nHR, “A basic introduction to fixed-effect and random-effects models for meta-analysis,”
\nResearch Synthesis Methods 1(2), 97–111, 2010).
\nWelz et al. (2022) proposed improved confidence intervals for the Fisher z approach in
\nrandom-effects meta-analysis of correlations, showing that standard intervals can be
\nunsatisfactory and providing enhanced variance estimators (Welz T, Viechtbauer W, Pauly
\nM, “Fisher transformation based confidence intervals of correlations in fixed- and random-
\neffects meta-analysis,” British Journal of Mathematical and Statistical Psychology 75(1), 1–
\n38, 2022). DerSimonian and Laird (1986) provide the most widely used random-effects
\nestimator for between-study variance (DerSimonian R, Laird N, “Meta-analysis in clinical
\ntrials,” Controlled Clinical Trials 7, 177–188, 1986). Higgins et al. (2003) introduced the I²
\nstatistic for quantifying heterogeneity (Higgins JPT, Thompson SG, Deeks JJ, Altman DG,
\n“Measuring inconsistency in meta-analyses,” BMJ 327, 557–560, 2003).
\nSimpson’s paradox in the physical sciences
\nExamples of Simpson’s paradox in hard sciences remain rare, making instances in materials
\nscience especially noteworthy. Selvitella (2017) demonstrated the paradox in quantum
\nmechanics (quantum harmonic oscillator, nonlinear Schrödinger equation) and in geometric
\nand linear algebraic settings (Selvitella A, “The ubiquity of the Simpson’s paradox,” Journal
\nof Statistical Distributions and Applications 4, Article 2, 2017). Chuang et al. (2009)
\n"}