{"id":"references","title":"References & Intellectual Lineage","subtitle":"The 35 external works we build on — each with what it is and why we cite it.","category":"references","tags":["references","bibliography","lineage"],"source":"articles/docs/references.md","lang":"en","words":939,"readMinutes":4,"toc":[{"depth":2,"text":"Sloppy-model theory — foundational","id":"sloppy-model-theory-foundational"},{"depth":2,"text":"Sloppy-model theory — geometric / information-theoretic","id":"sloppy-model-theory-geometric-information-theoretic"},{"depth":2,"text":"Sloppy models applied to interatomic potentials","id":"sloppy-models-applied-to-interatomic-potentials"},{"depth":2,"text":"Modern networks — same sloppy manifold","id":"modern-networks-same-sloppy-manifold"},{"depth":2,"text":"Simpson's paradox & the ecological fallacy","id":"simpson-s-paradox-the-ecological-fallacy"},{"depth":2,"text":"Correlation pitfalls","id":"correlation-pitfalls"},{"depth":2,"text":"Meta-analysis methodology","id":"meta-analysis-methodology"},{"depth":2,"text":"Benchmark infrastructure & reference data","id":"benchmark-infrastructure-reference-data"}],"html":"<h1 id=\"references-amp-intellectual-lineage\">References &amp; Intellectual Lineage</h1><p>The external literature Lupine builds on — what each work is, and <em>why we cite it</em>. This\nis the annotated counterpart to the IMMI paper&#39;s <code>references.bib</code>: the bibliography tells\nyou the source, this tells you the load it bears in the argument.</p>\n<p>Threads: sloppy-model theory → its application to interatomic potentials → the\nSimpson&#39;s-paradox / ecological-fallacy machinery → meta-analysis → the benchmark\ninfrastructure and reference data.</p>\n<hr>\n<h2 id=\"sloppy-model-theory-foundational\">Sloppy-model theory — foundational</h2><ul>\n<li><strong>Brown &amp; Sethna (2003)</strong> — <em>Statistical mechanical approaches to models with many\npoorly known parameters.</em> Phys. Rev. E 68, 021904.\n<a href=\"https://doi.org/10.1103/PhysRevE.68.021904\">doi</a>\n— The origin of &quot;sloppiness&quot;: most parameter combinations barely affect predictions.\nThe conceptual seed of the hyper-ribbon.</li>\n<li><strong>Waterfall et al. (2006)</strong> — <em>Sloppy-model universality class and the Vandermonde\nmatrix.</em> Phys. Rev. Lett. 97, 150601.\n<a href=\"https://doi.org/10.1103/PhysRevLett.97.150601\">doi</a>\n— Establishes sloppiness as a <em>universality class</em> — why we expect the same geometry\nacross unrelated potentials.</li>\n<li><strong>Gutenkunst et al. (2007)</strong> — <em>Universally sloppy parameter sensitivities in systems\nbiology models.</em> PLoS Comput. Biol. 3, e189.\n<a href=\"https://doi.org/10.1371/journal.pcbi.0030189\">doi</a>\n— Cross-domain evidence that the eigenvalue spectrum is generic, not model-specific.</li>\n</ul>\n<h2 id=\"sloppy-model-theory-geometric-information-theoretic\">Sloppy-model theory — geometric / information-theoretic</h2><ul>\n<li><strong>Transtrum, Machta &amp; Sethna (2010, 2011)</strong> — <em>Why are nonlinear fits so challenging?</em>\n/ <em>Geometry of nonlinear least squares.</em> PRL 104, 060201; PRE 83, 036701.\n<a href=\"https://doi.org/10.1103/PhysRevE.83.036701\">doi</a>\n— The model manifold as a bounded <strong>hyper-ribbon</strong> with a hierarchy of widths — the\nobject Lupine measures in error space.</li>\n<li><strong>Machta et al. (2013)</strong> — <em>Parameter space compression underlies emergent theories.</em>\nScience 342, 604. <a href=\"https://doi.org/10.1126/science.1238723\">doi</a>\n— Why low effective dimensionality is <em>expected</em>, not coincidental.</li>\n<li><strong>Transtrum &amp; Qiu (2014)</strong> — <em>Model reduction by manifold boundaries.</em> PRL 113, 098701.\n<a href=\"https://doi.org/10.1103/PhysRevLett.113.098701\">doi</a>\n— The Manifold Boundary Approximation Method — the route from &quot;sloppy&quot; to a reduced\npredictive model (the long-term retraining target).</li>\n<li><strong>Transtrum et al. (2015)</strong> — <em>Perspective: Sloppiness and emergent theories.</em>\nJ. Chem. Phys. 143, 010901. <a href=\"https://doi.org/10.1063/1.4923066\">doi</a>\n— The synthesis we treat as the canonical statement of the paradigm.</li>\n<li><strong>Quinn et al. (2019, 2023)</strong> — <em>Chebyshev approximation and the global geometry of\nmodel predictions</em> / <em>Information geometry for multiparameter models.</em> PRL 122, 158302;\nRep. Prog. Phys. 86, 035901. <a href=\"https://doi.org/10.1088/1361-6633/aca6f8\">doi</a>\n— Rigorous mathematical foundations for the manifold&#39;s geometry and the origin of\nsimplicity.</li>\n</ul>\n<h2 id=\"sloppy-models-applied-to-interatomic-potentials\">Sloppy models applied to interatomic potentials</h2><ul>\n<li><strong>Frederiksen, Jacobsen, Brown &amp; Sethna (2004)</strong> — <em>Bayesian ensemble approach to\nerror estimation of interatomic potentials.</em> PRL 93, 165501.\n<a href=\"https://doi.org/10.1103/PhysRevLett.93.165501\">doi</a>\n— The direct ancestor: sloppiness applied to <em>potentials</em>. Lupine is the\ncross-potential, data-driven generalization of this idea.</li>\n<li><strong>Mortensen et al. (2005)</strong> — <em>Bayesian error estimation in DFT.</em> PRL 95, 216401.\n<a href=\"https://doi.org/10.1103/PhysRevLett.95.216401\">doi</a>\n— Extends error estimation to the DFT reference level — relevant to ground-truth\nuncertainty.</li>\n<li><strong>Wen et al. (2017)</strong> — <em>Force-matching Stillinger–Weber potential for MoS₂ + Fisher\ninformation sensitivity.</em> J. Appl. Phys. 122, 244301.\n<a href=\"https://doi.org/10.1063/1.5007842\">doi</a>\n— Fisher-information sensitivity analysis on a real potential — the per-potential\nanalogue of our cross-potential manifold.</li>\n<li><strong>Kurniawan et al. (2022)</strong> — <em>Bayesian, frequentist, and information-geometric\napproaches to parametric UQ of classical empirical interatomic potentials.</em>\nJ. Chem. Phys. 156, 214103. <a href=\"https://doi.org/10.1063/5.0084988\">doi</a>\n— The closest prior art on potential UQ; we cite it to position the\ncross-potential corpus as the missing complementary view.</li>\n</ul>\n<h2 id=\"modern-networks-same-sloppy-manifold\">Modern networks — same sloppy manifold</h2><ul>\n<li><strong>Mao et al. (2024, 2026)</strong> — <em>The training process of many deep networks explores the\nsame low-dimensional manifold</em> / <em>Analytical characterization of sloppiness in neural\nnetworks.</em> PNAS 121, e2310002121; PRE 113, 015306.\n<a href=\"https://doi.org/10.1073/pnas.2310002121\">doi</a>\n— The bridge to MLIPs: the sloppy manifold appears in deep networks too — the\ntheoretical reason to expect the <a href=\"#/read/hyp-hyper-ribbon-mlip-transfer\">hyper-ribbon to transfer classical → MLIP</a>.</li>\n</ul>\n<h2 id=\"simpson-39-s-paradox-amp-the-ecological-fallacy\">Simpson&#39;s paradox &amp; the ecological fallacy</h2><ul>\n<li><strong>Simpson (1951)</strong>, <strong>Blyth (1972)</strong> — the paradox and the sure-thing principle.\n<a href=\"https://doi.org/10.1080/01621459.1972.10482387\">doi</a></li>\n<li><strong>Bickel, Hammel &amp; O&#39;Connell (1975)</strong> — <em>Sex bias in graduate admissions: Berkeley.</em>\nScience 187, 398. <a href=\"https://doi.org/10.1126/science.187.4175.398\">doi</a>\n— The canonical worked example; structurally identical to pooling elastic-constant\nerrors across elements.</li>\n<li><strong>Robinson (1950)</strong> — <em>Ecological correlations and the behavior of individuals.</em> Am.\nSociol. Rev. 15, 351. <a href=\"https://doi.org/10.2307/2087176\">doi</a>\n— The ecological fallacy proper — why cross-element pooling is not optional to avoid.</li>\n<li><strong>Pearl (2014)</strong> — <em>Understanding Simpson&#39;s paradox.</em> Am. Stat. 68, 8.\n<a href=\"https://doi.org/10.1080/00031305.2013.857687\">doi</a> — and <strong>Pearl (2009)</strong>,\n<em>Causality</em> (2nd ed.) — the causal-graph criterion the Lean spec encodes to prove the\npaper&#39;s Simpson&#39;s claim <a href=\"#/read/formal-proof-ledger\">cannot arise</a>.</li>\n<li><strong>Kievit et al. (2013)</strong>, <strong>Selvitella (2017)</strong> — practical guides to its ubiquity;\ncited to justify treating element identity as a confounder by default.</li>\n</ul>\n<h2 id=\"correlation-pitfalls\">Correlation pitfalls</h2><ul>\n<li><strong>Jackson &amp; Somers (1991)</strong> — <em>The spectre of &#39;spurious&#39; correlations.</em> Oecologia 86,<ol start=\"147\">\n<li>— <strong>Archie (1981)</strong> — <em>Mathematic coupling of data.</em> Ann. Surg. 193, 296.\n<a href=\"https://doi.org/10.1097/00000658-198103000-00008\">doi</a>\n— Why reference↔prediction correlations need the <a href=\"#/read/methodology\">matched-n discipline</a>:\nshared terms manufacture correlation.</li>\n</ol>\n</li>\n</ul>\n<h2 id=\"meta-analysis-methodology\">Meta-analysis methodology</h2><ul>\n<li><strong>DerSimonian &amp; Laird (1986)</strong> — <em>Meta-analysis in clinical trials.</em> Control. Clin.\nTrials 7, 177. <a href=\"https://doi.org/10.1016/0197-2456(86)90046-2\">doi</a>\n— The random-effects estimator we use to aggregate heterogeneous potential\nperformance correctly.</li>\n<li><strong>Higgins et al. (2003)</strong> — <em>Measuring inconsistency in meta-analyses (I²).</em> BMJ 327,<ol start=\"557\">\n<li><a href=\"https://doi.org/10.1136/bmj.327.7414.557\">doi</a>\n— The heterogeneity statistic; our corpus shows extreme I² ≈ 98.6 %, which is <em>why</em>\nrandom-effects is mandatory.</li>\n</ol>\n</li>\n<li><strong>Hedges &amp; Olkin (1985)</strong>, <strong>Borenstein et al. (2009, 2010)</strong> — standard references\nfor fixed- vs random-effects models.</li>\n<li><strong>Welz, Viechtbauer &amp; Pauly (2022)</strong> — <em>Fisher-transformation CIs of correlations in\nmeta-analysis.</em> Br. J. Math. Stat. Psychol. 75, 1.\n<a href=\"https://doi.org/10.1111/bmsp.12242\">doi</a>\n— The correct confidence intervals for pooled correlations.</li>\n</ul>\n<h2 id=\"benchmark-infrastructure-amp-reference-data\">Benchmark infrastructure &amp; reference data</h2><ul>\n<li><strong>OpenKIM</strong> — Open Knowledgebase of Interatomic Models.\n<a href=\"https://openkim.org\">openkim.org</a> — primary prediction source (559 potentials).</li>\n<li><strong>NIST Interatomic Potentials Repository.</strong>\n<a href=\"https://www.ctcms.nist.gov/potentials/\">ctcms.nist.gov/potentials</a> — cross-reference\nand potential-identity anchor.</li>\n<li><strong>Simmons &amp; Wang (1971)</strong> — <em>Single Crystal Elastic Constants and Calculated Aggregate\nProperties: A Handbook.</em> MIT Press. — the experimental ground truth for the elastic\nconstants the corpus is benchmarked against.</li>\n</ul>\n<hr>\n<p>The machine-readable bibliography is <code>paper/references.bib</code> (35 entries). See\n<a href=\"#/read/methodology\">Methodology</a> for how these tools are used, and the\n<a href=\"#/read/conjecture-ledger\">Conjecture Ledger</a> for what they were used to test.</p>\n"}