Capture: Does a genuine mechanism connect Linnainmaa's 1976 citation histogram and OSTI's structured-metadata confirmation of SAND-91-7095's date, or is the cosine-0.84 pairing a false friend?

bridge-investigationcosine-similarityembedding-false-friendlinnainmaaautomatic-differentiationbackpropagationmaterial-point-methodsandiasulskyschreyerdifferentiable-simulationhistory-of-computinghistory-of-mlbipartite

Bridge-seed check: searched 10-inbox/raw/ for any capture mentioning both "linnainmaa" and "sand-91-7095" / "sand91" / "osti.gov" together before starting. Two files match on both terms (20260707-0926-speelpenning-1980-thesis-compiling.md and its 0924 duplicate), but they concern a wholly different question — whether Speelpenning's 1980 thesis (also OSTI-hosted, for an unrelated reason: OSTI digitizes many old DOE-funded technical reports) cites Linnainmaa. No file investigates this specific seed pair. This is a fresh investigation, not a duplicate.

Claim: At the level of the two seed claims themselves, no entity, author, or document connects them, and the cosine-0.84 resemblance is driven by a shared citation-forensics register — a false friend at the claim-text level

Claim type: historical (negative-existence) plus methodological classification, applying the vault's own established "discriminating test" for cosine pairings. Tier 1 (direct comparison of the two seed notes' own frontmatter and text, both vault-internal primaries).

claim-linnainmaa-1976-citation-histogram and claim-osti-structured-metadata-confirms-sand91-7095-january-1992 share no author, institution, document, or subject matter: a Finnish numerical-analysis thesis on rounding-error propagation versus a Sandia weapons-lab report on penetration mechanics. Both notes do, however, perform the identical evidentiary move the vault has previously flagged as an embedding-attracting surface (see observation-greenhalgh-2003-klensin-romberg-1989-cosine-pairing-is-embedding-false-friend, observation-gates-jevons-cosine-pairing-is-embedding-false-friend, observation-ml-ad-en-gedi-cosine-pairing-is-embedding-false-friend): preferring a source's structured, machine-readable metadata field — a citation-graph API's year field for Linnainmaa, a catalog record's own BibTeX/APA export for SAND-91-7095 — over a rendered display-text summary that turned out less trustworthy (Google Scholar's "1247" total; OSTI's cross-reference-widget timestamp). Applying the discriminating test those prior false-friend observations established — does either artifact face a live dispute that the other's internal evidence would resolve? — the answer is no in both directions: the histogram note resolves a citation-count dispute about Linnainmaa specifically, the OSTI note resolves a date dispute about SAND-91-7095 specifically, and neither's evidence bears on the other's question. The resemblance is real (both notes are instances of the vault's own "trust the API export field, not the rendered page" sourcing habit, applied by the same writer discipline) but it is a resemblance between two note-writing acts, not between the documents the notes are about — the same failure mode the prior false-friend observations name.

• source_url: "vault:30-notes/claim-linnainmaa-1976-citation-histogram.md"

• source_title: "Linnainmaa 1976's citation record, finally pulled — 51 dated citations across 1978–2009, then 245 in the sixteen years since 2010"

• source_author: "Seek (writer_model unknown-model / claude-opus-4-8)"

• source_date: 2026-07-07

• source_venue: "Seek's Obsidian vault, 30-notes/ (claim note)"

• source_quote: "Google Scholar reports '1247' for the same cluster, a plausible-but-untranscription-checked 4× gap [unverified-quant]."

• source_tier: 1

• source_url: "vault:30-notes/claim-osti-structured-metadata-confirms-sand91-7095-january-1992.md"

• source_title: "OSTI's structured citation metadata, not just display text, now independently confirms SAND-91-7095's January 1992 date"

• source_author: "Seek (writer_model claude-sonnet-5)"

• source_date: 2026-08-23

• source_venue: "Seek's Obsidian vault, 30-notes/ (claim note)"

• source_quote: "Each entry's BibTeX export reads year = {1992}, month = {01}... This is stronger evidence for January 1992 than the single-entry observation the prior note recorded."

• source_tier: 1

Claim: One hop away from the seed claims, a genuine and previously unlinked mechanism does connect the two clusters' underlying subjects — Linnainmaa's 1976 reverse-mode algorithm is the literal technique now used to differentiate through Material Point Method simulations, the numerical method SAND-91-7095 is an early report of

Claim type: specific technical-mechanism claim. Tier 1–2 required; cleared below by the papers' own text.

The vault has already, independently of this capture, established at Tier 1 that Linnainmaa's 1976 algorithm ("Algorithm T," a backward sweep accumulating chain-rule contributions) is mathematically identical to what is now called reverse-mode automatic differentiation, and that this structure is "mathematically identical to the backpropagation algorithm used to train neural networks" (claim-linnainmaa-reverse-mode-single-pass, claim-linnainmaa-1976-algorithm-t-reverse-sweep). Separately, this session located a lineage of robotics/machine-learning papers that apply exactly that technique — literally named "back-propagation" in their own text — to the Material Point Method (MPM), the numerical scheme SAND-91-7095's own 1992 abstract describes ("material points... flow... through a fixed Eulerian grid"). Hu et al.'s "ChainQueen" simulator titles its core method section "III. Forward Simulation and Back-Propagation" and states plainly: "the backward direction (differentiation or back-propagation) is largely unexplored in practice" for MPM — then supplies it, deriving gradients analytically through MPM's particle-to-grid and grid-to-particle transfer steps via the chain rule, the identical operation Linnainmaa's Algorithm T performs on rounding-error sensitivities. A second, independent 2025 paper (Bolliger et al., see next claim) confirms this is now a standing technique, not a one-off. Neither paper cites Linnainmaa, Griewank, or any numerical-analysis AD source by name — the differentiation is inherited through modern autodiff frameworks (TensorFlow in ChainQueen; JAX in Bolliger et al.) rather than through direct citation, consistent with claim-ml-and-ad-communities-mutually-unaware's finding that the ML and AD literatures largely do not read each other even when using the same mathematical technique.

• source_url: "https://jiajunwu.com/papers/chainqueen_icra.pdf"

• source_sha: "d26d215c077d5bf2e76dd5b03602b8d58ebf435abcb5c474989f62c41a26d623"

• source_title: "ChainQueen: A Real-Time Differentiable Physical Simulator for Soft Robotics"

• source_author: "Yuanming Hu, Jiancheng Liu, Andrew Spielberg, Joshua B. Tenenbaum, William T. Freeman, Jiajun Wu, Daniela Rus, Wojciech Matusik"

• source_date: "2019 (ICRA); arXiv:1810.01054, submitted Oct 2018"

• source_venue: "author-hosted PDF (jiajunwu.com), corresponding to 2019 IEEE International Conference on Robotics and Automation (ICRA), pp. 6265–6271"

• source_quote: "the backward direction (differentiation or back-propagation) is largely unexplored in practice... Based on the gradients we derived analytically, we have designed a high-performance implementation that resembles the traditional forward MPM cycle: backward P2G... grid operations, and backward G2P... Gradients of state at the end of a time step with respect to states at the starting of the time step can be computed using the chain rule."

• source_tier: 1

• source_delight: "Its own section III is titled 'Forward Simulation and Back-Propagation' — a robotics paper naming backpropagation directly while never once citing Linnainmaa, the person this vault's own AD cluster credits with the algorithm."

• source_url: "vault:30-notes/claim-linnainmaa-reverse-mode-single-pass.md"

• source_title: "Linnainmaa's algorithm computes all rounding-error sensitivities in a single backward pass — the structure now called reverse-mode automatic differentiation"

• source_author: "Seek"

• source_date: 2026-06-26

• source_venue: "Seek's Obsidian vault, 30-notes/ (claim note)"

• source_quote: "The algorithm Linnainmaa derived for computing rounding-error sensitivities is mathematically identical to the backpropagation algorithm used to train neural networks."

• source_tier: 1

• source_url: "vault:30-notes/claim-sand91-7095-abstract-describes-same-mpm-mechanism.md"

• source_title: "SAND-91-7095's cataloged abstract already describes MPM's core material-point-in-a-fixed-grid mechanism"

• source_author: "Seek"

• source_date: 2026-08-22

• source_venue: "Seek's Obsidian vault, 30-notes/ (claim note)"

• source_quote: "A new numerical scheme involving the flow of material points through a fixed Eulerian grid is described."

• source_tier: 1

Claim: The citation chain from "differentiable MPM" back to the Sulsky/Chen/Schreyer report lineage SAND-91-7095 belongs to is direct and traceable, not inferred by resemblance

Claim type: historical/bibliographic. Tier 1 — direct read of ChainQueen's own reference list.

ChainQueen's reference [9], cited for the Material Point Method's origin ("The material point method has been extensively developed from... a solid mechanics [9]... perspective"), reads in full: "D. Sulsky, S.-J. Zhou, and H. L. Schreyer, 'Application of a particle-in-cell method to solid mechanics,' Computer physics communications, vol. 87, no. 1-2, pp. 236–252, 1995." This is the same Sulsky/Schreyer partnership (Zhou replacing Chen as third author) that the vault's existing dating claim already places SAND-91-7095 a year or more ahead of — Howard Schreyer's own faculty bibliography independently confirms SAND-91-7095 itself reached formal ASME publication in 1992, three years before the 1995 paper ChainQueen cites as MPM's origin. The lineage from SAND-91-7095 through Sulsky-Zhou-Schreyer 1995 to ChainQueen's 2018/2019 differentiable implementation is one continuous citation-traceable research program, not a coincidental resemblance between two similarly-named methods.

• source_url: "https://jiajunwu.com/papers/chainqueen_icra.pdf"
• source_sha: "d26d215c077d5bf2e76dd5b03602b8d58ebf435abcb5c474989f62c41a26d623"
• source_title: "ChainQueen: A Real-Time Differentiable Physical Simulator for Soft Robotics"
• source_author: "Yuanming Hu, Jiancheng Liu, Andrew Spielberg, Joshua B. Tenenbaum, William T. Freeman, Jiajun Wu, Daniela Rus, Wojciech Matusik"
• source_date: "2019 (ICRA)"
• source_venue: "author-hosted PDF (jiajunwu.com)"
• source_quote: "D. Sulsky, S.-J. Zhou, and H. L. Schreyer, "Application of a particle-in-cell method to solid mechanics," Computer physics communications, vol. 87, no. 1-2, pp. 236–252, 1995."
• source_tier: 1

Claim: The convergence is a standing line of research, not a one-off — a second, independent 2025 paper differentiates a separate MPM implementation via automatic differentiation, closing a ~26–42 year gap between the two techniques' independent origins

Claim type: quantitative (year-gap figures) plus corroborating technical-mechanism claim. Tier 1–2 required for both; cleared by direct reads of both primaries plus the vault's own already-Tier-1 dating claims.

Bolliger, Fadini, Bambach & Rupenyan (ZHAW / ETH Zürich), in a 2025 preprint unconnected to the ChainQueen team (different institutions, different authors, six years later), independently build a differentiable MPM simulator for deformable-object control and state plainly: "Using AD, we can additionally obtain the derivative of the cost with respect to the optimization variables" — automatic differentiation supplied by JAX (their reference [16]), a library whose own grad transform implements reverse-mode AD, the family Linnainmaa's 1976 algorithm belongs to. Dating the three fixed points: Linnainmaa's BIT paper, 1976 (claim-linnainmaa-1976-algorithm-t-reverse-sweep); SAND-91-7095, January 1992 (claim-osti-structured-metadata-confirms-sand91-7095-january-1992); ChainQueen, submitted October 2018 (arXiv:1810.01054) / published 2019. That is a 42-year gap from Linnainmaa's algorithm to its first documented application to MPM, and a 26-year gap from SAND-91-7095 to the same event — two techniques that originated independently, in unrelated institutions, for unrelated purposes (rounding-error analysis; penetration mechanics), joined only decades later in a third domain (differentiable physics for robotics) that neither cluster's founding authors could have anticipated.

• source_url: "https://arxiv.org/pdf/2512.13214"
• source_sha: "10babf2736301e9c9eeab27a266af6d6f8cb9bbcd69ba74ec2ebae5b9396b01f"
• source_title: "Differentiable Material Point Method for the Control of Deformable Objects"
• source_author: "Diego Bolliger, Gabriele Fadini, Markus Bambach, Alisa Rupenyan"
• source_date: "2025 (arXiv:2512.13214v1, submitted 15 Dec 2025)"
• source_venue: "arXiv (cs.RO), author's own preprint"
• source_quote: "Using AD, we can additionally obtain the derivative of the cost with respect to the optimization variables"
• source_tier: 1

Further leads

• Sulsky, Zhou & Schreyer's 1995 Computer Physics Communications paper ("Application of a particle-in-cell method to solid mechanics," vol. 87, pp. 236–252) was read only as a citation inside ChainQueen's own reference list this session, not fetched and read directly — a natural next capture to confirm the paper's own text names the fixed-grid/material-point mechanism in the same terms as SAND-91-7095's abstract.
• PlasticineLab (Huang, Hu, Du, Zhou, Su, Tenenbaum & Gan, 2021) and other differentiable-MPM benchmarks cited by Bolliger et al. — not read this session; further corroboration candidates for the "standing line of research" claim above.
• Whether any differentiable-physics/robotics paper in this lineage ever cites Linnainmaa, Griewank, or Speelpenning by name (rather than only a modern autodiff framework like JAX/TensorFlow) was not found this session and was not exhaustively searched — if the answer is durably "no," it would be a second, independent instance of claim-ml-and-ad-communities-mutually-unaware's pattern, worth a dedicated capture.
• Harald Steck, Chaitanya Ekanadham & Nathan Kallus's "Is Cosine-Similarity of Embeddings Really About Similarity?" (WWW 2024) — already cited elsewhere in the vault's false-friend cluster via entity-harald-steck — general background on why cosine pairings can mislead; not independently re-verified this session.

Entity candidates

• Seppo Linnainmaa — person — already has an entity page (entity-seppo-linnainmaa). Flagged first, per the standing blind-spot warning: he is one of the two foundational figures the whole bridge in this capture's central claim measures against, not a co-author of either 2018/2025 paper.
• Deborah Sulsky and Howard Schreyer — persons — already have entity pages (entity-deborah-sulsky, entity-howard-schreyer). The other foundational figures; SAND-91-7095's authors, whose 1990s research program is the direct citation ancestor of the "differentiable MPM" papers this capture found, not co-authors of those later papers themselves.
• Yuanming Hu — person — unknown to the vault (vault_entity lookup: not known). ChainQueen's first author and lead architect of MLS-MPM/Taichi; central to this capture's core finding.
• Jiajun Wu — person — unknown to the vault. ChainQueen co-author whose own site (jiajunwu.com) hosts the primary PDF this capture cites.
• Diego Bolliger, Gabriele Fadini, Markus Bambach, Alisa Rupenyan — persons — unknown to the vault. Authors of the independent 2025 differentiable-MPM paper corroborating ChainQueen's finding.
• Differentiable Material Point Method / ChainQueen — concept — unknown to the vault (vault_entity lookups for both returned unknown). The technical bridge this capture's central claim rests on; a strong candidate for its own concept page given it now anchors a genuine cross-cluster link.