Does ChainQueen (2019) or Bolliger et al. (2025) — or the wider differentiable-physics/robotics literature around them — ever name Linnainmaa, Griewank, or Speelpenning directly, or does the inheritance route entirely through JAX/TensorFlow with the founding names erased?

linnainmaagriewankspeelpenningautomatic-differentiationchainqueenbolligerdifftaichimaterial-point-methoddifferentiable-simulationroboticscitation-historyhistory-of-mlbridge-investigation

Capture: Does ChainQueen, Bolliger et al., or their kin papers ever name Linnainmaa, Griewank, or Speelpenning?

Method note: this capture resolves a specific open lead, not a fresh topic. The vault's two existing claim-notes on ChainQueen and Bolliger et al. (claim-chainqueen-differentiates-mpm-via-linnainmaas-reverse-mode-technique, claim-bolliger-2025-second-independent-paper-confirms-standing-differentiable-mpm-technique) each already state, as an unchecked lead, that neither paper names Linnainmaa, Griewank, or any numerical-analysis AD source. Both PDFs were re-fetched via extract_pdf (both TLS-verified) and their full reference lists read end to end — 34 entries for ChainQueen, 20 for Bolliger et al. — to convert that lead into a checked claim. A third paper, DiffTaichi (Hu et al., ICLR 2020), was added because it is not a distant member of "the wider literature": it shares ChainQueen's first author (Yuanming Hu), cites ChainQueen directly, and is the direct technical ancestor of the Taichi-language ecosystem ChainQueen's footnote credits. Reading it changed the answer from a clean "erased" to a more exact one.

Claim: ChainQueen's (2019) full 34-item reference list contains no citation of Linnainmaa, Griewank, Speelpenning, Werbos, Rumelhart, or any dedicated automatic-differentiation source

Claim type: historical/bibliographic (negative-existence), load-bearing given it is surprising. Tier 1 — direct read of the primary document, author-hosted PDF.

ChainQueen's reference list runs from [1] (Todorov, Erez & Tassa's MuJoCo paper) to [34] (Fang, Hu, Hu & Jiang on adaptive time stepping), closing with the Sanchez-Gonzalez et al. 2018 ICML paper on graph-network physics engines as item [34]: "A. Sanchez-Gonzalez, N. Heess, J. T. Springenberg, J. Merel, M. Riedmiller, R. Hadsell, and P. Battaglia, 'Graph networks as learnable physics engines for inference and control,' in ICML, 2018." Every entry was read; none names Seppo Linnainmaa, Andreas Griewank, Bert Speelpenning, Paul Werbos, or Rumelhart/Hinton/Williams, and none cites a dedicated automatic-differentiation text (e.g. Griewank & Walther's Evaluating Derivatives). The paper's own autodiff tooling — TensorFlow, used for its reference implementation and benchmarked against the CUDA version in Table II — is used and discussed in-text but is not even given its own citation; it enters the paper as infrastructure, not as attributed prior work. This confirms, exhaustively rather than as a lead, what claim-chainqueen-differentiates-mpm-via-linnainmaas-reverse-mode-technique already flagged as unchecked.

• 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), 2019 IEEE International Conference on Robotics and Automation (ICRA), pp. 6265–6271"
• source_quote: "A. Sanchez-Gonzalez, N. Heess, J. T. Springenberg, J. Merel, M. Riedmiller, R. Hadsell, and P. Battaglia, "Graph networks as learnable physics engines for inference and control," in ICML, 2018."
• source_tier: 1

Claim: Bolliger et al.'s (2025) reference list likewise omits Linnainmaa, Griewank, and Speelpenning; the paper names its AD inheritance explicitly as JAX and cites ChainQueen directly as prior art

Claim type: historical/bibliographic (negative-existence) plus a positive citation-chain fact. Tier 1 — direct read of the primary document, arXiv preprint.

Bolliger, Fadini, Bambach & Rupenyan's 20-item reference list (ending at [20], Sánchez-Banderas & Otaduy 2018 on strain-rate dissipation) likewise contains no citation of Linnainmaa, Griewank, or Speelpenning. The paper is explicit about where its automatic differentiation comes from: "several computational frameworks have been developed to compute the gradients algorithmically through automatic differentiation (AD)," naming PyTorch, TensorFlow, Taichi-Lang, and JAX as the candidate libraries, before stating "Among these AD alternatives, we decided to use JAX [16] because it is mature, under active development, and can be easily integrated with other machine learning libraries." Reference [16] is the JAX GitHub/paper (Bradbury et al., 2018) — not a numerical-analysis AD source. Reference [12] is ChainQueen itself, cited directly as one of two prior differentiable-MPM simulators ("[12], [13] introduce differentiable MPM simulators"), making this the paper's own documented link back to ChainQueen — a citation chain that runs MPM-paper to MPM-paper and library-to-library, never touching the 1976/1980s AD-discovery literature.

• 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: "several computational frameworks have been developed to compute the gradients algorithmically through automatic differentiation (AD)."
• source_quote_2: "Among these AD alternatives, we decided to use JAX [16] because it is mature, under active development, and can be easily integrated with other machine learning libraries."
• source_tier: 1
• independent_roots: note — Bolliger et al. is already the subject of one prior claim-note (claim-bolliger-2025-second-independent-paper-confirms-standing-differentiable-mpm-technique); this is the second note resting on it, under the three-note single-source concentration cap (sources.md).

Claim: DiffTaichi (Hu et al., ICLR 2020) — ChainQueen's direct kin paper, sharing its first author and citing it by name — explicitly names Griewank (and Wengert) as the AD lineage, but even there Linnainmaa and Speelpenning go unnamed

Claim type: historical/bibliographic, both positive (Griewank is named) and negative (Linnainmaa/Speelpenning are not). Tier 1 — direct read of the primary document, arXiv preprint published as an ICLR 2020 conference paper.

DiffTaichi is not an arbitrary data point in "the wider literature": it is written by Yuanming Hu (ChainQueen's first author) and collaborators, explicitly follows "ChainQueen's implementation (Hu et al., 2019b)" for its own differentiable-MPM demo, and sits in the same Taichi-language lineage ChainQueen's own footnote credits ("Based [on] the Taichi [30] open source computer graphics library"). Unlike ChainQueen and Bolliger et al., DiffTaichi's design-decision section names a specific AD-tradition figure by name, twice: "Source Code Transformation (SCT) (Griewank & Walther, 2008) and Tracing (Wengert, 1964) are common choices when designing AD systems," and later, describing its own reverse-mode compiler pass, "which it achieves with reverse-mode AD, using a standard source code transformation (Griewank & Walther, 2008)." The full bibliography entry reads: "Andreas Griewank and Andrea Walther. Evaluating derivatives: principles and techniques of algorithmic differentiation, volume 105. Siam, 2008." The same bibliography separately credits R. E. Wengert's 1964 CACM paper by name for the tracing/tape alternative — consistent with claim-wengert-list-named-for-forward-mode-inventor's pattern of the forward-mode originator retaining citation life. Seppo Linnainmaa and Bert Speelpenning do not appear anywhere in DiffTaichi's bibliography (34 entries read in full) or body text, despite the paper implementing exactly the reverse-mode operation Linnainmaa's 1976 algorithm describes.

• source_url: "https://arxiv.org/pdf/1910.00935"
• source_sha: "88837aafd34badf97251153106530ce546705e64630fa7f1d94849b391a6e4be"
• source_title: "DiffTaichi: Differentiable Programming for Physical Simulation"
• source_author: "Yuanming Hu, Luke Anderson, Tzu-Mao Li, Qi Sun, Nathan Carr, Jonathan Ragan-Kelley, Frédo Durand"
• source_date: "arXiv:1910.00935v3, 14 Feb 2020; published as a conference paper, ICLR 2020"
• source_venue: "arXiv (cs.LG) preprint; International Conference on Learning Representations (ICLR) 2020"
• source_quote: "Source Code Transformation (SCT) (Griewank & Walther, 2008) and Tracing (Wengert, 1964) are common choices when designing AD systems."
• source_quote_2: "Andreas Griewank and Andrea Walther. Evaluating derivatives: principles and techniques of algorithmic differentiation, volume 105. Siam, 2008."
• source_tier: 1

Claim: Across this specific cluster, the inheritance route runs predominantly through modern AD frameworks with founding names erased — but the erasure is not uniform, and Griewank's name travels further into this literature than Linnainmaa's or Speelpenning's does

Claim type: synthesis of the three claims above; historical, Tier 1 (rests entirely on the three direct primary reads above, no new source).

The direct answer to the core question, for the specific papers named: ChainQueen and Bolliger et al. both name their AD inheritance only as a software library (TensorFlow; JAX) and cite no AD-discovery figure. But "the wider differentiable-physics/robotics literature around them" is not uniform on this point — DiffTaichi, the closest possible kin paper (same lead author, direct citation of ChainQueen), does name Andreas Griewank, twice, via the standard textbook citation for source-code-transformation AD. What does not happen, in any of the three papers read in full, is a citation of Seppo Linnainmaa or Bert Speelpenning by name. This sharpens rather than overturns the vault's existing finding in claim-speelpenning-1980-does-not-cite-linnainmaa that Griewank's name circulates in AD-adjacent literature in a way Linnainmaa's does not: here, in a cluster of papers computing the literal operation Linnainmaa's 1976 algorithm first formalized, Griewank surfaces as the attributed textbook reference for the technique, while the man who first described that technique, and the man who first built it in software, remain unnamed even in the one paper of the three that cites anyone from the AD-discovery tradition at all.

Further leads

• PlasticineLab (Huang, Hu, Du, Zhou, Su, Tenenbaum, Gan, 2021) and GradSim (Jatavallabhula et al., 2021) — both cited directly by Bolliger et al. as prior differentiable-MPM/physics work, and both plausible further links in this same citation cluster — not read in full this session; check their reference lists for Linnainmaa/Griewank/Speelpenning next.
• Baydin, Pearlmutter, Radul & Siskind's "Automatic Differentiation in Machine Learning: A Survey" (JMLR 2018) is the paper that does name Linnainmaa, Griewank, and Speelpenning together in one place; none of the three papers read this session cites it — worth checking whether any paper in the wider differentiable-physics cluster uses the Baydin survey as its bridge to the AD-discovery literature, as opposed to going straight to a software library.
• JAX's own paper/documentation (Bradbury et al., 2018, cited by Bolliger et al. as ref [16]) was not read this session for its own citation practice — unknown whether JAX's own materials name Linnainmaa, Griewank, or the Baydin survey.
• DiffTaichi's bibliography also names Pearlmutter & Siskind (2008, "Lambda the ultimate backpropagator") and Hascoet & Pascual (2013, Tapenade) as AD-tool citations distinct from Griewank & Walther — a second compiler-AD-tradition citation inside the same paper, not explored further here.

Entity candidates

• Andreas Griewank — person — already a vault hub (entity-andreas-griewank); this capture adds a new sourced instance of his name surfacing in ML/robotics literature (DiffTaichi) rather than only in AD-history literature, worth a dated ## Learned addition to that entity note.
• Seppo Linnainmaa — person — already a vault hub (entity-seppo-linnainmaa); this capture extends its existing 2026-10-04 ## Learned entry (which covers ChainQueen and Bolliger) to a third, more closely related paper (DiffTaichi) that still does not name him.
• Bert Speelpenning — person — no entity page exists yet despite being the subject of a Tier-1 claim-note (claim-speelpenning-1980-does-not-cite-linnainmaa); this capture is the third note finding his name absent from a differentiable-physics paper's bibliography, strengthening the case that his citation reach is narrower than Griewank's.
• DiffTaichi / Taichi (programming language) — concept/term — no entity page; the direct technical bridge between ChainQueen and the wider differentiable-MPM ecosystem (PlasticineLab builds on it), and the one point in this cluster where an AD-discovery-tradition name (Griewank) surfaces explicitly.
• Griewank & Walther, Evaluating Derivatives (2008 SIAM textbook) — concept/document — the specific citation vehicle by which Griewank's name (not Linnainmaa's) reaches ML/robotics papers; worth its own note if more instances of this citation turn up.