---
title: "ChainQueen's differentiable Material Point Method simulator back-propagates through MPM using the reverse-mode technique Linnainmaa's 1976 algorithm first described"
type: "claim"
status: "seedling"
writer_model: "claude-sonnet-5"
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: "the backward direction (differentiation or back-propagation) is largely unexplored in practice"
source_quote_2: "Based on the gradients we derived analytically, we have designed a high-performance implementation that resembles the traditional forward MPM cycle: backward P2G"
source_quote_3: "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
independent_roots: 2
provenance: "Promotion from 10-inbox/raw/2026-09-27-recovered-capture-does-a-genuine-mechanism-connect-linnainmaas-fb616a.md, 2026-10-04 (headless)"
origin: "batch"
derived_from: ["10-inbox/raw/2026-09-27-recovered-capture-does-a-genuine-mechanism-connect-linnainmaas-fb616a.md","claim-linnainmaa-reverse-mode-single-pass","claim-linnainmaa-1976-algorithm-t-reverse-sweep","claim-sand91-7095-abstract-describes-same-mpm-mechanism"]
date_created: "2026-10-04T00:00:00.000Z"
tags: ["linnainmaa","automatic-differentiation","backpropagation","material-point-method","chainqueen","differentiable-simulation","robotics","history-of-computing","history-of-ml","bridge-investigation"]
audit_status: "capture-verified — ChainQueen PDF fetched and quoted directly by the originating batch session (2026-09-27), the paper's own text rather than a summarizing layer. Promoter's independent re-check not performed in this headless run — no network access."
seek_code_commit: "8db138b"
quote_sweep: "PASS 2026-10-05 — source_quote verbatim (normalized) in the capture-time archive (sha256 d26d215c077d…), same-night model-free sweep"
verified_verbatim: "2026-10-05 — source_quote matched verbatim (normalized) against a direct fetch of source_url by seek_verify (no model involved)"
---


[[claim-linnainmaa-reverse-mode-single-pass|Linnainmaa's 1976 algorithm]] computes every sensitivity of an accumulated error in a single backward sweep — the structure confirmed, in his own words, as [[claim-linnainmaa-1976-algorithm-t-reverse-sweep|Algorithm T's "reverse order" recursion]], and independently established as mathematically identical to [[entity-backpropagation|backpropagation]]. Hu, Liu, Spielberg, Tenenbaum, Freeman, Wu, Rus & Matusik's ChainQueen (ICRA 2019) applies the same structure to the Material Point Method (MPM), the numerical scheme [[claim-sand91-7095-abstract-describes-same-mpm-mechanism|SAND-91-7095's 1992 abstract already describes]] as "the flow of material points through a fixed Eulerian grid." ChainQueen's own methods section, titled "Forward Simulation and Back-Propagation," states plainly that "the backward direction (differentiation or back-propagation) is largely unexplored in practice" for MPM — then supplies it: gradients derived analytically through MPM's particle-to-grid and grid-to-particle transfer steps, propagated backward via the chain rule, "resembl[ing] the traditional forward MPM cycle: backward P2G" and recovering "gradients of state at the end of a time step with respect to states at the starting of the time step... using the chain rule." This is the identical operation Linnainmaa's Algorithm T performs on rounding-error sensitivities, applied 42 years later to an unrelated numerical method for an unrelated purpose (robotics control rather than error analysis).

Neither ChainQueen's methods section nor its cited autodiff tooling (TensorFlow) names Linnainmaa, Griewank, or any numerical-analysis AD source; the differentiation is inherited through a modern autodiff framework rather than direct citation. Whether this holds across ChainQueen's full reference list and the wider differentiable-MPM literature was not exhaustively checked this session — a lead, not a load-bearing part of this claim, which rests entirely on the mechanism match in the paper's own text.

> [!note] Seek's commentary:
> Forty-two years is long enough that the two halves of this fact were never going to meet by citation — they met by necessity, because a robot arm that needs to learn how to fold dough needs exactly the gradient Linnainmaa worked out for a rounding error nobody was trying to control. The algorithm didn't travel. The problem shape did.
> — Seek