한
← All rivers
River of knowledge

The River of Numerical Analysis

How did computation learn to measure its own error when exact answers were unavailable?

From Babylonian approximations of √2 through Newton-Raphson iteration, Gaussian quadrature, floating-point error analysis, the 1979 *LINPACK Users’ Guide*, and modern GPU computing — a history of obtaining useful approximations while controlling error.

1800 BCE–20245 city-to-city segments

Read the current

What changed between the cities?

  1. 1

    1800 BCE–1669

    independent comparison

    BabylonCambridge

    From a Clay-Tablet Approximation of √2 to Iterative Calculation

    The approximation of √2 on Babylonian tablet YBC 7289 and seventeenth-century Newton-style iteration are not one documented chain of transmission. Placing them side by side exposes a durable question: when no convenient closed answer is available, which procedure gives a useful approximation and how good is it?

    The sexagesimal approximation on YBC 7289 · Newton-style iteration
    Place this segment on the map
  2. 2

    1669–1815

    problem reformulation

    CambridgeGöttingen

    Iteration and Observational Error Demanded General Procedures

    Newton–Raphson-style iteration refined roots step by step, while Gauss's work estimated orbits and unknowns amid astronomical measurement error. The line from Cambridge to Göttingen marks a reformulation around approximation, error, and procedure—not the travel of one book.

    Root-finding iteration · least squares · Gaussian quadrature
    Place this segment on the map
  3. 3

    1944–1955

    team computation

    PrincetonLos Alamos

    Electronic Computation Turned Approximation into Large-Scale Experiment

    Mathematical and computing work around Princeton met transport and weapons problems at Los Alamos in team efforts using ENIAC and Monte Carlo methods. Repeated random samples made a computable experiment stand in for an unavailable closed formula.

    ENIAC calculation · Monte Carlo methods · error estimation
    Place this segment on the map
  4. 4

    1948–1965

    error analysis

    Los AlamosTeddington

    Ask for Trustworthy Computation, Not Speed Alone

    Set large-scale Los Alamos computation beside James Wilkinson's work at the NPL in Teddington and the central questions become error amplification and sensitivity to small input changes. Numerical analysis became a discipline for explaining when an answer is trustworthy, not merely producing one.

    Floating-point error · condition numbers · backward error analysis
    Place this segment on the map
  5. 5

    1961–2024

    software and hardware ecosystem

    TeddingtonMountain View

    Matrix Algorithms Became a Shared Engine for Accelerated Computing

    Many libraries, standards, and hardware teams stand between stable QR work at Teddington and today's large matrix calculations on accelerators. This compressed route does not skip them into a lone transmission claim; it highlights the recurring redesign of stability, portability, and speed across LINPACK, LAPACK, and parallel computing.

    QR algorithms · LINPACK/LAPACK · parallel matrix computation · GPUs
    Place this segment on the map

How to read the lines

Each line is an editorial route through problems, texts, and practices. It does not imply one book moving in a straight line, a sole invention, or identical adoption everywhere.