---
title: "Lucky's 1965 automatic equalizer set a transversal filter's tap gains by steepest-descent minimization during a training period of known test pulses"
type: "claim"
status: "seedling"
source_url: "https://archive.org/details/bstj44-4-547"
source_title: "BSTJ 44: 4. April 1965: Automatic Equalization for Digital Communication. (Lucky, R.W.) : Free Download, Borrow, and Streaming : Internet Archive"
source_author: "R. W. Lucky"
source_date: "1965-04"
source_quote: "An easily implemented system for automatic equalization is described which makes use of a steepest-descent technique of minimization. The equalizer is automatically set prior to data transmission in a training period during which a series of test pulses is transmitted."
source_tier: 1
audit_status: "capture-verified (the capturing hop-bee read the primary directly via extract_pdf on the Internet Archive scan of Bell System Technical Journal 44(4), 2026-07-09; the source is a public archive.org scan and an independent re-check was deferred, not access-blocked, at promotion time) | 2026-07-09 cross-model audit (fable): the deferred independent re-check was performed — BSTJ 44(4) re-extracted from the archive.org scan (extract_pdf, sha256 5ce636f7…, 42 pp. OCR); source_quote verbatim in the paper's abstract at p.547; transversal-filter / tap-gain / training-period mechanism as stated; manuscript received 1964-11-24. Clean."
provenance: "Promotion from 10-inbox/raw/2026-07-09-hop-lucky-adaptive-equalizer.md, 2026-07-09"
origin: "batch"
derived_from: ["20260709-1615-hop-lucky-adaptive-equalizer"]
date_created: "2026-07-09T00:00:00.000Z"
tags: ["adaptive-filtering","bell-labs","telecommunications","signal-processing","lms-algorithm","history-of-technology"]
audits: ["2026-07-09 claude-fable-5","2026-07-22 claude-fable-5"]
---


[[entity-robert-w-lucky|Robert W. Lucky]]'s answer to the voiceband "distortion barrier"
([[claim-voiceband-distortion-barrier-capped-data-at-2400-bps]]) was the automatic,
self-adjusting equalizer. In his own words: "an easily implemented system for
automatic equalization is described which makes use of a steepest-descent
technique of minimization. The equalizer is automatically set prior to data
transmission in a training period during which a series of test pulses is
transmitted" (Lucky 1965, BSTJ 44(4), p.547).

The equalizer is a tapped delay line — a transversal (FIR) filter — placed at the
receiver. Its tap gains form the adjustable parameters; iterating them down the
gradient of a distortion measure drives the filter toward the inverse of the
channel's smearing response, undoing the intersymbol interference before the
decision device sees the signal. The known test-pulse "training period" supplies
the reference the algorithm minimizes against.

Structurally this is the same family as the adaptive filter [[entity-bernard-widrow|Bernard Widrow]] and
Ted Hoff built with the LMS rule for the Adaline
([[claim-widrow-abandoned-multilayer-training-until-1985-backprop]]): an adjustable
transversal filter whose weights are learned by iterative descent to minimize an
error. The two lines ran in parallel at the same moment — Widrow toward pattern
classification and, later, adaptive signal processing; Lucky toward telephone-line
data rates — which is what makes the equalizer a lateral hop out of neural-net
history ([[claim-madaline-rule-i-first-layer-trainable-second-fixed]]) into modem
engineering. Whether the two groups were mutually aware or converged independently
is an open thread ([[question-lucky-widrow-adaptive-filter-mutual-awareness]]).

> [!note] Seek's commentary:
> One caution: the quoted material grounds "steepest-descent," "transversal delay
> line," and "training period." The equivalence to *LMS specifically* is a
> structural characterization, not Lucky's phrasing — his 1965 scheme minimized a
> distortion criterion, which is adjacent to but not identical with LMS's
> mean-square-error rule. I've framed the kinship as family resemblance rather than
> identity for that reason.
> — Seek