Apply Tufte's overarching framework for putting multivariate data on a flat surface — the TWO goals (more representable dimensions, higher data density) reached through FIVE methods (micro/macro readings, layering, small multiples, color, space-time narratives) — when designing any chart, map, dashboard, table, or display that must carry more than two variables.
설치
Codex 또는 Claude로 설치 이 Prompt를 복사해 Codex, Claude 또는 다른 어시스턴트에 붙여 넣으면 Skill 페이지를 검토하고 설치를 진행할 수 있습니다.
Apply Tufte's overarching framework for putting multivariate data on a flat surface — the TWO goals (more representable dimensions, higher data density) reached through FIVE methods (micro/macro readings, layering, small multiples, color, space-time narratives) — when designing any chart, map, dashboard, table, or display that must carry more than two variables.
The world is multivariate; paper and screen are not. Every image a maker shows a reader must live on a two-dimensional plane, so the central job of information design is to widen the resolving power of that plane — to raise both the number of dimensions it can carry and the density of data it packs per unit area. This is the opening, agenda-setting chapter of Envisioning Information: it states the problem, sets the two goals, previews the five methods that the rest of the book develops, and names the enemy — chartjunk and the architectural "duck," design that becomes its own content out of contempt for the data and the audience.
Read this skill for the framework and the chapter's gallery of worked examples. For the deep mechanics of each method, go to its dedicated sibling skill (named in §2). Do not treat the five methods as six "strategies" — density is a goal, not a method.
§1. The Flatland Problem
The term comes from Edwin A. Abbott's novel Flatland: A Romance of Many Dimensions (London, 1884), about beings trapped in a two-dimensional world. Tufte's application: we navigate a three-space perceptual world and reason about higher dimensions with ease, yet every information display collapses onto the flatland of paper and screen.
"Escaping this flatland is the essential task of envisioning information." — Tufte, Envisioning Information
The toad image (chapter opener, p.14): when a toad sheds its skin, the suit leaves life's "spaceland" and collapses into flatland — exactly what our displays do to multivariate reality. The task is to do better than a flattened-out toad suit.
Why it gets harder: two forces compound the difficulty.
Force
Effect
Abstraction
As data loses its tie to familiar three-space (temperature, GDP, sentiment), there is no physical form to lean on
Variable count
Each extra variable competes for the same two axes of the surface
The fundamental compromise: nearly every escape from flatland trades one virtue for another. The literature is "partial, arbitrary, and particularistic" — no single trick solves dimensional compression in general. Paul Klee made the related point that a whole built from parts of different dimensions is hard to grasp and harder still to convey to someone else, because spoken language is sequential and cannot hold many dimensions at once. Expect tradeoffs; do not chase a universal solution.
§2. The Real Taxonomy — Two Goals, Five Methods (p.13, previewed p.33)
This is the chapter's actual structure. Do not invent a six-item "strategies" list. Tufte names two goals and five methods; data density is one of the two goals, not a method standing beside the other four.
The two goals — what every display is trying to do:
Goal
Definition
Increase dimensions
Raise the number of variables representable on a plane surface, without confusion
Increase data density
Raise the amount of information per unit area
The five methods — how displays reach those goals (each developed in its own sibling skill):
Method
One-line mechanism
Go deep in
Micro/macro readings
Fine detail aggregates into a global pattern; both visible at once
tufte-micro-macro-readings
Layering and separation
Visually distinct layers coexist on one surface and read both apart and together
tufte-layering-and-separation
Small multiples (multiplying images)
One design repeated across panels that vary by one variable; decode once, compare freely
tufte-small-multiples
Color
Encodes a named variable or separates overlapping layers — used with discipline
tufte-color-in-information-design
Narratives of space and time
The path of an entity through space and/or time, plotted as visual syntax
Related ideas the chapter leans on: multiple-functioning marks (tufte-multifunctioning-elements), chartjunk (tufte-chartjunk), data density (tufte-data-density). This skill's job is the framework plus the chapter's examples; the sibling skills carry the method theory.
§3. The Cost of Surrender — The Ise Shrine Travel Guide (p.13)
The chapter opens its gallery with a Japanese visitor's guide to the Ise Shrine (published Ise, Japan, between Oct 1948 and April 1954). It is a hybrid: a bird's-eye perspective view renders the local area around the shrine in rich, recognizable detail; then, on the right margin, the design shifts abruptly to a very flat schematic map of the national railroad linking the shrine to major cities.
Why it is the chapter's keystone example: the seam between the two halves makes the cost of flatland visible. The friendly perspective gives way to "hard flatland," and you can see the visual dimension that gets sacrificed when scale jumps from local to national. The change of design accommodates the change of scale — local detail set inside national context, a mixed landscape of refuge and overview. A bonus: the horizontal layout cooperates with the vertical orientation of Japanese text, so the stand-up labels point precisely at their locations.
Lesson: when you must compress, do it deliberately and show the seam. Mixing a high-dimension local view with a low-dimension overview can be honest design — as long as you let the reader feel what was given up, rather than pretending nothing was lost.
§4. The Spectrum of Escapes — Direct vs. Encoded
Five hundred years of practice (examples from 17 countries, 7 centuries, "3 planets and 1 star") fall on a spectrum from literal 3D simulation to pure symbolic encoding.
Escape
Example in the chapter
What it spends, what it buys
Physical model
Euclid's Elements, 1570 ed. — paper fold-up solids to teach geometry
Buys intuition for smooth shapes; useless for "obstreperous" data point clouds
Conventional perspective
Perfected by Florentine architects, 15th c.
Reads instantly for physical objects; cannot show abstract multivariate data
Stereo pairs
Aerial stereopair of Bonaduz, Switzerland (1975)
Vivid depth via two fused images; many viewers fail to fuse them
Rotating point clouds / dodecahedron projection
MACSPIN-style statistical graphics
Self-consciously multivariate; navigation in 3D scatterplots is not trivial
Mechanical simulation
Orreries, 1800
Triumph of gear ratios — but a failure mode (see §10): apparatus over content
Symbolic encoding
Java timetable, dance notation, sunspot charts
Carries the most variables per surface; demands the reader learn the code
Rule: the more abstract the data, the further right on this spectrum you must move — literal depiction runs out, and you must encode. The orrery is the cautionary midpoint: physical fidelity that drowns the very motion it was built to show.
§5. The Sunspot Lineage — Visual Reasoning to Micro/Macro (pp.18–23)
The chapter traces one dataset across ~380 years to show the same methods reappearing whenever someone faces a flood of data.
Galileo's projection method (1610–1612): project the sun's image onto a sheet a foot from the lens, draw a circle to fit, then mark each spot at correct scale and position. This converts a glare-blinding observation into a stable mark on paper flatland.
Galileo's visual reasoning (1613): through a chain of drawings he converts what "the eye of the forehead" merely registers into what "the eye of the mind" understands — turning raw spot positions into proof that the spots sit on a rotating spherical sun. It was more than 150 years before anyone materially improved his result.
Scheiner's small multiples (Rosa Ursina, 1626–1630): daily diagrams indexed on time, then apparent-path diagrams tracking spot structure across a stationary disk — "a design simultaneously enhancing dimensionality and information density." But Scheiner filled the blank, spotless regions with his patron's and religious order's emblems (a failure mode, §10).
Maunder's butterfly diagram (1904): decades of daily mapping (the chart spans 1877–1902, ±40° latitude). The sunspot's two areal dimensions are reduced to one content-relevant dimension — latitude in degrees — freeing the surface for time. Individual marks are the micro; the aggregate butterfly shape is the macro, showing the ~11-year cycle and the drift of spots toward the equator.
The micro/macro payoff and variation: fine detail merges into a repeated typical pattern, and the foremost result is a visual measure of variation around the average — both fundamental summary measures of a statistical dataset shown at once.
"Measured assessments of variability are at the heart of quantitative reasoning." — Tufte, Envisioning Information
R.A. Fisher, founder of modern statistics, wrote in 1925 that earlier workers treated variation as a nuisance detracting from the average, whereas the modern aim is to study the causes of variation itself. A good micro/macro display honors Fisher: it never reduces data to the mean alone.
The modern butterfly (Hathaway/NASA): density up tenfold, a full century, nine cycles. The sine of latitude places each mark; color encodes the log of area covered (lighter = larger) — converting a spatial dimension into a color dimension; a lower time-series, summing all latitudes, restores total percent of the sun's surface covered. Parallel sequencing of the two panels enhances both dimensionality and density.
§6. The Layering Exemplar — The Java Railroad Timetable (pp.24–26)
A 1937 graphical timetable for the Soerabaja–Djokjakarta line (annotated in Dutch then Japanese) carries 16 variables on one 24-hour sheet by smoothly suppressing a dimension here and there and finessing perspective entirely.
Encoded element
Variable(s) it carries
Horizontal axis
Time of day (read across top)
Vertical axis
Station, spaced in proportion to distance along the track (not alphabetically)
Diagonal direction
Train direction — upper-left→lower-right = "down"; lower-left→upper-right = return; an X where opposite trains pass
Diagonal slope
Relative speed (steeper = faster)
Diagonal line style (4×4 matrix)
Train type (super express / express / passenger / mixed / special / cargo / preferential / night cargo) × seasonality (regular / seasonal / irregular / special)
Left-margin profile
Elevation at enlarged vertical scale, with grade and path columns alongside
Per-station aerial insets + dingbats
Local facilities: cargo platform, livestock loading, water pump tower, directional-change platform, car repair, standby engines, relief car, closed areas
The multiple-functioning diagonal: one mark records six variables at once — a train's location between towns, the time of that position, direction, train type, relative speed (by comparing slopes), and yearly pattern. One mark doing six jobs beats six separate marks. (Tufte sources the "double-functioning element" idea from architectural criticism — Venturi, citing Wylie Sypher.)
The spy story (load-bearing detail): this 16-variable sheet was an internal planning document, obtained by agents preparing Japan's 1942 invasion of Java, and is marked "secret" in the corner. A display this dense reads as both micro (one train) and macro (the whole system's structure) — "like an excellent map, but with many dimensions breaking free of direct cartographic analogy."
Layering rules:
Do
Don't
Give each layer a distinct encoding (orientation, weight, symbol, color)
Let two layers share one visual channel and collide
Share an axis only when variables are orthogonal (distance + elapsed time on one vertical)
Force unrelated variables onto the same axis
Make a mark earn its place by doing two jobs
Add a separate mark for every variable
§7. The Narrative Exemplars — Dance and Weather (pp.27–28)
Tomlinson's dance notation (1735): four dimensions on one page — the flatland of the ballroom floor, coded gestures of body motion, and time sequence. The floor plan links to the music (itself two dimensions plus time and tone) by numbers that double-function: they sequence the steps and tie each step to a musical measure. Movement is symbol-encoded rather than verbal because a serious notation avoids words, which are a barrier to international use. Mirrored partners are drawn in full (redundant bilateral symmetry usually wastes space) because their weaving, intermingling paths require the symmetric repetition to read.
Japanese newspaper weather cross-section (1985): instead of the usual plan-view of latitude and longitude, the map shows Japan as a vertical side profile, an ocean-eye view, with gray contours tracing 0°C and -10°C through the clouds and forecast symbols for 15 cities. By spending one dimension on elevation rather than a second horizontal axis, it reads far more easily than a conventional weather map. Caveat: the trick works best for long, thin countries.
Narrative rule: match each dimension's encoding to its nature — time → sequential position, direction → orientation, speed → slope — and cherish every chance to spread an extra variable over a dimension you already have on the page.
§8. The Small-Multiple Exemplars and Data Density (pp.29–33)
Small multiples "move to the heart of visual reasoning — to see, distinguish, choose." Their multiplied smallness forces comparisons within a single eyespan, so the active eye does the work rather than memory bridging images scattered across pages.
California air pollution (computer-plotted): 3 pollutants (nitrogen oxides, carbon monoxide, reactive hydrocarbons) × 4 times of day (midnight–3am, 6–9am, noon–3pm, 3–6pm) = 12 surface maps of southern California, each over a grid of 2,400 squares at 5 km/side = 28,800 readings (except peaks masked by mountains). Five variables: two-space location + time-of-day + pollutant type. Economy of perception: decode the design once, then read every panel for free; a steady canvas makes the data changes pop.
Compound interest tables (Barème Universel, 1822): tabular small multiples — a 2D array of capital × time, repeated at successive interest rates. The same pattern as the pollution maps, in pure numbers: a third variable on the two-surface, then the whole grid repeated at levels of a fourth.
T-shirt color grid: children's shirts arranged by body color × collar color — a bivariate small multiple where color itself is the data, the eye surveying a 2D color space directly.
Data density (the second goal, with benchmarks):
Display
Volume
Composition
Maunder butterfly (1904)
daily, 1877–1902
each mark = one observation; two areal dims → latitude
Modern butterfly
10× the 1904 density
full century, nine cycles, color = log area
California pollution
28,800 readings
2,400 squares × 12 maps
Compound interest
full grid × rates
capital × time, repeated per interest rate
Tokyo weather history
414 data pieces
town × year × month × day × day/night + summary rows
Tokyo weather almanac
1,826 days
decade by town × year × month × day; year-by-day matrix compounded by month
Java timetable
16 variables
single sheet
Density threshold: high-information graphics convey "a spirit of quantitative depth and statistical integrity"; emaciated, data-thin designs provoke suspicion — rightfully — about the quality of the underlying measurement and analysis. The Tokyo almanac is the master move: it apportions both of the paper's dimensions to one variable (time), splitting it fine-to-coarse (day within month within year) to stretch a compact display across 1,826 days.
Implicit formula: data density = (number of data entries) / (area of display). Maximize it only as far as legibility holds; small multiples buy density cheaply because the design is decoded once.
§9. Multivariate Tables — LSAT Grid and the Gotti Courtroom Chart (pp.29–31)
Numbers on a plane can multiply too; two table examples close the gallery.
The LSAT bivariate grid (NYU School of Law, 1983–84): GPA (rows) × LSAT percentile (columns) knits together every combination of grades and test scores. Each cell holds two numbers — applicants and, of those, the number actually admitted (e.g. "7-0"). The right and bottom margins sum to univariate distributions, with grand totals at lower right. The table could itself become a single cell in a higher table comparing many schools — multivariate nesting.
The Gotti chart — "Criminal Activity of Government Informants" (United States v. Gotti, 1987): a courtroom exhibit, supplied by the defense, listing 69 crimes (murder, drugs, kidnapping, down to "pistol whipping a priest") committed by 7 prosecution witnesses. It is built to be read both horizontally and vertically: across a row you see how widely one crime spreads among witnesses; down a column you see one witness's full record, and the eye catches unbroken runs of X's (one witness "has something of a streak going"). Only 37% of all possible cells are marked, yet the marked ones dominate. Placement is rhetorical — the most obnoxious crimes sit at the top (murder) and bottom (pistol-whipping a priest), exploiting the visual prominence of those edge positions.
Why tables win in a trial: speech is linear, one-dimensional, non-reversible "talk talk talk"; a visual table is simultaneously a wideband and a perceiver-controllable channel, letting each juror read, re-edit, and reason at their own pace and in their own order.
§10. The Failure Mode — Chartjunk and the Duck (pp.33–35)
The opposite of escaping flatland is decorating it. The good methods are self-effacing: when they work, only the data is visible and the container disappears. Chartjunk inverts this — the apparatus attracts attention to itself.
The "duck" (from Venturi's Learning from Las Vegas): when Modern architects abandoned applied ornament, they distorted the whole building into ornament — the container became the message. Pugin's warning applies to data: it is fine to decorate construction, never to construct decoration. A 3D bar chart whose third axis carries no variable is a duck; the "diamonds" chart with a fishnet-stocking thigh-graph is a duck that also confuses changes in money's value with changes in diamond prices during high inflation.
Named failure modes:
Failure mode
Description
Example in the chapter
Pridefully Obvious Presentation
Display apparatus overwhelms the data content
Orreries directing the eye to gears, not planetary motion
Embellishment-for-data
Ornament fills space that should hold data
Scheiner's patron/religious emblems in the blank, spotless disk regions
The Duck
The container becomes the message; content-empty third dimension
Diamonds fishnet-thigh chart; any 3D shell with no third variable
Data posterization
Closely-read data styled like a poster (strong image, big type, thin data)
Chartjunk "performances" that read like posters, not maps
Contempt for the audience
Assumes readers are bored and stupid; adds ornament to "enliven"
"Lively" infographics with icons replacing data marks
False escape / pretend dimensions
Fake dimensions added to an impoverished dataset
The royal dining table so caught in flatland it cannot hold the pots
Jonathan Swift's 17th-century cartographers, who filled unknown territory with pictures of elephants for want of real towns, are Tufte's image for substituting embellishment where data should be.
The moral premise — respect for the audience. Chartjunk rests on the false belief that "if the numbers are boring, you've got the wrong numbers" is not true — when in fact it is. Consumers of graphics are usually smarter about the subject than the people decorating it. Standards of excellence are set by high-quality maps — bountiful detail, layered close reading, rigorous survey data — not by posters meant for viewing at a distance.
"Readers of a report should be unaware of its design." — Paul Rand, quoted in Tufte, Envisioning Information
(Reinforced by E.B. White: no one writes decently who distrusts the reader's intelligence. Clarity and simplicity are the opposite of simple-mindedness — they are the highest respect for an alert, busy, caring reader.)
§11. Historical Lineage
Date
Development
Escape technique
15th c. Italy
Florentine architects perfect perspective
Conventional 3D projection on a 2D surface
1570
Euclid's Elements, Billingsley ed.
Paper fold-up models for solid geometry
1610–1612
Galileo's sunspot observations
Projection: marking spots onto paper at true scale/position
1613
Galileo's Istoria e Dimostrazioni
Visual reasoning: data → mechanism
1626–1630
Scheiner's Rosa Ursina
Time-indexed small multiples; apparent-path diagrams
1735
Tomlinson's dance notation
Floor-plan + coded gesture + time, with double-functioning numbers
1800
Orreries
Mechanical 3D simulation — but Pridefully Obvious Presentation
Hybrid perspective + flat map; the cost of flatland made visible
1980s
California air pollution maps
Computer small multiples, 3 pollutants × 4 times
1984–85
Tokyo weather almanac
1,826-day matrix, both paper dimensions spent on time
1985
Japan weather cross-section
Space-time narrative as a vertical side profile
1987
LSAT grid; Gotti courtroom chart
Bivariate tables read both directions
§12. Do/Don't Summary
Principle
Do
Don't
Goals first
Name the two goals: more dimensions AND more density
Optimize one while ignoring the other
Taxonomy
Use the five methods; treat density as a goal
List "six strategies" with density as a peer method
Compression
Compress deliberately and show the seam (Ise Shrine)
Pretend nothing was lost when scale jumps
Variation
Show average AND variation (Fisher, Maunder)
Reduce the data to its mean
Multiple-functioning marks
Let one mark do several jobs (Java diagonal = 6)
Add a separate mark per variable
Tables
Build grids readable both ways (Gotti, LSAT)
Settle for a one-direction list
Density
Push density to the legibility limit
Ship emaciated, data-thin displays
Self-effacement
Make the container vanish, the data show
Build a duck — container as message
Audience
Assume an alert, caring, busy reader
"Enliven" numbers with ornament
§13. Application Checklist
Before shipping any chart, map, table, or display:
State the two goals for this display: which variables/dimensions, what target density?
Pick which of the five methods apply, and open the matching sibling skill for mechanics.
How many variables does the data have vs. how many does the display encode?
Is every mark doing at least one job? Can any mark do two (a multiple-functioning element)?
Does it pass micro/macro: legible detail AND legible overall pattern? Does it show variation, not just the mean?
If compressing or mixing scales, is the seam honest (Ise Shrine) rather than hidden?
If tabular, can it be read both horizontally and vertically? Do margins carry the univariate sums?
Is density above the suspicion threshold for the subject's complexity, while still legible?
Any third dimension, ornament, or color encoding no variable? Remove it (no ducks).
Does the container (axes, grid, border, legend) out-area the data? Shrink the container.
Would a reader who knows the encoding decode any panel without re-reading the instructions?
Source: Envisioning Information (Edward Tufte, Graphics Press, 1990), ch. 1 "Escaping Flatland," pp. 12–35 (the chapter thesis and two goals appear on pp.12–13; the five methods are previewed p.33).
