Cézanne 1864–1906

A color and structural kernel reading of Cézanne

Russell Parrish · Artist Influencer · artistinfluencer.com · Images: WikiArt CDN · deduplicated

What This Shows

Five measurements extracted from 427 unique Cézanne paintings across four decades, plotted against time and subject. Color panel: how broadly distributed hue is across each painting, and how often color draws edges where light does not. Structural panel: how cohesive the gradient field is (one dominant edge mass, or many disconnected fragments) and how dispersed visual energy is across the canvas.

These metrics measure pixel structure, not intention. With Cézanne that gap is real: passage, warm-cool planar modulation, and the deliberate assertion of the canvas surface operate in registers the kernel sees obliquely or not at all. The findings should be read as pixel observations first, then translated into Cézanne's language. They are not 1:1.

The canonical Cézanne, geometric and building toward resolution, shows up in the still lifes. The late mountain paintings are the most structurally dispersed work he made. The kernel found the seam. He put it there.

Four Periods

Early 1864–1872	HE 0.623 · CBI 0.103	Palette knife dominates. High mu (structural cohesion peak): unified gradient slabs, no passage, no breathing room between planes. Dark tonal ground, aggressive. Color largely follows luminance. n=79.
Impressionist 1873–1882	HE 0.622 · CBI 0.128	Working with Pissarro. CBI rises, color begins loosening from luminance at edges. High-key rises toward career high (0.180). Still life practice intensifies; mu holds in studio subjects. CI/CBI begins its career range (0.683). n=129.
Constructive 1883–1895	HE 0.649 · CBI 0.113	Parallel hatching at full development. HE career peak. CBI dips: color is not leading at edges; it is distributed across the whole field. CI/CBI rises (0.715): field-wide decoupling, not edge-event autonomy. High-key peaks (0.250). Still life mu peaks; the studio is the laboratory. n=139.
Late 1896–1906	HE 0.661 · CBI 0.131	HE and CBI both return toward Impressionist levels. CI/CBI holds high (0.730), color continues working field-wide. mu reaches career low (0.302): passage is fully operational, contours dissolved. Chateau Noir forest paintings show highest subject CBI. Portrait of the Gardener Vallier: HK 0.0004, the darkest painting in the dataset. n=80.

Career Arcs

Top: Hue Entropy. Middle: CBI. Bottom: mu (structural cohesion). Constructive period marked ↑ for HE peak; Late period ↓ for mu career low.

Early

0.623

0.103

0.381

→

Impressionist

0.622

0.128

0.326

→

Constructive

0.649↑

0.113

0.354

→

Late

0.661

0.131

0.302↓

high_key (tonal lightness): rises through the career, peaking at Constructive: 0.130 → 0.180 → 0.250↑ → 0.240

CI/CBI ratio (field-wide vs. edge-event color autonomy): consistently elevated across all periods: 0.745 → 0.683 → 0.715 → 0.730

Subject Findings

Still life	CBI 0.096 · mu 0.404↑	Lowest CBI of any subject. Highest mu. The pixel structure is most cohesive here: one dominant gradient field. In Cézanne language: the studio table is where form is most completely argued through tonal structure. Color does not lead at edges. It modulates within the plane, where this kernel cannot follow.
Portrait	CBI 0.113 · CI/CBI 0.723	Highest CI/CBI ratio of any major subject in late decade (1900s: 0.722). Color decoupled field-wide from luminance across the whole figure, not at its edge. The Gardener Vallier (1906): HK 0.0004, near-total darkness, CBI 0.172. Color holding the figure against the ground with no tonal scaffolding left.
Provence	CBI 0.131 · CI/CBI 0.845*	CI/CBI rises to 0.845 in the 1900s, the highest subject-decade value in the dataset. CBI falls simultaneously. Color is not creating edge events; it is occupying planes. The warm ochre and cool green are no longer describing a hillside. They are two planes making an argument. (*1900s decade mean.)
Mont Sainte-Victoire	HE 0.704↑ · mu 0.221↓	Highest HE of any subject. Lowest mu of any subject. The mountain paintings use the broadest palette and have the most fragmented gradient field. In Cézanne language: passage is most fully developed here. The edge between mountain and sky is held open deliberately. The kernel reads disconnection. He intended ambiguity.
Card Players	CBI 0.078↓ · HK 0.068	Lowest CBI subject mean. The early large versions (Barnes, Met, Orsay) approach near-zero CBI, color entirely following luminance at edges. Color serves the figure's volume through conventional tonal modeling. As the series progresses toward smaller versions, CBI rises: the chromatic grip loosens. The most formally celebrated paintings are the most conventionally structured by measurement.
Chateau Noir	CBI 0.163↑ · mu 0.270	Highest CBI of any subject. The forest and rock paintings suppress luminance hierarchy through dark canopy and fragmented stone surfaces. Color is forced to carry edge structure. Same atmospheric mechanism as fog-suppressed luminance in atmospheric painting: different surface, same measurement.

What the Axes Measure

Color Panel

Hue Entropy

Shannon entropy of a chroma-weighted 36-bin hue histogram. High = palette spread across the full hue wheel. Low = palette concentrated in one or two families. Career range: 0.21 to 0.92.

Color Panel

CBI

Chromatic Boundary Independence. Color drawing edges where light does not draw them. Top 15% chroma gradient vs. bottom 30% luminance gradient. Warm-cool planar modulation within a plane is largely invisible to CBI; it works at edges.

Structural Panel

mu (μ)

Cohesion: largest connected edge component as fraction of total edge mass. High = one dominant gradient structure. Low = gradient dispersed across many disconnected regions. In Cézanne, low mu is passage working as intended, not structural failure.

Structural Panel

CI/CBI ratio

color_independence ÷ CBI. When ratio is high, color decoupling is distributed field-wide, not concentrated in edge hotspots. Runs 0.683–0.745 across all four periods, field-wide by construction, not a late development.

Navigation & Controls

Click + dragPan within panel

Double-clickReset zoom and pan

HoverThumbnail, title, year, HE, CBI, CI/CBI, high-key, mu

Period togglesIsolate periods; stats bar updates

Subject togglesFilter by subject series

Fixed / rescaleRescale axes to visible range

Tooltip Metrics

Metric	What it measures
`hue_entropy`	Angular breadth of color use, 0–1. Career median: 0.648. Mont Sainte-Victoire has the highest subject mean (0.704); Card Players the most concentrated palette within a major series.
`CBI`	Color-edge autonomy from luminance. Chateau Noir forest paintings lead all subjects (0.163). Card Players are lowest (0.078). CBI is an edge metric; it cannot see warm-cool planar modulation happening within a plane. Moderate CBI rising over time (r = +0.109 with year, weak) means controlled structural reassignment, not color autonomy for its own sake. Subject type predicts CBI more than time does.
`CI/CBI`	Field-wide color decoupling relative to edge-event autonomy. Runs 0.683–0.745 across all four periods, elevated above the career range for most atmospheric painters, where color autonomy concentrates in edge hotspots rather than distributing field-wide. Late Provence 1900s peaks at 0.845: color occupying planes, not drawing edges. CI/CBI values above 1.0 likely reflect denominator instability when CBI approaches zero; cross-check raw CBI before interpreting as chromatic liberation.
`high_key`	Fraction of pixels with OKLab L ≥ 0.72. Rises across the career (0.130→0.250), peaking in the Constructive period before a slight late pullback. Portrait of the Gardener Vallier (1906): 0.0004, the darkest painting in the dataset.
`mu`	Structural cohesion. Early palette knife work peaks (0.381). Declines with constructive partial recovery (0.354). Late career low (0.302). Low mu does not indicate visual incoherence; in Cézanne it reflects distributed passage and unresolved boundary sharing across planes. Still life mu highest of any subject (0.404): the studio is the laboratory.

What to Look For

Early period: high mu, low CBI. The palette knife paintings are structurally unified and chromatically conventional. Uncle Dominique as a Monk (1866) has mu 0.851: one connected gradient mass, no passage, no ambiguity. The problem he would spend forty years solving had not yet been found.

Constructive period: CBI dips, CI/CBI rises. Color autonomy moves away from edges and distributes across the whole surface. The parallel hatching decouples color from luminance field-wide, not through dramatic edge events but through distributed marks. This is not color doing less structural work. It is color doing structural work differently, in a register CBI cannot fully see.

Late period: mu at career low. The mountain paintings and forest scenes show the most dispersed gradient fields. Passage is fully operational. Toggle subject = mont_sainte_victoire and compare its mu against still_life. Same artist, same decade, opposite structural reading.

Card Players: CBI near zero across all versions. The most formally celebrated paintings are the most conventionally luminance-led. Watch CBI rise from the early large-format versions (Barnes, Orsay) toward the later smaller ones, the chromatic grip loosening across the series.

What the Measurements Show

Cézanne does not simply reduce structure. The measurements suggest a transfer from centralized form toward distributed color relationships. That is what the data shows across 427 paintings and four decades. mu declines. CBI rises moderately. HE holds. Palette complexity does not contract when structural cohesion breaks. He increases the structural demand on the same palette. That is the system finding.

The still life is his laboratory and the data confirms it structurally. Still life mu rises to 0.467 in the 1880s, the highest subject-decade cohesion value in the dataset, before pulling back in the 1890s. The studio table is where gradient structure is most completely resolved. Not the landscapes. Not the mountain. The apples on a table are where form is being worked out in the most controlled terms the measurement can find. And the palette breadth of those paintings is angular, not saturated: HE and chroma correlate negatively across the full dataset (r = −0.363): more of the hue wheel requires spreading chromatic intensity thinner. The ochre-dominant, low-saturation quality of the still lifes is a structural consequence of palette breadth, not a tonal choice made independently.

The CI/CBI ratio runs 0.683–0.745 across all four periods, consistently elevated, not a late development. Field-wide color decoupling was a baseline characteristic, present from the Impressionist period onward. The parallel hatching of the constructive period distributes color-luminance separation across the whole surface rather than concentrating it at edges. CI/CBI is a supporting indicator here, not the headline; it confirms the field-wide character of the system but should not be over-read. CBI itself is moderate and rising weakly with time (r = +0.109): subject type predicts color-edge autonomy far more than career progression does. Chateau Noir forest paintings and Bathers show the highest CBI; Card Players and still lifes the lowest. The subject is doing most of the work.

The Card Players invert the canonical reading. The early large-format versions (Barnes, Metropolitan, Orsay) have CBI values approaching zero. Color is following luminance at every edge. The figure structure is built through conventional tonal modeling; color serves form rather than asserting itself. The later smaller versions show higher CBI. The chromatic grip loosens as the series progresses. The paintings art history calls his greatest formal achievement are, by measurement, his most conventionally structured. The later ones are where color starts making independent decisions. Meyer Schapiro identified the tension between Cézanne's formal ambitions and his subject choices. The measurement finds it too.

Mont Sainte-Victoire has the broadest palette of any subject (HE 0.704) and the most fragmented gradient field (mu 0.221). These two readings translate directly into Cézanne's method. The hue breadth captures the mountain's simultaneous presence of sky blue, warm limestone ochre, dark pine green, and distant violet atmosphere, all held in one frame. The low mu captures passage: the edge between mountain and sky is not completed. The plane of rock and the plane of air share the boundary. The kernel reads this as disconnection; Cézanne intended it as ambiguity held open. The measurement is correct. The interpretation requires translation.

The late Provence landscapes produce the highest CI/CBI ratio of any subject-decade in the dataset: 0.845 in the 1900s. CBI falls simultaneously. Color is no longer creating edge events; it is occupying planes. The warm ochre and cool green are not describing a hillside through light and shadow. They are two planes making a structural argument. The kernel cannot name that argument. It can confirm that color has stopped following luminance and is operating independently across the whole field.

Portrait of the Gardener Vallier (1906): high_key 0.0004, the darkest painting in the dataset. CBI 0.172, moderate color-edge autonomy in near-total darkness. He is holding the figure against the ground with no tonal scaffolding remaining. The measurement gives you a painting that is almost entirely dark ground with color quietly carrying what structure there is. That is not a decline reading. It is a consistency reading: the same problem he was solving in 1866, finally without the crutch of value contrast. The counterevidence is in the same year: Portrait of Vallier (different view, 1906) has mu 0.727, high enough to rank among the early palette knife portraits. The dispersal in the Gardener version was a choice, not a capacity limit.

Dataset & limitations. 427 paintings deduplicated from WikiArt CDN. Four periods follow the standard Cézanne periodization; subject tags assigned before kernel computation. CBI is an edge metric and cannot observe warm-cool planar modulation occurring within a plane, a central Cézanne technique. Low mu should be read as passage working as intended, not structural incoherence. Blank canvas between brushstrokes inflates local gradient density; paintings with significant bare canvas may show artificially elevated gradient counts at stroke boundaries. CBI values are measured at thumbnail resolution (200px long edge) and carry a slight upward bias, diminishing toward zero for high-CBI works. Period rankings and relative comparisons are not affected. The ≥ 0.20 individual threshold was calibrated against full-resolution images; at thumbnail resolution the equivalent is approximately 0.23.

Form to Color