Start with one solid equilateral triangle. Remove its central inverted triangle, repeat inside every surviving corner triangle, and continue indefinitely. At stage five, just 243 small triangles remain and the total filled area has fallen to 243/1024 of the original. In the mathematical limit, the area becomes zero even though a structured, uncountable set remains.
The Sierpiński triangle illustrates why area, number of pieces and geometric dimension must be answered separately. We will derive the exact formulas rather than treating an increasingly detailed picture as proof, distinguish each finite stage from the limiting set, and show how the same logic informs digital images and experiments without pretending a physical object is infinitely resolved.
Find the question you need
The twelve chapters are open and readable from beginning to end. Choose a route for a particular question or use the expandable index.
Learn the exact construction
Chapters 1–2. Begin this route.
Derive the count and area
Chapters 3–6. Begin this route.
Understand fractal dimension
Chapters 7–8. Begin this route.
Check real-world and image limits
Chapters 9–10. Begin this route.
Practise and connect
Chapters 11–12. Begin this route.
Complete chapter index · Advanced worked lab · Exercises with answers · Fractals subject root
Chapter index
Learn the exact construction · Chapters 1–2
Derive the count and area · Chapters 3–6
Understand fractal dimension · Chapters 7–8
Check real-world and image limits · Chapters 9–10
Practise and connect · Chapters 11–12
Begin with an equilateral triangle of side length one. Connect the midpoints of its three sides. Four smaller equilateral triangles of half the original side length appear; remove the open central triangle and retain the three corner triangles. Repeat the procedure independently inside each surviving triangle. We call the uncut triangle stage zero and the first removal stage one.
The geometric instruction is unusually precise. Every retained piece is scaled by a factor of 1/2 in side length, and each gives rise to three retained pieces at the next stage. This is not the process of repeatedly shrinking one triangle; it is a simultaneous replacement rule acting on all the remaining triangles.
The stage-n object is a finite union of closed triangles that includes their boundaries. The limiting object is their nested intersection as n tends to infinity. Keeping this distinction prevents an important misconception: a finite picture with 243 filled triangles is not literally a zero-area mathematical fractal.
A drawing tool may use solid pixels to show the retained regions and pale pixels for removed holes. Those colours are a display choice. The exact construction defines which points belong to the set, regardless of which software or screen is used to illustrate it.
Let Nₙ be the number of surviving small triangles at stage n, sₙ their side length measured relative to the original unit side, and Aₙ the total occupied area. Start with N₀=1, s₀=1 and A₀=√3/4 square units. For any other initial side a, A₀=√3a²/4.
The replacement rule makes the count triple: Nₙ=3ⁿ. Each stage halves every side: sₙ=(1/2)ⁿ. Since area scales as the square of a length factor, each small triangle has 4⁻ⁿ times the original triangle’s area.
Multiply count by area per piece: Aₙ=3ⁿ×4⁻ⁿA₀=(3/4)ⁿA₀. These three formulas are linked but not interchangeable: 3ⁿ is a count, 2⁻ⁿ a linear scale and (3/4)ⁿ an area fraction.
If a calculation says the occupied area halves at every iteration, it has confused the side-length rule with the area rule. If it says area triples, it has retained the count increase while forgetting that each individual triangle becomes four times smaller in area.
Contents · Previous chapter · Next chapter · Fractals subject root
At stage one, three retained triangles each have side 1/2 and area A₀/4, leaving A₁=3A₀/4. At stage two, nine triangles have side 1/4 and each area A₀/16, leaving A₂=9A₀/16. The pattern continues without a new geometric rule.
At stage three, N₃=27, s₃=1/8 and A₃=27A₀/64≈0.421875A₀. At stage four, N₄=81, s₄=1/16 and A₄=81A₀/256≈0.316406A₀. The number of pieces increases even while the total area falls.
Use these results to audit any construction diagram. Count the major triangular groups, notice each scale change and ask whether the total dark area seems to shrink by three-quarters. A drawing that contains four retained corner triangles at the first step implements a different rule.
For a unit-side initial triangle, stage-four area is (81/256)(√3/4)≈0.1370 square units. A scientific explanation should state whether it reports the absolute area, with square units, or a fraction of the original, which is dimensionless.
Contents · Previous chapter · Next chapter · Fractals subject root
The formula Aₙ/A₀=(3/4)ⁿ follows from the recursion Aₙ=(3/4)Aₙ₋₁. We multiply by three-quarters once per iteration, not once per triangle. This ratio arises because each old triangle is partitioned into four equal-area subtriangles and precisely three remain.
Suppose a stage-five drawing is required. It contains 3⁵=243 retained triangles, each with side 1/32 and area A₀/1024. Its total area is 243A₀/1024≈0.2373047A₀. The fraction can be obtained either from counting or from (3/4)⁵; independent routes are a useful accuracy check.
Stage ten leaves 3¹⁰=59,049 conceptual triangles, each 1/1024 of the initial side, while the total area fraction is (3/4)¹⁰≈0.0563135. So under six percent of the original filled area remains, despite the rapidly multiplying count.
This is a geometric decay problem. A finite positive area at every stage does not stop the sequence of areas converging to zero. Conversely, a limiting area of zero does not imply that one of the earlier finite stages had no occupied area.
| Stage n | Retained triangles | Side / original | Area / original |
|---|---|---|---|
| 0 | 1 | 1 | 1 |
| 1 | 3 | 1/2 | 3/4 |
| 2 | 9 | 1/4 | 9/16 |
| 3 | 27 | 1/8 | 27/64 |
| 4 | 81 | 1/16 | 81/256 |
| 5 | 243 | 1/32 | 243/1024 |
Contents · Previous chapter · Next chapter · Fractals subject root
Because 0<3/4<1, the factor (3/4)ⁿ approaches zero as n becomes arbitrarily large. It follows that lim Aₙ=0. The decreasing closed triangular stages have finite area at the beginning, so the measure of their nested limit is zero under the usual planar-area interpretation.
This is not an argument that the set contains no points. All three original triangle vertices remain present, as do many points generated by the infinitely repeating corner construction. In fact the resulting set is uncountable and carries a nontrivial spatial arrangement.
The common intuition error is to use one everyday phrase—’there is nothing left’—for two different claims. Zero two-dimensional area means the limit occupies no positive plane area; it does not mean the set is empty. The original outer vertices provide an immediate counterexample to emptiness.
A practical parallel arises in geometry exercises where a line segment has zero plane area but certainly contains points. The Sierpiński object is more elaborate because its structure remains spread across a triangle at infinitely many scales.
Contents · Previous chapter · Next chapter · Fractals subject root
At stage one, the removed central triangle occupies one-quarter of the original area. At stage two, three newly removed middle triangles each have area A₀/16, removing 3A₀/16 in that stage. Stage three removes 9A₀/64, and so on.
The amount newly removed at stage n≥1 is ΔAremoved,n=(1/4)(3/4)ⁿ⁻¹A₀. The infinite geometric series of removed area therefore totals (A₀/4)/(1−3/4)=A₀.
Exactly the original area has been removed in the limit, although no single finite step removes everything. This complements the remaining-area expression: A₀−Aₙ is the cumulative area removed through stage n.
When constructing an Excel, Python or hand-computed table, place remaining and removed fractions in adjacent columns. For every n, they should sum to one. That check detects off-by-one iteration errors that a convincing graphic might hide.
Contents · Previous chapter · Next chapter · Fractals subject root
The infinite triangle consists of three self-similar copies of itself, each scaled to half size. If a dimension exponent D captures this ideal scaling, consistency requires 3(1/2)ᴰ=1. Solve for D to obtain D=log(3)/log(2)≈1.5849625.
This number lies between one and two. It is larger than the dimension of a smooth line segment, but smaller than the dimension of a filled planar region. The value describes how the quantity of detail grows as the observing scale becomes finer, not a claim that space has developed one and a half coordinate axes.
At the natural stage-n scale ε=2⁻ⁿ, the model has 3ⁿ distinct retained pieces. The ratio log(3ⁿ)/log(2ⁿ) gives the same exponent. For this self-similar set with the usual separation conditions, its Hausdorff and box-counting dimensions agree.
Compare our earlier Cantor-set guide, where two pieces appear at one-third scale and the resulting dimension is log 2/log 3≈0.63093. The replacement rule determines the exponent.
Contents · Previous chapter · Next chapter · Fractals subject root
CHAPTER 8 OF 12
8. Do not confuse the dimensions of the finite and infinite objects
For any fixed finite n, the retained figure is a union of triangles with nonempty interiors. Therefore its standard two-dimensional Hausdorff and box-counting dimension is two, even if the occupied area has become very small. Its many triangular holes do not change this fact.
For the limiting Sierpiński triangle, every would-be interior region is eventually interrupted by a removed central triangle. The limit has zero area and dimension log 3/log 2≈1.585. That is a property of the infinite nested intersection, not the fixed stage-five polygon.
This difference is easy to lose when using digital pictures. A computer stores a finite grid, so its displayed version has a pixel scale and finite resolution. At scales below that pixel size, the mathematical scaling relationship is not represented by new information.
To report a fractal dimension estimated from image data, name the scaling range and explain how the binary mask was created. Merely drawing a few iterations and appending the exact mathematical dimension is not a measurement of a real sample.
Contents · Previous chapter · Next chapter · Fractals subject root
CHAPTER 9 OF 12
9. Explain why counting pixels is not the same as counting pieces
Consider a rasterised stage-six triangle. Many of its smallest pieces may occupy only a few pixels. When the image is enlarged, interpolation can blur edges and make removed holes appear partly filled. When it is reduced, two nearby corner pieces can merge visually even though the ideal construction keeps their mathematical distinction.
Pixel counts depend on threshold, antialiasing, cropping and output resolution. A single row of dark pixels might be a real fine line, a display artefact or an edge created by a thresholding rule. Those possibilities require different interpretations.
For exact classroom mathematics, use the algebraic count 3ⁿ and area fraction (3/4)ⁿ. For an empirical figure, keep the original scale and investigate how results change under reasonable grid translations and thresholds.
A helpful diagnostic is to plot log of occupied box count against log of inverse box size. A stable near-linear interval can support an empirical scaling exponent over that interval, while curvature or a fine-scale plateau warns that one apparent ‘dimension’ may be dominated by finite resolution.
Contents · Previous chapter · Next chapter · Fractals subject root
The Koch snowflake repeatedly replaces each boundary segment with four segments one-third as long. Its perimeter grows by 4/3 per iteration, while its enclosed area tends to a finite positive limit. The Sierpiński triangle instead repeatedly removes part of a filled region; its remaining area shrinks by 3/4 toward zero.
Both objects have noninteger boundary or set dimension under their respective ideal mathematical constructions, but the quantity being analysed is different. For the Koch snowflake the scaling object in the dimension formula is its boundary curve. For the Sierpiński triangle it is the surviving triangular gasket.
The comparison is a warning against treating ‘fractal’ as one mechanical definition. Different replacement rules yield different counts, dimensions, measures and topology. The same word cannot determine whether area converges, stays positive or diverges.
Read Infinite Koch Snowflake Perimeter, Finite Enclosed Area after this chapter, and write down the scaling factors in each construction. That is a more reliable comparison than comparing decorative illustrations.
Contents · Previous chapter · Next chapter · Fractals subject root
Problem A: A learner starts at stage zero and asks for the stage-six piece count. Compute N₆=3⁶=729. The side length of each retained triangle is 1/64 of the original and the total area fraction is 729/4096≈0.17798.
Problem B: A second learner says, ‘Stage six has almost zero area, so its dimension is 1.585.’ The word ‘almost’ does not solve the mathematical issue. Each stage-six triangle still has interior, so the finite filled stage has dimension two. The limit of the nested sequence has dimension about 1.585.
Problem C: Replace the construction rule with four retained copies at half scale. The area ratio becomes 4×(1/2)²=1, so area no longer shrinks. The similarity-dimension equation gives D=log 4/log 2=2 under the appropriate configuration. Changing the number of retained pieces changes the result.
Problem D: If the initial side is two metres instead of one, stage-three side length is 2/8=0.25 metre. Its total area is (27/64)×√3 square metres. The relative area fraction stays 27/64. This checks that an absolute unit and a dimensionless fraction must not be confused.
Contents · Previous chapter · Next chapter · Fractals subject root
Begin every recursive-geometry question with a drawing of one valid replacement step. Identify the number of self-similar pieces N, the linear scale r, whether pieces overlap in area, and which measure you want: count, length, area, volume or scaling dimension. Write a recurrence before jumping to a logarithm.
Second, compute small stages exactly. A stage table reveals whether the exponent comes from one application of the rule or from n iterations, and helps distinguish the number of retained pieces from the number newly removed. Only then take the infinite-stage limit, using a justified sequence or series convergence argument.
Third, test what the mathematical model licenses. An exact ideal fractal calculation does not establish the dimension of a leaf, cloud or medical image. An empirical application needs sampling resolution, classification choices and scale sensitivity. Transfer the method, not an unsupported exact exponent.
For connected reading, start again at The Gold Standard of Fractals, then compare the Cantor set and Koch snowflake. These three worked constructions clarify the distinct roles of count, area, dimension and limiting process.
Common questions
Does zero area mean the limiting triangle disappears?
No. Plane area measures the size of a two-dimensional region; the limiting gasket is a nonempty uncountable set with zero plane area and a noninteger Hausdorff dimension.
Why is the finite drawing not already dimension 1.585?
Every finite stage still contains ordinary filled triangles, so its dimension is two. The exponent log 3/log 2 belongs to the infinite limit under the exact self-similar construction.
Can I assign this exact dimension to a photograph?
Not by appearance alone. A photograph has finite pixels, and threshold choices, noise and image scale influence estimated box counts. An empirical estimate needs a justified scale interval and uncertainty checks.
Sources and connected reading
All numerical calculations in the worked example are derived from the stated ideal construction. The educational examples are not presented as measurements of natural surfaces.
- Wolfram MathWorld — Sierpiński Sieve
- The Gold Standard of Fractals | Cantor Set and Box-Counting
- The Gold Standard of Fractals | Koch Snowflake
Advanced worked lab: scale, boundaries and teaching checks
Boundary length is a different limit from filled area
One more quantity reveals why the same drawing can support two opposite-looking statements. At stage n there are 3ⁿ small retained equilateral triangles of side 2⁻ⁿ times the original. Each small triangle has a three-sided boundary. Because distinct retained pieces meet only at corner points, the total length of all their edges at that finite stage is 3ⁿ×3×2⁻ⁿ times the original side length.
For an original side of one length unit, the sum of these finite-stage edge lengths is 3(3/2)ⁿ. That quantity grows without bound as n increases. At stage five it is 3×(3/2)⁵=22.78125 length units, whereas the remaining filled area is only 243/1024≈23.73% of the initial triangle’s area.
Neither fact negates the other: a length and an area use different units and respond differently to subdivision. A learner should not infer that ‘area approaching zero’ forces all geometric lengths to become small. Equally, an infinite-stage boundary must be discussed carefully in the terminology of fractal measure rather than identifying one finite polygon with the final set.
Notice that these edge-length calculations apply to the finite collection of triangles. The mathematical limit itself is not a conventional filled polygon with a smooth enclosing boundary. Its noninteger dimension, previously derived, gives the more useful description of how the surviving set occupies scale.
How many iterations are needed to leave less than one percent of the area?
Consider a school investigation that asks for the first stage where the total filled area is strictly less than 1% of the original. Begin with (3/4)ⁿ<0.01. Take natural logarithms: n ln(3/4)<ln(0.01). Because ln(3/4) is negative, dividing by it reverses the inequality: n>ln(0.01)/ln(3/4).
The ratio is a little greater than sixteen, so the first integer stage satisfying the question is n=17. At stage sixteen the remaining fraction is approximately 0.01002, still slightly above 1%. At stage seventeen it is approximately 0.00752, below 1%.
This is a useful lesson in checking borderline numerical conditions. Rounding stage sixteen’s fraction to 1.0% prematurely can disguise that ‘less than’ is a strict inequality. Reporting ‘about one percent’ is a different claim from satisfying a strict mathematical threshold.
The stage-seventeen drawing would involve 3¹⁷ retained pieces in the ideal construction. You do not need to render every triangle to answer the threshold question; using the recurrence and logarithms is both more reliable and more transparent.
A different route to the same triangle pattern
The Sierpiński pattern also appears when Pascal’s triangle is coloured according to whether binomial coefficients are odd or even. This is a numerical construction rather than the midpoint-removal drawing. Rows generated by addition modulo two form a triangular pattern with repeating empty and occupied regions.
One can see the first few rows as single black cells, then small solid triangular groups, then gaps that repeat at doubled scales. The connection is rooted in properties of binomial coefficients modulo two, and exact correspondence requires specifying how the row pattern is embedded and scaled.
This is not evidence that every parity array has the same physical area as a geometric triangle. The combinatorial representation and geometric limiting object have different immediate definitions. Their shared self-similar organisation creates the surprising connection.
For a useful classroom exercise, draw the first sixteen Pascal rows modulo two and compare their black-cell pattern with a stage-four Sierpiński illustration. Then ask which statements are about number parity, which are about geometric side lengths, and which depend on treating cells as finite squares.
A worked comparison of three constructions
Suppose three students investigate self-similar figures. Student A keeps three copies at one-half scale. Student B keeps two copies at one-third scale. Student C replaces each boundary segment by four segments at one-third scale. The three scaling exponents are log3/log2≈1.585, log2/log3≈0.631 and log4/log3≈1.262.
The formulas look similar because each is derived from N copies at linear scale r using N rᴰ=1. But the counted structures differ: A is a triangular gasket, B is a Cantor-type dust on an interval, and C is a replacement boundary curve. These are not interchangeable physical models.
Ask which construction can have a fixed positive planar area at every finite stage and zero planar area in its limit. Both A and the finite-stage interpretation of B fit that type of statement, but their ambient spaces and scaling dimensions are different. Student C instead represents a boundary whose growing length coexists with a bounded enclosed snowflake area.
The transferable skill is not memorising three logarithms. It is writing the rule that each logarithm describes, choosing the correct dimension or measure, and checking the limiting object separately from the finite drawing.
How to teach the concept without a misleading graphic
A careful teaching sequence starts with a tangible paper triangle. Students shade and remove the centre once, then twice, then three times. Before showing a computer rendering they predict how the number of pieces, individual side length and remaining area will change.
Next they complete a stage table independently. One group may count every surviving triangle; another calculates the total area removed; a third measures side-length ratios. Compare the tables at stage three to establish that apparently different observations all obey one replacement rule.
Then ask the uncomfortable question: ‘At stage n, does any triangle still have area?’ The answer is yes for every finite n. Follow it immediately with ‘What happens to the sequence of those total areas?’ The answer is that the limit is zero. Students must articulate why these statements are compatible.
Finally bring in logarithmic dimension and resolution limits. This order moves from concrete geometry to recurrence, limiting process and scale analysis. It addresses a common error—using a visually intricate drawing as a substitute for stating which mathematical quantity was measured.