Integral geometry studies shapes by integrating measurements made with families of lines, planes and other geometric probes. Instead of tracing an entire boundary directly, we may count how often lines cross it, average its projected widths, or integrate the material encountered along each line. The gold standard is to define the probe, its sampling measure and the information recorded before interpreting the resulting average.
This guide develops the subject through line segments, rectangles, circles, random needles and simple projection data. It explains how crossings recover length, how widths recover convex perimeter, and how line integrals lead to reconstruction problems. The examples are mathematical models with stated assumptions; they are not instructions for clinical imaging or claims that a finite set of measurements determines every possible object.
The Gold Standard master connects the collection. The Education parent develops learning that transfers, while How X Works connects mechanisms. For dimensional measurement and rough sets, continue with Geometric Measure Theory.
Choose your integral geometry route
Chapters 1–3: Choose geometric probes
Chapters 4–6: Recover length and perimeter
Chapters 7–9: Build projections and reconstruction
Chapters 10–12: Sampling, uncertainty and applications
Chapter index
Choose geometric probes · Chapters 1–3
1. Start with a probe and an observable
Recover length and perimeter · Chapters 4–6
4. Derive the crossing formula for a segment
Build projections and reconstruction · Chapters 7–9
7. Turn line lengths into area and total mass
Sampling, uncertainty and applications · Chapters 10–12
10. Separate sampling error from model error
11. Use Buffon’s needle as a controlled probability model
12. Apply the ideas without overstating what the model knows
Transfer and continue · Chapters 13–14
Imagine a closed curve drawn on paper. A line placed across it can tell us several different things. We can record whether the line touches the curve at all, count its intersections, measure the portions lying inside the enclosed region, or accumulate a density along the line. These observations are not interchangeable.
For a circle, a typical line passing through its interior meets the boundary twice. Its intersection with the filled disk is a chord, whose length changes with the line’s offset. The same geometric probe therefore produces a boundary count of two and a continuously varying chord length.
A single line cannot usually determine the whole object. Integral geometry asks what happens when we combine information over a precisely specified family of lines. Some aggregates recover a scalar quantity such as perimeter. Richer collections of measurements may support reconstruction of a function or shape.
Three questions to ask before calculating
First, what is the target: boundary length, enclosed area, a density field or something else? Second, what does each probe record: a count, a length or a weighted integral? Third, how are probes distributed? An average without its distribution is incomplete.
Consider two curves with the same total length but very different shapes. A suitable averaged crossing count can agree for both. That does not mean the curves are identical. The measurement was designed to retain length, not every detail of shape.
This distinction prevents a common overclaim. Recovering one geometric invariant is not equivalent to solving an inverse problem for the entire object. Before announcing that measurements “recover the shape”, list which unknowns are actually determined and which remain unconstrained.
The elementary examples below keep these questions visible. Their purpose is to make the mathematics checkable rather than to attach an impressive name to an unexplained numerical procedure.
CHAPTER 2 OF 14 · Choose geometric probes
2. Give each planar line an unambiguous description
Write a line as x cosθ+y sinθ=p. The vector nθ=(cosθ,sinθ) is a unit normal, and p is the signed perpendicular offset from the origin. We use 0≤θ<π and all real p. This describes each unoriented line once, apart from the conventional identification at the angular endpoint.
The angle is the direction of the normal, not the direction of travel along the line. At θ=0, the equation is x=p, so the line is vertical. At θ=π/2, it is y=p, so the line is horizontal. Checking these two cases prevents a ninety-degree mistake in later calculations.
An alternative convention allows θ over a full turn but restricts the offset differently. Another uses a full turn and signed offsets, describing every unoriented line twice. All can be useful, but the normalising constant in an integral must match the convention.
Why use dp dθ?
The natural Euclidean line measure is proportional to dp dθ. Translating the coordinate origin shifts p by an angle-dependent amount without changing the area element in line-coordinate space. Rotating the plane shifts the angular coordinate. The measure therefore treats rigidly moved copies of a configuration consistently.
This is not a probability distribution on all lines in the plane: the offset extends without bound, so the total measure is infinite. To make a random-line experiment, restrict to a finite family, such as lines meeting a chosen disk, and normalise there.
The University of Utah introduction to integral geometry develops these coordinates and their motion-invariant measure. Our calculations use the half-turn, signed-offset convention throughout. Keeping that convention fixed makes the factors of two transparent.
A useful notebook entry is therefore the equation, angular range, offset range and integration weight. Write these before collecting counts. Otherwise an apparently harmless change in parametrisation can double the result.
CHAPTER 3 OF 14 · Choose geometric probes
3. Make “random line” a reproducible instruction
Suppose the object lies inside a disk of radius R centred at the origin. Choose θ uniformly from [0,π), and independently choose p uniformly from [−R,R]. The resulting line intersects the disk, apart from tangencies of probability zero. This samples the invariant line measure conditioned on meeting that disk.
The joint density in these coordinates is 1/(2πR). It is a statement about the sampling procedure, not a claim that every imaginable way of drawing a random chord produces the same distribution.
For example, choosing two independent uniform points on a circle and joining them is a different experiment from choosing a uniform signed offset and normal angle. Choosing a midpoint uniformly inside a disk is different again. Each rule induces its own distribution of chord distances and lengths.
A circle illustrates the distinction
For a line at distance |p|<R from the centre of a disk, the chord length is 2√(R²−p²). If p is uniform on [−R,R], near-centre offsets and near-edge offsets receive equal probability per unit offset. A midpoint chosen uniformly by area does not have a uniform radial distance.
Thus “random chord” is not enough information to determine an average chord length. The sampling rule must be part of the problem. This is a geometric version of the statistical warning that a sample cannot be interpreted without knowing how it was selected.
If a simulation is intended to approximate invariant line averages, generate the stated θ and p directly. Do not substitute a visually appealing chord generator and assume equivalence. First derive its induced distribution or compare it against known analytic cases.
A simulation should also preserve independence between angle and offset when that is part of the intended design. Reusing an offset as a function of angle can create a systematic pattern of probes rather than the stated product distribution. Plotting the sampled pairs in line-coordinate space is a useful preliminary check, although a plot alone cannot prove independence.
The window also matters. Enlarging the sampling disk introduces more lines that miss a fixed target. The mean crossing count falls, while the normalising factor grows. A correctly designed estimator balances those changes; an unnormalised count does not.
CHAPTER 4 OF 14 · Recover length and perimeter
4. Derive the crossing formula for a segment
Take a horizontal segment of length L centred at the origin. For a fixed normal angle θ, the offsets of lines that cross the segment form an interval of length L|cosθ|. Except for negligible endpoint and collinear cases, each such line contributes one intersection.
Let N(θ,p) be the number of intersections. Integrating N over p therefore gives L|cosθ|. Integrating over the half-turn gives L∫₀^π|cosθ|dθ=2L. Hence L=(1/2)∫₀^π∫R N(θ,p)dp dθ for this segment.
The calculation supplies both the mechanism and the constant. Projecting the segment onto the normal direction identifies which offsets hit it. Averaging that projected span over all directions produces twice its length under our coordinate convention.
From segments to curves
For polygonal curves, add contributions from their segments. Lines through vertices or lying along an edge form exceptional families that do not alter the integral. With appropriate regularity and counting conventions, the Cauchy–Crofton formula extends the relationship to rectifiable curves.
For a simple piecewise smooth curve C, its length is one-half the integral of its intersection count over our line space. If a parametrised path retraces an arc, one must decide whether to count traversal multiplicity or the image only. The two choices correspond to different length questions.
The Ohio State presentation of Crofton’s formula gives the theorem and segment-based intuition. A beginner should reproduce the horizontal-segment calculation before applying the result to an intricate curve.
Notice what has not been done: we did not estimate length from the number of points in a digital drawing. We integrated crossings against a particular geometric measure. Pixel density and line-space density are different quantities and require different normalisations.
CHAPTER 5 OF 14 · Recover length and perimeter
5. Check the formula on a circle and a polygon
For a circle of radius a, every non-tangent line with |p|<a meets the boundary twice. At each θ, the integral over p is therefore 2·2a=4a. The angular integral is 4πa, and multiplying by one-half gives 2πa, the familiar circumference.
This is an especially useful calibration case because the count is simple. Counting only whether a line hits the circle would give one instead of two and halve the recovered perimeter. Tangent lines do not repair that error: their set has zero measure in the integral.
A rectangle gives a directional check
For an axis-aligned rectangle with side lengths a and b, its projection width in normal direction θ is w(θ)=a|cosθ|+b|sinθ|. A typical line whose offset lies within that projection crosses the convex boundary twice. Thus ∫N(θ,p)dp=2w(θ).
Integrating over θ gives perimeter ∫₀^πw(θ)dθ=2a+2b. For sides 3 and 2, the answer is 10. At θ=0 the width is 3; at θ=π/2 it is 2. These endpoint checks also confirm that θ denotes the normal direction.
A rotated rectangle has the same averaged result, although its directional width function shifts. If an implementation reports a different perimeter after rotation, inspect angular sampling, grid effects and coordinate conventions before blaming the geometric theorem.
For a nonconvex boundary, a line can cross four or more times. Retaining the full count is essential. A yes-or-no hit indicator loses the additional crossings that record indentations and separated components.
These examples form a compact test suite: segment for the constant, circle for multiplicity, rectangle for angle conventions, and a nonconvex shape for repeated crossings. A method that fails one of these tests should not be trusted merely because it produces plausible values on an unfamiliar drawing.
CHAPTER 6 OF 14 · Recover length and perimeter
6. Recover convex perimeter from mean width
For a compact convex planar body with nonempty interior, Cauchy’s perimeter formula states that perimeter P equals ∫₀^πw(θ)dθ. If mean width is defined as the uniform angular average, then P=π times mean width.
The convexity assumption explains why this follows so neatly from crossing counts. For almost every line hitting the interior, there is one chord with two boundary endpoints. Projection width tells us the entire offset interval of such lines. No hidden extra components need to be counted.
For the 3-by-2 rectangle, mean width is 10/π, approximately 3.1831. It is neither simply the arithmetic mean of the two side lengths nor the diagonal. It averages projections over every direction, including oblique ones.
Why the convexity condition cannot disappear
A body with an inward notch can have the same extreme projection spans as its convex hull, while its boundary is longer. Widths record supporting extremes. They do not record every excursion into a concavity. Integrating widths therefore recovers the convex hull’s perimeter, not generally the nonconvex boundary’s full length.
Likewise, a region with a hole has an inner boundary that width alone cannot detect. Crofton intersection counts can register extra crossings of that boundary, while outer projection spans cannot. The observable has changed, so the theorem’s conclusion must change too.
This gives a practical distinction between two measurement tasks. If only directional spans are available, the data describe convex-envelope information. If all boundary intersections are available, more detailed length information can survive. Neither should be advertised as a universal reconstruction of the interior.
A good diagnostic is to draw two shapes with the same convex hull but different internal holes or indentations. Ask which measurements remain the same. This makes information loss visible before a complicated inverse problem hides it behind notation.
CHAPTER 7 OF 14 · Build projections and reconstruction
7. Turn line lengths into area and total mass
Fix an angle θ and introduce coordinates x=p nθ+t vθ, where vθ is a perpendicular unit vector along the line. This is a rotation of Euclidean coordinates, so the area element is dp dt. Integrating chord length over all offsets therefore recovers the area of a measurable planar region of finite area.
For a disk of radius a, the chord-length function is 2√(a²−p²) for |p|≤a and zero elsewhere. Its integral over p is πa². One way to see this is to recognise the graph as twice the upper semicircle; the area under it is the area of the full disk.
This identity holds at each fixed angle. Unlike perimeter from projected widths, recovering area by integrating actual section lengths does not require convexity. A line may meet several disjoint intervals; their lengths must all be included.
Weighted material along a line
Replace the region’s indicator by a nonnegative integrable density f(x,y). Define Rf(θ,p)=∫R f(p nθ+t vθ)dt. This is the planar Radon transform: a function of angle and offset obtained by integrating along each line.
Integrating Rf(θ,p) over p recovers ∫f over the plane for every θ, under the stated integrability conditions. This is a consistency check. If different angles produce substantially different total integrals in a supposedly complete dataset, investigate truncation, calibration or an incorrect coordinate factor.
For a disk filled with constant density c, the line integral is c times its chord length. Its central projection is 2ca, and the integral of the projection is cπa². Density units matter: if f has units of mass per area, a line integral has units of mass per length.
Do not confuse this operation with measuring the outline’s intersections. A boundary count and an integral through a filled region may use the same line coordinates but represent different observables. Their formulas, units and normalising factors need not coincide.
CHAPTER 8 OF 14 · Build projections and reconstruction
8. Work a complete projection example
Let f equal 2 inside a disk of radius 3 and zero outside. The centre is initially the origin. For every angle, Rf(θ,p)=4√(9−p²) when |p|≤3, and zero when |p|>3.
At p=0, the chord length is 6, so the projection value is 12. At p=√5, the chord length is 4, so the projection value is 8. Integrating over all offsets gives total mass 2·9π=18π. These three checks test the centre, an off-centre chord and the global integral separately.
Now move the disk centre to c=(1,−2). Its normal coordinate becomes c·nθ=cosθ−2sinθ. The projection profile shifts to that offset; its shape and total integral remain unchanged. At θ=0 the peak occurs at p=1. At θ=π/2 it occurs at p=−2.
Why a point produces a sinusoidal trace
An ideal point at (x₀,y₀) lies on the measured line when p=x₀cosθ+y₀sinθ. As θ varies, its location in angle-offset coordinates follows a sinusoidal curve. A plot of projection data over angle and offset is therefore called a sinogram.
For an extended object, contributions overlap rather than appearing as isolated traces. The shifted disk produces a moving band with a rounded cross-section, not a single sharp sinusoid. Distinguishing a point model from an extended density prevents misleading interpretations of the plot.
A useful numerical exercise generates the centred disk analytically, shifts it, and verifies that each projection’s integral stays constant. Then intentionally reverse the sign of p. The peak locations move incorrectly, providing an immediate diagnostic for a coordinate mismatch.
This example does not require reconstructing an image to test the forward model. Known shapes are valuable because expected measurements can be calculated independently. Forward-model tests should come before trusting a complicated reconstruction algorithm on unknown data.
CHAPTER 9 OF 14 · Build projections and reconstruction
9. Explain what projection data can reveal about an image
The Fourier slice theorem connects line integrals to frequency information. For a sufficiently well-behaved planar function, taking the one-dimensional Fourier transform of Rf(θ,p) in the offset variable gives the two-dimensional Fourier transform of f restricted to the radial line in direction nθ, using consistent Fourier conventions.
A short derivation explains why. Substitute the definition of Rf into the Fourier integral. The exponential depends on p=x·nθ. Changing from (p,t) to x turns the double integral into the two-dimensional Fourier transform evaluated at a frequency vector parallel to nθ.
Each projection angle therefore supplies a different radial slice of frequency space. Many angles provide complementary information. A single angle collapses spatial structure along the lines and cannot generally identify how material is distributed in that direction.
Backprojection is not automatically inversion
Backprojection spreads each line measurement back along its corresponding line. It combines evidence but also spreads a local feature across many candidate locations. Simply adding those contributions produces blur. Filtered backprojection includes a frequency weighting that corrects the ideal transform relationship, with normalisations depending on convention.
In practice, high-frequency weighting also makes noise important. Filtering, finite sampling and regularisation affect the resulting estimate. A visually sharp image is not by itself proof of accurate reconstruction.
The inverse-problems lecture on tomography develops this connection. The existing CT Scans, Projections and Image Reconstruction guide applies the mathematical chain to an imaging context, including physical assumptions and reconstruction limitations.
Here the central lesson is informational: the forward map determines what data mean, and the inverse method must respect that map. Reconstruction does not manufacture independent evidence. It combines measurements with mathematical structure and, often, additional assumptions about the unknown object.
CHAPTER 10 OF 14 · Sampling, uncertainty and applications
10. Separate sampling error from model error
Suppose a simple curve lies entirely inside a disk of radius R. Sample independent lines with uniform θ in [0,π) and uniform p in [−R,R]. Crofton’s formula gives length L=πR E[N], where N is the number of intersections of one sampled line with the curve.
For M independent sampled lines, the estimate is L̂=πR times the sample mean of their intersection counts. If the sample standard deviation of counts is sN, an estimated standard error is πR sN/√M, under the usual finite-variance assumptions. This describes sampling variability, not every possible error.
A numerical uncertainty calculation
Take R=5, M=400, mean count 0.8 and sample standard deviation 0.9. The estimated length is 4π, approximately 12.566. The estimated standard error is 5π·0.9/20=0.225π, approximately 0.707.
Multiplying M by four would approximately halve this standard error if the distribution and independence assumptions remain appropriate. It would not repair a systematic failure to detect crossings or an incorrect line distribution.
If the image truncates part of the curve, more lines measure the truncated curve more precisely. If only horizontal probes are used, more samples do not create angular coverage. If tangent or near-tangent events are mishandled systematically, the resulting bias can persist.
This is why uncertainty should be decomposed. Sampling variability, discretisation, calibration and model mismatch are different sources. Reporting only a Monte Carlo standard error can make a biased result look more trustworthy than it deserves.
For reproducibility, record the target window, random seed, number of probes, crossing rule and handling of boundary events. Then repeat the calculation on a segment, circle and rotated rectangle. The purpose is not to eliminate every numerical imperfection but to identify which assumptions the claimed precision actually depends on.
CHAPTER 11 OF 14 · Sampling, uncertainty and applications
11. Use Buffon’s needle as a controlled probability model
Place parallel lines a distance d apart. Drop a needle of length ℓ≤d with a centre position uniform modulo the line spacing and an orientation independent and uniform over a half-turn. Let α be its acute angle to the parallel lines and let X be the distance from its centre to the nearest line.
Then X is uniform on [0,d/2], and α is uniform on [0,π/2]. The needle crosses a line when X≤(ℓ/2)sinα. Conditional on α, the crossing probability is (ℓ/d)sinα. Averaging gives P=2ℓ/(πd).
The derivation makes the assumptions visible. A preferred orientation changes the average of sinα. A nonuniform centre distribution changes the distribution of X. If ℓ>d, the simple short-needle derivation no longer applies unchanged because the conditional probability must respect the upper bound of one.
A worked prediction
For ℓ=2 and d=5, the probability is 4/(5π), approximately 0.25465. In 1,000 independent ideal drops, the expected number of crossings is about 254.65. An expected count need not be an integer; an observed count must be.
The standard deviation of the count under this model is √(1000P(1−P)), approximately 13.78. A result differing from the expectation by a few crossings is ordinary random variation. A repeated large discrepancy calls for checking the physical or simulated sampling procedure.
One may rearrange the formula to estimate π from a crossing fraction. That is an educational illustration, not an efficient numerical method for computing π to high precision. The reciprocal transformation also introduces its own finite-sample behaviour; the estimate is not magically exact because the theoretical formula is exact.
Buffon’s needle provides a bridge between geometry and probability. The important skill is deriving the event from the geometry, then averaging under a specified distribution. Memorising the final fraction without those steps makes the result difficult to transfer safely.
CHAPTER 12 OF 14 · Sampling, uncertainty and applications
12. Apply the ideas without overstating what the model knows
Integral-geometric methods help connect partial probes to global quantities. In stereological reasoning, lower-dimensional sections or test lines can inform geometric quantities of a larger structure, provided the sampling design and relevant assumptions are satisfied. Directional bias is especially important when the object has a preferred orientation.
A collection of aligned fibres illustrates the issue. Lines parallel to the fibres may produce few transverse crossings, while perpendicular lines produce many. Treating a fixed-direction count as though it came from isotropic probes can misrepresent total length. Randomised directions or an explicitly orientation-aware estimator address different versions of that problem.
Reconstruction has a physical forward model too
In an idealised transmission model, logarithms convert an exponential attenuation ratio into an accumulated line integral. Real instruments involve additional effects, so the mathematical Radon model is an approximation whose suitability must be checked. The geometry alone does not establish calibration or application-specific accuracy.
The existing Seeing Inside Without Cutting Open article connects forward modelling, hidden structure and validation. Read it alongside the more specific CT guide when moving from geometric examples to imaging systems.
For industrial inspection or scientific imaging, a sensible validation sequence starts with objects of known shape or density. Compare predicted measurements with observations before judging the reconstructed image. Otherwise a flexible inverse method may hide errors in the forward model by producing a plausible-looking explanation.
Do not treat an integral-geometric identity as a guarantee that a particular sensor estimates its inputs correctly. The theorem concerns specified mathematical quantities. A sensor supplies noisy and discretised observations that may only approximate them.
The most useful practical output is often a clearly scoped estimate: what was measured, how probes were selected, which model connects data to geometry, and what checks were passed. That is more informative than a single unexplained “accuracy” percentage or an attractive image without a measurement contract.
CHAPTER 13 OF 14 · Transfer and continue
13. Test understanding by changing the measurement rule
First, sample lines meeting a disk of radius 10 instead of radius 5 while keeping the same enclosed curve. The family of available offsets doubles, so the expected count halves. In L=πR E[N], the two changes cancel. Keeping the old normalising radius would halve the estimated length incorrectly.
Second, record only whether a line hits a circle rather than how many times it crosses the boundary. Almost every hitting line contributes one instead of two. The usual Crofton crossing estimator now gives half the circumference. Restore multiplicity rather than trying to compensate by drawing more lines.
Third, apply the mean-width perimeter formula to a region containing a hole. Outer projection widths cannot reveal the inner boundary, so the result does not give total boundary length. Use the observable and theorem suited to the requested quantity.
Change coordinates, then check units
Fourth, double every spatial coordinate of an object and the sampling window. Boundary length doubles, while area quadruples. A crossing count is dimensionless; its conversion factor contains one power of length. An integrated chord-length calculation contains two powers. Units expose which scaling rule is appropriate.
Fifth, shift the density-two, radius-three disk to (1,−2). The total projection integral remains 18π, but the peak offset becomes cosθ−2sinθ. A method that preserves mass while leaving every peak at zero has lost location information.
Sixth, collect only horizontal and vertical sums of a two-by-two array. The arrays with rows (1,0),(0,1) and (0,1),(1,0) have identical row and column sums. Those measurements do not uniquely recover the array. More precision on the same four sums does not distinguish the two possibilities.
For each exercise, state the assumption that changed before calculating. The correct repair may concern the sampling law, multiplicity, convexity, coordinates or identifiability. Repeating a formula without diagnosing that change is exactly how a useful theorem becomes a misleading measurement procedure.
CHAPTER 14 OF 14 · Transfer and continue
14. Follow a learning route and resolve common questions
Begin with coordinate geometry, trigonometry and integration. Derive the projected span of a segment and the chord length of a circle. These calculations reveal how a line’s angle and offset control its interaction with a shape.
Next, learn probability and change of variables. Define a random-line experiment, compute an expected crossing count, and test the effect of changing the window. Work the rectangle and needle examples independently, checking constants against a familiar perimeter or probability bound.
Then study measure theory, Fourier analysis and inverse problems. Distinguish a theorem about continuous complete data from a numerical reconstruction using finite noisy measurements. The former guides the latter, but it does not erase discretisation or uncertainty.
Is integral geometry just integration of area?
No. The defining idea is integration over families of geometric probes or motions. The result may concern length, perimeter, area, intersection counts or transform data. Ordinary integration remains part of the machinery, but the domain of integration may be a space of lines rather than the original plane.
Is one random-line distribution the natural choice for every task?
No. Motion-invariant line measure is a central Euclidean choice, and conditioning it on a window gives a clear probability model. Other applications may deliberately use different sampling laws. Their estimators and weights must match those laws; calling them random does not make them equivalent.
Does the same mean width imply the same shape?
No. Mean width determines perimeter for a convex planar body under the stated formula, not its entire geometry. A scalar average discards substantial information. Even richer measurements can leave ambiguities if angles, offsets or other parts of the data are missing.
Can many measurements overcome every reconstruction problem?
Only if they supply the needed information under a suitable model. Repeating the same insufficient measurements can reduce noise without resolving nonuniqueness. The two-by-two array example separates precision from identifiability in a form that can be checked without advanced mathematics.
What should I be able to explain after studying this subject?
Name the probe, its coordinates, the observable, the measure and the target quantity. Derive the normalisation on a simple shape. Identify the assumptions that permit the conclusion, then test a changed case. These habits connect geometric intuition to reliable quantitative work.
Continue with the Utah integral-geometry notes, the Crofton presentation, and the tomography lectures. Return to the Geometric Measure Theory root when the unresolved question is what length or area means for the object being measured.