The shape of everything.

Why does the world hold together at many sizes at once?

A cell is made of molecules, a person of cells, a society of people. At each size there are things stable enough to name and predict. Where do those levels come from, and is there a single piece of mathematics that says which ones are allowed?

Most of science studies one level at a time. Quantum Biodynamics (QBD) asks a cross-level question. Is there a shared structure that decides which patterns can be stable at each level? We think a branch of geometry built around shapes-inside-shapes might describe it. We also think that idea is far more likely to be wrong than right until it passes tests. So this page has two jobs:

1. Build the mathematics you need to state the idea precisely. It starts from a single point, in the order you would meet it at school.
2. Show, in plain language, exactly how we test the idea, and what happened when we did.

An idea only counts as science if some result could show it is wrong. We write that result down before we look.

It is easy to find patterns after the fact. The discipline is to say, in advance and in writing, what you expect, how you will measure it, and what outcome would mean you were wrong. Then you run the test and report the answer.

A single point has no size, only a location.

Start with the simplest thing in mathematics. A point. It has no length, no width, no thickness. It is not made of anything. It is a location — a where — with nothing there. This sounds like a trick, and it is not. Every deep structure in geometry begins here.

Because a point has no extent, we cannot draw one. Any dot you sketch on paper is already too big. What we draw is a marker for a location. In mathematics we write a point as a name — P — or as a set of numbers that pick it out from every other possible location.

A single point on its own is zero-dimensional. There is nowhere to move. No left, no right, no up. Just here.

All of geometry — every curve, every surface, every high-dimensional shape we will build — is made of points. Everything is a set of locations. The whole game is: which sets of locations mean something?

A line is an infinite family of points laid end to end.

Now let us make a line. Take one point and imagine placing another point beside it. Then another. And another. Keep going, without ever leaving a gap, forever in both directions. What you get is a number line — the first place where infinity enters the picture, quietly.

Each location on the line is labelled by a single real number: 0, 1, 2, –1, ½, √2, π, and so on. Every possible real number is a point on the line. This is what mathematicians mean when they say the line is one-dimensional: one number is enough to say exactly where you are.

An interval, like all the points from 0 to 1, contains uncountably many real numbers. This is the first thing that will feel strange and important: even a tiny piece of the line already contains infinitely many points, and not in a countable way — you cannot list them all. Cantor proved this in the 1870s. The line is genuinely thicker with points than the whole numbers are.

What we call the length of an interval — for instance, the interval from 0 to 1 has length 1 — is a way of measuring an infinite family of points with a single number. That will matter enormously when we get to calculus.

Length is a property that appears only when infinitely many points come together.

Here is the first taste of emergence. A single point has no length. Two points have no length between them either — just a gap. Even a thousand points, laid out neatly, do not have length in themselves. Yet a whole line, made of infinitely many points, plainly has length. Where does the length come from?

The answer is that length is not a property of individual points. It is a property that only makes sense once you have a continuous run of them. Length is the first thing in this walk that emerges: a real, measurable feature of a collection that no single member possesses.

You can think of a line as being built from string-bits — infinitely many tiny pieces, each individually of zero length, that together add up to a definite total. This picture will show up again when we do integration in Part III. Integration is exactly the mathematics of summing infinitely many infinitesimal pieces.

Two axes, two numbers. Every point in the plane has a unique address.

Now take a second number line and lay it across the first at right angles. The horizontal one we call the x-axis, the vertical one the y-axis. Their crossing is the origin, at (0, 0). Every point in the flat plane is now pinned down by two numbers — its x-coordinate and its y-coordinate. Written together as (x, y).

This is the Cartesian plane, and it is the single most important idea in school mathematics. It joins geometry (shapes) to algebra (equations) in one framework. Once shapes can be described by numbers, all of them can be manipulated by the rules of algebra. Descartes gave us this in the 1600s and everything downstream — from Newton's calculus to modern physics — depends on it.

Just as the line held infinitely many points, so the plane holds infinitely many points — in a much richer way. Between any two points on the plane there are infinitely many others, along infinitely many possible paths. The plane is two-dimensional because you always need exactly two numbers, no more and no fewer, to pin down a location. One number is not enough (you would be stuck on a line). Three would be redundant.

Add one more axis, add one more number. This is the space we live in.

Add a third axis at right angles to the first two — call it the z-axis, pointing out of the page. Every point in the volume of space around you now takes three numbers: (x, y, z). This is ordinary three-dimensional space, and this is what you actually inhabit.

The pattern is now clear. Each new dimension adds a new number and a new independent direction of movement. There is nothing special about three. Three just happens to be the number of independent directions in the physical space we experience with our eyes and bodies. The mathematics does not stop here.

A point can be a list of any length. The mathematics does not care whether we can picture it.

Once you have accepted that a point in 2D is a pair of numbers and a point in 3D is a triple, there is no reason to stop. A point in N-dimensional space is simply a list of N numbers, written (x₁, x₂, …, x_N). We cannot picture four dimensions, or ten, or ten thousand. It does not matter. The mathematics runs identically.

This sounds abstract until you realise how often you already use it.

• A pixel in a colour photograph has three numbers — red, green, blue. A whole image is a point in a space of very many dimensions (three per pixel, millions of pixels).
• A person's medical profile — age, weight, blood pressure, glucose, sodium, potassium, cholesterol, and so on — is a point in a space of maybe a hundred dimensions. Two patients close together in that space are clinically similar.
• The state of a single human cell, described by which genes are on and to what level, is a point in a space of roughly twenty thousand dimensions. One number per gene.
• The state of a market portfolio, one number per instrument held, might live in a space of thousands of dimensions.

This is the beginning of what mathematicians call a state space. Every possible configuration of a system is a single point in this space; the whole life of the system is a curve traced through it. When we later talk about which configurations are stable, we are asking: which points, or families of points, in this enormous space are the ones the system can actually stay at?

Notice how much has already been said and how little was needed. Points, coordinates, and the idea that the number of coordinates equals the number of dimensions. Nothing more.

A point can also be a whole function.

There is one more leap to make in this part. If a point can be a list of N numbers, it can also be a list of infinitely many numbers. And if we let the list be labelled continuously rather than discretely, then a point becomes a function — an assignment of a value to every real input.

The temperature everywhere in a room, right now, is one such object: at every location (x, y, z) there is a temperature. That single, whole temperature-field is a point in an infinite-dimensional space of possible temperature fields.

The same is true of a sound wave (a function of time), an electric field (a function of space and time), the shape of a heartbeat trace, or the probability distribution of where an electron is. In all these cases, the object of interest lives in a space with infinitely many independent directions.

A function is a rule that turns one number into another.

A function f takes an input, does something to it, and returns an output. We write it f(x) — read as "f of x" — and we can name the output y. So y = f(x).

Feed this rule the input 3 and you get 9. Feed it −2 and you get 4. Feed it every real number in turn, and you get every non-negative real number back. If we plot the pair (x, f(x)) for every input x, we get a shape in the plane — the graph of the function.

A function is more than an equation. It is a rule. It is the same rule whether we write it as f(x) = x², or as a table of pairs, or as a computer program. The graph is one visual face of it. Every well-behaved shape you have ever met — parabola, sine wave, exponential — is the graph of some function.

Functions can go in any dimension. A function of two inputs, f(x, y), gives a single output — a surface hovering above the plane. A function of three inputs gives one number for every point in 3D space (temperature, pressure, density). The idea is the same at every level.

A transformation moves every point of space to a new location, all at once.

A function takes a number to a number. A transformation takes a whole point to a whole point, and does it simultaneously for every point in the space at once. Rotations, reflections, stretchings, shearings — all of these are transformations of the plane, and each has a precise mathematical description.

Consider rotation by 90° about the origin in the plane. The point (1, 0) moves to (0, 1). The point (0, 1) moves to (−1, 0). Every other point moves in the same rotational way. The rule that describes this move is a single object — the transformation — that maps (x, y) to (−y, x).

The important idea is that the whole space is moved by one rule. We can compose transformations (do one, then another). We can invert them (undo the move). We can ask which points, if any, are left alone — those are the fixed points. For a rotation about the origin, only the origin itself is a fixed point.

Every symmetry in nature — the rotational symmetry of a molecule, the mirror symmetry of a face, the periodic symmetry of a crystal — is a transformation that leaves the shape looking the same. The mathematics of transformations is the mathematics of symmetry.

A matrix packages a linear transformation into a rectangular array of numbers.

Not every transformation, but a very important class of them — the linear ones — can be written as a rectangular grid of numbers called a matrix. Linear means: the transformation takes straight lines to straight lines and leaves the origin fixed. Rotations, reflections, stretches and shears are all linear.

Here is a 2×2 matrix acting on a point in the plane:

Every 2×2 matrix corresponds to one linear transformation of the plane. Different matrices do different things. For example:

• The rotation by 90° we just met is the matrix with rows (0, −1) and (1, 0).
• Reflection in the x-axis is (1, 0) and (0, −1).
• Stretch by 2 in the x-direction is (2, 0) and (0, 1).
• Shear that leans everything to the right is (1, 1) and (0, 1).

This bookkeeping matters for two reasons. First, matrix multiplication lets us compose transformations by simply multiplying the matrices. Second, and more importantly, matrices generalise to any number of dimensions. The description of how a 20,000-dimensional cell state changes over one moment of time is a matrix — a huge one, but the same object at heart. Every deep transformation we will eventually meet — including the ones acting on the shapes-inside-shapes we build in Part VI — is either a matrix or something built from matrices.

The derivative is the slope of a curve at a single point — a number that captures how fast the function is changing right there.

The graph of a function is a curve. At most points on that curve, the curve is going up or down at some rate. The derivative of the function is a new function that tells you, at every input x, exactly how steep the graph is.

To get at it precisely, pick two points on the curve — one at x, one at x + h for some small h. Draw the straight line joining them; its slope is

Now shrink h towards zero. The chord becomes shorter and shorter, and the two points slide together. In the limit — if the function is smooth enough — the chord becomes the tangent line at x, and its slope is the derivative:

The derivative is a machine that turns one function into another. There are standard rules — the product rule, the chain rule, the quotient rule — that let you differentiate any polynomial, any trigonometric function, any exponential, in a few lines. What matters here is what a derivative measures: instantaneous rate of change. Velocity is the derivative of position with respect to time. Acceleration is the derivative of velocity. The rate at which a tumour is growing is the derivative of tumour size with respect to time. The steepness of a hill is the derivative of altitude with respect to horizontal distance.

Integration accumulates infinitely many infinitely small contributions into a single total.

Look again at the string-bits picture from Chapter 5. A line has length because infinitely many zero-length points add up in a controlled way. Integration is the general machinery for doing that adding.

Suppose we want the area under the curve y = f(x) between x = a and x = b. One way to approximate it: chop the interval into narrow vertical strips of width Δx, treat each strip as a rectangle of height f(x), sum the areas, and get a rough answer. Then make the strips narrower. Then narrower again. In the limit, the sum becomes exact, and we call it the integral:

The elongated S in the notation ∫ is a stretched-out capital S for "sum." The dx is a reminder that we are summing over infinitesimal widths in the x-direction. Every integral is, in spirit, the sum of infinitely many infinitely small contributions.

Integration is not only about areas. In higher dimensions it computes volumes, and in still more general form it computes the amount of a physical quantity — mass, charge, energy — spread through a region. When we later meet Hodge theory, the pairing between a shape and a form is done by integration: you integrate the form over the shape and get a number. That number carries deep information.

Differentiation and integration are inverse operations.

This is the single most important theorem in first-year calculus. It says that if you integrate a function and then differentiate the result, you recover the original function. And if you differentiate a function and then integrate over an interval, you recover the change in the original function over that interval.

The theorem knits the two operations into a single working system. It is why, in practice, we can evaluate integrals by finding antiderivatives instead of computing infinite sums by hand every time.

The deep reason it matters here is more subtle. Integration turns "small local information" (the values of a function at every point) into "big global information" (a total over a region). Differentiation goes the other way. The bridge between local and global is the same bridge we will need at every step below — from local rules to global stable structures, from an individual cell's dynamics to the whole tissue that emerges, from a single equation to the shape it cuts out. Calculus is the local-to-global bridge in its simplest incarnation.

Two ideas from probability we will need in the next chapter: entropy, and mutual information.

The next chapter uses two quantities from information theory. They are less abstract than they sound, and one page of set-up is enough to earn them.

A probability is a share of possible outcomes. A fair coin has two outcomes; each gets share ½. A biased coin that lands heads 90 % of the time has shares 0.9 and 0.1. Probabilities are numbers between 0 and 1, and if you list every possible outcome they must add up to 1.

Given a probability distribution, its Shannon entropy is one single number that measures how much uncertainty is packed into it. The formula is:

The minus sign is there because probabilities are between 0 and 1 and their logs are negative, so the formula flips them back to positive. The fair coin gives H = 1 (one bit of uncertainty). The heavily biased 90/10 coin gives H ≈ 0.47 bits — less uncertainty, because you can already guess pretty well. A coin that always lands heads gives H = 0. Entropy measures ignorance.

Now consider two things at once — say the day's weather (X) and whether you carry an umbrella (Y). The mutual information I(X; Y) measures how much learning one reduces your uncertainty about the other. If umbrella and weather are unrelated, I = 0. If knowing one perfectly tells you the other, I equals the entropy of either. In general:

Entropy is uncertainty within a single quantity. Mutual information is shared structure across two quantities. That is the whole vocabulary the next chapter needs.

Entropy is not a property of the world alone. It depends on which part of the world you can see.

Take two coins that are perfectly linked: whenever one shows heads, so does the other. Someone holding both coins knows exactly what they will see, because the pair is one of just two possibilities. Someone holding only one coin sees a fair coin: fifty-fifty, one full bit of uncertainty. Same world, different access, different entropy.

Quantum physics makes this sharper. A pair of particles can be in a single, completely definite joint state, with zero entropy for the pair. Yet each particle on its own looks completely random. The uncertainty is not in the pair. It is in the cut: the choice of which part is 'the system' and which part is 'everything else'. Physicists call the choice of cut a factorisation. Choosing one is the most basic thing an observer does.

That energy-like operator matches ordinary physical energy only in special situations: systems in thermal equilibrium, and some idealised observers accelerating through empty space. Everywhere else, each observer's 'effective energy' is different. This is precisely what 'observer-dependent entropy' means. It means dependent on access, not on a mind.

A scale is predictable when the pattern between its units is strong compared with the noise inside them.

Here is the puzzle. I can predict where Jupiter will be tomorrow night to within a few kilometres — using nothing more than its mass and its current position and velocity, treated as if Jupiter were a single point. I do not track its 10⁵⁰ atoms. I do not know where any of them are, and I do not need to. Jupiter is one thing, a unit of 1, and Newton's equations are enough.

Why does this work? At the scale of the solar system, Jupiter has crystallised out of the underlying stochasticity into a single stable object. The atoms below it jitter enormously — but the jitter averages out into properties (mass, centre of mass, angular momentum) that behave exactly as if Jupiter were a point. The Planck-scale, atomic-scale, and molecular-scale details are still there. From the solar system, they are simply invisible.

This is coarse-graining. At each scale of nature, only certain patterns are stable, and where they are stable we can treat each pattern as a single unit — a node. Between the nodes there is discrete structure we can measure. Below each node there is stochastic microstate detail we cannot.

So far this is standard physics and information theory. It says when a level of description works, but not why nature offers the particular levels it does. QBD's hypothesis, stated in Part VIII, is that the geometry of cycles answers the 'why'. That is kept separate on purpose, so that the established part of this chapter cannot lend its authority to the untested part.

There is a physical version of this criterion that has been in the folklore of chemistry and biology for a century. It is called Kramers' escape formula. A particle sitting in a well of depth ΔE, buffeted by thermal noise of scale kT, takes an expected time proportional to exp(ΔE/kT) to escape. When the barrier is much larger than the noise, the well is a stable node on human timescales. When the barrier is comparable to the noise, the well is not a node — the particle drifts freely and no coarse-graining at that scale is stable. Kramers' law is the microscopic reason nodes exist at all: nodes are wells too deep for the noise to escape on the timescale that matters.

That is why we can predict Jupiter without knowing about its atoms. The same ratio, measured on real networks, is one of the quantities our tests use.

If you could not look at a space, you could still count how many neighbours are within one step, two steps, three steps. How fast that count grows tells you its dimension.

On a line, the number of points within distance r grows like r. On a flat sheet it grows like r², and in a room like r³. The exponent is the dimension. The same trick works on a network of dots and lines, where you step from dot to dot instead of measuring distance with a ruler.

Physicists who think space itself is built from discrete events use a version of this idea. In a quick check of our own, we scattered 1,500 random events in a known 2-, 3-, 4- and 5-dimensional spacetime. Using only which event could influence which, the standard estimator returned 2.01, 3.00, 4.01 and 4.96.

That is a replication of a known result, not a discovery. Our review panel also pointed out a trap. A carefully built fake, with only two layers, can make that single estimator read 'four' when the true structure is nothing like four-dimensional space. One number is never enough.

The shape and the equation are the same thing, seen two ways.

Take the equation x² + y² = 1. This is an algebraic relation between two numbers. Now ask: which points (x, y) in the plane satisfy it? The answer is exactly the set of points whose distance from the origin is 1 — the unit circle.

This is the doorway to algebraic geometry — the study of shapes defined by polynomial equations. A polynomial is a sum of terms in which each term is a constant times variables raised to whole-number powers. So x² + y² − 1 is a polynomial, and setting it equal to zero picks out the circle.

The move is: choose any polynomial in several variables; the set of points where it equals zero is a variety. Different polynomials cut out different varieties. Straight lines are varieties (they are the zero set of a linear polynomial). Parabolas, ellipses, hyperbolas, conic sections — all varieties. And so are much more elaborate shapes in higher dimensions.

In N-dimensional space, one polynomial equation cuts out an (N−1)-dimensional shape.

In the plane, one equation (like the circle's) cuts out a 1D curve. In 3D space, one equation cuts out a 2D surface. In 4D, a 3D "hypersurface." The rule is clean: each equation costs one dimension.

For example, the equation x² + y² + z² = 1 in three-dimensional space cuts out the unit sphere — a 2D surface. If we also impose z = 0, the two equations together carve out the equator, a 1D circle.

This is what mathematicians mean by an algebraic variety — the solution set of one or more polynomial equations. Once we allow the ambient space to be very high-dimensional and the equations to be genuinely complicated, the shapes we can build in this way are enormous, subtle, and beautiful. The Hodge conjecture — which we will meet in Part VII — is a conjecture about the internal structure of exactly this kind of shape.

A manifold looks flat if you zoom in far enough, however curved it is globally.

The surface of the Earth is curved on the whole, but if you stand outside and look around, the ground is flat. A small enough patch of the sphere looks like a small enough patch of the plane. A shape with this property — flat when you look closely, potentially very curved globally — is called a manifold.

A one-dimensional manifold is a curve that, near any of its points, looks like a piece of the number line. A two-dimensional manifold is a surface that near any point looks like a piece of the plane. And so on for higher dimensions.

The advantage of manifolds is that calculus works on them locally, exactly as it does in flat space, because they look flat close up. The subtlety is that the global structure of a manifold can be very rich — it can have holes, twists, and higher-dimensional cavities that no local patch reveals.

All the shapes we will care about in the rest of this walk are manifolds — often complex manifolds, which have an even stronger structure (they carry a compatible notion of "complex direction," where directions can be multiplied together in a two-dimensional way). Complex manifolds are the setting for Hodge theory.

Two adjectives that will do a lot of work: bounded, and expressible in a large uniform ambient space.

Two words are going to come up again and again from now on. Let us define them plainly.

Compact. A shape is compact if it is closed and bounded — it fits inside some big enough box, and it includes its own edges. A sphere is compact. An open interval like (0, 1), missing its endpoints, is not compact. The whole number line is not compact — it runs off to infinity in both directions.

Compactness matters because it lets us use limits properly. On a compact shape, every infinite sequence of points has a subsequence that converges to a point that is still on the shape. Nothing escapes. This is what makes it possible to build well-behaved global objects on the shape (like Hodge decompositions) rather than have them slip through cracks.

Projective. A shape is projective if it can be embedded — placed as a subset — in a special ambient space called projective space. Projective space is what you get from ordinary Euclidean space by adding, at every direction, a single "point at infinity" where all parallel lines going in that direction meet. Projective space is compact, and shapes that live inside it inherit some of its good behaviour.

These two adjectives — compact and projective — together pick out the setting where the deepest theorems of algebraic geometry, including everything to do with Hodge theory, take place. Not every shape is compact and projective. But those that are, behave extraordinarily well, and it is precisely for such shapes that the Hodge conjecture is even stated.

A shape usually contains many smaller shapes inside it, of every lower dimension.

Take a 2D surface — a sphere, a torus, or the surface of some more elaborate variety. Draw a smooth closed curve on it. That curve is a 1D shape sitting inside a 2D shape. Draw another. And another. Very quickly you have a family of curves living on the surface, each one a genuine geometric object in its own right, each one contained inside the ambient surface.

Now do the same in 3D. A solid volume contains countless 2D surfaces (spheres of various sizes, sheets, folded planes). Each surface, in turn, contains 1D curves. And each curve contains 0D points. Every shape has an internal ecology of lower-dimensional shapes contained within it.

Some of these sub-shapes are more disciplined than others. A perfectly smooth polynomially-defined curve on a surface is one kind. A wildly wiggly curve you draw freehand is another. In algebraic geometry we care only about the disciplined ones — the sub-shapes that themselves come from polynomial equations. Those are the algebraic sub-shapes. Everything that follows is about them.

An algebraic cycle is a sub-shape of a variety that is itself defined by polynomial equations.

Inside a variety X, an algebraic cycle is a sub-variety — a lower-dimensional shape that, like X itself, is cut out by polynomials. Curves on a surface, surfaces in a 3-fold, hypersurfaces in a 4-fold. Cycles come in every dimension from 0 up to just below the dimension of X.

Why cycles and not any old sub-shape? Because polynomial equations are the only tool we have for describing shapes with the tools of algebra. If a sub-shape is algebraic, we can compute with it, add it, count it, and (in Part VII) pair it against forms via integration. If a sub-shape is just some free-drawn set of points, none of that machinery applies.

Cycles are how we describe "what is inside" a variety in a way that is compatible with algebra. They are the internal parts of a shape that carry mathematical weight.

Cycles can be added together, subtracted, and compared. The result is a group.

Given two cycles on the same variety, you can consider their formal sum — literally the pair (Z₁, Z₂) treated as if they were two things being added. You can also multiply a cycle by a whole number (say, count the same cycle twice), and you can consider a cycle with a minus sign in front of it. Under these operations the cycles form what mathematicians call an abelian group.

The group of all algebraic cycles of a given dimension on a variety, taken up to a natural equivalence called rational equivalence (two cycles are treated as the same if one can be continuously deformed into the other through a family of cycles), is called a Chow group. It is named after Wei-Liang Chow.

The Chow group is enormous but tractable. It is the natural bookkeeping device for asking questions like: how many essentially different lines are there on this surface? How many independent curves? How many hypersurfaces? Every classical enumerative question in algebraic geometry translates into a question about a Chow group.

The rational-equivalence quotient is a precise mathematical operation. QBD's hypothesis (Part VIII) is that it is a good model of the operation that turns a jittering microstate ensemble into a single stable unit — the same move by which Jupiter's 10⁵⁰ atoms become one number-with-mass in Newton's equations. Chapter 18 introduced the physical version of this move.

Two cycles inside a variety either meet or they do not. The count of how many times they meet is a topological invariant.

Adding cycles gave us the Chow group. There is a second, equally fundamental operation on cycles: intersection. Two curves on a surface can cross at a point. Two surfaces inside a 3-fold can meet along a curve. In general, if A and B are cycles inside a variety X and they meet cleanly, the dimension of their intersection follows a simple rule:

When A and B have complementary dimensions — their dimensions add up to the dimension of X — this formula gives dim(A ∩ B) = 0. Their intersection is a finite set of points. We can count them. That count, written A · B, is called the intersection number.

The intersection number lives at the level of Chow classes. If you replace A by another representative of the same rational-equivalence class, or B by another representative of its class, A · B does not change. Rational equivalence quotients out exactly the deformations that would move a crossing without creating or destroying one.

This gives cycles a second structure alongside addition. Chow classes can be added — and they can be paired with each other to give integers. Together, these two operations make the direct sum of Chow groups a ring: the Chow ring of X. Its multiplication is the intersection product.

Every cycle of a given kind is itself a point of a new, higher-dimensional variety.

Here is where the picture gets genuinely startling. Take all the algebraic cycles of a fixed dimension on a variety X. Each such cycle can be thought of as a single object. Now: is there a way to parametrise them all — to arrange the whole family of cycles as points of some other shape?

Yes. The answer is a variety in its own right, called a Chow variety. Its points are exactly the cycles of a given kind on X. Moving from one point on the Chow variety to a neighbouring point corresponds to deforming one cycle into a nearby cycle. The Chow variety is the space of cycles.

This is the key move. Cycles were parts of X. Now they are points of a new variety. The dimension of the Chow variety is typically much larger than the dimension of X, because there are so many possible cycles.

What was internal structure of X at one level has become external, first-class geometry at the next level. Deforming a cycle inside X is now the same thing as moving along a curve inside the Chow variety. Adding cycles is now addition of points. Every property we have — dimension, cycles inside it, Chow variety of it — is available all over again.

Chow variety of the Chow variety. And so on. The construction stacks upward without limit.

Nothing prevents us from doing the trick again. Take the Chow variety of X — a new variety, call it X⁽¹⁾. It has its own cycles, its own Chow variety X⁽²⁾. And so on: X⁽³⁾, X⁽⁴⁾, forever.

What we have is a tower of varieties. At each level, the cycles of that level become the points of the next level up. The tower captures how disciplined internal structure at one level becomes external structure at the next.

Why does this matter? Because most complex systems we care about — biological, physical, economic — have exactly this kind of hierarchical structure. A single cell has internal states (level 0). Which states a whole tissue can adopt depends on which cell states are available (level 1: patterns of cells). Which tissue configurations an organism supports depends on which tissue patterns are available (level 2). And so on up to the whole organism, and beyond into whole populations and ecosystems.

The Chow tower is not an analogy for this hierarchy. It is the mathematical structure that best captures it. When the tower is well-behaved, the whole hierarchy is coherent; when it collapses or truncates unexpectedly, only certain patterns can be stable across scales.

A differential form is the object you integrate over a shape. Forms are what turn geometry into numbers.

In Part III we integrated a function of one variable over an interval on the line. When the underlying space is a higher-dimensional manifold, the object we integrate is no longer just a function — it is a differential form. Forms come in different degrees: a 1-form is integrated over curves, a 2-form over surfaces, a 3-form over volumes, and in general a k-form is integrated over k-dimensional cycles.

Think of a 1-form as a rule that assigns, to each direction at each point of the manifold, a small number. Integrating it along a curve accumulates all these small numbers as you travel — very much as in first-year calculus, except the "small numbers" now depend on which direction you go, not just where you are.

A 2-form assigns a number to each pair of directions — an oriented area element at each point. Integrating it over a surface accumulates the area-elements. And so on. The mechanics generalise cleanly from the calculus you already know.

Two facts about forms will matter here. First, forms can be added and multiplied. Second, there is a differential operator called d that takes a k-form to a (k+1)-form — an "exterior derivative" that generalises ordinary differentiation. A form is called closed if dω = 0, and exact if it is itself the d of some lower form. Every exact form is automatically closed. The interesting question is: which closed forms are not exact?

The closed forms that are not exact are exactly the ones that detect holes in the shape.

On a shape without holes, every closed form is exact — the two notions coincide, and there is nothing more to say. On a shape with holes, there are closed forms that are not exact, and these forms carry the information about the holes.

The cohomology of a shape, written H^k(X), is a vector space whose elements are equivalence classes of closed k-forms, two forms treated as equal if they differ by an exact form. Cohomology counts, in a precise sense, the independent k-dimensional holes in the shape.

A sphere has one 0-dimensional hole (it is one connected piece) and one 2-dimensional hole (it is hollow). A torus has one 0-dim hole, two 1-dim holes (around the tube, and around the donut hole), and one 2-dim hole. The Betti numbers of a shape are the dimensions of these cohomology spaces. They are the numerical fingerprint of a shape's topology.

What Hodge did in the 1930s and 40s was more than just count holes. He showed that when the shape is a compact complex manifold (a shape that is not just curved but complex — carrying an internal notion of multiplication of directions), the cohomology has a much richer, decomposed structure.

On a compact complex manifold, cohomology splits into a two-index grid — the (p,q) decomposition.

On a compact complex manifold X, the k-th cohomology space splits — cleanly, canonically, uniquely — into a sum of pieces indexed by two numbers p and q with p + q = k. This is the Hodge decomposition.

The piece H^{p, q} consists of cohomology classes that can be represented by forms with p holomorphic and q anti-holomorphic differentials. On a complex manifold, directions come in two flavours — holomorphic and anti-holomorphic — and the (p, q) index tracks how many of each are involved.

What matters here is not the technical details of holomorphicity, but the picture: each level of cohomology, one for each degree k, splits into a diamond of pieces. The diamond is symmetric (H^p,q and H^q,p are conjugates). The whole thing is called the Hodge diamond, and it is the fingerprint of the shape at a much deeper level than the Betti numbers alone.

The middle row of the diamond — where p = q — is special. Elements of H^{p, p} are called Hodge classes. They are the cohomology classes with a balanced Hodge type. And it is these Hodge classes that the Hodge conjecture is about.

Every algebraic cycle gives a Hodge class. The Hodge conjecture asks whether every Hodge class comes from an algebraic cycle.

Here is the key connection. Every algebraic cycle Z on a compact complex projective manifold X has an associated cohomology class [Z] in H^2k(X, ℚ). That class always lies in the middle row of the Hodge diamond — that is, it is always a Hodge class. This half of the story is a theorem: algebraic cycles give Hodge classes.

The Hodge conjecture is the reverse question. If you hand me a Hodge class on X — an element of H^p,p(X) that happens to be rational — does it necessarily come from an algebraic cycle? Or could there be Hodge classes that are not made by cycles?

In one direction: cycles → Hodge classes (proved). In the other direction: Hodge classes → cycles (conjectured). The conjecture asserts that the two pictures agree.

This is one of the seven Millennium Prize Problems for a reason. It is a bridge between two very different worlds — the algebraic world of polynomial equations and the topological world of shapes and holes. If it is true (in full generality), it says the disciplined internal structure of a shape (its cycles) exactly matches its topological content (its Hodge classes). If it is false in general, the truth is subtler and more interesting.

The conjecture is known for classes of codimension 1 (Lefschetz's (1,1) theorem), and therefore for every shape of dimension up to three. It is also known for special families in dimension four, including fourfolds covered by rational curves (Conte and Murre, 1978). That family includes the Gushel–Mukai fourfolds we look at in Chapter 35. The general case is open. A version with whole-number coefficients instead of fractions is known to be false (Atiyah and Hirzebruch, 1962).

A proof is a chain of statements, each following from the last by an accepted rule, ending at what you set out to show.

Proofs come in three main flavours. It helps to see them side by side before we look at the Hodge case.

Direct proof. Start from the assumptions. Apply valid steps. Arrive at the conclusion. Nothing tricky about the structure.

Proof by contradiction. Assume the opposite of what you want to prove. Show that this assumption leads to something impossible. Therefore the assumption is false, and what you wanted was true.

Proof by induction. Show the statement holds for the base case (usually n = 1). Then show that if it holds for some n, it also holds for n + 1. It follows that it holds for every natural number, like knocking down an infinite row of dominoes.

A real proof at the level of a Millennium problem uses these building blocks over and over, often combining thousands of steps and drawing on decades of earlier theorems. But every step in the chain, if you zoom in far enough, is one of these three moves. There is nothing else. This is what makes mathematics unlike almost any other kind of knowledge: every claim can, in principle, be reduced to a chain of small, checkable steps.

How the Hodge conjecture was settled for one family of four-dimensional shapes: reduce to an easier case, then carry the answer back.

Gushel–Mukai fourfolds (GM fourfolds) are a well-studied family of four-dimensional compact complex projective shapes. For them the Hodge conjecture is not an open problem: it is a theorem. We include the case because it shows the shape of a real proof, and because an earlier revision of this page wrongly described it as an unfinished programme of ours.

1. Find the interesting part. The middle hole-count of a GM fourfold has Hodge numbers (0, 1, 22, 1, 0). All the action is in the (2,2) piece.
2. Rational version: already known. GM fourfolds are covered by rational curves. For all such fourfolds, the conjecture with fractional coefficients was proved by Conte and Murre (1978).
3. Split off a K3-like piece. The derived category of a GM fourfold contains a component, the Kuznetsov component, that behaves like the category of a K3 surface: a far better understood two-dimensional shape. Only special ('Hodge-special') GM fourfolds have an actual associated K3 surface. The K3-like component exists for every one of them.
4. Prove it for K3-like pieces. Perry (2022) proved an integral Hodge conjecture for such two-dimensional Calabi–Yau categories. The proof deforms them to genuine K3 surfaces, where Lefschetz's (1,1) theorem already settles the question.
5. Carry it back. Transporting the result through the component gives the integral Hodge conjecture in degree 2 for every GM fourfold.

Notice the structure. A hard problem is mapped to an easier one whose answer is already known. The real work is in the bridge. This is a very common shape of proof, and it is the same shape as the bridge QBD tries to build between geometry and biology in Chapter 38. The difference is that a mathematical bridge can be proved, while ours can only be tested.

Hypothesis: a compact complex projective manifold M that contains the whole tower as sub-varieties.

Every shape we have talked about so far — a variety, its Chow variety, the tower of Chow varieties — is itself a candidate to be embedded inside something larger. QBD's working hypothesis is that nature's stable levels behave as if there were such a larger object. That is a hypothesis about nature, not a theorem of mathematics.

M is required to be:

• Compact (bounded, closed — as in Chapter 23)
• Complex (it carries the internal complex structure that Hodge theory needs)
• Projective (it can be embedded in projective space, so the Hodge conjecture is stated for it in the strongest form)
• Large enough to contain the whole Chow tower as sub-varieties of ascending dimension

None of these adjectives is decorative. Drop any one and the structure below can fail. Voisin (2002) showed exactly why the projectivity requirement matters: on Kähler-but-not-projective manifolds, the Hodge conjecture can fail. If the hypothesised outer M were only Kähler, even the mathematical half of the hypothesis would lose its footing.

If the hypothesis is right, the outer shape's geometry would limit which inner patterns can be stable.

This is the chain of logic that ties the geometry to the hypothesis. Each step is labelled with what kind of claim it is.

1. Mathematics A compact complex projective M has a full Hodge decomposition (Chapter 32).
2. Mathematics, for proven cases Where the Hodge conjecture is proved for M, its Hodge classes are exactly the classes of algebraic cycles. Classes that are not Hodge classes cannot come from cycles.
3. Hypothesis H1 Stable patterns in nature correspond to algebraic cycles of some such M.
4. Hypothesis H2 The levels of description we use in science correspond to levels of the Chow tower inside M.
5. Consequence, if H1 and H2 hold Which patterns can be stable at each level would be limited by which Hodge classes M has.

Steps 1 and 2 are rigid: within mathematics they are prohibitions, not preferences. Steps 3 and 4 are where all the risk lives. They cannot be proved. They can only be tested, indirectly, through the network measurements of Chapter 38. If H1 or H2 fails, the mathematics stays true and the hypothesis simply stops describing nature.

Geometry lives in exact shapes. Biology arrives as noisy networks. The bridge between them is a set of explicit choices, and each choice is a place the idea could fail.

No one can measure a Hodge class in a cell. What we can do is take a real biological network (for example, which yeast proteins interact), compute features inspired by the geometry, and ask a plain question. Do those features predict something real (which genes the cell cannot live without) better than the standard network measures already used for this?

1. Atoms. Find the smallest tightly knit clusters in the network. These play the role of the basic cycles.
2. Gluing. Measure how those clusters overlap and share boundaries. This plays the role of the intersection structure of Chapter 27.
3. Symmetry checks. Renaming the proteins must not change any result.
4. Zoom checks. Merging nodes into bigger groups and then measuring should give the same answer as measuring and then merging. If it does not, the feature does not survive coarse-graining, which is the whole point of the idea.
5. Prediction. Compare a standard model with the same model plus the geometric features, using fixed code and a pass mark set in advance.

Because English has no existing word for "the outer geometric object whose cycle structure decides what patterns can be stable below," we borrow a placeholder.

In our internal working notes we call the outer shape M a univa. This is not a discovered word; it is a working label. It stands in for "the top-down geometric object whose Hodge structure constrains what is stable at every scale within it."

The reason we introduce a placeholder at all is that no existing English word carries the right meaning. "Space" is too broad (any set of points is a space). "Universe" is loaded with cosmological connotations that we do not intend. "Manifold" describes the local structure but not the top-down role. "Container" implies a passive vessel; the hypothesis is that its cycle structure actively constrains what can exist inside it.

Everywhere else on this page we have used only standard mathematical vocabulary. This one term is the exception. It appears once for what it is, and then we return to speaking about M as a compact complex projective manifold whose Hodge structure constrains its cycles.

Every test says in advance what result would count against us. Here is each one, what happened, and what it means.

Six tests have been designed so far. Four were run on the yeast protein-interaction network, with each protein labelled by whether the cell can survive without it. One passed, three failed, and one of those failures was also underpowered. Two further tests are designed or drafted but not run. We list them all, with the failures first in importance.

Proposed tests, not yet frozen

Places the idea might apply, each labelled with how far it has actually got.

If the hypothesis were right, it should show the same structure across very different-looking fields. Here are nine candidates. Only one has been tested at all, and that test failed (Chapter 40). None of these is a result.

What we are not claiming.

Every framework has a boundary beyond which it is speculation. Here is ours, stated as plainly as we can.

• The Hodge conjecture is open in general. We use only the cases where it is proven. Nothing on this page depends on the general case being true.
• The link from geometry to nature is a hypothesis. Hypotheses H1 and H2 (Chapter 37) are not proven, and could not be proven by mathematics alone. Our first network test went against the tested version.
• We have not shown that the idea fits any biological system. Earlier revisions said it fitted Wnt signalling and gene-knockout data. That was an overstatement, and it has been removed.
• The observer is not a mind. Observer-dependence here means dependence on which part of a system you can access (Chapter 17).
• No consciousness, cosmology or teleology is claimed. Those threads are philosophically interesting, and are kept apart from anything testable.
• Nothing here is medical, financial or investment advice.

Plain answers to the most common questions, and every technical word on the page in one sentence each.

Some of these questions came from readers, and some from our own review panel. Several of them changed the page.