Symmetry Groups: Crystals, Conservation Laws, and Neural Networks

A group on its own is abstract. It becomes useful when it acts: each element is a transformation of some set, and composing elements composes transformations.

Two objects come out of an action. The orbit of a point is everywhere the group can send it. The stabiliser is the subgroup of elements that leave it fixed. They are linked by a counting result you can prove with cosets, exactly as in the previous lesson:

$|\text{orbit}(x)| \cdot |\text{stabiliser}(x)| = |G|$

Try it: Take the 24 rotations of a cube acting on its 6 faces. One face can be sent to any of the 6, so its orbit has size 6. Its stabiliser is the 4 rotations about the axis through it. And $6 \times 4 = 24$. The theorem is really just Lagrange again, wearing the clothes of geometry.

How many distinct dice can you paint with two colours, if two paintings are the same when one can be rotated into the other? Enumerating gives $2^6 = 64$ paintings, but that overcounts wildly.

Burnside's lemma answers it: the number of orbits is the average number of arrangements each group element leaves fixed.

$\#\text{orbits} = \frac{1}{|G|}\sum_{g \in G} |\mathrm{Fix}(g)|$

For the cube's 24 rotations, grouped by type:

The total is 240, and $240 / 24 = 10$. Ten distinct dice, not 64. With three colours the same computation gives 57.

Now let the group act on an infinite periodic lattice. Something remarkable falls out: the rotations compatible with translational periodicity are only those of order 1, 2, 3, 4 and 6.

The argument is short. Take the shortest translation vector in the lattice. If a rotation of order $n$ exists, applying it to that vector generates other lattice vectors, and for $n = 5$ or $n \geq 7$ the combination produces a translation shorter than the one you started with, contradicting its minimality.

Key idea: No periodic tiling of the plane can have five-fold rotational symmetry. Not "none has been found": none can exist. This is the crystallographic restriction theorem, and it is a group-theoretic result with an experimental consequence, which is a rare and powerful combination.

That is why you see hexagons and squares in tilings and honeycombs, and why regular pentagons never tile a floor.

Combining the allowed rotations with reflections, glides and translations gives a finite list of possible symmetry types, in every dimension.

Seven ways to decorate a strip. Seventeen ways to tile a wall, a fact the Alhambra's builders arguably exhausted by hand centuries before it was proved. And exactly 230 ways for atoms to repeat in three dimensions.

That last number is not a survey of known crystals. It is a complete enumeration: every crystal that exists, and every crystal that ever will, belongs to one of 230 groups. New materials appear constantly and new space groups cannot.

Key idea: A theorem is only as strong as its hypotheses, and the useful reading of a contradiction is to ask which hypothesis nature declined to satisfy. The mathematics was never in question; what changed was the class of objects it was being applied to.

The groups so far have been finite or discrete. Rotating a sphere can be done by any angle, so its symmetry group is continuous: a Lie group, one that is also a smooth manifold.

$SO(3)$ is where robotics lives, and its structure has practical consequences: it is not simply connected, which is the deep reason why three Euler angles cannot cover all orientations without gimbal lock, and why quaternions, which live in $SU(2)$, are used instead.

Continuous groups have infinitely many elements but a finite number of independent directions of motion, which is what makes them tractable and is the subject of Lie theory.

In 1918 Emmy Noether proved that if a physical system's laws are unchanged by a continuous symmetry, some quantity is exactly conserved, and the correspondence is precise rather than poetic.

This inverted the logic of physics. Conservation of energy stopped being an empirical regularity and became a consequence of the observation that experiments give the same results next week.

Key idea: Modern physics is largely built by choosing a symmetry group and asking what theory is consistent with it. The Standard Model is specified by $SU(3) \times SU(2) \times U(1)$, and the forces are what those symmetries require. Group theory is not a description applied afterwards; it is the input.

The same idea appears in machine learning under the name equivariance. A function is equivariant to a group when transforming the input transforms the output the same way:

$f(g \cdot x) = g \cdot f(x)$

A convolution is equivariant to translation by construction: shift the image and the feature map shifts identically. That is not a learned property, it is built into the operation, and it is why a convolutional network needs far fewer examples than a fully connected one to learn that a cat is a cat wherever it appears in the frame.

• Translation equivariance gives convolutional networks
• Rotation and reflection equivariance gives group-equivariant networks, used in medical and satellite imagery where there is no canonical orientation
• Permutation invariance gives graph networks and set-based models, where reordering the input must not change the answer
• $SE(3)$ equivariance gives molecular and protein models, where a rotated molecule is the same molecule

In practice: Building a symmetry in is a hard constraint; learning it from data is a soft one that costs examples and can fail quietly. Where the symmetry is genuinely exact, encoding it is nearly always the better trade.

A symmetry of the laws need not be a symmetry of the outcome, and the gap is where most of the interesting physics and chemistry lives.

A pencil balanced on its tip is rotationally symmetric. It falls in one direction. Nothing in the setup preferred that direction, and the result has less symmetry than the situation that produced it. This is spontaneous symmetry breaking, and it explains why a magnet has a north pole despite the underlying physics having no preferred axis, and why water freezing into ice picks one of the 230 space groups out of a rotationally symmetric liquid.

Gotcha: "The system is symmetric" and "the state is symmetric" are different claims. Confusing them produces predictions that a symmetric setup must give a symmetric answer, which is false in general and is exactly what makes phase transitions possible.

The same failure mode appears in optimisation: a loss function symmetric under permuting hidden units has many equivalent minima, and gradient descent picks one arbitrarily.

The recurring pattern in this lesson is that identifying a group answers a question no direct calculation was going to reach.

1. It rules things out completely. No five-fold periodic crystal, no 231st space group. Impossibility results are the strongest kind and are almost always symmetry arguments.
2. It counts things up to sameness. Burnside turned an intractable enumeration into an average over 24 rotations.
3. It produces conserved quantities. Noether converts a statement about invariance into a number that does not change.
4. It constrains what is possible before you know the details. Selection rules in spectroscopy forbid transitions on symmetry grounds alone, without solving anything.
5. It can be built into a model. Equivariant architectures get a property for free that a general model would have to learn.

The next lesson pushes this one step further, to a case where a symmetry group settles a question that resisted mathematicians for three centuries.