A few weks ago I posted this tweet
Even if that could be said about a number of Witten’s papers, the one I was referring to in particular is his 1982 «An anomaly». I was introduced to it not so long ago, during a course on anomalies which I am following, and found it so interesting that I felt like I needed to share it. The paper itself is a work of art. It has everything, from physical intuition to a wide range of mathematical ideas and concepts which are masterfully sewn together. So, strap on, because I am about to deliver a (probably mediocre) recollection of the aforementioned paper, with some heuristic explanations and some fangirling sprinkled throughout.
The key concept which Witten introduces, and which makes this paper so important, is that of global anomalies. For those unfamiliar, an anomaly occurs when a symmetry displayed by a classical system is broken upon quantization. When this happens to non-gauge symmetries, it merely indicates that the symmetry is somehow unphysical, and an artifact of the classical treatment of the theory. However, when it happens to a gauge symmetry, it’s nothing short of a catastrophe, which indicates that our theory as a whole is inconsistent. This is due to the fact that gauge symmetries are not really symmetries in the usual sense, but rather redundancies in our description of a theory which are of great physical relevance, and have very real physical consequences.
Apart from gauge and non-gauge anomalies, there is one further classification which is absolutely key in setting up the stage for Witten’s paper. Symmetry transformations are encoded into group actions. Oftentimes, they are given by Lie groups, on which there is a notion of (path) connectedness. We therefore make the distinction between transformations which are connected to the identity element, and those which are not. For the former, we can follow a perturbative approach, by expanding the transformation near the identity element and using our fancy Lie algebra machinery. However, this cannot be done for the latter. This is indeed a problem, since both types of transformations are perfectly capable of developing anomalies upon quantization, but our techniques are only sensitive to anomalies coming from transformations which are connected to the identity. We call these perturbative anomalies, whereas the remaining ones are known as global anomalies. Because of the seminal paper which we are about to discuss, they are also sometimes called Witten anomalies.
Without further ado, let’s see what these global anomalies are all about. The setup is that of a gauge theory in , where the gauge group is
, and the theory contains
Weyl (chiral) fermions in the fundamental (often denoted by
) representation on
in Euclidean signature. To understand how a gauge anomaly may develop, we need to look closely at the gauge transformations themselves. At the mathematical level, the action of choosing a gauge is the same as choosing a section
on a principal fiber bundle
, with
being our base space (in this case Euclidean four-space
), and
the principal fiber bundle that sits on top, with structure group
, A gauge transformation is then a map that takes us from one section to another. Because each fiber
is homeomorphic to the gauge group
, and hence the transition functions are themselves elements of
, then a gauge transformation (in our particular setup) is nothing but a map
Furthermore, we will consider gauge transformations that fall off at infinity, which means that
for
This last condition allows us to identify (at the level of the gauge transformation) all points at infinity to a single one, a one-point compactification. Hence, this defines a map
This is a key point of the argument, because maps of into a generic topological space
are precisely what define the homotopy groups
of said space. In particular, if we use that as a manifold
, we have that the relevant homotopy group for the above gauge transformations is nontrivial, and given by
This tells us that gauge transformations in this setup are classified into two distinct homotopy classes. In particular, maps that belong to one of the classes cannot be continuously deformed to maps in the other class. Because the identity element can only belong to one of the classes, we are led to the conclusion that there are some gauge transformations that cannot be deformed into the identity, and thus cannot be treated perturbatively.
All of the above may sound like abstract nonsense, so it would be of use to come up with some (rather crude) mental picture of the situation. We have said that mathematically, a choice of gauge is a choice of section from the base manifold to the principal bundle
sitting «on top» of it. Thus, a gauge transformation is a map that takes us from one choice of section to another. If such a transformation can be deformed to the identity, we would picture this as if the two sections that it connects could be deformed to one another. An obstruction to this would be if the principal bundle
had more than one connected component. All in all, we imagine the situation as follows:

This is a (rather very crude and possibly flawed) way of visualizing the situation. In this case, we have three sections (or gauge choices), labelled . Two of them,
and
can be continuously deformed into one another, and therefore the gauge transformation which would take us from (say)
to $s_2$ can be deformed to the identity. This is to say that this transformation is homotopic to, or in the same homotopy class as, the identity. However, this is not true for
, which cannot be deformed into either
or
, as they live in different connected components. Therefore, transformations that take us from one component to the other cannot be deformed to the identity. This mental picture illustrates the main consequence of the fact that
. Let
be some gauge transformation which is not connected to the identity. Then, for each gauge configuration (or, strictly speaking, choice of connection)
, there is a gauge equivalent connection given by
which appears at the level of the (Euclidean) path integral
Because the path integral runs over all possible connections, there is a double counting: for each we also have to count
.
Now, to make things interesting, suppose that we include fermions into the picture. In particular, we will include a single Weyl (chiral) fermion. »Why so?», you might ask. If you are not familiar with gauge anomalies, the answer is, at a surface level, quite straightforward: Gauge fields couple in equal and opposite ways to left- and right- handed fermions (provided they are of the same species), and it turns out that the anomalous variation of the former exactly cancels the anomalous variation of the latter. Thus, a theory which contains Dirac fermions, which can be each decomposed as the combination of a left- and a right- handed fermion, will always be anomaly free. In a more direct way, Dirac fermions admit a mass term, and thus a choice of gauge invariant (Pauli-Villars) regulator, defined in terms of said term. In summary, we want to add an uneven amount of left- and right- handed fermions. Because that can always be reduced to a sum of cases in which there is one chiral fermion, we will only add one chiral fermion.
After this fermion-adding discussion, we would like to integrate out all matter, to leave the partition function (path integral) such that it only depends on the connection . This integrating out essentially means that we need to perform the path integral over the fermionic fields. Usually, for a Dirac fermion, what we get is the following functional determinant:
If we instead consider a single Weyl fermion, we run over half of the modes, and hence the result is
and because of this square root, there is a potential sign ambiguity in the paritition function.
Suppose we make a particular choice of , and assign to it some sign for
. Because it is a sign, it is discrete, and therefore invariant under infinitessimal (continuous) gauge transformations. In turn, this implies that it will be the same in each of the individual
components. However, it may not be the same between the two of them. Namely, it may change under
, which was the transformation that took us from one component to the other. If that were to be the case, we would have
This is a phase appearing on the (supposedly gauge-invariant) partition function under a gauge transformation: an Anomaly. But it gets worse! Since each has one counterpart in the other component of the gauge group given by
, each contribution from the former to the partition function gets cancelled by the latter, and in the end the whole
vanishes identically. This is a catastrophe! Without a partition function we cannot even define expectation values, and we conclude that the theory would be ill defined. This is the famous global anomaly.
Up to here, our task is only halfway done, because we still have to check that the above sign ambiguity does indeed happen. In here lies a big portion of the brilliance of Witten’s original paper, and also the germ for a later, more modern generalization of global anomalies, of which I will (if my laughably short attention span permits) write more in the future. What follows will be somewhat more technical than the setup of the anomaly, and I do not claim to be an expert in some of the things which I will mention. Therefore, some of the explanations will either be incredibly subpar or simply nonexistent. To be fair, I am not actually an expert on anything, so everything is okay.
To make our lives somewhat easier, we will take spacetime to be a 4-sphere (whereas up to now we had only identified it to be so at the level of gauge transformations). This implies that the Dirac operator has real eigenvalues, as well as a choice of real regulator. Furthermore, we assume that has no zero modes (i.e. zero eigenvalues). Otherwise the sign difference between the two components of the gauge group would not matter, as the equation would be
. As we said before, for our present choices, the eigenvalues of
are real, but we further have that for each eigenvalue
, there is another eigenvalue
, since
.
Here we have used that is a Weyl spinor, and therefore an eigenvector of $\gamma^5$ itself, and that $\gamma^5$ anti-commutes with $\gamma^{\mu}$. We conclude that, because every eigenvalue comes in a pair
, and each sign is associated with one chirality,
is a product over half of the eigenvalues. That is, for each
, we pick one sign out of
, and chug it in the product.
Because of the above discussion, we have the freedom of choosing a sign for each eigenvalue for a particular configuration. Again, to make our life easier, we choose all positive for some
. Now, we define a continuous map between the two (which we could understand as a homotopy) as follows
While this is of course possible (we have just done it), note that only the initial and final gauge configuration are gauge equivalent. This is not necessarily the case for any intermediate value of . In fact, this is precisely where the magic happens. Since the initial and final configurations are gauge equivalent, they must give rise to the same eigenvalues of
, which have to match between the two. However, the intermediate configurations are no man’s land, where anything (that respects the Geneva conventions) can happen. The situation may be visualized as follows

Not only can the eigenvalues reshuffle among themselves, but we will in fact show that it is the case that there is an odd number of crossings between them, which will give rise to the anomaly. The way in which Witten does it is simply brilliant.
To show that the anomaly does take place, we will increase the dimension by , and work in five-dimensional space. We do this in order to make use of a result of one of the famous Atiyah-Singer index theorems, in particular the so-called «mod 2 index theorem». To set the stage up, consider
gauge theory in 5 dimensions with a doublet of fermions which obey the Dirac equation:
Now strap on because here come a bunch of representation theory facts which I have not checked nor understand pretty well myself. We will have to trust Witten on this one. The spinor transforms first as the spinor representation of
(namely
, the double cover of $SO(5)$), which is pseudo-real, and secondly as an
doublet, which is also pseudo-real. Here, pseudo-real means that a representation and its conjugate are related by a unitary matrix, as opposed to a real representation, where the matrix is in particular the identity, or a complex representation, where the representation and its conjugate are not equivalent. In any case, while these representations are pseudo-real on their own, their tensor product is in fact real, so this allows us to choose the
matrices to be real and symmetric
matrices. On the other hand, the infinitesimal generators
of $SU(2)$ are taken to be real, anti-symmetric matrices. These two facts combine to make
a real, antisymmetric operator acting on an infinite dimensional space of functions. The eigenvalues of an operator with these characteristics are either
or come in conjugate pairs
.
The big conclusion to take away from the previous paragraph is that, if we vary the which defines
(not to be confused with the
from a few paragraphs back, which was four dimensional), only two things can happen: Either a pair of conjugate eigenvalues «annihilate» and become
, or a pair of two (previously
) eigenvalues become a nonzero pair
. As a consequence, the number of zero eigenvalues of
mod
is conserved, which leads us to interpret them as topological invariants. This is precisely the mod
index of the Dirac operator.
So, how do we use this information for the task at hand? Witten’s idea was to consider as a background, with coordinates
, with
. This seems a priori very logical, since it contains both our original four-dimensional background
, plus some additional parameter living in the real line. Here is a mental picture:

Admittedly, I should have drawn it like a straight cylinder, but this has a reason which will hopefully become apparent at the end of the post. On this space, we define the following gauge configuration:
Essentially, plays more or less the same role as the parameter
from a few paragraphs ago, which connected both gauge configurations. The crucial difference is that this setup allows us to use the Atiyah-Singer index theorem. Upon doing so, we obtain that with the above configuration, the mod
index of
is 1. In other words, the Dirac operator in five dimensions in this case has an odd number of zero modes. We will now see that these zero modes are related to the crossings in four dimensions.
Now that we know that has an odd number of zero modes, let us investigate them further. For each zero mode, we have that
If we interpret as a sort of time parameter, this is an evolution equation whose source term is governed by the (four-dimensional) Dirac operator. Furthermore, because
varies over all of
, we make the evolution as slow as possible, which allows us to work in the so-called adiabatic approximation. In it, we assume that this evolution is slow enough so that, if the system is initially defined by some eigenstate (and there is a gap between eigenstates), then it will stay in said eigenstate for the whole evolution. First we perform a variable split (which is a common technique in PDE theory) by writing
There is still a notion of a dependence on
, because we for each value of
we take it to be a smoothly evolving solution to the eigenvalue equation
which, again, is defined for each . Now, the adiabatic variation of the system in
, allows us to assume that the eigenfunctions above will always have the same eigenvalues
associated to them. The subtlety is that this does not mean that the eigenvalues themselves cannot change (and, in fact, they will change), but rather that they will do so while still being defined by the same eigenfunction. In other words, they will evolve in the curves which we plotted in the last schematic figure above. To see the particular way in which these eigenvalues evolve, we use the ansatz above to rewrite the zero-mode equation as
As you might know from your undergraduate days, this ODE is solved by
And, now for the grand reveal: this solution is only normalizable if is positive for
and negative for
. Tracing the definition of this eigenvalue back, this translates to the fact that for each
eigenvalue of
, we get a zero-crossing of an eigenvalue of
, and vice-versa. Because of the Atiyah-Singer index theorem, we conclude that
has an odd number of zero-crosings, and finally that the
anomaly does indeed take place.
I know that this post was maybe way too long (and I appreciate it if you stuck through!), so this might call for a small recap and summary of what we have seen, before the closing paragraph. At the beginning, we discussed the difference between perturbative and global anomalies. The latter of the two are those related to anomalies which cannot be connected or deformed to the identity, but can still develop anomalies. Because of our inability to expand these transformations perturbatively, we required new methods for studying them. In particular, we saw that if , a theory with one Weyl fermion could develop one such global anomaly. After that, we translated the existence of said anomaly to a change of sign in the square root of the functional determinant of the (four-dimensional) Dirac operator. In turn, this led us to consider the amount of its eigenvalues which changed from positive to negative. To study them, Witten proposed jumping one dimension up, and using the Atiyah-Singer index theorem to deduce that in five dimensions, the Dirac operator possessed an odd number of zero modes. The argument was complete when we discovered that these zero modes were in one-to-one correspondence with the zero-crossings of the four-dimensional Dirac operator.
Many years have passed since the publication of Witten´s groundbreaking paper. The world still had to wait for another 19 years for the release of Shrek, the event that still gives hope to a lot of young students around the world, such as myself. In the meantime, people have applied and expanded on these ideas, in ways in which either ecape me or I cannot summarize in this paragraph. There is, however, one of these generalizations which I would like to touch on in the future. You would have noticed that, in order to study a four-dimensional anomaly, we resorted to a five-dimensional setup. If you have seen perturbative anomalies before, you might know that the most natural way to treat them is by means of the anomaly polynomial, which is defined in two dimensions more than the original theory (i.e. for we would have defined the polynomial in
). This is another manifestation of the difference between global and local anomalies. Now, I would like to draw your attention to the particular five-dimensional setup which we have used. We could understand it as a cylinder, whose two bases are the original (four-dimensional) manifolds which we had as backgrounds for our theory. The choice of having a cylinder was taken out of convenience, but we could think of other, more esoteric manifolds which somehow connect our backgrounds. For example, we could try punching holes in the middle, or modifying its topology in other ways. Fortunately, our mathematician friends had already been working on exactly this kind of idea for a few decades: they called them cobordisms. Given two (closed and compact)
-dimensional manifolds
and
, a cobordism (or bordism) bewteen them is a
-manifold, which has
and
as its boundary. The introduction of cobordisms into the mix led people to what is now known as Dai-Freed anomalies, and also their usage in other contexts, such as topological quantum field theories, or quantum gravity conjectures. If time allows, this will hopefully be one of the next topics of a future post, as it is also related to the kind of research that I currently find myself into.

