Lectures on the Bethe Ansatz

Fedor Levkovich-Maslyuk

Mathematics Department, Kings College London

The Strand, WC2R 2LS London, United Kingdom

Abstract

We give a pedagogical introduction to the Bethe ansatz techniques in integrable

arXiv:1606.02950v2 [hep-th] 24 Jul 2016

QFTs and spin chains. We first discuss and motivate the general framework of
asymptotic Bethe ansatz for the spectrum of integrable QFTs in large volume,
based on the exact S-matrix. Then we illustrate this method in several concrete
theories. The first case we study is the SU (2) chiral Gross-Neveu model. We
derive the Bethe equations via algebraic Bethe ansatz, solving in the process the
Heisenberg XXX spin chain. We discuss this famous spin chain model in some
detail, covering in particular the coordinate Bethe ansatz, some properties of
Bethe states, and the classical scaling limit leading to finite-gap equations. Then
we proceed to the more involved SU (3) chiral Gross-Neveu model and derive the
Bethe equations using nested algebraic Bethe ansatz to solve the arising SU (3)
spin chain. Finally we show how a method similar to the Bethe ansatz works in
a completley different setting, namely for the 1d oscillator in quantum mechanics.

This article is part of a collection of introductory reviews originating from lectures

given at the YRIS summer school in Durham during July 2015.

E-mail: fedor.levkovich@gmail.com
Contents
1 Introduction 2

2 Asymptotic Bethe ansatz equations in 2d integrable QFTs 3

3 Algebraic Bethe ansatz: building the transfer matrix 7

4 Algebraic Bethe ansatz: solving the SU(2) chiral Gross-Neveu model 11

4.1 Coordinate Bethe ansatz for the XXX Hamiltonian . . . . . . . . . . . . . . 15
4.2 Bethe equations for the SU (2) chiral Gross-Neveu model . . . . . . . . . . . 17

5 Exploring the XXX spin chain 18

5.1 The Bethe states in-depth . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
5.2 Spectral curve and finite-gap equations . . . . . . . . . . . . . . . . . . . . . 21

6 Bethe ansatz for the SU(3) chiral Gross-Neveu model 24

6.1 Nested Bethe ansatz for the SU(3) chain . . . . . . . . . . . . . . . . . . . . 25

7 Bethe ansatz for the harmonic oscillator 30

1
1 Introduction
Integrable QFTs in 1+1 dimensions have a vast array of remarkable properties. In particular,
it is often possible to exactly compute the S-matrix of the theory (an extensive discussion of
this aspect is given in another part of the present collection [1]). The S-matrix captures a lot
of the theorys dynamics but only describes the scattering of asymptotic states, i.e. particles
starting from infinitely far away and then flying off to infinity again. In contrast, another
interesting setup to consider is when our theory is put into a spatial box of finite size L. In
finite volume the spectrum of the Hamiltonian becomes discrete, so a natural question to
ask is what are the energies of the states.
It turns out that for integrable theories this energy spectrum can be computed to a large
extent using only the scattering data. For large volume L one can write down equations for
the spectrum in terms of the exact S-matrix. These equations are known as the asymptotic
Bethe ansatz equations and they will be the main topic of this article. Like the S-matrix
itself, they are exact at any value of the coupling constants. These equations are only valid
when L is large (in a sense that will be made more precise later), but still provide a lot
of important information. They are also the first step towards formulating the so-called
Thermodynamic Bethe ansatz (TBA) equations which give the energies exactly at any L
including all corrections. The TBA approach is covered in detail in a different part of this
collection [2] (see also the introductory article of this collection).
Importantly, both the asymptotic Bethe ansatz and the TBA have been crucial for the
recent applications of integrability to several AdS/CFT dualities between gauge and string
theories [3]. A key problem in this setting is computing the energies of multiparticle string
states in finite volume, which are mapped to operator conformal dimensions in gauge theory
(in fact the volume L corresponds to the number of elementary fields in the operator).
Integrability methods have led to great success in exploring this problem, and in particular in
the computation of superstring energies on the AdS5 S 5 space which coincide with operator
dimensions in the dual N = 4 supersymmetric Yang-Mills theory in four dimensions. We
hope that several simpler examples discussed in this article will serve as a starting ground
for understanding how Bethe Ansatz works in the AdS/CFT context.
The name Bethe ansatz equations originates from the famous solution of the XXX spin
chain by Hans Bethe in [4]. While our main goal is to study integrable QFTs we will see that
often computing the spectrum of some QFT model leads to an auxiliary spin chain which
should be solved first. We will see several examples and discuss the Bethe ansatz solutions
of these spin chains as well. At the same time, various spin chain models are interesting on
their own as many of them find important applications in condensed matter physics.
We will first discuss the asymptotic Bethe equations in a general setting and then cover
several examples for particular models. The presentation is structured as follows. In section
2 we give physical motivation for the asymptotic Bethe ansatz in integrable QFT and write
the Bethe equations in a generic form as a periodicity condition on the wavefunction. In
section 3 we present in a general form the algebraic Bethe ansatz approach allowing to
greatly simplify the periodicity constraint by reducing it to a transparent diagonalization
problem. In section 4 we demonstrate the method in action on the example of the SU (2)

2
chiral Gross-Neveu model. In the process we obtain the solution of the celebrated XXX
SU (2) Heisenberg spin chain. In section 5 we discuss some features of the XXX chain and
its Bethe eigenstates, as well as the classical limit leading to finite-gap equations. In section
6 we proceed to the more complicated case of the SU (3) chiral Gross-Neveu model which we
solve via nested algebraic Bethe ansatz. Finally in section 7 we illustrate the versatility of
Bethe ansatz by applying a Bethe-like method to solve the 1d quantum mechanical oscillator.
Some exercises for the interested reader are also included throughout the text.
There is certainly a large literature on the subject available, in particular we would
like to point out several reviews discussing various aspects of the Bethe ansatz methods
[5, 6, 7, 8, 9, 10]. For reasons of presentation clarity, only some selected references are
included in this pedagogical article1 .

2 Asymptotic Bethe ansatz equations in 2d integrable

QFTs
In this section we will discuss the Bethe equations for the spectrum of a generic 1+1 dimen-
sional integrable theory. We will always consider a theory on a circle2 of length L, i.e. we
impose periodic boundary conditions.
As discussed in the part of this collection of articles dedicated to integrable S-matrices
[1], in an integrable theory the scattering has several remarkable features:

The number of particles is conserved in any scattering event

The momenta do not change but can only be redistributed between particles

The S-matrix for multiparticle scattering factorizes, i.e. the S-matrix for any number
of particles is a product of the 2 2 S-matrices

Although in general one cannot introduce a wavefunction in QFT due to the production
of virtual particles, for integrable theories these special features (most importantly the first
one) make it possible to do this at least in some regions of the configuration space. Then from
the periodicity of the wavefunction one can derive quantization conditions which determine
the spectrum.
Let us first discuss a toy model a theory with only one particle in its spectrum. An
example is the sinh-Gordon model for some values of the parameters. Then an intuitive
picture which gives the correct equations for the spectrum is the following one (for a more
rigorous discussion see e.g. [11]). Since the number of particles is conserved we can speak of
a wavefunction as in quantum mechanics. If we have n particles on a circle the wavefunction
must be periodic. Imagine that we take the first particle around the circle once, eventually
bringing it back to its place again.
1
We also tried to make the notation in this article maximally consistent with other parts of the present

3
Figure 1: Deriving the periodicity condition on a circle. We take the first particle with
momentum p1 around the circle, scattering it through all the other particles.

If there were no other particles (or if the theory was non-interacting) the wavefunction
would acquire a phase factor eip1 L where p1 is the particles momentum. Then from period-
icity of the wavefunction
eip1 L = 1 , (2.1)
i.e. the momentum would be quantized according to
2k
p1 = , kZ . (2.2)
L
However we need to take into account the interaction with other particles. If L is large
compared to the interaction range between particles (e.g. the inverse mass scale of the
theory), the particles are almost always well separated from each other. Because of this
their interaction is described by the asymptotic S-matrix which we know. The number of
particles does not change in this interaction. Thus when we take a particle around the circle,
it will scatter through all the other particles (see Fig. 1), and all that happens is that for
each scattering the wavefunction is multiplied by the S-matrix which is just some phase
factor, S(p1 , p2 ) = ei(p1 ,p2 ) . This product of S-matrix phases will combine with the eip1 L
phase acquired due to free propagation. Thus periodicity of the wavefunction will be ensured
if
eip1 L S(p1 , p2 )S(p1 , p3 ) . . . S(p1 , pn ) = 1 (2.3)
Similarly for any particle we get
n
Y
ipj L
e S(pj , pk ) = 1 (2.4)
k=1, k6=j

This is a set of n algebraic equations for n variables p1 . . . , pn which thus allow us to fix the
values of the momenta. The system (2.4) are the asymptotic Bethe ansatz equations for our
toy model.
review volume.
2
By a circle we actually mean a straight line segment [0, L] whose endpoints x = 0 and x = L are
identified.

4
 Once the momenta are found from (2.4), one can compute the energy of the state. In
regions where the particles are well separated they propagate freely, so the energy of the state
should always be equal to simply the sum of individual particles energies. If the energy of
a single particle with momentum p is (p) we will have
X
E= (pi ). (2.5)
i

From the derivation it is clear that the asymptotic Bethe equations will describe the
energy only in a large volume L. It turns out that, more precisely, the Bethe equations
capture all terms in the large L expansion of the energy which scale as 1/Lk (with integer
k), i.e. the powerlike corrections. However they miss the exponential corrections of the kind
emL (where m is the particles mass) which physically correspond to the effects of virtual
particles propagating around the circle.
In a more general theory one could have different types of particles, and different types
could transform into each other during scattering. In this case the S-matrix would have
some matrix structure with indices labeling the incoming and the outgoing particles. Then
in (2.4) the product in the r.h.s. would actually be a product of matrices, and should be
understood as acting on a wavefunction which also carries indices corresponding to particle
types.
Let us first present without derivation how Bethe equations will look like in this case. We
will need to introduce some important notation. Let us consider a model with K possible
particle flavours, or in other words K particle types. For each particle we should also allow
linear combinations of different flavours so its flavour state can be though of as an element
of CK . For n particles, their state which we denote by A is then the element of a tensor
product
A H1 H2 Hn (2.6)
with each Hi ' CK . Choosing a usual basis ei in CK we can also write
K
X K
X
A= Aj1 ...jn ej1 ejn . (2.7)
j1 =1 jn =1

Each S-matrix is a linear operator acting on the tensor product of two of the spaces Hi ,
Sij End(Hi Hj ) (2.8)
where we put a hat on S to underline that it now has a matrix structure. This of course
matches the notation from the part of this collection focussed on S-matrices [1] where the
S-matrix has four indices two for incoming particles and two for outgoing ones. A typical
example that we will discuss in the next section is
S12 (p1 , p2 ) = f (p1 , p2 )I + g(p1 , p2 )P12 (2.9)
where I is the identity operator and P is the permutation operator, i.e. P12 (ea eb ) = eb ea ,
while f and g are some explicit functions. As another illustration, we can write the unitarity
condition as
S12 (p1 , p2 )S12 (p2 , p1 ) = 1 (2.10)

5
and the highly important Yang-Baxter equation as

S12 (p1 , p2 )S13 (p1 , p3 )S23 (p2 , p3 ) = S23 (p2 , p3 )S13 (p1 , p3 )S12 (p1 , p2 ) (2.11)

and it is satisfied over the space H1 H2 H3 .

Now we are ready to write the Bethe equations for a theory with several particle types.
They are similar to (2.4) but the product of S-matrices now acts on a state A H1 H2
Hn ,
eipk L Sk,k+1 Sk,k+2 . . . Sk,n Sk,1 . . . Sk,k1 A = A, k = 1, . . . , n (2.12)
The energy is again the sum of individual energies.
Our main goal is to understand how to solve this equation. Notice that the ordering in
the product is also important since S is a matrix3 . So, the r.h.s. of (2.12) is an operator
acting in the full space H1 H2 Hn . To solve the Bethe equations (2.12) we need
to find its eigenvalues and eigenvectors. Fortunately, this is possible to do in a very efficient
way using the fact that the S-matrix satisfies the Yang-Baxter equation. In the next several
sections we will demonstrate how this works in concrete examples.
To finish the discussion, let us outline the derivation of the Bethe equations (2.12). First
we will need to write the wavefunction in a more explicit form. This was already discussed
to some extent in the article on S-matrices [1] (for more details on this see [11] and the
review [10]). To automatically take care of the (anti-)symmetrization for identical particles,
let us introduce creation and annihilation operators aj , aj , whose index j labels the different
particle types. Then we can describe the wavefunction as

i1 ...in (x1 , . . . , xn ) = h0|ai1 (x1 ) . . . ain (xn )|(p1 , . . . , pn )i (2.13)

where we use the product of annihilation operators to extract the part of the wavefunction
corresponding to particle flavours i1 , . . . , in . The state |(p1 , . . . , pn )i is defined as
Z n
!
X Y
|({p})i = dn y APj1 ...jn ({p}) eipPm ym (y1   yn ) (2.14)
PSn m=1

aj1 (y1 ) . . . ajn (yn )|0i

where (x) is the Heaviside step function4 and we assume summation over repeated indices.
The coefficients A are related to each other as
0
AP = SPi ,Pi+1 AP (2.15)
3
These equations again match well the picture of taking one particle around the circle. E.g. after we
scatter the 1st particle through the 2nd one they both change flavours, then the 2nd is untouched and we
scatter the 1st (with the flavour now different) through the third, etc. So once we take the first particle
around the circle it can change the flavour and also all other particles can change the flavour. The result
has to match the initial wavefunction which is exactly the statement in (2.12).
4
I.e. (y1   yn ) is equal to 1 when y1   yn and is equal to zero otherwise. Let us mention
that this expression for the wavefunction is valid only in the regions when the particles are well separated,
so it makes sense to consider the condition y1   yn .

6
if P 0 is obtained from P by permutation of elements i, i+1 corresponding to particles Pi , Pi+1 .
The flavour indices of the S-matrix in (2.15) are understood to be appropriately contracted
with those of AP . As any permutation can be written as a sequence of permutations that
affect only two elements, any AP can be related to AI (with I being the identity permutation)
via a sequence of multiplication by the S-matrices. The Yang-Baxter equation satisfied by
the S-matrix ensures this relation is unambiguous.
As an example, for a state with two particles in a theory with only one particle type, we
would get
x1 x2 = eip1 x1 +ip2 x2 + S(p1 , p2 )eip2 x1 +ip1 x2 (2.16)

x1 x2 = eip2 x1 +ip1 x2 + S(p1 , p2 )eip1 x1 +ip2 x2 (2.17)

Exercise: Derive these equations from (2.14).

Let us now recall that we are considering the theory on a circle where the absolute ordering
of particles is meaningless, so the ordering is only important up to cyclic permutations. Let
us consider for example the wavefunction for x1   xn . If we define y1 = x1 + L then
since the separation between particles cannot be larger than L we have x2   xn  y1 .
For this ordering we would get from (2.14) a different expression for the wavefunction, but
it should coincide with the first one as on a circle x1 is indistinguishable from x1 + L. This
leads to
eip1 L AI = S1,2 S1,3 . . . S1,n AI (2.18)
and in general (notice that the ordering in the product is important since S is a matrix)

eipk L AI = Sk,k+1 Sk,k+2 . . . Sk,n Sk,1 . . . Sk,k1 AI (2.19)

These are precisely the equations (2.12) that were announced above, where A is identified
with AI .
In conclusion, the crucial problem is to diagonalize the product of S-matrices in (2.12). In
the next section we will describe a general procedure for doing this based on the Yang-Baxter
equation, and then we will see how it works for concrete examples.

3 Algebraic Bethe ansatz: building the transfer matrix

The method we are going to use for solving the periodicity condition (2.12) goes under the
name of the algebraic Bethe ansatz. In this section we will discuss its part which is common
for all models the construction of the so-called transfer matrix and later we will specialize
to concrete examples.
The key insight which allows to diagonalize the product of S-matrices in (2.12) is to
introduce an unphysical particle with momentum p in an auxiliary space Ha ' CK and
scatter it through all our particles. That is, we define the monodromy matrix

Ta (p) = Sa1 (p, p1 )Sa2 (p, p2 ) . . . San (p, pn ) (3.1)

7
which acts in Ha H where H = H1 Hn is our physical Hilbert space. From the
Yang-Baxter equation for the S-matrix it follows that the monodromy matrix satisfies a
similar condition:
Sab (p, p0 )Ta (p)Tb (p0 ) = Tb (p0 )Ta (p)Sab (p, p0 ) (3.2)
Then we define the transfer matrix by taking a trace over the auxiliary space

T (p) = Tra Ta (p), (3.3)

and it is now an operator on the physical space only,

T (p) End(H1 Hn ) . (3.4)

Remarkably, the transfer matrices for different values of p commute,

[T (p), T (p0 )] = 0 . (3.5)

This follows from the RTT relation (3.2)5 and is the main point of the construction6 . This
commutativity means that they have a common set of eigenvectors.
Moreover, T (p) is also related to the product of S-matrices that we want to diagonalize.
To show this we need an extra property

S12 (p, p) = P12 (3.6)

which holds in many theories including all examples we consider below. Then

T (p1 ) = Tra Pa1 Sa2 (p1 , p2 ) . . . San (p1 , pn ) (3.7)

= Tra S12 (p1 , p2 ) . . . S1n (p1 , pn )Pa1
= S12 (p1 , p2 ) . . . S1n (p1 , pn )

and the result is exactly the operator in the r.h.s. of the periodicity condition. We have
used
Pa1 Sai = S1i Pa1 , Tra Pab = Ib (3.8)

2
Pab =1 (3.9)

Pab = Pba (3.10)

Pa1 Pa2 = Pa2 P12 = P12 Pa1 (3.11)

5
the name RTT relation is due to the fact that in the literature usually one has the R-matrix in place
of the S-matrix
6
Here is one way to explain the construction informally. We want to build a commuting set of operators
on the Hilbert space H, and for this we uplift it to H Ha Hb , then on that space we have Ta and Tb
which almost commute up to multiplication by S-matrices as in (3.2). Then the operators obtained from
Ta , Tb by tracing over the auxiliary space will really commute with each other on the physical Hilbert space.

8
 i2

S
i1 i1

i2

Figure 2: Graphical representation of the S-matrix. The S-matrix S12 acting in H1 H2 is

shown as an intersection of two lines. Each of the two lines corresponds to one of the two spaces
H1 , H2 . The four ends of the lines correspond to the four indices of the S-matrix.

j1 jn

a a

j1 jn

Figure 3: Graphical representation of the monodromy matrix Ta . The monodromy matrix

is a product of several S-matrices. The horizontal line corresponds to the auxiliary space Ha , while
vertical lines are associated with the physical spaces H1 , . . . , Hn .

Using cyclicity of the trace we can also show that for any k the transfer matrix T (pk ) gives
the operator that we want to diagonalize,

T (pk ) = Sk,k+1 Sk,k+2 . . . Sk,n Sk,1 . . . Sk,k1 (3.12)

So our goal is to solve the eigenvalue problem for T (p),

T (p)A = (p)A (3.13)

and then the periodicity condition (2.12) reduces to just an algebraic equation,

eipk L = (pk ) ! (3.14)

In the next sections we will attack the problem of diagonalizing the transfer matrix T (p) for
several models.
Let us mention that a pictorial representation is often used for the transfer matrix and
the S-matrix, as already discussed in part in other chapters of this collection [12, 1]. To

9
 S13 S13
S12 S12
=
S23 S23

Figure 4: Graphical representation of the Yang-Baxter equation (2.11). The equation

means that we can move the vertical line across the intersection point of the two other lines.

a Sab Sab a
=
b b

1 n 1 n

Figure 5: The RTT relation (3.2). The two horizontal lines correspond to auxiliary spaces
Ha , Hb , while the vertical lines correspond to physical spaces H1 , . . . , Hn (labels on the picture next
to the lines show which spaces the lines are associated with). Due to the Yang-Baxter equation we
can move all the vertical lines one by one to the other side of the intersection, leading to the RTT
relation (3.2).

10
understand how it works, let us write the S-matrix S12 in index notation. As it acts on
the tensor product of two spaces H1 H2 , its index structure is Sii01ii02 where the first upper
1 2
index and the first lower index correspond to the H1 space and the second pair of indices
corresponds to the H2 space. As shown on Fig. 2, we can represent this structure a pair of
intersecting lines, with the ends of one line corresponding to the H1 space and the ends of
the other line to the H2 space. The convenience of this notation is that contraction of indices
in any expression with a product of operators such as (3.1) is simply represented as joining
the corresponding lines. For example the monodromy matrix can be depicted as shown on
Fig. 3. The Yang-Baxter equation in graphical form is shown on Fig. 4. Notice that using
the pictorial representation one can easily prove the RTT relation using the Yang-Baxter
equation, as shown on Fig. 5.

4 Algebraic Bethe ansatz: solving the SU(2) chiral

Gross-Neveu model
We will now specialize to a concrete example: the chiral SU (2) Gross-Neveu model (in section
6 we will study the more complicated SU (3) case). This theory has already been discussed
and introduced in the part of this collection about S-matrices [1]. We will parameterize the
particles energy and momentum7 in terms of the rapidity u,
u u
E = m cosh , p = m sinh (4.1)
2 2
In the SU (2) chiral Gross-Neveu model there are effectively two massive particles, so in
terms of the previous notation K = 2 and the S-matrix acts in C2 C2 . Explicitly, the
S-matrix reads
1
S12 (p1 , p2 ) = S f f (u1 u2 )R12 (u1 u2 ) (4.2)

u
(1 )( 12 + u
)
S f f (u) = 4i
u
4i
(4.3)
(1 + 4i
)( 12 u
4i
)
where
1  
R12 (u) = uI + 2iP12 . (4.4)
u + 2i
This is the usual R-matrix of the rational type, satisfying the Yang-Baxter equation

R12 (u v)R1a (u)R2a (v) = R2a (v)R1a (u)R12 (u v) (4.5)

as well as
R12 (u)R12 (u) = 1 . (4.6)
7
here we choose an unconventional prefactor in front of u for a better match with the usual spin chain
notation

11
As we discussed we need to diagonalize the transfer matrix built as a product of these S-
matrices with a trace over the auxiliary space. For that the important thing is the matrix
structure, so let us drop for some time the prefactor S f f (u), and then
Nf Nf
Y Y
1
T (u) = Tra Rai (u ui ) = Tra Rai (ui u) (4.7)
i=1 i=1

where Nf is the number of particles on the circle.

For some time now we will concentrate on the problem of diagonalizing this transfer
matrix. Later we will return to the Gross-Neveu model and assembling all the ingredients
we will write equations for its spectrum.
As we saw, the operators T (u) commute for different values of u. As the transfer matrix
in our case is a polynomial, its operator coefficients in front of the powers of u all commute
with each other. In particular, if we set all ui to zero, a particular combination of these
operators gives the famous XXX spin chain Hamiltonian a system of Nf interacting spins
s = 1/2. It is defined as
Nf
1X
H = (~i~i+1 1) (4.8)
4 i=1
where i are the Pauli matrices acting on the i-th site and we identify i = Nf + 1 and i = 1.
The same operator is expressed via the transfer matrix as

d 
H = i log T (u) . (4.9)
du u=0

The Hamiltonian can be also written in terms of permutation operators,

Nf
1 X 
H = 1 Pk,k+1 . (4.10)
2 k=1

The importance of this Hamiltonian was first recognized in condensed matter applications
where it serves as a model for a ferromagnetic material. In the AdS/CFT context the XXX
Hamiltonian is also directly relevant as it describes the leading order anomalous dimensions
for operators in a simple subsector of the N = 4 supersymmetric Yang-Mills theory (see the
review [13]).
If we keep ui nonzero, the relation (4.9) gives the Hamiltonian for the spin chain for
which ui are called the inhomogenieties. The fact that all coefficients of T (u) commute with
the Hamiltonian, i.e. represent a large number of conservation laws, is a strong sign for the
integrability of this model and ultimatley leads to its solution.
To construct the eigenstates of T (u) we will use operators originating from the mon-
odromy matrix Ta (u). We can write the monodromy matrix explicitly as a 2 2 matrix in
auxiliary space,  
A(u) B(u)
Ta (u) = (4.11)
C(u) D(u)

12
where the entries act on the physical space,
A(u), B(u), C(u), D(u) End(H1 Hn ). (4.12)
In this notation we have
T (u) = A(u) + D(u) . (4.13)
The entries of T (u) satisfy important commutation relations following from the identity (3.2)
which takes the form8
R12 (v u)T1 (u)T2 (v) = T2 (v)T1 (u)R12 (v u) . (4.14)
In particular,
[B(u), B(v)] = 0 (4.15)

v w + 2i 2i
A(v)B(w) = B(w)A(v) B(v)A(w) (4.16)
vw vw
w v + 2i 2i
D(v)B(w) = B(w)D(v) + B(v)D(w) (4.17)
wv vw
Exercise: Derive these relations.
Let us introduce the vacuum state |0i, in which all spins are up9 :
   
1 1
|0i = | . . . i = ... (4.18)
0 0

Using the explicit form of the R-matrix we find that the vacuum is an eigenstate of T and
Y uj u
A(u)|0i = |0i, D(u)|0i = |0i, C(u)|0i = 0 . (4.19)
j
uj u + 2i

The idea is to view B as a creation operator and build the transfer matrix eigenstates as
|w1 , . . . , wNa i = B(w1 )B(w2 ) . . . B(wNa )|0i (4.20)
where w1 , w2 , . . . are parameters known as Bethe roots. Lets see how T = A + D acts on
this state. If the second term in the r.h.s. of (4.16) was absent, we could just commute A(u)
through all Bs until it hits the vacuum which is its eigenstate. Similarly if there was no
second term in the r.h.s. of (4.17) we could commute D(u) through all Bs and then again
arrive at the vacuum. However due to the presence of these extra terms in (4.16), (4.17) we
will find extra unwanted contributions, and the full result is
Y u wj + 2i
A(u)|w1 , . . . , wNa i = B(w1 ) . . . B(wNa )|0i (4.21)
j
u w j

X
+ Mj B(u)B(w1 ) . . . B(wj1 )B(wj+1 ) . . . B(wNa )|0i
j

8
the R-matrix has argument v u as Ta is built from R1
9
Sometimes this state is called pseudovacuum rather than vacuum, since e.g. the ground state in this
model is actually degenerate, for instance the state with all spins down has the same energy.

13
 Y wj u + 2i Y uk u
D(u)|w1 , . . . , wNa i = B(w1 ) . . . B(wNa )|0i (4.22)
j
w j u k
u k u + 2i
X
+ Mj B(u)B(w1 ) . . . B(wj1 )B(wj+1 ) . . . B(wNa )|0i
j

The unwanted terms in these two equations are those that include Mj , Mj . Explicitly, by
Mj we denote the coefficient of the term where A was commuted with B using the second
term in the commutation relation (4.16), i.e. the one which exchanges the arguments (and
similarly for Mj ). For instance, it is easy to compute
2i Y w1 wk + 2i
M1 = (4.23)
u w1 k6=1 w1 wk

2i Y um w1 Y wk w1 + 2i
M1 = (4.24)
u w1 m um w1 + 2i k6=1 wk w1

And since all Bs commute, Mj , Mj are trivial generalizations of these expressions,

2i Y wj wk + 2i
Mj = (4.25)
u wj k6=j wj wk

2i Y um wj Y wk wj + 2i
Mj = (4.26)
u wj m um wj + 2i k6=j wk wj
We see that for any j we can cancel the unwanted terms against each other! This will happen
if Y wj wk + 2i Y um wj Y wk wj + 2i
= (4.27)
w j wk m
um w j + 2i w k wj
k6=j k6=j

Its convenient to relabel the Bethe roots as wk = wk + i, then dropping the tilde we get
Y wj um + i Y wj wk + 2i
= . (4.28)
m
w j u m i k6=j
w j wk 2i

The equations (4.28) are known as Bethe equations for the XXX chain, and they are one of
the key results of this section. The eigenvalue of the transfer matrix then reads
Y u wm + i Y u uk Y u wm 3i
SU (2) (u) = + (4.29)
m
u w m i k
u uk 2i m
u w m i
This is the main outcome of our discussion. In particular, we can extract from it the
eigenvalue of the XXX Hamiltonian (4.8). To do this we use the relation (4.9) which links
the Hamiltonian to the transfer matrix in which all ui should be set to zero. We find the
simple result10
X 1
E=2 . (4.30)
j
wj2 + 1
10
In the spin chain literature the Bethe roots wk and the inhomogenieties um are usually rescaled by a factor
Q w um +i/2 Q wj wk +i
of two compared to our notation so that the Bethe equations would read m wjj um i/2 = k6=j wj wk i
and the energy of the XXX Hamiltonian would be E = 21 j w2 +1/4 1
P
j

14
Also, it is important that the XXX spin chain Hamiltonian has an SU (2) symmetry, which
will be discussed in more detail in section 5.1, together with its implications for the structure
of the eigenstates and eigenvalues.
Let us finally underline that the Bethe equations we have just obtained give the spectrum
of the transfer matrix and of the spin chain Hamiltonian exactly at any length of the chain,
i.e. any Nf . This is in contrast with the Bethe ansatz for the 2d field theory we started
with, which captures only powerlike and not exponential corrections in the volume L.
One can ask whether these equations provide all eigenstates and eigenvalues of the trans-
fer matrix, i.e. whether the algebraic Bethe ansatz solution is complete. While the answer
is certainly expected to be positive, a fully rigorous proof has not been found so far (the
proofs which are available rely on some conjectures, see [5] for an initial discussion and also
[14] as well as references therein for a more recent summary). A related issue is that the
Bethe ansatz could have some singular solutions which do not correspond to eigenstates.
This issue as well as the question of completeness become more tractable if one introduces
twisted boundary conditions for the spinj chain (see e.g. the recent discussion in [15] and
references therein). It is also expected that it is sufficient to consider only solutions where
the Bethe roots are pairwise distinct in order to get all eigenvalues of the Hamiltonian.
Let us mention that there is a shortcut to the Bethe equations for our transfer matrix.
Suppose we forget about the unwanted terms in the commutation relations, then we would
still arrive at the same expression for the eigenvalue (4.29). This eigenvalue however appears
to have poles when u = wj . The poles cannot be really there as the transfer matrix is
not singular at these points11 . Demanding that the residue of the poles vanish we obtain
equations on the roots wk which are nothing but the Bethe ansatz equations (4.28) ! This
is not a rigorous derivation of the Bethe equations, but this trick is very useful. We will
apply it in section 6 for the SU (3) case.
Notice that if we had only the XXX Hamiltonian it would be very hard to guess the
transfer matrix and the algebraic Bethe ansatz procedure! Historically the XXX chain was
solved first by another method which we will discuss in the next section.

4.1 Coordinate Bethe ansatz for the XXX Hamiltonian

The exact solution of the XXX chain was originally obtained by a more intuitive method
known as the coordinate Bethe ansatz [4] (see e.g. [5] for a review and more details on
this model). Let us forget about the transfer matrix and consider just the XXX chain
Hamiltonian,12
L
1 X 
H = 1 Pk,k+1 (4.31)
2 k=1
where as usual we identify the (L + 1)-th and the 1st sites (we consider the case when there
are no inhomogenieties, ui = 0). The method involves making a clever guess (ansatz ) for the
11
its only singular at u = uj + 2i which corresponds to the pole in our R-matrix
12
In this subsection as well as in section 5 below we denote the length of the chain as L (to simplify
notation) rather than as Nf which was used in the discussion above.

15
explicit form of the eigenstates. We start with the ground state in which all spins are up,

H| . . . i = 0 . (4.32)

Let us look for the first excited state as a combination of terms

|ni = | . . . . . . i (4.33)

where one spin at the nth site is flipped. Writing

X
|i = eipn |ni , (4.34)
n

and noting also the periodic boundary conditions n n + L we find that its an eigenstate
if
eipL = 1 . (4.35)
The corresponding energy is
2
E(w) = , (4.36)
w2 +1
where we parameterize the momentum as
w+i
eip = . (4.37)
wi
So it is natural to understand |i as a 1-particle state, and the momentum of the particle
is quantized according to (4.35). Also, notice that the energy (4.36) matches the general
formula (4.30) for the case with only one root wj .
Let us further write a two-particle state as
X
|i = (n, m)|n, mi , (4.38)
1n<mL

where |n, mi is the state with nth and mth spins flipped. We make an ansatz for the
wavefunction as
(n, m) = eip1 n+ip2 m + S(p1 , p2 )eip1 m+ip2 n , (4.39)
where the coefficient S is to be understood as a phase acquired by the wavefunction when
the two particles scatter through each other. We find that this will be an eigenstate for
w1 w2 + 2i
S(p1 , p2 ) = , (4.40)
w1 w2 2i
with the eigenvalue E(u1 ) + E(u2 ). What is truly remarkable is that this construction still
works for more than two particles. E.g. for three excitations, denoting
X
|p1 , p2 , p3 i = eip1 n1 +ip2 n2 +ip3 n3 |n1 , n2 , n3 i , (4.41)
1n1 <n2 <n3 L

16
we can write the wavefunction as

|i = |p1 , p2 , p3 i + S12 |p2 , p1 , p3 i + S23 |p1 , p3 , p2 i + S13 S12 |p2 , p3 , p1 i (4.42)

+ S13 S23 |p3 , p1 , p2 i + S12 S13 S23 |p3 , p2 , p1 i

(with Sij = S(pi , pj )) and it is still an eigenstate provided the Bethe equations
 L Y wj wk + 2i
wj + i
= (4.43)
wj i k6=j
wj wk 2i

are satisfied13 . Notice that the wavefunction is built using only the same two-particle scatter-
ing S-matrix, so in this sense multiparticle scattering is reduced to only 2 2 interactions.
This is in full analogy with the factorization of scattering in 2d integrable QFTs. This spin
chain wavefunction is in fact one of the inspirations for writing the QFT wavefunction in the
form that we did before.

4.2 Bethe equations for the SU (2) chiral Gross-Neveu model

Let us get back to the SU (2) Gross-Neveu model. Now we are ready to write the full set of
Bethe equations for its spectrum. To do this we should plug the explicit expression (4.29)
for the transfer matrix eigenvalue into the periodicity condition (3.14). Adding the scalar
factor S f f (u) that we dropped before, we get
Y Y uj wk + i
eipj L S f f (uj um ) = 1 . (4.44)
m
uj wk i
k

These equations should be supplemented by the Bethe equations (4.28) which fix the pa-
rameters wj ,
Y wj um + i Y wj wk + 2i
= . (4.45)
m
wj um i k6=j wj wk 2i

Then the energy is given by X uj

E= m cosh . (4.46)
j
2

We expect this to be the exact result for the energy, up to corrections that are exponentially
small in L. This concludes our solution for the spectrum of the Gross-Neveu model at large
L.
In the next section we will discuss the XXX chain in some more detail, and then in section
6 derive a generalization of these equations to the SU (3) Gross-Neveu model.
13
These equations are of course obtained from (4.28) by setting all ui to zero.

17
5 Exploring the XXX spin chain
The SU (2) XXX spin chain which we already encountered in the previous section is a very
important and widely used model, and deserves a deeper look. In this section we will discuss
several of its features in more detail. We will consider the case without any inhomogeneities
for clarity.

5.1 The Bethe states in-depth

In the previous section we saw that the XXX spin chain Hamiltonian
L
1 X 
H = 1 Pk,k+1 (5.1)
2 k=1

can be diagonalized via algebraic Bethe ansatz and its eigenstates are built as14

|i = B(w1 )B(w2 ) . . . B(wM )| . . . i , (5.2)

where the Bethe roots are determined by

 L M
wj + i Y wj wk + 2i
= . (5.3)
wj i k6=j
w j wk 2i

Importantly, this Hamiltonian commutes with the operators Sx , Sy , Sz giving the total spin
of the system, which are defined as a sum of the individual spins, i.e.

[H, Sx ] = [H, Sy ] = [H, Sz ] = 0, (5.4)

L
X
S = S(i) , = x, y, z . (5.5)
i=1

These operators Sx , Sy , Sz are the generators of the global SU (2) symmetry algebra under
which the Hamiltonian is thus invariant. First, this means that we can choose eigenstates
of H to be eigenstates for Sz as well. Also, if we have an eigenstate we can generate more
eigenstates (with the same energy) by acting on it repeatedly with the spin-lowering or
raising operators S = Sx iSy .
Since there is an SU (2) algebra acting at each site of the chain, the whole 2L -dimensional
Hilbert space is a tensor product of L copies of the fundamental representation of SU (2). It
can be thus decomposed into a direct sum of irreducible representations (irreps) V of the
global SU (2) symmetry,
C2 C2 = V . (5.6)
14
To simplify notation compared to the discussion of the Gross-Neveu model above, in this section we
denote the length of the chain as L and the number of excitations as M .

18
Due to the relations (5.4) above, each V is an invariant subspace for the Hamiltonian and
all states there have the same energy. For instance, when L = 2 we have a system of two
spin-1/2 particles, so the Hilbert space decomposes into a spin-0 and a spin-1 representation.
The full space is spanned by the three states
1
| i, | i, (| i + | i) (5.7)
2
which form the spin-1 irrep of the global SU (2), together with the state
1
(| i | i) (5.8)
2
in the singlet (spin-0) representation. The three states (5.7) all have zero energy like the
ground state (which is one of them), while the singlet state has the energy E = 2.
In each of the representations V there is a highest weight state from which all other
states are obtained by acting with the lowering operator S . These other states are known
as descendants. In our example above with L = 2, the states | i and 12 (| i | i) are
highest weight. In fact, the highest weight states are always precisely the Bethe states of the
algebraic Bethe ansatz (5.2).15
Exercise. What are the solutions of Bethe equations corresponding to the highest weight
states for L = 2 ?
There is also a very handy descripton for the descendants, as the lowering operator is
expressed through the operator B(w) in the limit w ,

B(w) = const wL1 S + . . . , w . (5.9)

Its easy to see that one can add roots wk = to any solution of the Bethe ansatz and the
Bethe equations will still be satisfied. Thus for finite roots the Bethe state (5.2) is a highest
weight state, and if some roots are at infinity it is a descendant. Also, due to

[Sz , B(u)] = B(u) (5.10)

the Bethe state (5.2) is always an eigenstate of Sz , with eigenvalue L/2 M , and we can
think of this state as having M spins flipped from | i to | i.
As we said above the Bethe ansatz is expected to give the complete spectrum for this
model. Also, all highest weight eigenstates can be constructed as Bethe states (5.2). In fact
to get all highest weight states its enough to consider only M L/2 in the Bethe equations
(this is clear from the fact that the eigenvalue of Sz for a Bethe state is L/2 M ).
Let us also mention that the algebraic and coordinate Bethe ansatz solutions both rely on
the existence of a simple reference eigenstate (|0i in our case), on top of which the other states
are constructed. In some integrable models like the anisotropic XYZ chain such a simple
reference state is not known in the generic situation, and one has to use other methods to
solve them (see [16] for some recent discussion).
15
And conversely, the states built as (5.2) with finite wk are highest weight states.

19
 Another curious fact is that the Bethe equations have a potential . More precisely there
exists a function F ({wj }) such that the Bethe equations are obtained from its derivative.
To see this let us take the logarithm of the Bethe equations, finding
J
wj + i X wj wk + 2i
L log = + 2inj , (5.11)
wj i k=1,k6=j wj wk 2i

where the integers nj Z are known as mode numbers16 . Then we can write these equations
for all values of k as a derivative of a single function,17
F
= 2ink , nk Z, k = 1, . . . , M (5.12)
wk
where
M
X
F = L [(wj + i) log(wj + i) (wj i) log(wj i)] (5.13)
j=1

M
X
+ [(wk wj 2i) log(wk wj 2i) (wk wj + 2i) log(wk wj + 2i)] .
k<j

The function F is known as the Yang-Yang function (its analog first appeared in [17]) and
can be generalized to almost any other quantum integrable model. In particular it plays
an important role in the relation between N = 2 supersymmetric gauge theories in four
dimensions and integrable systems [18]. Some of its properties are further discussed in the
lecture course of this collection devoted to Thermodynamic Bethe Ansatz [2].
Let us finally mention for completeness the celebrated Gaudin formula for the norm of
the Bethe states [20, 19]. It can be derived almost solely from the commutation relations
between elements of the transfer matrix. The result is written as18
Y wj Y  2F

M 4
h|i = (2i) 1+ 2
det (5.14)
j
w j + 2i j6=k
(w j w k ) m,n wm wn

where the key part is the determinant involving the Yang-Yang function. There is also
a generalization of this formula, again in determinant form, for a scalar product of two
Bethe states in one of which the uk are off-shell, i.e. do not necessarilly satisfy the Bethe
equations. Finding a compact extension of that formula to a higher rank (e.g. SU (3)) chain
is a famous and longstanding open problem.
16
Sometimes not n but n is called the mode number in the P literature
17
One could alternatively include in F an extra term 2i k nk wk with some specific nk , then the Bethe
equations would read just F/wk = 0 but F would explicitly depend on the specific choice of the mode
numbers nk which are in general different for different states.
18
This formula is valid if the set of Bethe roots is invariant under complex conjugation, i.e. {wj } = {wj }

20
5.2 Spectral curve and finite-gap equations
Let us now discuss some interesting features which emerge in the classical limit of the XXX
chain namely the limit when the number of excitations and the length are very large while
their ratio is finite. We will see that in this limit the Bethe equations reduce to a set of
discontinuity conditions known as the finite-gap equations which are ubiquitous in classical
(rather than quantum) integrable systems. These equations also define a Riemann surface
known as the classical spectral curve, which encodes the conserved charges of the system.
For a particular simple example we will show how to solve these equations and compute the
energy using only various analyticity constraints. For a more detailed discussion we refer the
reader to e.g. [21]19 . Our discussion here complements the description of classical integrable
systems in another part of the present collection [22].
It is convenient to use the Bethe equations in logarithmic form (5.11), i.e.
M
wj + i X wj wk + 2i
L log = + 2inj (5.15)
wj i k=1,k6=j wj wk 2i

In our limit L and M are very large. The roots will scale as wj L, and let us use the
rescaled roots defined as
wj = Lxj , xj 1 (5.16)
then we have
1 2X 1
= + nj . (5.17)
xj L k6=j xj xk
The number of roots is very large and they will get close to each other, so instead of a
discrete set one can describe them as a continous distribution. The roots will form several
cuts in the complex plane. Let us introduce the density of the roots
1X
(x) = (x xj ) (5.18)
L j

and the resolvent Z

1X 1 d()
G(x) = = , (5.19)
L j x xj C x
where C is the union of cuts Ci on which the roots condense, C = Ci . Introducing the
filling fraction
= M/L (5.20)
we have the normalization condition
Z
()d = , (5.21)
C

so that

G(x) = + ..., x. (5.22)
x
19
due to our notation there are various factor of 2 differences with [21]

21
In our new variables the Bethe ansatz equations (5.17) can be written as
Z
d() 1
G(x + i0) + G(x i0) = 2 = nj , x Cj (5.23)
x x
where the dashed integral sign means that we should take the integrals principal value (there
is a singularity at x = ). These relations are known as finite-gap equations20 . The resolvent
has branch cuts formed by the Bethe roots and is thus a multivalued function whose Riemann
surface is known as the classical spectral curve of the model. We derived it from a limiting
case of the Bethe ansatz for the quantum XXX chain, but similar curves arise in various
other situations, e.g. in classical integrable systems and in matrix models.
Let us also further restrict the solution by imposing an extra condition
Y wj + i
=1 (5.24)
j
wj i

where each term in the product is nothing but eipj with pj being the momentum of a single
excitation. This condition (known as the zero-momentum requirement) in fact means that
the Bethe state is invariant under cyclic shifts of the sites. In our limit we can write it as
1X 1
P = m, m Z , (5.25)
L j xj

or Z
()
d = m . (5.26)

Since at small x we have
Z Z
d() d()
G(x) = x + O(x2 ) , (5.27)
C C 2
we can equivalently write
G(0) = m . (5.28)

Let us now consider in detail an example when there is only one cut, whose mode number
we denote as n. We will fix the resolvent purely from analyticity constraints. From the
definition of the resolvent we see that it is an analytic function with 1/x asymptotics at
x and the only singularity being the branch cut formed by the Bethe roots. The
discontinuity on the cut is

G(x + i0) G(x i0) = 2i(x), xC (5.29)

Furthermore from (5.23) we see that the values of the resolvent above and below the cut
sum up to a meromorphic function, thus the cut is of the square root type. Combining all
this we can write p
G(x) = f (x) + g(x) Q(x) (5.30)
20
the word finite refers to the fact that we consider the case where the number of cuts Ci is finite

22
where f, g are meromorphic functions and Q is a polynomial. Since there are only two branch
points, Q has degree 2. From (5.23) we get
 
1 1
f (x) = n . (5.31)
2 x

With this f (x) the resolvent G(x) has an apparent singularity at x = 0 which can only
be compensated by g(x). Recalling the asymptotics of G(x) at large x, the only choice is
g(x) = c/x with some constant c. Let us fix the normalization of Q(x) by choosing its free
coefficient to be 1, i.e. Q(x) = ax2 + bx + 1. Then the remaining constants a, b, c are fixed
from finiteness of G(x) at x = 0 together with (5.22), (5.28). This gives
 
1 1 1p
G(x) = n + (nx)2 + (4m 2n)x + 1 (5.32)
2 x 2x

and
= m/n, (5.33)
so we must have m < n.
Im x

0.10

0.05

Re x
0.02 0.04 0.06 0.08

-0.05

-0.10

Figure 6: The branch cut of the resolvent (5.32) connecting the two branch points (plot generated
from numerical solution of Bethe equations). Notice the bending of the cut.

Its important to understand a subtlety with the choice of branches of the square root
in (5.30). Naively one might think that at x = 0 the square root in (5.32) will be equal to

1 = +1, but then the pole at x = 0 wouldnt cancel. To clarify this let us consider as an
example n = 3, m = 1. Then the branch points are at

1 + 2i 2 1 2i 2
x1 = , x2 = (5.34)
9 9

23
and the cut should connect them. First, strictly speaking the cut is not a straight line in
this case. If we take a small part of the cut approximated by the segment x C, then
L(x)x is the number of Bethe roots inside this part of the cut. This quantity should be
real, so the condition (x)dx R determines how exactly the cut will bend. Qualitatively
it is shown onp Fig. 6. Second, far away from the cut, in particular for large positive x, we
should have Q(x) nx > 0. However if we start from large positive x and go along the
real axis to the point x = 0,pwe will cross the cut. This means that when evaluating G(x) at
x = 0 we will have to take Q(x) = 1, not +1. Now its clear that (5.32) indeed satisfies
all the constraints we discussed21 .
Finally, to extract the energy, notice that in the classical limit
X 2 Z
1 X 1 2 d()
E=2 2
' 2 2
= , (5.35)
j
wj + 1 L j xj L C 2

i.e. the energy is proportional to the linear coefficient of the resolvents Taylor expansion in
(5.27). From (5.32) we thus get

2 2 m(n m)
E= . (5.36)
L
We see that we have found the energy by just imposing the correct analyticity, and not
solving the Bethe equations directly at all!
The spectral curve in this case consists of two Riemann sheets joined by a single cut,
i.e. it is a sphere. Due to this the solution we found is known as a rational solution. The
solution for the case with e.g. two branch cuts, corresponding to a torus, would be called an
elliptic solution.
Let us finally note that the finite-gap equations played an important role in the devolp-
ment of integrability in the AdS/CFT context. In particular, the classical limit of Bethe
equations derived from gauge theory (conceptually similar to the discussion above) matches
the classical spectral curve of the AdS5 S 5 integrable string sigma model, providing a nice
demonstration of the AdS/CFT duality (see [24] and the review [23]).

6 Bethe ansatz for the SU(3) chiral Gross-Neveu model

Now our goal is to study the SU (3) chiral Gross-Neveu model. There are now more particle
types, and we will first discuss only the particles in the fundamental representation of SU (3).
That is, we have three particle flavors and the flavor space for each particle is C3 . Our goal
is to derive Bethe equations for these excitations. As other particles are their bound states,
these Bethe equations are in fact enough to decribe the full spectrum (see e.g. [9]).
21
p In Wolfram Mathematica the proper choice of branch cut in the square root
(nx) 2 + (4m 2n)x + 1 appearing in (5.32) would be given for n = 3, m = 1 by
p p
3i i(x x1 ) i(x x2 ) (this expression however doesnt take the bending of the cut into ac-
count). The extra factors of i ensure that both square root factors have cuts going off to infinity in the
same direction, so in their product the cut will disappear except between the branch points.

24
6.1 Nested Bethe ansatz for the SU(3) chain
The method we will use is a more advanced version of the algebraic Bethe ansatz from section
4, known as the nested algebraic Bethe ansatz. 22 .
It will be more convenient to use a slightly different parameterization of the energy and
momentum compared to what we had (Eq. (4.1)) in the SU (2) case, namely we take23
u u
E = m cosh , p = m sinh . (6.1)
3 3
Then extracting from the S-matrix the scalar prefactor (see e.g. [9] and references therein),
1
S12 (p1 , p2 ) = S su(3) (u1 u2 )R12 (u1 u2 ) (6.2)

u
(1 )( 23 + u
)
S su(3) (u) = 6i
u
6i
, (6.3)
(1 + 6i
)( 23 u
6i
)
we have the R-matrix
1
R(u) = (u + 2iP ) , (6.4)
u + 2i
which has the same form as in the SU (2) case but now acts in C3 C3 . The main goal is to
diagonalize the transfer matrix
 
T (u) = Tra Ta (u) = Tra R1a (u1 u)R2a (u2 u) . . . Rna (un u) . (6.5)

Let us write out its structure in the auxiliary space explicitly, in the following notation:

T00 (u) B1 (u) B2 (u)
Ta (u) = C1 (u) T11 (u) T12 (u) . (6.6)
C2 (u) T21 (u) T22 (u)
The commutation relations between the entries follow from the RTT relation as usual. In
particular,
[B1 (u), B1 (v)] = 0, [B2 (u), B2 (v)] = 0 , (6.7)

vu 2i
B1 (u)B2 (v) = B2 (v)B1 (u) + B1 (v)B2 (u) . (6.8)
v u + 2i v u + 2i
While one can do the calculation in a more abstract way we will use index notation to ensure
full clarity. We will have Greek indices , , . . . and Latin indices a, b, . . . , all of which take
values 1 and 2. In this notation the relations for commuting the T s with the Bs read
v u 2i 2i
T00 (u)B (v) = B (v)T00 (u)+ B (u)T00 (v) (6.9)
vu vu
22
Let us note that the SU (3) spin chain and many other models with higher rank symmetry group can
also be solved by a coordinate rather than algebraic version of the nested Bethe ansatz which also proved
useful in AdS/CFT (see e.g. [25] and references therien).
23
We use this notation so that the R-matrix in (6.4) that we get here is the same as we had before in the
SU (2) case.

25
and
v u + 2i 2i
T0 (u)B (v) = R0 (v u)B (v)T (u)+ B0 (u)T (v) , (6.10)
vu uv
where we marked by red the unwanted terms, i.e. those that later will cancel when we
construct the eigenstate and impose the Bethe equations. Remarkably, the SU (2) R-matrix
which we denote as R(u), appears in these equations. Explicitly its nonzero elements are, as
before,
R11 22
11 (u) = R22 (u) = 1, (6.11)

u 2i
R21 12
21 (u) = R12 (u) = , R21 12
12 (u) = R21 (u) = (6.12)
u + 2i u + 2i

While in the SU (2) case we had B(u) as a creation operator, here we have two candidates
B1 (u) and B2 (u). Let us try to build the eigenvectors as
X
|i = Ba1 (v1 )Ba2 (v2 ) . . . Ban (vn )F a1 a2 ...an |0i (6.13)
{a}

with a1 , a2 , . . . taking the values 1 or 2, and the vacuum is as usual

1 1
|0i = | . . . i = 0 0 .
(6.14)
0 0

Lets assume for now that there are no unwanted (i.e. red) terms in the commutation
relations, and act with T (u) on this state. From T00 we get

T00 (u)|0i = |0i (6.15)

and Y vk u 2i
T00 (u)|i = |i . (6.16)
k
vk u
ThenPlets see what we get acting on our state |i with the rest of the trace of T (u), i.e.
with T . Lets take for a start a state |i with only two excitations,

|i = Ba1 (v1 )Ba2 (v2 )F a1 a2 |0i . (6.17)

Then
v1 u + 2i 1 b1
T (u)|i = Ra1 (v1 u)Bb1 (v1 )T1 (u)Ba2 (v2 )F a1 a2 |0i (6.18)
v1 u
v1 u + 2i v2 u + 2i 1 b1
= Ra1 (v1 u)R21 ba22 (v2 u)
v1 u v2 u
Bb1 (v1 )Bb2 (v2 )T2 (u)F a1 a2 |0i .

26
Finally, the operators T act on the vacuum in a very simple way. That is,
Y uj u
T (u)|0i = |0i . (6.19)
j
uj u + 2i

So we get
v1 u + 2i v2 u + 2i Y uj u
T (u)|i = (6.20)
v1 u v2 u j
uj u + 2i

R12 ba11 (v1 u)R21 ba22 (v2 u)Bb1 (v1 )Bb2 (v2 )F a1 a2 |0i .
In the indices of the R-matrices there is now a clear pattern, so we see that for any number
of excitations as in (6.13) we would get
n n
Y vk u + 2i Y uj u Y
T (u)|i = Bb (vk )|0i (6.21)
k=1
vk u j
uj u + 2i k=1 k

R21 ab11 (v1 u)R23 ba22 (v2 u) . . . R1nabnn (vn u)F a1 a2 ...an .
But this product of R-matrices is the transfer matrix of an SU (2) spin chain! It has free
indices a1 , . . . , an and b1 , . . . , bn , so it acts on the product of vector spaces V1 Vn with
each Vj ' C2 . And F a1 a2 ...an is the set of coordinates of a vector in that tensor product, or
in other words the wavefunction of an SU (2) chain with n sites. The first upper index and
the first lower index of each R-matrix are the indices in the auxiliary space and they are
contracted with the neighbouring R-matrices. So we have a product of R-matrices along the
auxiliary space and we take a trace over this space. This is precisely the transfer matrix of
the SU (2) spin chain on n sites. Notice that v1 , . . . , vn are the inhomogenieties in this chain.
The ak indices are contracted with F , so if F is an eigenstate of this transfer matrix we
see that T (u) will act diagonally on |i which is what we want. So we have reduced the
initial SU (3) problem to an SU (2) one. This is why the approach we are discussing is called
nested algebraic Bethe ansatz.
Of course we already solved the SU (2) spin chain, so we know how to diagonalize this
transfer matrix. States will be created by its off-diagonal element, and are parameterized by
yet another set of Bethe roots, which we call wm . (Its the third one in addition to uj and
vk .) Then we take as F a1 ...an the wavefunction of this state:
F a1 ...an = (| SU(2) eigenstate i)a1 ...an . (6.22)
For example, F 122 is the coefficient of the term | i in the eigenstate of the SU (2) chain.
The new Bethe roots satisfy the SU (2) Bethe equations (4.28) with inhomogenieties set to
be vk :
Y wm vk + i Y wm wm0 + 2i
= . (6.23)
w m vk i 0
w m wm0 2i
k m 6=m

The corresponding eigenvalue of the SU (2) transfer matrix is given by (4.29),

Y u wm + i Y u vk Y u wm 3i
SU (2) (u) = + . (6.24)
m
u w m i k
u vk 2i m
u w m i

27
 Then, we can assemble all our calculations to write the result for the full transfer matrix
eigenvalue,
(T00 (u) + T11 (u) + T22 (u))|0i = (u)|0i (6.25)
where
Y vk u 2i Y vk u + 2i Y uj u
(u) = + SU (2) (u) . (6.26)
k
vk u k
vk u j
uj u + 2i

During this whole discussion we of course ignored the extra terms in the commutation
relations. One can show that they will cancel provided the roots vk satisfy a set of constraints
which are the Bethe equations for this model. Showing this requires some careful and lengthy
work, and we will not do this here. Some guidance for a similar model can be found in [26].
Rather than going into details of this derivation we will follow instead a shortcut which
gives the same Bethe equations. Recall that for SU (2) one way to derive the Bethe equations
is to require cancellation of spurious poles in the transfer matrix eigenvalue. We can do the
same here. We see that (u) appears to have poles when u = vk which in fact should be
absent as the transfer matrix is not singular at those points. Demanding that the residue at
u = vk0 vanishes we find
Y vk vk0 2i Y vk vk0 + 2i Y uj vk0
0 = (2i) + SU (2) (vk0 ) 2i (6.27)
k6=k0
vk vk0 k6=k0
vk vk0 j
uj vk0 + 2i

(note that SU (2) is clearly not singular at this point). Plugging in SU (2) (vk0 ) from (6.24),
where only the first term is nonzero, we get
Y vk vk0 2i Y vk0 wm + i Y uj vk0
0= + . (6.28)
k6=k0
vk vk0 + 2i m
vk0 wm i j uj vk0 + 2i

Shifting vk = vk + i, wm = wm + i and dropping the tildes we get

Y vk vk0 2i Y vk0 wm + i Y uj vk0 i
0= + (6.29)
k6=k0
vk vk0 + 2i m
vk0 wm i j uj vk0 + i

Note that the Bethe equations (6.23) for ws do not change under this shifting. So, we find
Y vk0 uj + i Y vk0 vk + 2i Y vk0 wm i
= , (6.30)
vk 0 uj i vk 0 vk 2i vk 0 wm + i
j k6=k 0 m

while the ws satisfy the same equations as before in (4.28) but with inhomogeneities vk ,
Y wj um + i Y wj wk + 2i
= . (6.31)
m
wj um i k6=j wj wk 2i

The equations above determine the eigenvalue of the SU (3) transfer matrix and in par-
ticular of the SU (3) version of the XXX spin chain Hamiltonian. Like in the SU (2) case this
Hamiltonian is expressed in terms of the transfer matrix as in (4.9)

d 
H = i log T (u) , (6.32)
du u=0

28
and can be written explicitly as
Nf  
X
H = const 1 Pk,k+1 , (6.33)
k=1

the only difference with the SU (2) case being that it now acts in the tensor product of C3
rather than C2 spaces.
Let us also note that for the SU (2) chain we had one type of Bethe roots (wk ) parame-
terizing the eigenstates, while here we have two types vk and wk . In fact one should think
of each Bethe root type as associated to the nodes on the Dynkin diagram of the symmetry
group. In our case, accordingly, SU (2) has a Dynkin diagram with only one node, while for
SU (3) there are two nodes. One can also consider spin chains with higher rank symmetry
groups G beyond the SU (2) and SU (3) cases. Accordingly, instead of C2 or C3 one one
would have at each site of the chain a higher-dimensional complex space. At least for most
of the compact simple Lie groups G, the corresponding spin chain again can be solved by
Bethe ansatz, and the Bethe equations are written in a uniform way in terms of the groups
Cartan matrix as well the representation of the group chosen at each site (see e.g. the review
[9] and references therein). A similar story continues to hold even for super Lie algebras,
in particular for the algebra psu(2, 2|4) which underlies the structure of Bethe equations
describing the spectrum of long operators in N = 4 supersymmetric Yang-Mills theory [3].
Finally we can assemble the equations for the spectrum of the SU (3) chiral Gross-Neveu
model. Using the expression above for the eigenvalue of the SU (3) transfer matrix, we can
write the periodicity condition as
Y Y uj vk + i
eipj L S su(3) (uj um ) = 1 , (6.34)
m
u j vk i
k

together with equations for auxiliary Bethe roots

Y vk0 uj + i Y vk0 vk + 2i Y vk0 wm i
= , (6.35)
vk 0 uj i vk 0 vk 2i vk 0 wm + i
j k6=k 0 m

Y wj um + i Y wj wk + 2i
= . (6.36)
m
wj um i k6=j wj wk 2i

The energies are as usual a sum of single particle energies (notice we use the parameterization
(6.1)) X uj
E= m cosh . (6.37)
j
3

Thus we have completed the solution for the spectrum of the SU (3) Gross-Neveu model in
large volume L.

29
7 Bethe ansatz for the harmonic oscillator
Let us discuss in this last section a completely different setting where Bethe-like equations
also appear. Namely, one can use a kind of Bethe ansatz to get eigenstates of the very
well-studied one-dimensional harmonic oscillator in quantum mechanics (for a more detailed
discussion of this case see e.g. [27]). Thus, we are studying the Schrodinger equation

~2 00
(x) + V (x)(x) = E(x) (7.1)
2m
with the potential
m 2 x2
V (x) = . (7.2)
2
Let us introduce the so-called quasimomentum

~ 0 (x)
p(x) = , (7.3)
i (x)

in terms of which the Schrodinger equation takes the form

p2 i~p0 = 2m(E V ) . (7.4)

As (x) is regular, the only singularities of p(x) are at the zeros x = xj of the wavefunction,
where the quasimomentum has simple poles with residue ~i .
In the classical limit, i.e. for highly excited states, we get from (7.4)
p
p ' pcl = 2m(E V ) (7.5)

so now the quasimomentum has a branch cut. This cut can be understood as a collection
of poles at xj , which become denser and denser, eventually forming a smooth distribution
giving rise to a cut. The situation is very similar to the classical spectral curve of the XXX
model we discussed in section 5.2
Let us now see that for any state (not necessarily semiclassical) we can derive a simple
set of equations fixing the positions of these poles. At large x we have

p(x) = imx + O(1/x) , (7.6)

so we can write
N
~X 1
p(x) = imx + . (7.7)
i j=1 x xj

From the large x asymptotics of (7.4) we can already find the spectrum! It reads
 
1
E = ~ N + . (7.8)
2

30
From (7.4) we also get a set of Bethe-like equations for the roots,

~ X 1
xj = . (7.9)
2m k6=j xj xk

They are clearly reminiscent of the usual Bethe ansatz form, with one root in the l.h.s. and
interaction between roots in the r.h.s. As one can expect from the usual form of the oscillator
wavefunctions, the solutions to this equation are the roots of the N -th Hermite polynomial,
r !
2m
HN xj = 0 . (7.10)
~

Knowing the roots xj we can also reconstruct the wavefunction from (7.3), (7.7). Equations
similar to (7.9) frequently arise as limiting cases of the Bethe ansatz equations for other
models, e.g. in the limit of large L and fixed number of excitations in the XXX spin chain.

Acknowledgements
This article arose out of a series of lectures gven by the author at the Young Researchers
Integrability School at Durham University in July 2015. I wish to thank all the organizers
and the other lecturers for their effort in making the event run successfully and creating a
great atmosphere. I would like to also thank the Department of Mathematical Sciences at
Durham University for hospitality. I am particularly grateful to Z. Bajnok, D. Bombardelli,
N. Gromov, A. Sfondrini, S. van Tongeren and A. Torrielli for discussions related to the
material in these notes. I am also grateful to the student participants for their interest
in the school, excellent questions and lots of feedback. My work is supported in part by
funding from the People Programme (Marie Curie Actions) of the European Unions Seventh
Framework Programme FP7/2007-2013/ under REA Grant Agreement No 317089 (GATIS).

References
[1] D. Bombardelli, Lectures on S-matrices and Integrability, in: An integrability primer
for the gauge-gravity correspondence, ed.: A. Cagnazzo, R. Frassek, A. Sfondrini, I.
M. Szecsenyi and S. van Tongeren, J. Phys. A 49 (2016) 323003, doi:10.1088/1751-
8113/49/32/323003 [arXiv:1606.02949]

[2] S. van Tongeren, An integrability primer for the gauge-gravity correspondence: ther-
modynamic Bethe ansatz, in: An integrability primer for the gauge-gravity correspon-
dence, ed.: A. Cagnazzo, R. Frassek, A. Sfondrini, I. M. Szecsenyi and S. van Tongeren,
J. Phys. A 49 (2016) 323005, doi:10.1088/1751-8113/49/32/323005 [arXiv:1606.02951]

[3] N. Beisert et al., Review of AdS/CFT Integrability: An Overview, Lett. Math. Phys.
99 (2012) 3 [arXiv:1012.3982 [hep-th]].

31
 [4] H. Bethe, On the theory of metals. 1. Eigenvalues and eigenfunctions for the linear
atomic chain, Z. Phys. 71 (1931) 205.

[5] L. D. Faddeev, How algebraic Bethe ansatz works for integrable model, hep-
th/9605187.

[6] B. Sutherland, Beautiful Models: 70 Years of Exactly Solved Quantum Many-Body

Problems. World Scientific, 2004.

[7] M. Karbach and G. Muller, Introduction to the Bethe ansatz I, Computers in Physics
11 (1997) 36, arXiv:cond-mat/9809162.

[8] F. H. L. Essler, H. Frahm, F. Goehmann, A. Kluemper, and V. E. Korepin, The One-

Dimensional Hubbard Model, Cambridge University Press (2005).

[9] D. Volin, Quantum integrability and functional equations: Applications to the spectral
problem of AdS/CFT and two-dimensional sigma models, J. Phys. A 44 (2011) 124003
[arXiv:1003.4725 [hep-th]].

[10] S. J. van Tongeren, Integrability of the AdS5 S5 superstring and its deformations,
J. Phys. A 47 (2014) 433001 [arXiv:1310.4854 [hep-th]].

[11] A. B. Zamolodchikov and A. B. Zamolodchikov, Factorized S Matrices in Two-

Dimensions as the Exact Solutions of Certain Relativistic Quantum Field Models,
Annals Phys. 120 (1979) 253.

[12] F. Loebbert, Lectures on Yangian Symmetry, arxiv:1606.02947, in: An integra-

bility primer for the gauge-gravity correspondence, ed.: A. Cagnazzo, R. Frassek,
A. Sfondrini, I. M. Szecsenyi and S. van Tongeren, J. Phys. A 49 (2016) 323002,
doi:10.1088/1751-8113/49/32/323002 [arXiv:1606.02947]

[13] J. A. Minahan, Review of AdS/CFT Integrability, Chapter I.1: Spin Chains in N=4
Super Yang-Mills, Lett. Math. Phys. 99 (2012) 33 doi:10.1007/s11005-011-0522-9
[arXiv:1012.3983 [hep-th]].

[14] W. Hao, R. I. Nepomechie and A. J. Sommese, Completeness of solutions of Bethes

equations, Phys. Rev. E 88 (2013) no.5, 052113 doi:10.1103/PhysRevE.88.052113
[arXiv:1308.4645 [math-ph]].

[15] R. I. Nepomechie and C. Wang, Twisting singular solutions of Bethes equations, J.

Phys. A 47 (2014) no.50, 505004 doi:10.1088/1751-8113/47/50/505004 [arXiv:1409.7382
[math-ph]].

[16] J. Cao, W. L. Yang, S. Cui, K. J. Shi and Y. Wang, Spin- 21 XYZ model re-
visit: General solutions via off-diagonal Bethe ansatz, Nucl. Phys. B 886 (2014) 185
doi:10.1016/j.nuclphysb.2014.06.026 [arXiv:1307.0280 [cond-mat.stat-mech]].

[17] C. N. Yang and C. P. Yang, Thermodynamics of one-dimensional system of bosons

with repulsive delta function interaction, J. Math. Phys. 10 (1969) 1115.

32
[18] N. A. Nekrasov and S. L. Shatashvili, Quantization of Integrable Systems and Four
Dimensional Gauge Theories, arXiv:0908.4052 [hep-th].

[19] V. E. Korepin, Calculation Of Norms Of Bethe Wave Functions, Commun. Math.

Phys. 86 (1982) 391.

[20] M. Gaudin, Journal de Physique 37, 1087 (1976); B. M. McCoy, T. T. Wu, and
M. Gaudin, Phys. Rev. D 23, 417 (1981).

[21] V. A. Kazakov, A. Marshakov, J. A. Minahan and K. Zarembo, Classical/quantum

integrability in AdS/CFT, JHEP 0405 (2004) 024 [hep-th/0402207].

[22] A. Torrielli, Lectures on Classical Integrability, arxiv:1606.02946, in: An integra-

bility primer for the gauge-gravity correspondence, ed.: A. Cagnazzo, R. Frassek,
A. Sfondrini, I. M. Szecsenyi and S. van Tongeren, J. Phys. A 49 (2016) 323001,
doi:10.1088/1751-8113/49/32/323001 [arXiv:1606.02946]

[23] S. Schafer-Nameki, Review of AdS/CFT Integrability, Chapter II.4: The Spec-

tral Curve, Lett. Math. Phys. 99 (2012) 169 doi:10.1007/s11005-011-0525-6
[arXiv:1012.3989 [hep-th]].

[24] N. Beisert, V. A. Kazakov, K. Sakai and K. Zarembo, Complete spectrum of long

operators in N=4 SYM at one loop, JHEP 0507 (2005) 030 doi:10.1088/1126-
6708/2005/07/030 [hep-th/0503200].

[25] N. Beisert, The SU(22) dynamic S-matrix, Adv. Theor. Math. Phys. 12 (2008) 948
doi:10.4310/ATMP.2008.v12.n5.a1 [hep-th/0511082].

[26] F. H. L. Essler and V. E. Korepin, Higher conservation laws and algebraic Bethe
ansatze for the supersymmetric t-J model, Phys. Rev. B 46 (1992) 9147.

[27] N. Gromov and P. Vieira, The AdS(5) x S**5 superstring quantum spectrum from the
algebraic curve, Nucl. Phys. B 789 (2008) 175 [hep-th/0703191].

33