Prepared for submission to JHEP

On Unitarity of Tree-Level String Amplitudes

arXiv:2201.11575v1 [hep-th] 27 Jan 2022

Nima Arkani-Hamed,1 Lorenz Eberhardt,1 Yu-tin Huang,2,3 Sebastian Mizera1

1 Institute for Advanced Study, Einstein Drive, Princeton, NJ 08540, USA
2 Department of Physics and Center for Theoretical Physics, National Taiwan University,
Taipei 10617, Taiwan
3 Physics Division, National Center for Theoretical Sciences, Taipei 10617, Taiwan
E-mail: [email protected], [email protected], [email protected],
[email protected]

Abstract: Four-particle tree-level scattering amplitudes in string theory are mag-

ically consistent with unitarity, reflected in the non-trivial fact that beneath the
critical dimension, the residues of the amplitudes on massive poles can be expanded
in partial waves with all positive coefficients. While this follows (rather indirectly)
from the no-ghost theorem, the simplicity of the statement and its fundamental im-
portance for the physical consistency of string theory begs for a more direct and
elementary understanding. In this note we take a step in this direction by present-
ing a new expression for the partial wave coefficients of string amplitudes, given by
surprisingly simple double/triple contour integrals for open/closed strings. This rep-
resentation allows us to directly prove unitarity of all superstring theories in D 6 6
spacetime dimensions, and can also be used to determine various asymptotics of the
partial waves at large mass levels.
Contents
1 Introduction 2

2 Unitarity of the superstring 9

3 Double-contour representation 14

4 Applications 18
4.1 Manifest unitarity in D 6 6 dimensions 18
4.2 Unitarity in D = 10 for low spin 20
4.3 Asymptotics 21
4.3.1 Fixed spin 21
4.3.2 Regge trajectories 23

A Some positive functions 25

A.1 Simple functions 25
A.2 Generating function of the D = 10, j = 0 coefficients 27

B Single-variable generating function 28

C Derivation of the triple-contour representation 30

D Some coefficients on the Regge trajectories 36

1 Introduction
The four-point tree-level amplitude in string theory is a famously miraculous object
[1–3]. Consider for instance the color-ordered four-tachyon amplitude of open bosonic
strings, where (with α0 = 1)

Γ(−s−1)Γ(−t−1)
Aopen,bos (s, t) = . (1.1)
Γ(2+u)

The above expression satisfies all the stringent consistency constraints on amplitudes
in one shot. It is consistent with causality, by being appropriately bounded by a
power of s for large centre-of-mass energy s, working in the Regge limit at fixed
momentum transfer t. It is famously UV complete, being exponentially soft for
physical Lorentzian kinematics (stu > 0 with s+t+u = −4), at large energy and
fixed angle.

–2–
 But by far the most miraculous property of the amplitudes is the way they are
consistent with unitarity. Consider the residue of Aopen,bos (s, t) as we approach a
1
pole at s → n. We have Aopen,bos (s, t) → s−n × Rnopen,bos (t) with Rnopen,bos (t) =
(2+t)(3+t)···(n+2+t)
(n+1)!
. As usual, we can express this result in terms of the scattering
angle x = cos θ with t = (s+4)(x−1)/2 = (n+4)(x−1)/2, so that the “residue
polynomial” Rnopen,bos (x) is
 n+1 n+1
Y  n−2i+2 
open,bos 1 n+4
Rn (x) = x− . (1.2)
(n+1)! 2 i=1
n+4

Unitarity demands that Rnopen,bos (x) can be expanded in Gegenbauer polynomials

(D) P D (D)
Gj (x) with all positive or zero coefficients: Rnopen,bos (x) = j Bn,j Gj (x) with
D D
Bn,j > 0 and spacetime dimension D. The positivity of the Bn,j ’s follows from
the fact that they are the mod-square of three-particle amplitude couplings to the
internal states of spin-j, or equivalently, represent the cross-section for the resonant
production of the intermediate spin-j states. Here the first interesting level is n = 1,
where we find
      
8 open,bos 1 1 2 1 1 1
R (x) = x− x+ = x− + − . (1.3)
25 n=1 5 5 D−1 D−1 25
1
The spin-2 Gegenbauer polynomial in D dimensions is proportional to (x2 − D−1 ),
so this shows that we are exchanging a spin-2 state with a positive norm. But
the remaining constant term indicates the presence of a spin-0 state with the norm
1
− 1 . This gives a negative-norm state for D > 26 spacetime dimensions. In this
D−1 25
way, the four-point tree-level amplitude—whose form is completely independent of
spacetime dimensionality—knows about the critical dimension, through unitarity.
The story is much more interesting for the scattering of spinning states. For the
four-gluon amplitude in type-I superstrings,
Γ(−s)Γ(−t)
AI (s, t) = F 4 AI (s, t) = F 4 . (1.4)
Γ(1+u)
Here, F 4 is a famous permutation-invariant polynomial in the field strengths that ap-
4
pears in the tree-level Yang–Mills amplitude: AYM = Fst . Unitarity at the first mas-
sive level is already non-trivial: as the residue is simply F 4 itself, the permutation-
invariant polynomial must be written as a positive sum of s-channel exchanges. With
x = p1 −p2 and y = p3 −p4 , we express
F 4 (i , x, y) = hx, 1 , 2 |O|y, 3 , 4 i , (1.5)
where |x, 1 , 2 i represents the “Hilbert space” of states with x, 1 , 2 . From these
states we can build irreps of SO(D−1), on which O can be expanded. We find that
in this form, F 4 is uniquely given as a combination of spin-0, -2 and a 3 form:
 
9
O= −1 · · + + . (1.6)
D−1

–3–
Positivity of the expansion now requires D 6 10. Note that this massive spectrum is
exactly the bosonic content of 11-dimensional supergravity, as it must from the view-
point of supersymmetry, i.e., massive irreps can be obtained from massless ones in
one dimension higher. Thus, in a precise sense, the tree-level amplitude of Yang–Mills
knows about the critical dimension of type-I superstring as well as supersymmetry!
Let us also mention that the exchanged three fields correspond exactly to the
field content expected from the worldsheet description of type-I strings. These fields
correspond to the first massive level. Letting ∂X µ and ψ µ be the free bosons and
fermions respectively, the unique two bosonic physical states in string theory at this
mass-level read

≡ µν ∂X µ ψ ν eip·X , ≡ µνλ ψ µ ψ ν ψ λ eip·X , (1.7)

where µν and µνλ are the transverse traceless symmetric and antisymmetric polar-
ization vectors respectively, i.e., pµ µν = 0 and pµ µνλ = 0. The scalar state is absent
in ten-dimensional string theory, as predicted from the coefficient in (1.6). In the
presence of an internal compactification manifold, one can generically write down
one more scalar by choosing (the coefficients are chosen in order to ensure a traceless
transverse polarization)
D−1
µν = ηab + pa pb − gij , (1.8)
D − 10
where a, b = 0, 1, . . . , D denote the flat directions and i, j are the indices of the
internal manifold with metric gij . We can identify this additional scalar with the
scalar appearing in the expansion (1.6) for D < 10.
1
Going beyond the first level we have AI (s, t) → s−n F 4 × RnI (t), where RnI (t) =
(1+t)(2+t)···(n−1+t)
n!
. Once again, in terms of the scattering angle it is given as:
n−1  
1  n n−1 Y n−2i
RnI (x) = x− . (1.9)
n! 2 i=1
n
P D (D)
Expanding on the Gegenbauer polynomials, RnI (x) = j Bn,j Gj (x), this can now
be combined with the expansion in eq. (1.6) to yield a “diagonalized” representation
of the residue:
F 4 RnI = hx, 1 , 2 |On |y, 3 , 4 i , (1.10)
where
  
X
D 9
On = Bn,j −1 · ⊗j · ⊗j + ⊗j ⊗j + ⊗j ⊗j .
j
D−1
(1.11)
The diagonalized is in quotes, since the states are constructed from tensor products
of either the spin-0, 2 or 3-form, with a symmetric spin-j irrep, i.e., the states are

–4–
 j=0 j=1 j=2 j=3 j=4
n=1 1
1
n=2 0 2(D−3)
10−D 3
n=3 24(D−1)
0 4(D−1)(D−3)
11−D 2
n=4 0 12(D+1)(D−3)
0 (D2 −1)(D−3)
9D2 −250D+1616 25(12−D) 125
n=5 1920(D2 −1)
0 96(D−1)(D2 −9)
0 16(D2 −1)(D2 −9)

D for the type-I amplitude with low n and j. Note that all of them
Table 1. Values of Bn,j
D indicates violation of unitarity for D > 10.
are non-negative for D 6 10 and, e.g., B3,0

not necessary orthogonal, see Section 2 for details. Nevertheless, if the coefficients
D
Bn,j are positive, the operator On is manifestly positive, establishing unitarity. We
already know that unitarity for the first massive level requires D 6 10. Unitarity can
D
be established if we can show that Bn,j > 0 for D 6 10. This is already non-trivial
for level n = 3,
      
8 I 1 1 2 1 1 1
R (x) = x− x+ = x− + − . (1.12)
3 n=3 3 3 D−1 D−1 9
D
Once again, we see that for B3,0 > 0 requires D 6 10.
The remarkable claim is that this pattern holds for all higher levels n: the residue
polynomial can be expanded as
X (D)
RnI (x) = D
Bn,j Gj (x), where D
Bn,j > 0 for D 6 10 . (1.13)
j

D
The coefficients Bn,j for the first few levels are given in Table 1, and can be manually
verified to be positive for D 6 10. Note also the clear pattern of zeroes. These zeroes
follow from the fact that Rn (x) is even in x for odd n and odd in x for even n. Thus
D
only even spins are exchanged for odd n and vice-versa, so that Bn,j vanishes when
n+j is even.
Note also that these fundamental positivity statements for open strings directly
prove that the residue polynomials for closed strings also have a positive expan-
sion. For example consider type-II superstring, heterotic Yang–Mills [4]1 and closed
1
Recall that the heterotic string contains the bosonic string in R1,9 × Γ16 as the left-moving part
and the superstring in R1,9 as the right moving part. The momenta on the bosonic side is now split
into 10 continuous, and 16 discrete momenta associated with the torus Γ16 . Here we consider four
gluons whose color indices are root vectors, and arranging the discrete momenta Ki , with Ki2 = 1,
satisfying

S = −(K1 +K2 )2 = 0, T = −(K1 +K4 )2 = 0, U = −(K1 +K3 )2 = −4 , (1.14)

such that in the low-energy limit one directly lands on the Yang–Mills amplitude with 1234 ordering.

–5–
tachyon amplitude are given by:

Γ(−s)Γ(−t)Γ(−u)
AII (s, t) = R4 AII (s, t) = R4 ,
Γ(1+s)Γ(1+t)Γ(1+u)
Γ(−s−1)Γ(−t−1)Γ(−u+3)
AHet (s, t) = F 4 AHet (s, t) = F 4 , (1.15)
Γ(1+s)Γ(1+t)Γ(1+u)
Γ(−1−s)Γ(−1−t)Γ(−1−u)
Aclosed,bos (s, t) = − ,
Γ(2+s)Γ(2+t)Γ(2+u)

where R4 = (F 4 )2 . The reason is simple: the residue for the type-II closed string
is R4 RnII (x) = (F 4 RnI (x))2 , and for the closed bosonic string is the square of the
open bosonic string Rnclosed,bos (x) = (Rnopen,bos (x))2 , while the residue for the heterotic
string, F 4 Rnhet (x), can be written as:

1

 (x+3)(x+5) n = 1,
8



1

het x(x+1)(x+2)(x+3) n = 2, (1.16)
Rn (x) =
 12
(1+x)(n2 x(x+2)+n2 −4)(n(1+x)+4) I


 open,bos
Rn (x)Rn−4 (x) n > 2.


16(n−2)(n−1)(n+1)


For n > 2, it is given by a product of type-I and open bosonic string residue with
a completely positive function (a polynomial in x with positive coefficients which
is guaranteed to have a positive Gegenbauer expansion in any dimension). But it
is obvious that if A(x) and B(x) have a positive expansion on the Gegenbauers, so
does the product A(x)B(x). Indeed, this simply follows from Gj1 (x)Gj2 (x) having
a positive expansion, since spin-j1 × spin-j2 representations of O(D − 1) obviously
decompose as a positive sum of irreps of various spins. Thus, given that Rn (x) has
a positive expansion for the open strings, the products associated with the closed
strings also have positive expansions. We should also notice that unitarity in higher
dimensions immediately implies unitarity in lower dimensions for essentially the same
(D) (D0 )
reason. Gj (x) decomposes into a positive sum of Gj 0 (x) for D0 < D according to
the branching rules of O(D − 1) into O(D0 − 1). This is also physically obvious since
we may always compactify a unitary theory in D spacetime dimensions to obtain a
unitary theory in D0 < D dimensions.
The Gegenbauer positivity of Rn (x) is a strikingly elementary statement. There
is an indirect proof of this positivity, via the famous no-ghost theorem [5]; a general
discussion of unitarity in closed string perturbation theory is given in [6]. On-shell
perspective on unitarity was previously explored in the context of high-energy expan-
sion of partial waves [7] and Hankel-matrix constraints on the Wilson coefficients [8].
There is also an interesting sort of converse to this statement, where the assumption
of positivity for the Gegenbauer expansion of four-particle amplitudes, for any the-
ory with an infinite tower of massive states, implies that the amplitude asymptotes

–6–
 0.0003

0.0002

0.0001

- 1.0 - 0.5 0.5 1.0

- 0.0001

- 0.0002

I
Figure 1. The residue coefficient Rn=21 (x) as a function of the scattering angle x = cos θ.

to the string amplitudes in the (unphysical) high-energy regime where s, t are both
large and negative [9].
But the basic statement of positivity is so simple, and of such fundamental
importance to the consistency of string theory, that it cries out for a more elementary
and direct understanding on the level of the amplitudes. Indeed, some extreme cases
are easy to understand. Most trivially, RnI (x) = n!1 ( n2 )n−1 xn−1 + . . . ; this tells us the
highest spin at level n is jmax = n−1, and since the coefficient of xn−1 is positive,
also establishes the coefficient of the spin-(n−1) Gegenbauer is positive.
D
For general spin, the coefficient Bn,j can be extracted using the orthogonality of
Gegenbauer polynomials as
Z 1
D−4 (D)
D
Bn,j ∝ dx (1−x2 ) 2 Gj (x) RnI (x) , (1.17)
−1

where we suppressed the (positive) normalization factor coming from the normal-
ization of the Gegenbauer polynomials. It is easy to see that for large n, RnI (x) is
peaked around x = ±1, and is crushed away from the ends of the interval, as in
D
Figure 1. This means that for fixed j and large n, the integral for Bn,j is dominated
D
by the regions near x = ±1, where the contribution to Bn,j is manifestly positive, so
D
it is clear that for fixed j and large enough n, Bn,j is positive.
D
But for general n and j, it is devilishly difficult to establish the positivity of Bn,j
D
by brute force. Almost all the “obvious” ways of extracting Bn,j are given as a sum
of terms with alternating signs, with no evident reason for positivity. This is ironic,
D
because it can be seen experimentally that the actual values of Bn,j rise from a small
√
positive value at j = 0, to a maximum near j ∼ 2 n, followed by a faster-than
D
exponential decrease at larger j: it is hardest to prove Bn,j is positive exactly where
it has the largest positive magnitude! This is illustrated in Figure 2. The maximum

–7–
 √
at j ∼ 2 n is reminiscent of a similar peak in the distribution of spins contributing
at a given mass level [10, 11].
In this note, we will revisit this basic question, by finding a remarkably simple
D D
new expression for the residues Bn,j as a double-contour integral. Recall that Bn,j =0
2
when n+j is even; instead for n+j odd we have, up to a positive constant

(v − u)j e−a(u+v)
I I
D du dv
Bn,j ∝ . (1.18)
u=0 2πi v=0 2πi (uv)j+ D−2
2 (ev − eu )n+2a

Here a = 0 and 1 labels the type-I and bosonic open string amplitude respectively.
D
We prove this formula in Section 3. Using this expression, the positivity of Bn,j in the
type I case for all spacetime dimensions D 6 6 follows easily, as shown in Section 4.1.
For 6 < D 6 10, the positivity is not as obviously manifest, though for any fixed spin
j can also be established for all n, with a finite amount of work. This is exemplified
in Section 4.2. Similar statements apply for the open bosonic string, where we can
prove unitarity in the range D 6 10, while unitarity in the range 10 < D 6 26 can
only be established for each fixed spin separately. The above expression also allows
D
us to quantitatively analyze the asymptotic limits of Bn,j , for fixed j at large n, and
also for fixed Regge trajectories, taking n large while keeping ∆ = n−j fixed, see
Section 4.3. These asymptotics are also indicated in Figure 2.
Quite apart from its use in proving positivity and understanding the asymp-
D
totics, we find this double-contour integral representation of Bn,j to be quite intrigu-
ing. While the string amplitudes themselves are clearly beautifully canonical objects,
there was no a priori reason to expect any nice structure in the partial wave expan-
sion of their residues! But such a structure does exist, and begs for a more direct
understanding than the straightforward but not especially transparent derivation we
will provide. Indeed as we will see, the most striking feature of this expression — its
evident symmetry between the u, v variables — will not be apparent in our deriva-
tion at all, only popping out as a mysterious surprise at the end of the computation.
There are analogous double-contour integral formulas for the open bosonic string,
and triple-contour integrals for closed string residues. But we will leave the explo-
ration of deeper origins for these formulae to future work; in this note we will content
ourselves with deriving these unusual and interesting expressions, and highlighting
their utility in understanding the fascinating positivity of the residues much more
directly than previously possible.
2
This formula as written only works for even D, but most of its consequences that we derive in
this paper also work for odd D.

–8–
 1060

1040

1020

100 200 300 400 500

D=10 for the type I superstring for five different values of n plotted
Figure 2. Coefficients Bn,j
in blue against j on a logarithmic scale. Only coefficients with n+j odd are plotted and
they have been rescaled by ( 4e )n to make the plot more readable. In each case, a peak
√
is reached near j ∼ 2 n, indicated with black dots, in qualitative agreement with the
asymptotic formula for density of states derived in [12]. The asymptotic formulas that are
derived in Section 4.3 for small-j and small-∆ = n−j are both plotted in red.

2 Unitarity of the superstring

Let us begin with the gluon amplitude for the type-I superstring, given in eq. (1.4).
The prefactor F 4 is a permutation invariant polynomial of polarization vectors and
momenta, which also appears as the numerator of the color-ordered 1234 tree-level
4
Yang–Mills amplitude as AYM = Fst , where

F 4 = (Fµν F µν )2 − 4(Fµν F να Fαβ F βµ ). (2.1)

Here it is understood that we expand Fµν = F1,µν + · · · + F4,µν , with Fi,µν = pi,µ i,ν −
pi,ν i,µ , and we extract only terms linear in each of the polarization vectors i,µ .
We will find it most convenient to consider the scattering process in the center

–9–
of mass frame, defined as

(2.2)

where x, y are unit (D−1)-dimensional vectors, and the polarization vectors are
chosen to be space-like i = (0, Ei ), satisfying E1 · x = E2 · x = E3 · y = E4 · y = 0. A
short computation then gives us F 4 in terms of these center-of-mass frame variables:

F 4 (Ei , x, y) = (E1 · E2 )(E3 · E4 ) (x · y)2 − 1

 
(2.3)
+ (E2 · E3 )(E1 · E4 )2 [1+(x · y)] + (E1 · E3 )(E2 · E4 )2 [1−(x · y)]
+ 2 [(E1 · E2 )(E3 · x)(E4 · x)+(E3 · E4 )(E1 · y)(E2 · y)−(E2 · E3 )(E1 · y)(E4 · x)
+(E1 · E3 )(E2 · y)(E4 · x)+(E2 · E4 )(E1 · y)(E3 · x)−(E1 · E4 )(E2 · y)(E3 · x)] .

The n-th level residue can then be written as

F 4 (Ei , x, y) · RnI (x · y) ≡ E1a E2b E3c E4d (F 4 )a,b,c,d (x, y) · RnI (x · y) . (2.4)

We will now introduce a Hilbert space |w, a b i, where w is a unit vector representing
the direction of the in/out scattering states in the center-of-mass frame, and a, b
denote the spin degrees of freedom. These states are orthonormal hw0 , a0 , b0 |w, a, bi =
δS D−2 (w, w0 )δa0 a δb0 b . This continuum of states transform as highly reducible represen-
tations of SO(D − 1), and we will build linear combinations of them by integrating
over the the sphere with various weights, that transform more conveniently, as fa-
miliar from the introduction of spherical harmonics.
Now the residue defines an operator Rn , by giving its matrix elements as:

hy, c d |Rn |x, a b i = (F 4 )a,b,c,d (x, y) · RnI (x · y) . (2.5)

We would like to demonstrate that Rn is a positive operator, i.e., that for all the
states in the Hilbert space |ψi, hψ|Rn |ψi > 0. We will do this by showing that the
P
operator can be written in the form are i ci |iihi| with ci > 0 and where |ii’s are
distinct but not necessarily orthogonal states.
Let us start with n = 1, where R1I (x · y) = 1. The goal is then to find a positive
operator R such that hy, c d |R|x, a b i = (F 4 )a,b,c,d . To this end, we will build SO(D−1)
irreps out of |w, a b i. For example, for spin-0, 2 and totally anti-symmetric 3-form

– 10 –
we can construct,
Z
· = |w, k k i,
w
Z Z  
(1) 1 δij (2) δij
i, j = (|w, i j i+|w, j i i) − |w, k k i, i, j = wi wj − |w, k k i,
w 2 D−1 w D−1
Z
i
j = w[i |w, j k] i , (2.6)
k
w
R R R
where w stands for the projective integral over a (D−2)-sphere, i.e., w ≡ hw dD−2 wi.
Note that we have two possible spin-2 structures.3 As it turns out, this will be suf-
ficient. The inner product with |x, a bi is given as

x, ab · = δab ,
(1) (δai δbj +δaj δbi ) δij δab (2) δij
x, ab i, j = − , x, ab i, j = (xi xj − )δab ,
2 D−1 D−1
i
x, ab j = x[i δaj δbk] . (2.8)
k

Using these irreps, we can construct the following rotationally-invariant operator,

2
X (α) (β) i i
O = c0 · · + cαβ i, j i, j + c3 j j . (2.9)
k k
α,β=1

a b
Sandwiched between |x, E1 , E2 i = |x, a b iE1 E2
and hy, E3 , E4 | we find
 
c11 +c12 +c21 +c22 2
hy, E3 , E4 |O|x, E1 , E2 i = (E1 ·E2 )(E3 ·E4 ) c0 − +(x·y) c22
D−1
hc i hc i
11 11
+(E1 ·E3 )(E2 ·E4 ) −(x·y)c3 +(E2 ·E3 )(E1 ·E4 ) +(x·y)c3
2 2
+(E3 ·E4 )(E1 ·y)(E2 ·y)c12 +(E1 ·E2 )(E3 ·x)(E4 ·x)c21
h
+ (E2 ·E4 )(E1 ·y)(E3 ·x)−(E2 ·E3 )(E1 ·y)(E4 ·x)
i
+(E1 ·E3 )(E2 ·y)(E4 ·x)−(E1 ·E4 )(E2 ·y)(E3 ·x) c3 . (2.10)

The coefficients can now be fixed by matching to F 4 in eq. (2.3), leading to:
 
9
c0 = −1 , c11 = 4, c12 = c21 = 2, c22 = 1, c3 = 2 . (2.11)
D−1
3
Other spin-2 structures will vanish when contracting with the polarization vectors. For example
(3) R
consider i, j = w w{i wk |w, j} k i. Then
(3)
E1a E2b x, ab i, j = (E2 · x)E1{i xj} (2.7)

which vanishes since E2 · x = 0.

– 11 –
 Since the states are irreps, they are orthogonal, and unitarity demands c0 > 0
which implies that D 6 10. Thus tree-level Yang–Mills amplitude “knows” about
critical dimension of type-I string! Note that the matrix of couplings for the spin-2
state has rank one:
     
c11 c12 42 T 2
= = vv with v = , (2.12)
c21 c22 21 1

which tells us that we have exactly one linear combination of the two spin-2 states
appearing here, which we can identify as = 2|i, ji(1) +|i, ji(2) . Thus we conclude
that (F 4 )a,b,c,d (x, y) = hy, c d |O|x, a b i with
 
9
O= −1 · · + + . (2.13)
D−1

Importantly, once it is established that (F 4 )a,b,c,d (x, y) is a positive operator, this

can be extended to (F 4 )a,b,c,d (x, y) · RnI (x · y) if RnI (x · y) has a positive expansion
on the Gegenbauer polynomial. To see this for each state in the Hilbert space |ψi, a
state |ψ; z, ji defined through
(D)
hx, ab |ψ; z, ji = hx, ab |ψi Gj (x · z) . (2.14)

In other words, |ψ; z, ji is the result of tensoring |ψi with symmetric spin-j irrep.
Note that even if the |ψi’s are irreps, |ψ; z, ji in general will not be and thus are not
orthogonal. Thus once we have
X (D)
RnI (x · y) = D
Bn,j Gj (x · y), Bn,j D
> 0, (2.15)
j

we can construct (F 4 )a,b,c,d (x, y) · RnI (x · y) as a positive operator:

" Z  
X
D 9
Rn = Bn,j −1 ·; z, j ·; z, j (2.16)
j z D−1
#
+ ; z, j ; z, j + ; z, j ; z, j .

We have already shown at n = 1 we must have D 6 10. Numerically, one observes

D
that Bn,j is indeed positive for D 6 10. In the next section we will prove its positivity
for D 6 6.
We note in passing that we can describe the usual Gegenbauer expansion for
external scalars, in exactly this same positive-operator-in-Hilbert-space language.
Here the states of the Hilbert space are simpler, just labeled by the unit vector on the
sphere |wi, and the residue Rn (x · y) defines an operator Rn via hx|Rn |yi = Rn (x · y).
The only irreps of SO(D−1) we can build from these are symmetric spin j tensors, as

– 12 –
 1
R R R
|·i = w |wi, |i1 i = w wi1 |wi, |i1 i2 i = w (wi1 wi2 − D−1 δi1 i2 )|wi etc. Since the states
are completely symmetric it is natural to define |z; ji = zi1 · · · zij |i1 · · · ij i. These
(D)
states define the Gegenbauer polynomials as hw|z; ji = Gj (w ·z). Any rotationally-
P P R
invariant operator O can be written as j cj |i1 · · · ij ihi1 · · · ij | = j cj z |z; jihz; j|.
P D R
In particular we can expand the operator Rn = Bn,j z |z; jihz; j|. Since these
D
states are irreps and orthogonal, for positivity we must have that Bn,j > 0, and so
the function Rn (x·y) must have an expansion as a sum over Gegenbauer polynomials
with all positive coefficients.
Let us finally turn to the closed superstring amplitudes. Having seem that the
open string residues are identified with positive operators, it is easy to see that via the
usual “closed = open2 ” connection [13], the closed string residues since R4 = (F 4 )2
and RnII = (RnI )2 are positive as well. Let us begin with

(R4 )a1 b1 a2 b2 a3 b3 a4 b4 = (F 4 )a1 a2 a3 a4 (F 4 )b1 b2 b3 b4 , (2.17)

and introduce the Hilbert space |x, a1 b1 a2 b2 i. Now given any two |ψi, |χi in the YM
Hilbert space, we can once again define |ψ, χi in the gravity Hilbert space, through

hx, a1 b1 a2 b2 |ψ, χi = hx, a1 b1 |ψihx, a2 b2 |χi . (2.18)

0
Similarly, for any two operators OYM , OYM on the YM Hilbert space,
X X
0
OYM = cψ |ψihψ|, OYM = cχ |χihχ| , (2.19)
ψ χ

gr P
we define an operator on the gravity Hilbert space OO 0
YM OYM
= χ,ψ cψ cχ |ψ, χihψ, χ|
with
gr
X
hy, a3 b3 a4 b4 |OO YM O
0 |x, a1 b 1 a2 b 2 i = cψ cχ hy, a3 a4 |ψihψ|x, a1 a2 ihy, b3 b4 |χihχ|x, b1 b2 i
YM
ψ,χ
0
= hy, a3 a4 |OYM |x, a1 a2 ihy, b3 b4 |OYM |x, b1 b2 i . (2.20)

Thus we have
!
X
4 0 0
(R )a1 b1 a2 b2 a3 b3 a4 b4 = hy, a3 b3 a4 b4 | cψ cψ0 |ψ, ψ ihψ, ψ | |x, a1 b1 a2 b2 i , (2.21)
ψ,ψ 0

9


where ψ = ( · , , ), c· = D−1 −1 and all the remaining cψ are 1. Thus
we see that for D 6 10, since all cψ are non-negative, once again we have a positive
operator. Note that for D > 10, the coefficients for ( ·, , ·, ) are negative.
This however, does not imply negativity since the states are not orthogonal, and may
have overlap. This is indeed the case, as seen from the fact that the 3-form ·,
can also be generated from , ). Thus generically the “critical dimension” for
closed string amplitudes—at least as directly inferred from our unitarity check for
four-point scattering—are higher than their open string counterparts.

– 13 –
3 Double-contour representation
We now turn to deriving our central result, the new double-contour integral rep-
resentation (1.18) for the coefficients of the partial wave expansion of the residue
polynomials defined above. In Appendix B, we give an alternative derivation of the
same formula.
To begin with, let us consider the type-I amplitude, for which

1 1
Z
Γ(−s)Γ(−t)
I
A (s, t) = =− dz z −s (1−z)−1−t . (3.1)
Γ(1−s−t) s 0

We would like to extract the residues on the massive poles of this expression at
1
s = n, AI (s, t) → s−n × RnI (t) for n = 1, 2, . . .. Of course we can trivially find from
the explicit form of the Gamma functions that RnI (t) = n!1 (1 + t)(2 + t) · · · (n − 1 + t),
but we would like to represent this slightly more naturally directly from the integral
representation. To whit, note that the the singularities in s all come from the region
of integration near z → 0. This motivates Taylor expanding the integrand around
z = 0, which as is familiar gives us the expansion of the amplitude as a sum over
P am (t)
poles in s. Putting (1 − z)−1−t = m am (t)z m , we have AI (s, t) = m s−m+1
P
, so that
I
the residue Rn (t) = an−1 (t). We can in turn represent an−1 (t) as n−1 derivatives
around z = 0 of the function (1 − z)−1−t , or what is the same, as a contour integral
around z = 0, as RnI (t) = n1 z=0 2πidzzn (1−z)−1−t . For the purpose of the partial wave
H

expansion, we would like to write this expression in terms of the x = cos θ, which is
related to t as usual t = s(x − 1)/2 = n(x − 1)/2 where we have used that on the
resonance s = n. Finally, since the dependence on x is exponential via (1 − z)−nx/2 ,
it is useful to introduce (1 − z) = e−u to manifest this dependence. The contour
integral around z = 0 translates into one around u = 0, and we have for the residue
polynomial RnI (x):
n n
du e 2 u e 2 ux
I
I 1
Rn (x) = . (3.2)
n u=0 2πi (eu − 1)n
Note that this expression transforms simply under u → −u, x → −x, and we dis-
cover that RnI (−x) = (−1)n+1 RnI (x), as we have already observed for the residue
polynomial.
We would now like to expand RnI (x) in the basis of Gegenbauer polynomials
(D)
Gj (x), which are orthogonal with respect to the usual measure dθ (sin θ)(D−3) =
D−4 (D) (α)
dx (1 − x2 ) 2 (compared to the classic conventions [14], we have Gj (x) = Cj (x)
with α = D−32
). For convenience given its ubiquitous appearance in our expressions,
(D)
we will define δ = D−42
. We will also use the Rodrigues formula for Gj (x):
(D)
Gj (x) = (−1)j αj,D (1 − x2 )−δ ∂xj (1 − x2 )δ+j . (3.3)

– 14 –
where αj,D = (2δ+1)j /(2j j! (δ+1)j ). Now by orthogonality, any function F (x) can
be expanded in Gegenbauers as F (x) = ∞
P (D)
j=0 Fj Gj (x), where

1
22δ (j+δ+ 12 )[Γ(δ+ 12 )]2
Z
2 δ (D)
Fj = nj,D dx (1 − x ) Gj (x)F (x) , nj,D = j! . (3.4)
−1 πΓ(j+2δ+1)

We will suppress these positive factors for now for simplicity, with the understanding
that an additional factor of αj,D nj,D is to be restored to the final answer.
Using the Rodrigues formula, and assuming that F (x) is regular at x = ±1, we
R1
can integrate by parts j times to arrive at Fj = −1 dx (1 − x2 )δ+j ∂xj F (x). We can
apply this form to extract the coefficients in the Gegenbauer expansion of the residue
polynomial RnI (x). We have
X (D)
RnI (x) = D
Bn,j Gj (x), (3.5)
j

where
1 n n
du e 2 u e 2 ux
Z I
D 2 δ+j
=Bn,j (1 − x ) u
∂xj n
−1 u=0 2πi (e − 1)
n 1
du e 2 u uj
I Z
1 n j
  n
= u n
dx (1 − x2 )δ+j e 2 ux . (3.6)
n 2 u=0 2πi (e − 1) −1
R1
Now we can easily perform the −1 dx integral above, using
1 1
ea − e−a
Z Z  
2 J ax
dx (1 − x ) e = (1 − ∂a2 )J dx e ax
= (1 − ∂a2 )J
−1 −1 a
= HJ (a) + HJ (−a), (3.7)

where we used the notation J = j + δ and

ea
 
HJ (a) ≡ (1 − ∂a2 )J . (3.8)
a
Note that here we need to specialize to even D so that J is an integer. Thus we find
n
du e 2 u uj 
I
D 1 n j n n




Bn,j = n 2 u n
H δ+j 2
u + Hδ+j − 2
u . (3.9)
u=0 2πi (e − 1)

Now, the same parity symmetry u → −u we used above in establishing the parity
of the residue polynomial RnI (x), tells us here that the integral with the first term
Hδ+j ( n2 u) is (−1)n+j+1 times that of the second term Hδ+j (− n2 u). We thus learn that
D
Bn,j vanishes when n+j is even, while while n+j is odd we have
n
du e 2 u uj
I
D n j−1 n



Bn,j = 2
Hδ+j 2
u . (3.10)
u=0 2πi (eu − 1)n

– 15 –
 Up to this point, all of the steps in our derivation have been completely straight-
forward, obvious and natural. But now something more slightly magical happens.
The object HJ (a) = (1−∂a2 )J (ea /a) we have introduced does not appear especially
simple, so that we still do not have a good way of integrating against this factor. But,
in fact, HJ (a) is essentially a J-th total derivative of a simple object, that will allow
us to proceed with the integration in a simple way. In fact we have a remarkable
identity,  a  2a 
2 J e −a J e
(1−∂a ) = J! e ∂a . (3.11)
a aJ+1
With this identity, we immediately arrive at our double contour-integral repre-
sentation. Using this representation for HJ ( n2 u) and integrating by parts J times we
have
du enu j+δ uj
I  
D n −2δ−2−j j+δ


Bn,j = 2 (−1) (j+δ)! ∂
j+δ+1 u
(3.12)
u=0 2πi u (eu −1)n
and we can finally write the (j+δ)-th derivative above as a contour integral, using
∂uJ F (u) = (−1)J J! v=0 2πidv
H
v J+1
F (u−v), to obtain our double-contour integral

du enu (u−v)j
I I
D n −2δ−2−j

2 dv
Bn,j = 2 [(j+δ)!]
2πi uj+δ+1 v=0 2πi v j+δ+1 (e(u−v) −1)n
Iu=0
(v − u)j
I
n −2δ−2−j

2 du dv
= 2 [(j+δ)!] j+δ+1 (ev −eu )n
. (3.13)
u=0 2πi v=0 2πi (uv)

Finally, restoring the extra positive factor of αj,D nj,D , we have

D 2D−2 Γ( D−3 )(j + D−3 )(j + D−4

)!
Bn,j = cD D
n,j βn,j , cD
n,j =
2
√ j+D−2 2 2
, (3.14)
πn
D
where βn,j is the double-contour integral displayed on the right-hand side of (1.18)
with a = 0.
All that remains is to naturally motivate and derive the fundamental identity in
eq. (3.11). The motivation is simple: in order to act with the operator (1 − ∂a2 )J ,
it is natural to work with Fourier/Laplace representations that diagonalize ∂a . The
R∞
Laplace transform is obviously suggested, since a1 = 0 dq e−qa . Differentiating J
R∞
times we also have that J!/aJ+1 = 0 dq q J e−qa . Putting q = r−p for any real p, we
find the Laplace representation
Z ∞
J! e−pa
= dr (r−p)J e−ra . (3.15)
aJ+1 p
R∞
Now, let us begin with this representation for e−a /a = 1 dt e−ta . In this form we
can trivially apply (1 − ∂a2 )J to get
 −a  Z ∞
e
2 J
(1 − ∂a ) = dt (1 − t2 )J e−ta . (3.16)
a 1

– 16 –
It is now obvious that we should recognize this object at a J-th total derivative. The
reason is that (1 − t2 )J = (1 − t)J (1 + t)J factorizes, so it is natural to put 1+t = r,
and discover our identity:
 −a  Z ∞
e
2 J
(1 − ∂a ) =e a
dr (−r)J (r − 2)J e−ra
a 2
Z ∞
a J
= e ∂a dr(r − 2)J e−ra
2
 −2a 
a J e
= J! e ∂a . (3.17)
aJ+1

Changing a → −a yields eq. (3.11).

We can also find a double-contour representation of the residues for scattering
the m2 = −1 tachyon states in the bosonic string, beginning with the Veneziano
amplitude Aopen,bos (s, t) = Γ(−1−s)Γ(−1−t)/Γ(−2−s−t). Following the same steps
as in the above derivation, we find

(v−u)j e−(u+v)
I I
open,bos 0D du dv
Bn,j = c n,j j+δ+1 (ev −eu )n+2
, (3.18)
u=0 2πi v=0 2πi (uv)
√
where c0 D
n,j = 2
D−2
Γ( D−3
2
)(j + D−3
2
)(j + D−4
2
)!/( π(n + 4)j+D−3 ). We will refer to
open,bos
the part stripped off from this constant as βn,j , just as in (1.18) with a = 1. See
Appendix C for a step-by-step derivation.
Similar formula for partial-wave expansion coefficients can be obtained for closed
strings. We parametrize their amplitudes with

|z|−2s |1 − z|−2t
Z
closed 1 a+b−2
Aa,b (s, t) := π s d2 z 2a , (3.19)
C z (1 − z)1+a z̄ 2b (1 − z̄)1+b

for a, b ∈ Z and d2 z = 2i1 dz ∧ dz̄. Explicitly, we have


Γ(−s)Γ(−t)Γ(−u)
for a = b = 0 (type-II),


Γ(1+s)Γ(1+t)Γ(1+u)





 Γ(−s−1)Γ(−t−1)Γ(−u+3)
closed
Aa,b (s, t) = for a = 0, b = 1 (heterotic),

 Γ(1+s)Γ(1+t)Γ(1+u)


 Γ(−1−s)Γ(−1−t)Γ(−1−u)
− for a = b = 1 (bosonic),


Γ(2+s)Γ(2+t)Γ(2+u)
(3.20)
2 2
where u = −s − t + 4m . We have m = − min(a, b), i.e., scattering of gravitons in
the superstring, gluons in the heterotic, and tachyons in the bosonic one.4 In this
4
Once again there will be an additional overall δ 2×8 (Q) for the full type-II amplitude, and δ 8 (Q)
for the heterotic amplitude.

– 17 –
notation, the spins range in j 6 2(n−1+a+b). Note that using Kawai–Lewellen–Tye
relations [13] we can express it in terms of the open-string amplitudes
1 sin(πs) sin(πt)
Aa,b (s, t) = Aa (s, t)Ab (s, t), (3.21)
π sin(π(s+t))
which implies that

Ress=n Aa,b (s, t) = (Ress=n Aa (s, t)) (Ress=n Ab (s, t)) . (3.22)

(An alternative, more direct derivation is given in eq. (C.20).) Therefore positivity of
the Gegenbauer expansion of closed-string amplitudes follows from that of the open
case. Nevertheless, it is still interesting to find a closed-string counterpart, which we
D
denote as βn,j , of the double-residue formula. We specialize to D ∈ 2Z.
Following a similar set of manipulations to those in the previous subsection, one
closed,II/bos
finds in the cases a = b that Bn,j = 0 for odd j (for any n and D), while for
even j we have
I I I
closed,II/bos 2(a−1) D−4 2 du dũ dv
Bn,j = 2n [(j+ 2 )!]
u=0 2πi ũ=0 2πi v=0 2πi
evn+(4v−u−ũ)a (u + ũ)j
× D−2 , (3.23)
[(1 − eu )(1 − eũ )]n+2a [v(v−u−ũ)]j+ 2
where the residue is taken around the origin in all the variables. In the heterotic
string case, a = 0, b = 1, there is no distinction between even/odd j and we have
closed,het
Bn,j = n1 [(j+ D−4
2
)!]2
((−1)j − e4ũ ) e−ũ+vn (u + ũ)j
I I I
du dũ dv
× . (3.24)
j+ D−2
u=0 2πi ũ=0 2πi v=0 2πi (1−eu )n (1−eũ )n+2 [v(v−u−ũ)] 2

Proofs of these formulae are given in Appendix C, where one can also find alternative
derivations and expressions in terms of quadruple-contour integrals.

4 Applications
In this section, we show that our double contour formula (1.18) is well-suited to
D
study questions of unitarity. From now on, we use βn,j to denote the double-contour
representation stripped off from the manifestly-positive constant according to (3.14),
and similarly for the bosonic case.

4.1 Manifest unitarity in D 6 6 dimensions

The double contour formula makes unitarity manifest in D 6 6 spacetime dimensions
for the case of the superstring. To see this, we simply have to change variables to

u = log(1 − x) , v = log(1 − y) , (4.1)

– 18 –
in terms of which the integral becomes
I I
D dx dy 1 1
βn,j = D−2
x=0 2πi y=0 2πi (1 − x)(1 − y)(log(1 − x) log(1 − y)) 2 (x − y)n−j
1 1
!j
log(1−x)
− log(1−y)
× . (4.2)
x−y

Each of the three factors that enter the integrand is a function with only positive
Taylor coefficients when expanded first around y = 0 and then around x = 0. We
call such a function (or rather power series) a positive function. Consequently, the
residue integral that just picks out one of the coefficients will be positive.
Indeed, the factors,
1
D−2 (4.3)
(1 − x)(− log(1 − x)) 2
are of the same form as the function hα for α = 2−D
2
discussed in the Appendix A.
They are positive for α > −2, i.e., D 6 6. The expansion of the function
∞  
−k
X k + m − 1 m −k−m
(x − y) = y x (4.4)
m=0
m

is obviously also positive (since k = n − j > 0). Finally, the positivity of the last
factor follows from the positivity of
∞
1 1 X
f (z) = + = cm x m (4.5)
log(1 − x) x m=0

with cm > 0. We demonstrate that also in Appendix A that this function is positive.
Writing
1 1 1 1
log(1−x)
− log(1−y) f (x) − x
− f (y) + y
=
x−y x−y
∞
1 X xm − y m
= + cm
xy m=0 x−y
∞ m−1
1 X X
= + cm xm−1−k y k (4.6)
xy m=0 k=0

shows that all coefficients are positive in the expansion. Thus in (4.2) the integrand
is a positive function and so in particular also the residue is positive. Thus our
formula makes unitarity manifest for D 6 6 spacetime dimensions.
The same arguments also apply for the bosonic string. The double contour
1
formula in its logarithmic form has an extra (1−x)(1−y) compared to (4.2) (as well

– 19 –
as n → n + 2 replaced). Our formula makes then unitarity manifest for D 6 10
spacetime dimensions, because the function
1
D−2 (4.7)
(1 − x)2 (− log(1 − x)) 2

2−D
is the square of the positive function hα discussed in Appendix A with α = 4
.
This function is positive for α > −2, i.e., D 6 10.

4.2 Unitarity in D = 10 for low spin

While our formula does not make unitarity manifest in ten spacetime dimensions,
one can still use it to demonstrate unitarity for low values of spin. We will give the
argument here for j = 0. Higher values of j and the corresponding analogue for the
bosonic string can also be checked on a case-by-case basis, but the computation gets
more and more involved.
For j = 0, starting with eq. (4.2) one can simply compute the y-residue and get

dx n ((n+2)(n+1)−3(n+1)x+x2 )
I
10
βn,0 = . (4.8)
x=0 2πi 6(1 − x)xn+3 log(1 − x)4
10
One can again reduce the statement βn,0 > 0 to the positivity of a particular generat-
ing function as follows. We can trade the n-dependence for derivatives and integrate
by parts
10
βn,0
I  
dx 1 1
∂x2 + 3∂x + 1 x−n−1



= ∂x 3
n(n+1) 2πi (n+1) 18 log(1 − x)
Ix=0  
dx 1 1
2
∂x x−n−1



= −∂x + 3∂x − 1 3
2πi (n+1) 18 log(1 − x)
Ix=0  
dx 2

1
= n+2
∂x − 3∂x + 1
x=0 2πi x 18 log(1 − x)3
(1 − x)2 log(1 − x)2 + (9x − 6) log(1 − x) + 12
I  
dx
= n+2
.
x=0 2πi x 18(1 − x)2 log(1 − x)5
(4.9)

Thus, it suffices to show that all Taylor coefficients starting from O(z) are positive
for the following function

(1 − z)2 log(1 − z)2 + (9z − 6) log(1 − z) + 12

f (z) = (4.10)
18(1 − z)2 log(1 − z)5
2 7 1 1 z z2 z3 3077z 5
=− 5 + − + + + + +
3z 18z 3 36z 2160 9072 20160 362880 79833600
3125z 6 13180873z 7
+ + + O(z 8 ) (4.11)
28740096 65383718400

– 20 –
Since unitarity is no longer manifest in our formula in D = 10 dimensions, the
10
singular coefficients in the expansion never enter the computation of βn,0 and can be
negative. We also remark that the z 4 term is absent, which indicates that D = 10
is the critical dimension. We prove the positivity of the non-singular terms of this
function in Appendix A.

4.3 Asymptotics
Finally, our double-contour formula is very useful to study asymptotics.

4.3.1 Fixed spin

We again start from the form (4.2). In the large n limit for fixed j, we can explicitly
compute the residue in y. When computing the residue, we get terms with different
powers of n. The greatest contribution is achieved by picking the highest possible
power from the term (x − y)−n+j , because these coefficients grow very rapidly with
n. This means that we pick up the leading coefficients of the other factors. We hence
get
D−4 D−4
nj+ 2 x−n−j− 2
I
D dx
βn,j ∼ D−2 , (4.12)
(j + D−4
2
)! x=0 2πi (1 − x)(− log(1 − x)) 2
where we further approximated the binomial coefficient appearing in the expansion
of (x − y)−n+j with its asymptotic value. We hence see in particular that the j-
dependence of the asymptotic behaviour of the coefficients is very simple.
It remains to determine the asymptotic value of the remaining integral. This can
be done in a very classical way as follows, see e.g. [15, Sec. VI]. We first deform the
contour that runs around 0 to the Hankel contour that runs from below once around
the branch cut as depicted in Figure 3. Since the integrand decays sufficiently fast
at infinity, the arcs at infinity are not contributing to the integral. We next perform
the change of variables x = 1 + nt . The t variables runs then from below the positive
real axis, encircles the origin and then runs back above the real axis. We get
D−4
−n−j− D−4
nj+ 2 1 + nt 2
I
D dt
βn,j ∼− (4.13)
(j + D−4


D−2
H 2πi t − log − n
)! t
2 ) 2
D−4
−n−j− D−4
nj+ 2 dt 1 + nt 2
I
=− D−2  D−2 . (4.14)
(j + D−4 )! log(n) 2 H 2πi  log(−t) 2
2 t 1 − log(n)

The remaining integrand converges uniformly for any bounded domain of the t-plane
and hence we can replace it with its value at large n, which gives
D−4
nj+ dt e−t
I
2
D
βn,j ∼− D−2 . (4.15)
(j + D−4
)! log(n) 2 H 2πi t
2

– 21 –
 x

Figure 3. Contour deformation to the Hankel contour H.

The contour integral finally simply picks up minus the residue, since we run clockwise
around the pole. Thus, we get the simple, but interesting asymptotics
D−4
D nj+ 2
fixed j, large n: βn,j ∼ D−4 D−2 . (4.16)
(j + 2
)! log(n) 2

For the actual residue Bjn , this means

2D−2 (j + D−3 D−3


D )Γ
Bn,j ∼ √ D 2 2
D−2 . (4.17)
π n 2 log(n) 2
In particular, the asymptotic values are positive. This estimate can be extended
to a whole asymptotic series in log(n)−1 by expanding the integrand in log(n)−1 ,
which gives
D−4 ∞  D−4
nj+ 2 dt e−t
 I
D
X + m 1
βn,j ∼ − D−4
2
D−2 log(−t)m (4.18)
(j + 2 )! m=0 m log(n)m+ 2 H 2πi t
D−4 ∞  D−4
nj+ 2 (−1)m dm
 I
X +m dt
∼ D−4
2
D−2 m
(−t)−α e−t (4.19)
(j + 2 )! m=0 m log(n)m+ 2 dα α=1 H 2πi
D−4 ∞
X D−4 + m
nj+ 2 (−1)m dm
 
2
1
∼ D−4 D−2 m
, (4.20)
(j + 2 )! m=0 m log(n)m+ 2 dα α=1 Γ(α)
where in the last step we made use of Hankel’s formula for the Gamma function.
By comparing the m = 0 and the m = 1 term, we can estimate that our asymp-
γ(D−2)
totic value becomes good in the regime n  e 2 , where γ ≈ 0.57 is the Euler–
γ(D−2)
Mascheroni constant. Hence we expect the couplings Bjn to be positive for n & e 2
for all dimensions. This only leads finitely many coefficients that can potentially be
negative and can be checked by hand. However, we were unable to turn this ap-
proximate inequality into a rigorous bound, which prevents us from mathematically
D
proving positivity of Bn,j also in D = 10 spacetime dimensions.

– 22 –
 Finally, we note that the corresponding formula for the bosonic string is
D−2
open,bos nj+ 2
fixed j, large n: βn,j ∼ D−4 D−2 , (4.21)
(j + 2
)! log(n) 2
2D−2 (j + D−3 )Γ D−3


open,bos 2 2
Bn,j ∼ √ D−4 D−2 . (4.22)
πn 2 log(n) 2

4.3.2 Regge trajectories

The second asymptotics we can look at is to keep ∆ = n − j fixed and consider large
n. This limit corresponds to moving in one fixed Regge trajectory. In this case, it is
convenient to start from the form
 −1 n
u − v −1
I I
D 1 du dv 1 −1 −1 −∆


βn,n−∆ = − D−2 u − v . (4.23)
2 u=0 2πi v=u 2πi (uv) 2 ev − eu
of the double contour formula. Notice that the inner integral runs around v = u
and not v = 0 as before. To obtain this equivalent form, one the usual relation of
contours with a singularity at u = v
I I I I I I
− = . (4.24)
v=0 u=0 u=0 v=0 u=0 v=u
The second contour on the left-hand side is the original contour, but both terms give
an equal contribution due to the symmetry of the integrand (for n + j odd). Hence
H H
the contour u=0 v=u leads to −2 times the original formula. This is also explained
in Appendix C, where it is also explained that in this form the formula is actually
also valid for odd D.
Since we want to get the dominating piece in the large n limit, we should pick
up as many powers of n from the second factor as possible when doing the integral
over v. This means that we can expand this term in v around u. It turns out that
one has to go to the second order and hence gets
u−1 − v −1 e−u (u + 2)(v − u) (u2 + 6u + 12) (v − u)2
 
+ O (v − u)3 .


v u
= 2 1− + 2
e −e u 2u 12u
(4.25)
In order for the second term in the expansion to make an appreciable contribution
to the integral in the large n limit, the term of order O(v − u) has to become small.
This means that the u-integral generically has a large degree of cancellation built in.
Note however that we can choose the radius of the u-integral to be 2, so that we are
integrating through the critical point u = −2, which becomes a saddle point in the
large n limit. This in turn implies that we can replace u with −2 in most places that
are not singular when u → v, u → −2 or u → 0. We can also replace v by u in the
first factor of (4.23). Hence we get the large-n approximation
n
e−un (u+2)(v−u) (v−u)2
I I 
D 1 du dv
βn,n−∆ ∼ − 1+ + .
2 u=0 2πi v=u 2πi u2n−2∆+D−2 (v−u)∆ 4 12
(4.26)

– 23 –
Let us now change variables to
√
t = (v − u) n . (4.27)

The integral then becomes

∆−1 n
e−un n 2 t2
I I 
D 1 du dt (u + 2)t
βn,n−∆ ∼ − 1+ √ + (4.28)
2 u=0 2πi t=0 2πi u2n−2∆+D−2 t∆ 4 n 12n
−un ∆−1
√
t2
I I  
1 du dt e n 2 (u + 2) nt
∼− exp + , (4.29)
2 u=0 2πi t=0 2πi u2n−2∆+D−2 t∆ 4 12
where we used that (1 + na )n ∼ ean in the second line. As a next step, we interchange
the order of the two integrals and perform the u integral. This leads to
3 7
D n2n− 2 ∆+D− 2
βn,n−∆ ∼
2(2n − 2∆+D−3)!
2n−2∆+D−3
1√ t2
I   
dt 1
× ∆
1− √ t exp nt + . (4.30)
t=0 2πi t 4 n 2 12
We can now use the Stirling approximation of the factorial to simplify the prefactor.
Since n is large, we can also remove the constant terms in the exponent of the first
factor of the remaining integral as follows
∆−2 I 2n
21−D−2n+2∆ e2n n 2 1√ t2
  
D dt 1
βn,n−∆ ∼ √ ∆
1− √ t exp nt + .
π t=0 2πi t 4 n 2 12
(4.31)
Next, we use that
2n
1√
    2
1 t  1
1− √ t exp nt = exp − + O n− 2 . (4.32)
4 n 2 16
The remaining t integral is then
 ∆−1
t2
I   
dt 1 1 2
exp = ∆−1

. (4.33)
t=0 2πi t∆ 48 2
! 48

Recall that ∆ = n − j is odd, so that this is non-vanishing. Putting the pieces

together leads then to the following asymptotics:

D 23−D−2n e2n  n  ∆−1

2
fixed ∆ = n − j, large n: βn,n−∆ ∼√ ∆−1

. (4.34)
πn 2 ! 3

D
For the actual residue Bn,n−∆ , this means
1−∆ 3 ∆−D+1
2 2 −2n en Γ D−3


D 3 2
2

n 2
Bn,n−∆ ∼ √ ∆+1
, (4.35)
πΓ 2

– 24 –
which decays exponentially with rate 4e . In particular, the asymptotic values are again
positive. We should also remark that the error of this approximation is actually of
1
order O(n−1 ). From our derivation it seems that it is of order O(n− 2 ). However,
this cannot be right, since for any fixed ∆, one can just compute the exact formula
by taking the residue in v in eq. (4.23). For example, we can compute explicitly for
∆ = 3 that
D nD+2n−8 (n−D+7) (2n+D−8)
βn,n−3 = (4.36)
48 (2n+D−7)!
and similarly for any other fixed ∆, see Appendix D for more explicit formulas. This
makes it clear that the square roots are just an artifact of our derivation.
The corresponding formula for the bosonic string can be derived analogously and
gives

open,bos 2−1−D−2n e2(n+4)  n  ∆−1

2
fixed ∆ = n − j, large n: βn,n−∆ ∼ √ ∆−1

, (4.37)
πn 2 ! 3
5
2−2n− 2 en+4 Γ D−3

  ∆−1
open,bos 2 n 2
Bn,n−∆ ∼ √ 1 (D−4) ∆−1
. (4.38)
πn 2 2
! 3

In the special case of D = 4 and ∆ = 1, this coincides with [16].

Acknowledgments
We thank Aaron Hillman, Dalimil Mazáč and Pinaki Banerjee for useful discussions.
N.A-H., L.E. and S.M. are partially funded by the grant DE-SC0009988 from the
U.S. Department of Energy. L.E. acknowledges funding from the Marvin Goldberger
fund. S.M. gratefully acknowledges the funding provided by Frank and Peggy Taplin.
Y.H. is supported by Taiwan Ministry of Science and Technology Grant No. 109-
2112-M-002 -020 -MY3.

A Some positive functions

In this appendix, we show that certain functions (power series) have only positive
Laurent coefficients when expanded around z = 0. We call such functions positive
functions.

A.1 Simple functions

We start with
1 1
f (z) = + (A.1)
log(1 − z) z
1 z z 2 19z 3 3z 4 863z 5 275z 6
= + + + + + + + O(z 7 ) . (A.2)
2 12 24 720 160 60480 24192

– 25 –
As one can see the first few Taylor coefficients are all non-negative. To prove that in
fact all Taylor coefficients are positive, we remark that we can write
Z 1
zf (z) = ds (1 − (1 − z)s ) . (A.3)
0
Multiplying the function by z does not change positivity and hence it suffices to
show that all coefficients of zf (z) are positive. In this representation one can easily
compute the m-th derivative at z = 0:
Z 1 m−1
Y
m
∂z (zf (z)) = ds s (k − s) . (A.4)
z=0 0 k=1
The integrand on the RHS is manifestly positive in the region 0 < s < 1, which
proves positivity of f (z).
A related positive function is
1 1
g(z) = − (A.5)
(z − 1) log(1 − z) z
1 5z 3z 2 251z 3 95z 4 19087z 5
= + + + + + + O(z 6 ) . (A.6)
2 12 8 720 288 60480
Positivity becomes manifest by writing
Z 1
ds (1 − z)s−1 − 1 ,


zg(z) = (A.7)
0
so that m
Z 1 Y
∂zm (zg(z)) = ds (k − s) > 0 . (A.8)
z=0 0 k=1
From the positivity of this function, we can also easily deduce that
 α
1 log(1 − z)
hα (z) = − (A.9)
1−z z
is a positive function for any α > −2. To see this, we first note that it is true for
α = −2, because
hα=−2 (z) = z 2 f 0 (z) + 1 , (A.10)
which is positive. Integrating g(z) implies that the function
Z z  
log(1 − z)
dy g(y) = log − (A.11)
z
has only positive coefficients. This then in turns gives positivity of
    β
log(1 − z) log(1 − z)
exp β log − = − (A.12)
z z
for any β > 0. Since
 2+α
log(1 − z)
hα (z) = hα=−2 (z) − (A.13)
z
is now the product of two positive functions for α > −2, it is also positive.

– 26 –
A.2 Generating function of the D = 10, j = 0 coefficients
Let us show using the same ideas that the regular coefficients of
(1 − z)2 log(1 − z)2 + (9z − 6) log(1 − z) + 12
f (z) = (A.14)
18(1 − z)2 log(1 − z)5
are positive. This is the function that appears in the unitarity for the superstring
in D = 10 for j =, see Section 4.2. The argument for this is very similar to the
argument above, but is less aesthetic. We write
Z 1
5
du (1 − z)u−2 − u4 z 4 + u3 3z 5 + 2z 4 − 4z 3


36z f (z) =
0
+ u2 −z 6 − 7z 5 + 8z 4 + 12z 3 − 12z 2

+ u 2z 6 + 3z 5 − 23z 4 + 6z 3 + 36z 2 − 24z

− z 6 + 2z 5 + 13z 4 − 28z 3 − 10z 2 + 48z − 24 . (A.15)

In this form, we can again compute derivatives of our function at zero, which gives
for n > 0
Z 1 n+1
Y
∂zn+6 36z f (z) = (n2 + 8n + 15)
5


du u (j − u)
0 j=1

× n4 u2 − n3 2u3 − 8u2 + 3u + 1 + n2 u4 − 14u3 + 24u2 − 7u − 1

+ n 6u4 − 28u3 + 26u2 − 2 + 4u 2u3 − 4u2 + u + 1 . (A.16)

The integrand is no longer manifestly positive and we have to work a bit harder to
show positivity from here. As a first step, we notice that at large n, the term n4 u2
will dominate over the other ones in the parenthesis. Hence we get an estimate from
below by minimizing the coefficients of n3 , n2 , n and n0 over u ∈ [0, 1]. We further
round them down to get nicer expressions. Thus, we have to show positivity of the
integral
Z 1 n+1   
Y
4 2 3 3 2
du u (j − u) n u − 2 n +n +n . (A.17)
0 j=1
4

Since 43 n3 + n2 + n 6 n3 for n > 5, it finally suffices to show positivity of the integral

Z 1 n+1
Y
(j − u) nu2 − 2


du u (A.18)
0 j=1

for sufficiently large n. The first few orders can then be checked directly on the
computer. Clearly, the integral is positive in the region u > √2n . Thus, we need to
show that
Z 1 n+1 Z √2 n+1
Y n Y
2 2



√ du u (j − u) nu − 2 > du u (j − u) 2 − nu (A.19)
2
n j=1 0 j=1

– 27 –
for sufficiently large n. Let us bound the left-hand side from below as follows:
Z 1 n+1
Y Z 1 n+1
Y
2
(j − 1)(nu2 − 2)


√ du u (j − u) nu − 2 > √ du u(1 − u) (A.20)
2 2
n j=1 n j=2
√ !
n 1 1 8 2
= n! − + − (A.21)
20 3 n 15n 32
(n + 1)!
> (A.22)
30
where the last inequality holds for n > 30. For the right-hand side, we have instead
Z √2 n+1 Z √2 n+1
n Y n Y
2
j 2 − nu2



du u (j − u) 2 − nu 6 du u (A.23)
0 j=1 0 j=1
(n + 1)!
= . (A.24)
n
Thus for n > 30, the left-hand side becomes greater than the right-hand side and
all Taylor coefficients with n > 30 have to be positive. It is then a simple matter to
check on the computer that also the first 30 coefficients are positive.

B Single-variable generating function

D
We can also write βn,j in terms of a single, more complicated contour integral. To
keep formulas simpler, we will focus on the supersymmetric case with a = 0, but
similar manipulations can be done for the bosonic formula. Starting from (1.18), we
first integrate by parts as follows
dv (v − u)j
I I
D 1 du 1
βn,j = − D−2 (∂u + ∂v ) (B.1)
n u=0 2πi v=0 2πi (uv)j+ 2 (ev − eu )n+2a
dv e−a(u+v) (v − u)j
I I
1 du
= (∂u + ∂ v ) D−2 (B.2)
n u=0 2πi v=0 2πi (ev − eu )n (uv)j+ 2
j + D−2 dv (v − u)j (u + v)
I I
2 du
=− . (B.3)
n j+ D
u=0 2πi v=0 2πi (uv) 2 (ev − eu )n

We next change variables to

uv v−u
x= , t= . (B.4)
u−v 2
This transforms the integral to

2
2−D
2
D
(−1) 2 (2j−2+D)
I
dt
I
dx √
D 1− 2x
βn,j = D−2 ent(1− t
)
.
j+ D
n t=0 2πi t 2 (1 − e2t )n x=0 2πi x 2

(B.5)

– 28 –
We recognize the exponential as the generating function of the modified Bessel func-
tion of the second kind Km− 1 (nt) and we have [17]
2

I
dx √ r
nm nt 2nt
nt(1− 1− 2x )
e t = e K 1 (nt) . (B.6)
x=0 2πi x
m+1 m! π m− 2
D
Hence we can write βn,j as
D 5 D D−3
(−1)n+ 2 2 2 − 2 −n nj+ 2
I dt Kj+ D−3 (nt)
D 2
βn,j = √ . (B.7)
π j + D−4
D−3


2
! t=0 2πi t 2 sinh(t)n
Finally, one can also rewrite this in terms of the modified Bessel functions of the first
kind I±(j+ D−3 ) (nt). We have in general the relation
2

π Ij+ D−3 (nt) − I−j− D−3 (nt)

Kj+ D−3 (nt) = − 2 2
(B.8)
sin π j + D−3



2 2 2

However, the two terms have opposite parity and using the constraint n+j odd, the
second term leads to an integrand of even parity which does not contribute to the
residue. This then leads to
√ 3 − D −n j+ D−3 I dt Ij+ D−3 (nt)
D π2 2 2 n 2 2
βn,j = . (B.9)
j + D−4
D−3


2
! t=0 2πi t 2 sinh(t)n

With some further knowledge about special functions, it is from here actually rather
D
simple to land again on the initial expression for the coefficients βn,j in terms of de-
composing the Pochhammer symbol into Gegenbauer polynomials. Using the integral
representation (see, e.g. [14, Ch. VI, Sec. 6.15])
Z 1
2−α z α 1
Iα (z) = √ 1 dx (1 − x2 )α− 2 exz , (B.10)
πΓ(α + 2 ) −1

and inserting back the normalization constant cD n,j , we get

(2j + D − 3)nj−1 Γ D−3

Z 1
dt tj exnt
I
D 2 2 j+ D−4
Bn,j = √ dx (1 − x ) 2 (B.11)
2j+n π j + D−4

n
2
! −1 t=0 2πi sinh(t)

(2j + D − 3)Γ D−3

Z 1
dt exnt
I
2
2 j+ D−4 j
= j+n √ dx (1 − x ) 2 ∂
x . (B.12)
2 n π j + D−4 2
! −1 t=0 2πi sinh(t)
n

We finally have upon substituting z = 1 − e−2t

dt exnt
I I
n
=2n
dt e(x−1)nt (1 − e−2t )−n (B.13)
t=0 2πi sinh(t) t=0
I
1−x
=2n−1
dz z −n (1 − z) 2 t−1 (B.14)
z=0
= 2 nRnI (x)
n−1
, (B.15)

– 29 –
where the residue polynomial was defined in (1.9). Thus, we recover
(2j + D − 3)Γ D−3

Z 1
D 2 j+ D−4
Bn,j = j+1 √ D−4
2

dx (1 − x ) 2 ∂xj RnI (x) , (B.16)
2 π j + 2 ! −1
which after integration by parts is the integral of the residue polynomial against the
(D)
Gegenbauer polynomial Gj (x) with the appropriate normalization. Hence following
these steps backwards gives an alternative derivation of the double contour formula.

C Derivation of the triple-contour representation

In this appendix we derive residue formulae for the Gegenbauer coefficients in the
case of closed strings. For the sake of variety, we will follow a different route to that
of Section 3.
For completeness and in order to exhibit parallels to the open-string case, let us
first revisit it. Setting α0 = 1, open-string tree-level amplitudes can be parametrized
by the following worldsheet integral with a ∈ Z,
Z 1
open
Aa (s, t) := −s a−1
dz z −s−2a (1 − z)−t−1−a , (C.1)
0

where z is the cross-ratio of the positions of vertex operators. We have analytically

extended the integrand as a holomorphic function of z ∈ C since contour deforma-
tions will be needed in the following steps.
The case a = 0 corresponds to gluon scattering in type-I superstring (the
polarization-dependent prefactors are positive and hence can be stripped-off without
affecting the conclusions), while a = 1 gives tachyon scattering in the bosonic string
case. This means m2 = −a. While the explicit form of the amplitudes will not be
needed for our purposes, let us spell them out for completeness:
Γ(−s)Γ(−t)


 for a = 0 (type-I),
open
 Γ(1−s−t)
Aa (s, t) = (C.2)
Γ(−1−s)Γ(−1−t)
−

 for a = 1 (bosonic).
Γ(−2−s−t)
1
Qn−1+3a
This easily gives Ress=n Aopen
a = (n+a)! k=1+a (t + k). Note that for a = 0, the
massless exchange with n = 0 is special because of the normalization in (C.1). Below
we focus on n > 0 since positivity of the massless pole can be easily checked by
inspection.
We first convert the cut in the kinematic s-variable into a worldsheet residue.
This can be achieved by compactifying the integration contour near z = 0, in a way
similar to the Hankel contour (see, e.g., [18, App. A.2.1]):
Z 1 I Z 1
1
= −2πis + , (C.3)
0 e − 1 |z|= 

– 30 –
where e−2πis is the monodromy around z = 0 of the integrand at hand. The starting
points of the two integrals on the right-hand side have to coincide and be on the
same sheet. The result has the effect of manifesting all the s-channel poles, as well
as making the integral converge for any value of s, which was not the case for (C.1).
Performing the cut leaves us with
I I
open
ds Aa (s, t) = na−1
dz z −n−2a (1 − z)−t−1−a . (C.4)
s=n z=0

Note that the integrand on the right-hand side is now single-valued around z since
n + 2a ∈ Z.
We can now plug in this result to the formula for the Gegenbauer coefficients
open,a
Bn,j , which expressed in terms of integrals over the Mandelstam invariants read
Z 0 I
open,a ds open
j+ D−4
Bn,j := (−1)j cm
n,j dt Aa (s, t) ∂tj −t(t+n−4m2 ) 2
. (C.5)
−n+4m2 s=n 2πi

The t-integral between −n + 4m2 and 0 corresponds to integrating over all scattering
angles in the s-channel (recall that cos θ = 1 + 2t/(s − 4m2 )), while the s-residue is
the unitarity cut. The overall mass-dependent constant cm n,j is a manifestly positive
and given by
2D−3 (j+ D−3 ) Γ( D−3 )
cm
n,j = 2 D+j−3
√ 2
D−2
2
> 0. (C.6)
(n − 4m ) π Γ(j+ 2 )
The coefficients still depend on D, but we suppress it from the notation for clarity.
Substituting (C.4) commuting the z and t integrals gives
Z 0
dz z −n−2a
I
open,a j m a−1 dt j

j+ D−4
Bn,j = (−1) cn,j n 1+a
∂
t t
−t(t+n+4a) 2
.
z=0 2πi (1 − z) −n−4a (1 − z)
(C.7)
Integrating by parts j times in t yields

dz z −n−2a
I
open,a m a−1
Bn,j = cn,j n 1+a
(− log(1−z))j (C.8)
z=0 2πi (1 − z)
Z 0
dt
j+ D−4
× t
−t(t+n+4a) 2
,
−n−4a (1 − z)
| {z }
a (0)−H a (−n−4a)
Hn,j n,j

since all the boundary terms vanish. As a sanity check we see that using − log(1−z) =
P∞ k
k=1 z /k all the terms with too high spin, j > n+2a, vanish identically.
At this stage, let us consider the t-integral in the indefinite form on its own,
Z t0
dt
j+ D−4
a
Hn,j (t0 ) = −t(t+n+4a) 2
. (C.9)
(1 − z)t

– 31 –
 a a
We are interested in the combination Hn,j (0) − Hn,j (−n−4a) and hence can ignore
the integration constant. From now on we assume that D is even. Notice that
introducing u = log(1 − z) we can rewrite the above integral as
D−4
Z t0
0 (n+4a)u j+ 2 D−4
a
Hn,j (t ) = e ∂u dt e−(t+n+4a)u tj+ 2 . (C.10)

At this stage we recognize that the integrand is a total derivative,


j+ D−4

Z t0 −(t+n+4a)u X 2 k
j+ D−4 e (tu) 
a
Hn,j (t0 ) = −(j + D−4
2
)! e(n+4a)u ∂u 2 dt ∂t  . (C.11)
j+ D−2 k!
u 2 k=0

Evaluated at t0 = 0, this is simply

e−(n+4a)u
 
a j+ D−4
Hn,j (0) = −(j + D−4
2
)! e(n+4a)u ∂u 2 D−2 . (C.12)
uj+ 2

a
In order to simplify the contribution from Hn,j (−n−4a), we consider the change of
variables z → z/(z − 1) together with t → −t−n−4a, giving
dz z −n−2a
I
1+a
(− log(1−z))j Hn,j
a
(−n−4a)
z=0 2πi (1 − z)
dz z −n−2a
I
n+j
= (−1) 1+a
(− log(1−z))j Hn,j
a
(0), (C.13)
z=0 2πi (1 − z)
a
which for even n+j cancels the Hn,j (0) term and for odd n+j it doubles it. We
therefore have
open,a
Bn,j = 0 for n + j ∈ 2Z, (C.14)
and from now on we focus on the case n+j ∈ 2Z + 1. Using (C.12) we have
 −u(n+4a) 
du uj eu(n+3a) j+ D−4
I
open,a j m a−1 D−4 e
Bn,j = 2(−1) cn,j n (j + 2 )! u n+2a
∂u 2 D−2 .
u=0 2πi (1 − e ) uj+ 2
(C.15)
Recall that this expression is a result of localizing in the forward limit (t = 0) and on
the cut (s = n), which in turn localized on the worldsheet OPE (z = 0 or equivalently
u = 0).
In the final step we use integration by parts j + D−4 2
times in u, followed by
rewriting the derivative as a residue integral using the identity
I
J J dv J!
∂u f (u) = (−1) J+1
f (u − v) (C.16)
v=0 2πi v

D−4
with J = j + 2
. This results in

(v − u)j e−(u+v)a
I I
open,a du dv
Bn,j = 2cm
n,j
a−1
n [Γ(j+ D−2
2
)]2 . (C.17)
u=0 2πi v=0 2πi (uv)j+ D−2
2 (ev − eu )n+2a

– 32 –
where the constant cm n,j was given in (C.6). This yields (3.14) and (3.18) for a = 0
and 1 respectively. Note the order of taking residues: exchanging u ↔ v picks up
a factor of (−1)j+n = −1, so the two residues do not commute, which can be seen
alternatively from the fact that the integrand has a pole at u = v.
Notice that the double-contour formula is not well-defined for odd D, since the
integrand is not single-valued around v = 0. Nevertheless, let us observe that for
even D one can use the commutator of residues familiar from two-dimensional CFT
(see, e.g., [19, Sec. 6.1.2]):
I I I I I I
− = (C.18)
v=0 u=0 u=0 v=0 u=0 v=u

together with the aforementioned anti-symmetry under u ↔ v to find an alternative

expression for the partial-wave expansion coefficients

(v − u)j e−(u+v)a
I I
open,a m a−1 D−2 2 du dv
Bn,j = −cn,j n [Γ(j + 2 )] (C.19)
j+ D−2
u=0 2πi v=u 2πi (uv) 2 (ev − eu )n+2a

for odd n+j. At this stage notice that after performing the v residue, non-analyticity
in u can only come through a factor of the form (u2 )−D/2 for odd D. We found that
in such cases the correct choice of branch is −u−D . In the following sections we only
consider the cases with D ∈ 2Z.
It is time to move on to the closed-string case (3.19) with the explicit expressions
spelled out in (3.20). Our starting point is compactifying the worldsheet integration
contour close to z = 0. For the integrand at hand, locally we have

1 sa+b−2
Aclosed
a,b (s, t) = − −2πis − 1
I 2πi e I
−s−2a −t−1−a
× dz z (1 − z) dz̃ z̃ −s−2b (1 − z̃)−t−1−b + . . . , (C.20)
z=0 z̃=0

where the ellipsis denotes non-singular terms around s ∈ Z. The additional minus
sign arose because we traded an anti-holomorphic residue for a holomorphic one by
replacing z̄ → z̃. Therefore, the cut at s = n gives directly the formula (3.22). We
take n > 0 from now on, since n = 0 can be checked by hand.
closed,a,b
In computing the coefficient Bn,j , the z and z̃ variables get mixed together
by the t-integration:

dz z −n−2a dz̃ z̃ −n−2b

I I
closed,a,b j m a+b−2
Bn,j = (−1) cn,j n 1+a 1+b
z=0 2πi (1 − z) z̃=0 2πi (1 − z̃)
Z 0 j+ D−4
dt j

2 2
× t
∂t −t(t+n−4m ) .
−n+4m2 [(1 − z)(1 − z̃)]
(C.21)

– 33 –
Recall that m2 = − min(a, b). Performing integration by parts j times in t gives

dz z −n−2a dz̃ z̃ −n−2b 

I I j
closed,a,b m a+b−2
Bn,j = cn,j n 1+a 1+b
− log(1−z)− log(1−z̃)
z=0 2πi (1−z) z̃=0 2πi (1−z̃)
Z 0 j+ D−4
dt 
2 2
× t
−t(t+n−4m ) , (C.22)
−n+4m2 [(1 − z)(1 − z̃)]
| {z }
Qa,b a,b 2
n,j (0) − Qn,j (−n+4m )

since boundary terms vanish. At this stage we are interested in evaluating Qa,b
n,j (0) −
a,b 2
Qn,j (−n+4m ), where
Z t0 j+ D−4
dt  2
Qa,b 0
n,j (t ) = t
−t(t+n−4m2
) . (C.23)
[(1 − z)(1 − z̃)]
Repeating steps identical to those above, for D ∈ 2Z we obtain
2 )(u+ũ)
!
(n−4m2 )(u+ũ) j+ D−4 e−(n−4m
Qa,b
n,j (0) = −(j + D−4
2
)! e ∂u+ũ 2 D−2 , (C.24)
(u + ũ)j+ 2

where we introduced u = log(1−z) and ũ = log(1−z̃). On the other hand, the

contribution from Qa,b 2
n,j (−n+4m ) can be evaluated by sending z → z/(z − 1), z̃ →
z̃/(z̃ − 1), and t → −t−n+4m2 :

dz z −n−2a dz̃ z̃ −n−2b 

I I j
1+a 1+b
− log(1−z) − log(1−z̃) Qa,b 2
n,j (−n+4m )
z=0 2πi (1 − z) z̃=0 2πi (1 − z̃)
z −n−2a dz̃ z̃ −n−2b 
I I
j+1 dz j
= (−1) 1+a 1+b
− log(1−z) − log(1−z̃)
z=0 2πi (1 − z) z̃=0 2πi (1 − z̃)
2 2
× (1 − z)4(a+m ) (1 − z̃)4(b+m ) Qa,b
n,j (0). (C.25)

In the final line, the exponents of (1 − z) and (1 − z̃) vanish for type-II and bosonic
string theories (where a = b = −m2 ), but not in the heterotic case. Therefore, in
closed,a,b
the former cases Bn,j is zero for odd j, while no simplification happens for the
heterotic string. In order keep the notation brief, let us introduce
(
2 2 2δj∈2Z for a = b (type-II, bosonic),
ga,b := 1+(−1)j e4[u(a+m )+ũ(b+m )] = j 4ũ
1 + (−1) e for a 6= b (heterotic),
(C.26)
labelled according to (3.20). To summarize, in terms of u and ũ, we have

eu(n−a) eũ(n−b)
I I
closed,a,b j+1 m a+b−2 D−4 du dũ
Bn,j = (−1) cn,j n (j+ 2 )! u n+2a ũ n+2b
u=0 2πi (1 − e ) ũ=0 2πi (1 − e )
2
!
j −4(u+ũ)m2 j+ 2
D−4 e−(n−4m )(u+ũ)
× ga,b (u + ũ) e ∂u+ũ D−2 . (C.27)
(u + ũ)j+ 2

– 34 –
At stage we can use the identity (C.16) to rewrite the (u+ũ)-derivative as a residue,
giving
I I I
closed,a,b j m a+b−2 D−2 2 du dũ dv
Bn,j = (−1) cn,j n [Γ(j+ 2 )]
u=0 2πi ũ=0 2πi v=0 2πi
2
ga,b e−ua−ũb+v(n−4m ) (u + ũ)j
× D−2 ,
(1 − eu )n+2a (1 − eũ )n+2b [v(v − u − ũ)]j+ 2
(C.28)

which gives (3.23) and (3.24) in special cases. One might in principle massage this
formula further, e.g., by a change of variables (u, ũ, v) → (u− v2 , ũ− v2 , −v), however
the expressions do not become significantly simpler.
Alternatively, we could have arrived at a quadruple-contour representation for
closed,a,b
Bn,j as follows. After eq. (C.27) we can split the derivatives in terms of ∂u and
∂ũ :
j+ D−4
j+ D−4
j+ D−4 X j+ D−4 
2
j+ D−4 −k
−j− D−4
∂u+ũ 2 = 1
∂
2 u
+ 1
∂
2 ũ
2
=2 2 2 ∂uk ∂ũ 2 . (C.29)
k=0
k

Integrating by parts in u and ũ for each term in the above k sum and writing the
resulting derivatives as residues we find:
I I I I
closed,a,b −j− D−4 m a+b−2 D−4 2 du dũ dv dṽ
Bn,j = −2 2 c
n,j n [(j+ 2 )!]
u=0 2πi ũ=0 2πi v=0 2πi ṽ=0 2πi
D−4
2 j+
ĝa,b ea(3v−u)+b(3ṽ−ũ)+4(v+ṽ)m (v + ṽ − u − ũ)j X2 1
× ,
j+ D−2 D−2
−k
(ev − eu )n+2a (eṽ − eũ )n+2b (u + ũ) 2
k=0 v k+1 ṽ j+ 2

(C.30)

where ĝa,b = ga,b |ũ→ũ−ṽ . Finally, we use the simplification

j+ D−4 D−2 D−2

X 2
1 v −j− 2 − ṽ −j− 2

D−2 =− . (C.31)
k=0 v k+1 ṽ j+ 2
−k v − ṽ

In summary, for the type-II (a = b = 0) and bosonic (a = b = 1) cases we have

closed,a,a
Bn,j = 0 for j odd and for j even:
I I I I
closed,a,a −j− D−6 m 2(a−1) D−4 2 du dũ dv dṽ
Bn,j =2 2 cn,j n [(j+ 2 )!]
u=0 2πi ũ=0 2πi v=0 2πi ṽ=0 2πi
−a(u+ũ+v+ṽ)
(v + ṽ − u − ũ)j
 
e 1 1
× D−2 D−2 − D−2 .
(v − ṽ)(u + ũ)j+ 2 [(ev − eu )(eṽ − eũ )]n+2a v j+ 2 ṽ j+ 2
(C.32)

– 35 –
Similarly, for heterotic string (a = 0, b = 1) we get
cm
I I I I
closed,0,1 −j− D−4 n,j D−4 2 du dũ dv dṽ
Bn,j =2 2 [(j+ 2 )!]
n u=0 2πi ũ=0 2πi v=0 2πi ṽ=0 2πi
3ṽ−ũ j 3ũ−ṽ
)(v + ṽ − u − ũ)j
 
(e + (−1) e 1 1
× D−2 D−2 − D−2 .
(v − ṽ)(u + ũ)j+ 2 (ev − eu )n (eṽ − eũ )n+2 v j+ 2 ṽ j+ 2
(C.33)

D Some coefficients on the Regge trajectories

D
In this appendix, we give explicit formulas for the βn,n−∆ for the superstring (in the
notation of Section 4.3.2) and for low values of ∆ = n − j.

D n2n+D−5
∆ = 1 : βn,n−1 = , (D.1a)
2(2n + D − 5)!
D n2n+D−8 (n − D + 7)(2n + D − 8)
∆ = 3 : βn,n−3 = , (D.1b)
48(2n + D − 7)!
D n2n+D−12 (2n + D − 10)(2n + D − 10)
∆ = 5 : βn,n−5 = 5D2 n + 2D2 (D.1c)
11520(2n + D − 9)!
− 10Dn2 − 92Dn − 40D + 5n3 + 78n2 + 415n + 198 ,

D n2n+D−16 (2n + D − 12)(2n + D − 14)(2n + D − 16)

∆ = 7 : βn,n−7 = (D.1d)
5806080(2n + D − 11)!
× − 35D3 n2 − 42D3 n − 16D3 + 105D2 n3 + 1239D2 n2 + 1542D2 n
+ 624D2 − 105Dn4 − 2100Dn3 − 14269Dn2 − 18630Dn − 8048D
+ 35n5 + 903n4 + 10211n3 + 53337n2 + 73994n + 34320 .

Higher orders can be easily computed, but the expressions become quickly unwieldy.
D
The expressions can be easily translated to the actual coefficients Bn,n−∆ by multi-
D
plying with the constant cn,n−∆ .

References
[1] G. Veneziano, Construction of a crossing-symmetric, Regge behaved amplitude for
linearly rising trajectories, Nuovo Cim. A 57 (1968) 190.
[2] M. A. Virasoro, Alternative constructions of crossing-symmetric amplitudes with
regge behavior, Phys. Rev. 177 (1969) 2309.
[3] J. A. Shapiro, Electrostatic analog for the virasoro model, Phys. Lett. B 33 (1970)
361.
[4] D. J. Gross, J. A. Harvey, E. J. Martinec and R. Rohm, Heterotic String Theory. 2.
The Interacting Heterotic String, Nucl. Phys. B 267 (1986) 75.

– 36 –
 [5] P. Goddard and C. B. Thorn, Compatibility of the Dual Pomeron with Unitarity and
the Absence of Ghosts in the Dual Resonance Model, Phys. Lett. B 40 (1972) 235.
[6] K. Aoki, E. D’Hoker and D. H. Phong, Unitarity of Closed Superstring Perturbation
Theory, Nucl. Phys. B 342 (1990) 149.
[7] M. Soldate, Partial Wave Unitarity and Closed String Amplitudes, Phys. Lett. B
186 (1987) 321.
[8] M. B. Green and C. Wen, Superstring amplitudes, unitarily, and Hankel
determinants of multiple zeta values, JHEP 11 (2019) 079 [1908.08426].
[9] S. Caron-Huot, Z. Komargodski, A. Sever and A. Zhiboedov, Strings from Massive
Higher Spins: The Asymptotic Uniqueness of the Veneziano Amplitude, JHEP 10
(2017) 026 [1607.04253].
[10] T. L. Curtright and C. B. Thorn, Symmetry Patterns in the Mass Spectra of Dual
String Models, Nucl. Phys. B 274 (1986) 520.
[11] T. L. Curtright, C. B. Thorn and J. Goldstone, Spin Content of the Bosonic String,
Phys. Lett. B 175 (1986) 47.
[12] J. G. Russo and L. Susskind, Asymptotic level density in heterotic string theory and
rotating black holes, Nucl. Phys. B 437 (1995) 611 [hep-th/9405117].
[13] H. Kawai, D. C. Lewellen and S. H. H. Tye, A Relation Between Tree Amplitudes of
Closed and Open Strings, Nucl. Phys. B 269 (1986) 1.
[14] G. N. Watson, A treatise on the theory of Bessel functions. Cambridge University
Press, 1995.
[15] P. Flajolet and R. Sedgewick, Analytic combinatorics. Cambridge University press,
2009.
[16] P. Maity, Positivity of the Veneziano Amplitude in D = 4, 2110.01578.
[17] H. L. Krall and O. Frink, A new class of orthogonal polynomials: the Bessel
polynomials, Transactions of the American Mathematical Society 65 (1949) 100.
[18] S. Mizera, Aspects of Scattering Amplitudes and Moduli Space Localization, Ph.D.
thesis, Perimeter Inst. Theor. Phys., 2020. 1906.02099. 10.1007/978-3-030-53010-5.
[19] P. Di Francesco, P. Mathieu and D. Senechal, Conformal Field Theory, Springer
(1997) .

– 37 –