Smoothed asymptotics: from number theory to QFT

Antonio Padilla∗ and Robert G. C. Smith†

School of Physics and Astronomy, University of Nottingham,

University Park, Nottingham NG7 2RD, United Kingdom and

The Nottingham Centre of Gravity, University of Nottingham, Nottingham NG7 2RD, UK

Inspired by the method of smoothed asymptotics developed by Terence Tao, we

introduce a new ultra-violet regularisation scheme for loop integrals in quantum field
theory which we call η regularisation. This reveals a connection between the elim-
ination of divergences in divergent series of powers and the preservation of gauge
invariance in the regularisation of loop integrals in quantum field theory. In partic-
arXiv:2401.10981v2 [hep-th] 20 Jun 2024

ular, we note that a method for regularising the series of natural numbers so that
it converges to minus one twelfth inspires a regularisation scheme for non-abelian
gauge theories coupled to Dirac fermions that preserves the Ward identity for the
vacuum polarisation tensor and other higher point functions. We also comment on
a possible connection to Schwinger proper time integrals.

I. INTRODUCTION

The problem of infinity dates back to the sixth century BCE when Anaximander, a tutor
of Pythagoras, conjured up apeiron, an indefinite and limitless source from which everything
is born and to which everything will return [1]. However, the ancient Greeks did not embrace
the infinite with Aristotle accepting only the potential for infinity, rejecting it in actuality.
Centuries later, Gauss, perhaps the greatest mathematician of the modern era, warned that
“the use of an infinite quantity as a completed one ... is never permissible in mathematics.
The infinite is only a façon de parler, where one is really speaking of limits where certain
ratios come as close as one likes, while others are allowed to grow without restriction” [2].
As physicists, how can we heed these warnings and at the same time make sense of
divergent series and integrals that emerge unapologetically in the mathematics we use to
describe fundamental physics? Perturbative quantum field theory (QFT) is well established
as a microscopic theory of the fundamental interactions, enjoying predictive power and
stunning experimental success [3], and yet the presence of divergences is well documented.
Indeed, as Dyson noted [4], the perturbative expansion used in quantum electrodynamics
does not converge, even after renormalisation. Although this does not hinder its predictive
power when the expansion parameter is small, it does raise concerns around the concepts
upon which the theory is built. Many notable advances have been made in this regard using
Borel summation, particularly under the heading of resurgence theory [5–7]. Here the usual
perturbative expansion is replaced with a trans-series expansion, including non-perturbative
instanton contributions.
Even at finite order in perturbative QFT, we encounter divergences from the integration
∗
[email protected]
†
[email protected]
 2

over momenta running through loops. These problems were first identified for quantum
electrodynamics by Oppenheimer [8] and later solved using the method of renormalisation
developed by Tomonoga [9], Schwinger [10] and Feynman [11], and extended to non-abelian
gauge theories by ’t Hooft and Veltman [12]. As for General Relativity, this works well as a
perturbative quantum field theory at scales below the Planck scale, but cannot be extended
to arbitrarily high energies as the theory is known to be perturbatively non-renormalisable.
Said another way, for General Relativity, the number of counter-terms required to renor-
malise the divergences arising from loops is itself divergent!
Adopting Aristotle’s philosophy in rejecting actual infinity, at least in a natural con-
text, one might hope that the correct theory of quantum gravity coupled to matter will be
free of all such divergences. If this is the case, it may point towards a preferred way of
regularising the divergences in the low energy QFT. String theory has certainly provided
multiple insights in this regard, particularly in terms of a natural exponential damping of
ultra-violet (UV) divergences, which can be seen, for instance, in the modular invariance
of the worldsheet theory [13, 14]. Further, in the Gross-Mende regime [15, 16] where the
high energy, fixed angle, behaviour of string scattering amplitudes is considered, the sum
over all Riemann surfaces is dominated by a saddle point, and we recover the exponential
damping from a stringy tower of states. With a view to understanding this from a simple
particle perspective, Abel and others [17, 18] have constructed UV complete particle theo-
ries based on the Schwinger representation and worldline formalism. This includes particle
theories with a tower of worldline internal degrees of freedom that mimic stringy behaviour
in the UV, from Gross-Mende saddle points to modular invariance [18]. Finally, string field
theory has revealed how such exponential damping comes partly from adding stubs to loop
diagrams [19], which can also be translated to the point particles of QFT [20].
Mathematically rigorous and axiomatic formulations of perturbative QFT have pro-
gressed significantly in recent decades, aiding our understanding on the nature of UV di-
vergences, with the view that all regularisation in QFT is really just an effort in obtaining
sensible results for products of distributions [21]. But, with this progress and insight, not
much has been said fundamentally about regularisation besides it being a useful, if not
altogether ad hoc, tool. This philosophy does not align with the hopes outlined in the pre-
vious paragraph and the quest to identify a preferred method of regularisation inherited
from string theory, or some other consistent quantum theory of gravity. Where else can
we take our inspiration? One possibility is through analytic number theory and the study
of divergent series. In particular, we ask if there are clever ways to regularise these series
that eliminate the divergences altogether and if so, can we connect them, via QFT, to the
softening of scattering amplitudes at high energies in a fundamental microscopic theory of
nature?
A particularly interesting approach to divergent series is the method of smoothed asymp-
totics, elegantly discussed by Tao [22]. For monomial series, this makes use of the Euler-
Maclaurin summation formula and as such is closely related to Ramanujan’s methods [23, 24].
The importance of smoothed asymptotics in understanding divergent P series is best illustrated
with an example. Consider the infinite series of natural numbers, ∞ n=1 n, which was fa-
mously assigned the seemingly absurd value of −1/12 by Ramanujan. The high school
technique for evaluating infinite P
series is to take the limit of partial sums, which for the
sum of natural numbers yields, N 1
n=1 n = 2 N(N + 1). There is no sign of −1/12 in this
expression, leading many to scoff at Ramanujan’s claim. However, this is simply an artefact
 3

of the discontinuities that occur as we increase the cut-off N. To remedy this, we note that
partial sums are really just infinite series where each term is weighted by a step function
θ(n/N) that equates to unity for n ≤ N and vanishes for n > N. However, suppose we
replace the step function with another regulator function, η(n/N), where η(x) is a bounded
smooth function with compact supportP on the non-negative real line and with η(0) = 1 and
η(x) → R0 at large x. Tao shows that ∞ 2
n=1 nη(n/N) = C1 [η]N − 1/12 + O(1/N), where
∞
C1 [η] = 0 dx xη(x) is the Mellin transform of the regulator function. As we will explicitly
show, the same method can also be extended to regulator functions that are Schwartz class.
Tao’s results have several interesting features. The first concerns the divergence as N →
∞. Unlike the partial sum, there is no linear divergence, while the quadratic divergence is
dependent on the choice of regulator. This is in stark contrast to the finite term which is
universal, returning Ramanujan’s famous result. These features extend to infinite series of
polynomials and are reminiscent of a well known result regarding UV divergences in QFT. In
particular, when we compute the one loop effective action for a theory cut-off at some scale,
Λ, the couplings associated with power law divergences are not universal. This is in contrast
to the logarithmic divergences where the couplings are universal, much like the finite terms
in regularising divergent series of polynomials using smoothed asymptotics. This overlap
suggests that one loop divergences in QFT may be connected to divergent series.
The power law divergences that emerge from regularising a series with smoothed asymp-
totics can also be eliminated with a suitable choice of regulator. We call these enhanced
regulators. For the series of natural numbers, an enhanced regulator is one for which the
corresponding Mellin transform is vanishing C1 [η] = 0. As we will show, such enhanced
regulators are relatively easy to find and we present several algorithms for finding them. A
n
particularly elegant choice is the enhanced regulator η(n/N) = e− N cos(n/N). When this
regulator is used for the series of natural numbers the result converges to −1/12 and there
is no divergence whatsoever!
To build these ideas into QFT, we introduce the concept of η regularisation. Here the
integrand in (Euclidean) loop integrals is weighted by a regulator function η(|k|/Λ), where
|k| is the norm of the loop momentum and Λ is the cut-off. It is convenient to work directly
with one fold irreducible loop integrals (ILIs) introduced by Wu [25] as the basic building
block of all one particle irreducible graphs. When we implement η regularisation we see, as
expected, that power law divergences are regulator dependent while logarithmic divergences
are universal. As with divergent series and smoothed asymptotics, the power law divergences
are multiplied by Mellin transforms of the regulator. By choosing enhanced regulators, we
can eliminate the power law divergences at will.
However, the real question is whether or not the elimination of power law divergences
by enhanced regulators has any deeper meaning from a QFT perspective. Perhaps as one
might have expected from dimensional regularisation, we find that the answer is yes. In
a series of papers [25–30], Wu and others have derived a set of consistency relations that
are necessary in order for the regulator to preserve gauge invariance. Requiring that these
hold for η regularisation, we find that certain regulators must be enhanced. In other words,
regulators that allow the infinite series of natural numbers to converge towards −1/12 are
intimately connected to the preservation of gauge invariance at one loop in a wide class of
non-abelian gauge theories coupled to an arbitrary number of Dirac fermions. This is the
main result of our work.
 4

The rest of the paper is organised as follows: in section II, we review several ideas
from analytic number theory and the study of divergent series. We introduce a number of
important concepts, including a brief review of the Euler-Maclaurin summation formula and
Ramanujan summation in section II A. In section II B, we review Tao’s work on smoothed
asymptotics, extending the analysis to include regulators that are Schwartz and to series of
general polynomials. We also introduce the concept of enhanced regulators that eliminate
divergences altogether and present several algorithms for finding them, including one that
is inspired by the Schwinger proper time formalism. In section III, we switch gears to
divergent integrals in QFT, drawing analogies between these and the results that emerged
from regularising divergent series with smoothed asymptotics. We formally introduce η
regularisation in section III A and study the implications for irreducible loop integrals in
section III A 1. In section III B we implement Wu’s consistency conditions, demonstrating
the role of enhanced regulators in preserving gauge invariance. We comment explicitly on
the connection to Schwinger proper time in section III B 1. In section IV, we conclude.

II. DIVERGENT SERIES: A PHYSICIST’S REVIEW

Let us begin with the following well known expressions for the finite sums of powers:
N
X 1 1
n = N + N 2, (1)
n=1
2 2
N
X 1 1 1
n2 = N + N 2 + N 3, (2)
n=1
6 2 3
N
X 1 1 1
n3 = = N 2 + N 3 + N 4 . (3)
n=1
4 2 4

These relationships date back more than two millennia. The first of them, corresponding to a
sum of natural numbers, can be traced to the Pythagorean school in the 6th century BCE [31,
32]. Archimedes of Syracuse (circa 287-212 BCE), considered the greatest mathematician of
antiquity, discovered the second relationship for a sum of the squares. The sum of cubes can
be found in the work of Nicomachus of Gerasa (circa 60 -120 CE), along with the remarkable
theorem that bears his name !2
XN XN
n3 = n .
n=1 n=1

The expressions for the three finite sums (1) to (3) are special cases of Faulhaber’s formula
[33]
N z  
X
z 1 X z+1
n = Bn N z−n+1 , (4)
n=1
z + 1 n=0
n
where z is a positive integer and Bn are the Bernoulli numbers (with B1 = 1/2), which are
 5

defined by the following recursion relation

z  
X z+1
Bn = z + 1, B0 = 1 (5)
n=0
n

or, equivalently, from the exponential generating function

∞
t X tn
= Bn . (6)
1 − e−t n=0
n!

In the limit where we take N → ∞,P Faulhaber’s formula will obviously break down. This
is because the infinite series S(z) = ∞ z
n=1 n is known to be divergent in the Cauchy sense
whenever ℜ(z) ≥ −1. According to Cauchy’s definition, an infinite series is said to be
convergent if the sequence of partial sums converges to some finite limit; otherwise, the
series is divergent. If the series is not convergent in the Cauchy sense, then it will generally
be considered as one of two types of divergent series [24, 34]: (i) a series that grows in
absolute value without limit, or (ii) a series that is bounded but whose sequence of partial
sums does not approximate any specific value.

At high school we are taught to think of an infinite series ∞

P
PN n=1 an as the limit of its
partial sums, limN →∞ n=1 an . Partial sums are favoured because they allow us to perform
standard arithmetic without issue. Of course, the method works well for convergent series,
less so for divergent series. With the latter the task is to develop a summation method that
shares the most important properties of partial sums but allows the result to be generalised
to infinity without giving a divergent answer. Alternative summation methods PN include Ce-
saro summation, where we compute the sequence of partial sums sN = n=1 an and find
1
PN
the limiting value of their average lim N →∞ N n=1 sn ; and Abel summation, where we com-
pute limt→1− ∞ n
P
n=1 a n t . Hardy has argued that any new summation method should satisfy
three properties: regularity, linearity and stability [24]. Regularity states that a summa-
tion method yields the known P∞ results for convergent
P∞ series obtained
P∞ using partial sums. For
linearity, we require
P∞ that λa
n=1 P n + µb n =λ a
n=1 n + µ n=1 n The third property of
b .
∞
stability, namely n=1 an = a1 + n=1 an+1 , is considered less crucial and known to fail in
some important cases, including Ramanujan summation [35].

The real question is whether any meaning can be extracted from the finite results ob-
tained by these alternative methods. This is more than just a mathematical curiosity. In
Physics, the mathematical description of physical phenomena is often given in terms of di-
vergent series and/or integrals. Although the formal treatment of divergent series remains
an open question, a commitment to truncated partial sums ignores the fact that divergences
can lead to inconsistent results in physical situations. Of course, such divergences may just
be a reflection of a breakdown in the relevant mathematical description. However, in some
cases, physically consistent results can be obtained by using alternative summation meth-
ods that yield finite as opposed to infinite answers. Indeed, the infinite series of natural
numbers appears in computations of the Casimir force [36] and in the critical dimension
of bosonic
P string theory [14], where it is necessarily interpreted as the seemingly absurd
formula ∞ n=1 n = −1/12 [23, 24].

Although ∞
P
n=1 n = −1/12 is often associated with Ramanujan [23], the result is best
 6

derived using zeta function regularisation.

P When ℜ(z) > 1, the Riemann zeta function is
defined by the convergent series ζ(z) = ∞ n=1 n −z
. Of course, when ℜ(z) ≤ 1 the series is
no longer convergent. However, we can analytically continue ζ(z) into ℜ(z) < 1 using the
following formula  πz 
z z−1
ζ(z) = 2 π sin Γ(1 − z)ζ(1 − z).
2
We may now identify the series given above, S(z) = ∞ z
P
n=1 n with ζ(−z), even when it is
P∞ ζ
divergent, ℜ(z) ≥ −1. In particular,Pwe obtain S(1) = n=1 n = ζ(−1) = −1/12. More
divergent series ∞
generally, to sum a P n=1 an using zeta function regularisation, we identify
∞ −z
a convergent series n=1 an for z in some complex domain, then analytically continue the
result to z = −1.
The majority of issues with divergent series come with the transition from partial sums
to infinity, as notably exposed by Ramanujan [23, 24]. Although it does not appear that
he fully understood the validity of his asymptotic expansions, he was able to extract some
remarkable results. As we will see in a moment, Ramanujan developed a study of divergent
series based upon the Euler-Maclaurin summation formula, often employing it in creative
ways. The Euler-Maclaurin formula also plays an important role in understanding the results
of smoothed asymptotic expansions, an intuitive approach to the regularisation of divergent
series beautifully elucidated by Tao [22]. Indeed, using Tao’s methodology we will begin to
understand how a divergent series of positive numbers could ever be identified with a finite
negative number in a meaningful way.

A. Ramanujan and the Euler-Maclaurin summation formula

In the early 1730s, Euler solved one of the most intriguing mathematical puzzles of the
time: the Basel problem [37]. First formulated by Leibniz and the Bernoulli brothers [31],
the Basel problem concerns the infinite series whose terms are the reciprocal squaresPof the
natural numbers. Using calculus to relate the discrete sum of an arbitrary function f (n)
n
π2
to an integral of the form dxf (x), Euler was able to prove that ∞ 1
R P
n=1 n2 = 6 . For a
historical review of the development of some of Euler’s key ideas, see [38, 39].
From elementary calculus, we know that a sum and an integral provide first approxima-
tions to one other. In particular, the sum can be interpreted as the total area of rectangles
forming a step graph along some curve, while the integral, evaluated between end points,
can be interpreted as the area under the same curve. More formally, the Euler-Maclaurin
b
P
summation formula provides an estimation of the sum f (n) in terms of the integral
n=a
Rb
a
dx f (x) and the derivatives of the function f (x). As an estimate of how much the
trapezoid rule fails, an error term is also included given by an integral involving Bernoulli
polynomials. The formula can be written as [35]
b b m
f (b) + f (a) X Bk (k−1)
X Z
f (n) = dxf (x) + + (f (b) − f (k−1) (a)) + Rm , (7)
n=a a 2 k=2
k!

where a, b are integers such that a ≤ b and the integer m ≥ 2. We denote the kth deriva-
 7

d k
tive f (k) (x) = dx k f . The sum on the right-hand side contains the non-vanishing Bernoulli

numbers B2n (recall that B2n+1 = 0 for √ n ≥ 1). As these numbers grow asymptotically for
large k as B2n ∼ (−1)n+1 4n2n (πe)−2n πn, this sum often diverges as m → ∞, and is best
treated as an asymptotic series. The corresponding error or remainder is given by
Z b
m+1 bm (x) (m)
Rm = (−1) dx f (x), (8)
a m!

where bm (x) is the periodic function Bm (x−⌊x⌋) given in terms of the Bernoulli polynomials.
The latter are defined by the generating function
∞
text X tn
= Bn (x) . (9)
et − 1 n=0 n!

As emphasized in [22], the level of approximation offered by (7) is heavily dependent on the
asymptotic behaviour of Rm . It is often the case that the expansion remains valid even after
taking the limits a → −∞ and/or b → ∞. Generally, the integral on the right-hand side
can be evaluated in closed form in terms of elementary functions, even though the sum on
the left-hand side cannot. In the best case scenario, all the terms in the asymptotic series
can be expressed in terms of elementary functions.

The idea behind Ramanujan’s summation method is to use the Euler-Maclaurin formula
in the following asymptotic form
N N ∞
f (N) X Bk (k−1)
X Z
f (n) ∼ dxf (x) + + f (N) + C(f, a), (10)
n=1 a 2 k=2
k!

where C(f, a) is the so-called “constant of the series”. This will depend on the choice of
lower limit on the integral, first noted by Hardy [24]. We will return to this ambiguity in a
moment. This constant is identified with the corresponding infinite series in the limit where
N → ∞ such that ∞
X (R,a)
f (n) = C(f, a). (11)
n=1

The R corresponds to Ramanujan, whereas the a reflects the ambiguity in choosing the
limits on the integral. To recover C(f, a), we first define a sequence of constants Cm (f, a)
via the Euler-Maclaurin formula. In particular, we write
N N m Z ∞
f (N) X Bk (2k−1) bm (x) (m)
X Z
m
f (n) = dxf (x)+ + f (N)+(−1) dx f (x)+Cm (f, a),
n=1 a 2 k=2
k! N (2m)!
(12)
so that
1 m ∞
f (1) X Bk (k−1) bm (x) (m)
Z Z
Cm (f, a) = − dxf (x) + − f (1) + (−1)m+1 dx f (x). (13)
a 2 k=2
k! 1 m!

Assuming that f ∈ C ∞ and the integral above is convergent for sufficiently large values of
 8

m, one is able to show that Cm (f, a) becomes independent of m at large m, settling down
to C(f, a). For f (n) = nz , where z is a positive integer, it is sufficient to take any m ≥ z + 1
and evaluate the constant of the series, giving
∞ z+1 
az+1

X
z (R,a)1 z+1 1 1 X z+1 Bz+1
n = (a − 1) + − Bk = − (14)
n=1
z+1 2 z+1 k z+1 z+1
k=2

where we have used the recursion relation (5) for the Bernoulli numbers. Given that
ζ(−z) = − Bz+1
z+1
for positive integers, z, we see the connection of this result to zeta function
regularisation in the limit where a = 0.

To better understand the role played by the arbitrary real number a in general, we first
note that Z b
C(f, b) − C(f, a) = dxf (x). (15)
a
P∞
Furthermore, when the series n=1 f (n) is convergent, it is easy to see from the asymptotic
formula (10) that [35]
∞
X Z ∞
C(f, 1) = f (n) − dxf (x). (16)
n=1 1
P∞
It immediately follows that n=1 f (n) = C(f, ∞), demonstrating the fact we should
Ps choose
a = ∞ for the case of convergent series. For a divergent series with f (n) = z=0 cz nz a
polynomial of degree s, it is clear that we must take a = 0 in order to match the result
obtained via analytic continuation. When we develop the method of smooth asymptotics
advocated by Tao [22] in the next section, we will see how C(f, 0) also picks out the finite
term that is independent of the choice of cut-off.

B. Smoothed asymptotics

When we use partial sums to sum an infinite series ∞

P
PN n=1 an , we truncate the series at
some finite value N and then compute the sum n=1 an before taking N → ∞. We can
think of the partial sum as modifying the infinite series with a step function
(
1 x≤1
θ(x) = (17)
0 x>1

such that
∞
X ∞
X n N
X
an → an θ = an . (18)
n=1 n=1
N n=1

In the case of a convergent series, the partial sum tends towards a unique finite value as
we take larger and larger values of N. For a divergent series, we have seen how there exist
alternative summation that yield finite answers. Unfortunately, there is no obvious trace of
those finite values anywhere in the partial sum. This is readily
P∞seen zin Faulhaber’s formula
(4) where we see no evidence for the finite value of S(z) = n=1 n obtained via analytic
continuation. In [22], Tao argues that this is an artefact of the jump discontinuity in θ(x)
 9

and can be remedied by generalising θ(x) to a smooth regulator function, as opposed to a

sharp cut-off.

To this end, we regularise the infinite series with a smooth function, η(x), such that
∞
X ∞
X n
an → an η . (19)
n=1 n=1
N

We assume that η(x) is a smooth bounded function, defined for the non-negative real
numbers, with η(0) = 1 and η(x) → 0 at large x. Indeed, Tao assumes that η(x) is a bump
function - a smooth function with compact support, vanishing whenever x ∈ / [0, 1]. Here we
will relax this condition a little, and allow η(x) to be a Schwartz function. That means that
η(x) and all its derivatives are rapidly decreasing at large x, going to zero at infinity faster
than xα for all α < 0.

For series that were already absolutely convergent, the smoothing by η(x) does not affect
the asymptotic value [22]. However, for divergent series, smoothing can significantly
P∞ improve
the convergence properties. For example, we can easily sum Grandi’s series n=1 (−1)n−1
using a regulator function η(x) = e−x , so that the corresponding series converges towards
one half.

P∞
PsLet usz consider the impact of smoothing on a divergent series n=1 f (n) where f (n) =
z=0 cz n is a polynomial of degree s. From the Euler-Maclaurin formula (7) applied to
the function gN (x) = f (x)η(x/N) with a = 0 and b = ∞, we readily obtain the following
expression
∞ Z ∞ m
X gN (0) X Bk (k−1)
gN (n) = dx gN (x) − − gN (0) + Rm , (20)
n=1 0 2 k=2
k!
where we have used the fact that gN (x) and all of its derivatives vanish at infinity and the
remainder term is given by
Z ∞
m+1 bm (x) (m)
Rm = (−1) dx g (x). (21)
0 m! N
(k) Ps
As gN (x) = f (k) (x)η(x/N) + O(1/N), and given the polynomial form for f (n) = z
z=0 cz n ,
we choose m ≥ s + 2 and find that
∞
X n s
X
cz Cz [η]N z+1 + ζ(−z) + O(1/N),
 
f (n)η = (22)
n=1
N z=0
R∞
where Cz [η] = 0 dx xz η(x) is the Mellin transform of the regulator function. Here we
have used the fact that ζ(−z) = − Bz+1 z+1
for natural numbers z. The appearance of the
zeta function follows in a very similar way to its appearance in equation (14). In choosing
m ≥ s + 2, we guarantee that the remainder term Rm = O(1/N). This is not immediately
 10

obvious. In particular we find that

s ∞
bm (x) dm h z  x i
X Z
m+1
Rm = (−1) cz dx xη (23)
z=0 0 m! dxm N
s ∞
bm (Ny) dm z
X Z
= (−1)m+1 cz N z+1−m dy m
(y η (y)). (24)
z=0 0 m! dy
m
Now since η is Schwartz and bm is bounded, it follows that the integrand bmm!
(N y) d
dy m
(y z η (y))
is also Schwartz, and so the corresponding integral is bounded[40]. Since m ≥ s + 2, and s
is finite, we immediately see that Rm = O(1/N).
For monomials f (n) = nz , there is a single divergence as N → ∞ in the corresponding
expression for the regularised series (22), and there is a unique finite piece given by ζ(−z).
This coincides with the result obtained using Ramanujan or analytic continuation. The
connection to Ramanujan summation (with a = 0) is understood via the following identity
derived directly from (22)
∞
"∞ Z ∞ #
X (R,0) X  n   x 
nz = C(f, 0) = lim nz η − dxxz η = ζ(−z). (25)
n=1
N →∞
n=1
N 0 N

Using the method of smoothed asymptotics, we see how Ramanujan performed a delicate
cancellation between infinities that can be rigorously understood. The procedure is reminis-
cent of renormalsation in perturbative QFT.
As we will explore in more detail in the next section, another feature of equation (22)
reminiscent of QFT is the regulator dependence in the power law divergences. The regu-
lator dependence in the divergences of (22) raises the possibility that there are families of
enhanced regulators for which the power law divergences vanish altogether, just as they do
for dimensional regularisation.

R ∞This isz indeed the case: an enhanced regulator is one for which the Mellin transform Cz,η =
0
dx x η(x) vanishes for integer values of z ≥ 0. More precisely, for all natural numbers
z,
R ∞ define
we an enhanced regulator of order z to be a regulator function η[z] (x) for which
z
0
dx x η[z] (x) = 0. We immediately see from (22) that P∞if wez use an enhanced regulator
of order z to regularise the monomial series S(z) = n=0 n , there are no divergences
whatsoever. Indeed, the regularised series converges to the finite value ζ(−z) as N → ∞,
without any need to renormalise.
An elegant example of an enhanced regulator of order z is given by the function
π −θ
 
−x cot 2
z+1 cos(x + θ)
η[z] (x) = e , (26)
cos θ

Rwhere 0 < θ < π2 and z is any natural number. One can readily check that we have
∞ z
P∞ z n


0
dx x η[z] (x) = 0 and so the smoothed series n=1 n η[z] N converges to ζ(−z). We
can actually take θ to be zero, except in the case of z = 0 where non-vanishing θ is required
in order to ensure that the regulator remains Schwartz. Note that for θ = 0 and z = 1 we
recover the rather beautiful enhanced regulator of order one, η[1] (x) = e−x cos x, from which
 11

we infer that[41]
X n
n 1
lim ne− N cos =− .
N →∞
n
N 12
R∞ 1
Since 0 dx xz e−x cos x = 2− 2 (1+z) z! cos π4 (z + 1) it just so happens that e−x cos x is


also an enhanced regulator of order z = 4m+ 1 for any natural number m. Likewise, it turns
out that η[z] (x) = e−x cot( 2(z+1) ) cos x is also enhanced regulator of order z ′ = z + 2(z + 1)m
π

for any natural number m and z ≥ 1.

More generally, we can employ a few simple algorithms for finding enhanced regulators
of any given order. The first assumes R ∞that we already know an enhanced regulator of order
zero, η[0] (x), with vanishing integral 0 dxη[0] (x) = 0. By changing variables from x → xz+1 ,
we see that η[z] (x) = η[0] (xz+1 ) is an enhanced regulator of order z. The second algorithm
requires no previous knowledge of enhanced regulators: take any Schwartz function χ(x)
defined on the positive real axis with the property that χ(0) = 0 and χ(z+1) (0) = 1, then
η[z] (x) = χ(z+1) (x) is an enhanced regulator of order z. To see this, we first note that
since χ(x) is a Schwartz function, the same is trivially true of χ(z+1) (x). Furthermore, since
χ(z+1) (0) = 1, it is clear that χ(z+1) (x) has all the properties of a regulator as defined at the
beginning of this section. To see that it is an enhanced regulator, we compute the relevant
Mellin transform
Z ∞ " z #∞
X z!
dx xz χ(z+1) (x) = (−1)k xz−k χ(z−k) (x) =0 (27)
0 (z − k)!
k=0 0

verifying the defining condition for an enhanced regulator of order z.

A third and final algorithm connects enhanced regulators to Schwinger’s proper time
formalism [42]. Indeed, for any natural number n, consider a regulator of the form
Z ∞
1 du 2
λn (x) = ρ(u)(ux2 )n e−ux , (28)
(n − 1)! 0 u

where the bounded function ρ(u) → 0 faster than any power as u → 0 and ρ(u) → 1 as
u → ∞. The properties of ρ(u) guarantee that this function is Schwartz with λn (x) → 1
as x → 0+ , as required for a regulator. After changing variables from x to ux2 , it is not
difficult to show that the Mellin transform is given by
z+1 Z ∞
Z ∞

Γ n +
du ρ(u)u−( 2 ) .
z+3
dx xz λn (x) = 2
(29)
0 2(n − 1)! 0

For any natural numbers n 6= m, we can now construct an enhanced regulator of order z as
follows
(n−1)!
Γ(n+ z+1
λn (x) − Γ (m−1)! λm (x)
2 ) (m+ z+1
2 )
η[z] (x) = (n−1)! (m−1)!
. (30)
Γ(n+ z+1
−
2 ) 2 )
Γ(m+ z+1

This is clearly normalised so that it is classed as a regulator. To Rsee that it is enhanced

we simply check that the corresponding Mellin transform vanishes dx xz η[z] (x) = 0 using
(29).
 12

Although the enhanced regulators described above work well for summing specific mono-
mial series, they are less suited to eliminating divergences for generic polynomial
P series. For
z
example, the regulators given in equation (26) will yield convergent series for n n η[z] (n/N)
but not for, say, n (nz−1 + nz )η[z] (n/N). To remedy these problems for all natural num-
P
bers r ≤ s, we introduce the notion of R ∞an enhanced regulator of order [r, s] defined to be a
regulator function η[r,s] (x) for which 0 dx xz η[r,s](x) = 0 for all natural numbers z ∈ [r, s].
It now follows from (22) that if we use an enhanced regulator of order [0, s] to regularise the
series over polynomials of degree P s, there are no divergences whatsoever. P The regularised
series for the polynomial f (n) = z=0 cz n converges to the finite value sz=0 cz ζ(−z) as
s z

N → ∞Pwithout any need to renormalise. For the example given above, we find that
limN →∞ n (nz−1 + nz )η[0,z] (n/N) = ζ(1 − z) + ζ(−z) for z ≥ 1.
We can generalise the second algorithm described above to find enhanced regulators of
order [r, s]: take any Schwartz function χ(x) defined on the positive real axis with the
property that χ(k) (0) = 0 for all natural numbers k ∈ [0, s − r] and with χ(s+1) (0) = 1. It
follows that η[r,s](x) = χ(s+1) (x) is an enhanced regulators of order [r, s]. The argument for
η[r,s] (x) being a regulator is the same as for η[z] (x). To see that it is enhanced at order [r, s],
we compute the Mellin transforms for all natural numbers z ∈ [r, s] and see that they all
vanish " z #∞
Z ∞
z (s+1)
X
k z! z−k (s−k)
dx x χ (x) = (−1) x χ (x) = 0. (31)
0 (z − k)!
k=0 0
xs+1
As an example, we consider the case where χ(x) = e−x (s+1)! which generates an enhanced
regulator of order [0, s] given by
s+1 
s + 1 (−x)k
X 
−x
η[0,s] (x) = e . (32)
k=0
k k!

The existence of enhanced regulators challenges the idea that divergent series of polyno-
mials must be renormalised in order to recover their corresponding finite values. Indeed,
they explicitly demonstrate a rigorous method of taking the series towards infinity without
encountering any divergence. Of course, there is a magic of sort: the enhanced regulators
must always change sign and in just the right way in order for large positive contributions
to be cancelled by large negative ones. It is here that we are reminded of the words of
Euler who said “the quantities greater than infinity are also smaller than nothing and the
quantities smaller than infinity also correspond to the quantities greater than nothing”[43].

III. DIVERGENT INTEGRALS IN QUANTUM FIELD THEORY

Divergent integrals are encountered with unsettling regularity in interacting QFTs. As

revealed by Tomonoga [9], Schwinger [10] and Feynman [11] in the context of quantum
electrodynamics and by ’t Hooft and Veltman for non-abelian gauge theories [12], for renor-
malisable theories, these divergences can be carefully regularised and absorbed into a renor-
malisation of a finite number of couplings. The origin of these divergences is understood to
be a consequence of pushing the corresponding theory too hard, and assuming it to apply
even at the shortest distance scales, or equivalently, at the highest energies. Thanks to the
Applequist-Carazzone decoupling theorem, the effects of very heavy particles on low energy
 13

scattering processes are hidden inside a renormalisaton of the low energy couplings [44].
Said another way, even though our QFTs encounter divergences that betray limitations on
their validity at all scales, they can still be made to work well as effective field theories at
low energies.
It is interesting to note parallels between divergences in effective field theory and the
characteristic behaviour we see whenever we regularise a divergent series with a smooth
cut-off. Indeed, it is well known that power law divergences in QFT are regulator dependent
in stark contrast to the universal prediction obtained from the logarithmic divergences.
Although not universal, power law divergences do play a role in the correct matching to
UV physics in the Wilsonian EFT. They can also indicate an unwelcome sensitivity to the
details of the UV completion, leading to problems with naturalness in high energy physics
and cosmology [45].
In order to develop the analogy with our divergent series, it is instructive to examine
power law and logarithmic divergences in QFT in a little more detail. Closely following
the discussion presented in [46], we consider the Wilson action SΛ [ϕ] for modes lighter than
some scale Λ defined according to the path integral [47]
Z
iSΛ [ϕ]
e = D[ϕ]eiS[ϕ] . (33)
k>Λ

This generically contains power law and logarithmic divergences, and can be written as
Z " #
X X
SΛ [ϕ] = d4 x Λ4−di gi (Λ)Oi (x) + ln(Λ/µ) ga (Λ)Oa (x) + . . . , (34)
i a

where Oi are relevant operators of dimension di = 0, 2 and Oa are marginal operators of

dimension 4. The ellipsis denote finite terms which may also depend on µ, an arbitrarily
chosen mass scale. Now consider the Wilson action, SΛ′ [ϕ], defined at some lower scale,
Λ′ < Λ, and given by Z
′
eiSΛ [ϕ] = D[ϕ]eiSΛ [ϕ] . (35)
Λ′ <k<Λ

This is obviously independent of the choice of Λ, but highly dependent on Λ′ and corre-
spondingly takes the form
Z " #
X X
SΛ′ [ϕ] = d4 x Λ′4−di gi (Λ′ )Oi (x) + ln(Λ′ /µ) ga (Λ′ )Oa (x) + . . . . (36)
i a

It follows that SΛ′ = SΛ +∆Γ where ∆Γ is the contribution from integrating out intermediate
modes Λ′ ≤ k ≤ Λ. Focussing on the relevant and marginal operators, we must have
Z " #
X X
∆Γ = d4 x ci (Λ′ , Λ)Oi (x) + ca (Λ′ , Λ)Oa (x) + . . . , (37)
i a

where the contribution from the intermediate modes ci (Λ′ , Λ) = Λ′4−di gi (Λ′ ) − Λ4−di gi (Λ)
completely cancels the power law divergences in passing from one Wilson action to another.
This is just Wilsonian renormalisation in action - the coefficients of power-like divergences
 14

in the momentum cut-off get cancelled in the regularisation procedure (see [48] for elegant
discussions on this point). In contrast, the cancellation of the ΛPdependence for the loga-
′ ′ ′
rithmic couplings yields ca (Λ , Λ) = ln(Λ /µ) a ga (Λ ) −ln(Λ/µ) a ga (Λ′ ). In order for the
P
µ to consistently drop out to leading order, we see that there must be a universal property
ga (Λ) ∼ ga (Λ′ ) ∼ ga .

In a similar vein, if we compute the one loop effective action

Z
iΓeff [ϕ]
e = D[δϕ]eiSΛ [ϕ+δϕ] (38)
k<Λ

so that Γeff [ϕ] = SΛ [ϕ] + ∆Γeff [ϕ], we find that any power law divergences generated in ∆Γ
are cancelled by the corresponding terms in the Wilson action, SΛ . In contrast, the couplings
associated with logarithmic divergences can survive as a universal feature of the effective
action, necessarily independent of Λ.

These universal features of logarithmic divergences, in contrast to power law divergences,

are strongly reminiscent of the results we saw in the previous section. When we regularised
a divergent series with a smoothed cut-off, we saw how it was the finite terms that were
universal, yielding the results obtained via analytic continuation, in contrast to the power
law divergences that were regulator dependent. How far can we push the similarities? For
example, having introduced the enhanced regulators to eliminate divergences appearing in
number theory, it is tempting to ask if they can also play an interesting role in the loop
divergences arising in QFT and whether this can be linked to the absence of divergences
in finite theories. Indeed, the reluctance of string theory to travel deep into the UV is a
direct manifestation of the smoothing or smearing effect of the string length scale as seen
in a study of modular invariance of the worldsheet. This was first observed in string theory
by Shapiro [49], but a more modern treatment can be found in [14].

A. η regularisation in QFT

To control the ultra-violet divergences that plague perturbative QFT, several regularisa-
tion schemes are included in the literature. A sharp momentum cut-off connects intuitively
with Wilson’s understanding of the renormalisation group [50–53] but is known to break
translational invariance and, perhaps more importantly, the gauge invariance of the underly-
ing theory. The standard way to regularise loop integrals whilst preserving gauge invariance
is dimensional regularisation [12]. However, in analytically continuing the dimensionality of
the spacetime, we run into difficulties when defining quantities unique to four dimensions
(such as the γ5 matrix).

For infinite series, the method of smoothed asymptotics amounts to replacing an infinite
sum with a weighted infinite sum
∞
X X∞ n
#→ η #. (39)
n=0 n=0
N

The generalisation of this to divergent loops is straightforward: working in Euclidean signa-

 15

ture, we simply replace the loop integral with a weighted loop integral
 
|k|
Z Z
4 4
d k# → dkη #, (40)
Λ

where |k| is the norm of the Euclidean four-momentum and Λ is a cut-off scale. Of course,
if η were just a step function, this would be nothing more than a sharp momentum cut-off.
However, inspired by the method of smoothed asymptotics discussed in the previous section,
we shall assume that η is a smooth regulator, corresponding to a Schwartz function that
equates to unity at the origin. We dub this η regularisation.

The idea of smoothing the cut-off is, of course, nothing new although in most applications,
this is implemented at the level of the propagator. For example, in his derivation of the
exact renormalisation group equation, Polchinski makes use of a modified propagator that
is exponentially damped at high momenta [54]. As a modification of the integral measure, η
regularisation is similar to dimensional regularisation albeit with the advantage that it does
not require us to analytically continue the dimension of spacetime. However, the proposed
scheme has most in common with the smooth operator regularisation method inspired by [55]
and developed in [56, 57] using Schwinger proper time [42]. Gauge invariance is also preserved
in this case, as it is with dimensional regularisation. This is because the regularisation is
transferred to the proper time integral, leaving the (gauge invariant) momentum integral
unchanged. For a recent application of the smooth operator regularisation and further
discussion on symmetry preservation, see [58].

Does η regularisation also preserve gauge invariance? Generically, it seems obvious that
it will not. Just as we saw for sharp momentum cut-offs, gauge invariance is typically broken
as soon as we regularise the momentum integral directly. However, if η regularisation has
any chance of opening up a better understanding of how divergences are absent in string
theory, it must, at the very least, admit a preferred set of regulators that preserve gauge
invariance. This may seem unlikely at first glance, although as we will see in a moment, it
will be possible to preserve gauge invariance and key to that are the enhanced regulators
that eliminate divergences. We will also see how this connects to regularisation of Schwinger
proper time integrals.

1. Irreducible loop integrals and eliminating divergences

To develop η regularisation in more detail, it is convenient to introduce the concept of

irreducible loop integrals (ILIs) [25, 26, 29]. In general, n-fold ILIs are defined as the n-loop
integrals for which there are no longer the overlapping factors (ki −kj +p) in the denominator
of the integrand and no factors of the scalar momentum k 2 in the numerator [29]. In this
work we focus on regularising ultra-violet divergences at one-loop, postponing a detailed
discussion of higher loops to future work [59]. It was shown in [25] that upon use of the
Feynman parameter method, all one-loop perturbative Feynman integrals of the one-particle
 16

irreducible graphs can be evaluated as the following one-fold ILIs in Minkowski spacetime:

d4 k 1
Z
2
I−2α (M ) = , (41)
(2π) (k + M2 )2+α
4 2

d4 k kµ kν
Z
µν 2
I−2α (M ) = , (42)
(2π)4 (k 2 + M2 )3+α
d4 k kµkν kρkσ
Z
µνρσ 2
I−2α (M ) = , (43)
(2π)4 (k 2 + M2 )4+α

where the subscript (−2α) labels the power counting dimension (of energy-momentum) with
α = −1 and α = 0 corresponding to quadratic and logarithmically divergent integrals. The
mass term M2 = M2 (m21 , p21 , . . . ) is a function of Feynman parameters, external momenta,
pi and corresponding mass scales, mi . Note that k 2 = gµν k µ k ν where the metric gµν is
written with mostly positive signature.
The ILIs of tensor type are related to scalar integrals in the usual way:

1 µν d4 k k2
Z
µν 2
I−2α (M ) = g , (44)
4 (2π)4 (k 2 + M2 )3+α
1 µνρσ d4 k k4
Z
µνρσ 2
I−2α (M ) = S , (45)
4! (2π)4 (k 2 + M2 )4+α

where S µνρσ = g µν g ρσ +g µρ g νσ +g µσ g νρ is a totally symmetric tensor of rank four. Notice that

these relations are merely a consequence of the rotational symmetries of the four dimensional
space and not to be confused with the gauge consistency relations derived in [25, 26, 29].
As we will see later, the latter can be accommodated by using different ηs for regularising
different ILIs.
When implementing the η regularisation, we Wick rotate the integrals to Euclidean signa-
··· ···
ture k0 → ik4 and insert a factor of η(|k|/Λ) in the integrand. This yields I−2α → iJ−2α [η],
where the form of η need not be universal for all one-fold ILIs, at least in principle. In
particular, we now have that

d4 k
 
1 |k|
Z
2
J−2α [η](M ) = 4 2 2 2+α
η , (46)
(2π) (k + M ) Λ
d4 k k2
 
1 µν |k|
Z
µν 2
J−2α [η](M ) = g 4 2 2 3+α
η , (47)
4 (2π) (k + M ) Λ
d4 k k4
 
1 µνρσ |k|
Z
µνρσ 2
J−2α [η](M ) = S η , (48)
4! (2π)4 (k 2 + M2 )4+α Λ

where the integration is over four dimensional Euclidean space in each case. Making use
of partial fractions, the regularised tensor ILIs can be written explicitly in terms of scalar
counterparts

µν 1 µν 
[η](M2 ) = g J−2α [η](M2 ) − M2 J−2(α+1) [η](M2 ) ,

J−2α (49)
4
µνρσ 1
[η](M2 ) = S µνρσ J−2α [η](M2 ) − 2M2 J−2(α+1) [η](M2 ) + M4 J−2(α+2) [η](M2)
 
J−2α (50).
4!
 17

To understand the role of divergences, it is now sufficient to study the scalar integrals of the
form (46). After integrating out the three-sphere, these can be written as
Z ∞
k3
 
2 1 |k|
J−2α [η](M ) = 2 dk 2 2 2+α
η . (51)
8π 0 (k + M ) Λ

For α > 0 the integrals are convergent as Λ → ∞ and one readily obtains
1
J−2α [η](M2 ) ∼ . (52)
16π 2α(1 + α)M2α

For α ≤ 0, the integrals diverge as Λ → ∞, where they take the following asymptotic form
 
2 1 1
J0 [η](M ) ∼ ln(Λ/|M|) + γ[η] − , (53)
8π 2 2
1  2
J2 [η](M2 ) ∼ 2

Λ C 1 [η] − M (ln(Λ/|M|) + γ[η]) , (54)
8π 2
and " s   #
2 1 X s 2(s−z) 2z+4
J2s+4 [η](M ) ∼ 2 C2z+3 [η]M Λ (55)
8π z=0 z
for any natural number s. These expressions are valid provided η is a regulator: a smooth
Schwartz function with η(0) = 1. Here we have also defined the finite integral
Z ∞
γ[η] = − dx η ′ (x) ln x, (56)
0
R∞
and recall that Cz [η] = 0 xz η(x) is the corresponding Mellin transform for any natural
number z. We immediately see the parallels with the divergent series in the previous section.
Power law divergences are regulator dependent and weighted by the corresponding Mellin
transform. These can always be eliminated with a judicious use of enhanced regulators.
Logarithmic divergences are independent of the regulator, just as the finite terms were in the
divergent series. This universal form for the log divergences agrees with the corresponding
terms obtained by other regularisation methods, such as dimensional regularisation. Of
course, for our QFT integrals, there are also finite terms, although unlike those that appear
in the divergent series, they are regulator dependent. This is a direct consequence of the
logarithm and the correction to the finite term that arises when we rescale the cut-off Λ.
Indeed, for any regulator η(x) and real number λ > 0, we can define another regulator
ηλ (x) = η(λx) which is equivalent to rescaling the cut-off. It then follows that γ[ηλ ] =
γ[η] − ln λ while Cz [ηλ ] = λ−1−z Cz [η]. These relations will be useful when we consider the
role of gauge invariance in the next section.

B. Gauge invariant η regularisation

To investigate how gauge invariance is affected by η regularisation, we follow [25, 26, 29]
and consider a general gauge theory where the gauge group has dimension dG and where Nf
Dirac spinors Ψn (n = 1, . . . , Nf ) are interacting with the Yang Mills field Aaµ (a = 1, . . . , dG ).
 18

Such a theory is described by a Lagrangian

1 a µν
L = ψ̄n (iγ µ Dµ − m)ψn − Fµν Fa , (57)
4
where
a
Fµν = ∂µ Aaν − ∂ν Aaµ − gfabc Abµ Acν , Dµ ψn = (∂µ + igT a Aaµ )ψn , (58)
and T a are the generators of the gauge group whose commutator [T a , T b ] = if abc T c defines
the structure constants f abc . A careful computation of the vacuum polarisation for the gauge
field at one-loop yields an expression of the form [25, 26, 29]

Πab (g)ab (f )ab

µν (p) = Πµν (p) + Πµν (p), (59)
(g)ab
where pµ is the external momentum. Here Πµν (p) are the pure Yang Mills contributions
(f )ab
coming from gauge field loops and ghost loops. Πµν (p) are the contributions from fermion
loops, arising from the interaction of the fermions with the gauge field. Gauge invariance is
understood in terms of the Ward identities pµ Πab ab ν
µν = Πµν p = 0. Requiring this to hold for
any gauge theory and with any number of fermions means that Ward identities should hold
separately for the gauge field and fermionic contributions

pµ Π(g)ab
µν = Π(g)ab ν
µν p = 0, pµ Π(f
µν
)ab
= Π(f )ab ν
µν p = 0. (60)
···
When the vacuum polarisation is computed in terms of the regularised ILIs, I−2α |regularised,
these generalized Ward identities impose strict consistency conditions on the regularisation
scheme. There are also generalised Ward identities and corresponding consistency conditions
associated with three and four point functions. Altogether, we find that for α = −1, 0, 1 we
must have (see Appendix A and [25, 26, 29, 58] for further details)

µν 1
I−2α |regularised ∼ g µν I−2α |regularised, (61)
2(α + 2)
µνρσ 1
I−2α |regularised ∼ S µνρσ I−2α |regularised, (62)
4(α + 2)(α + 3)

in the asymptotic limit. We now wish to apply these conditions in the context of η regu-
larisation. As mentioned earlier, it is important to note that, in principle, different ILIs use
different regulators. At this stage, we do this in order to keep things sufficiently general. As
we will see shortly, the Ward identities impose relations between different ηs for different
loop topologies, which may point to a deeper underlying structure. We will investigate the
meaning of these relations in more detail in our forthcoming work [59].

In the following we denote the regulators for scalar ILIs I−2α by η−2α and the regulators
for the corresponding tensor ILIs of rank two and four by θ−2α and κ−2α respectively. It
follows that

I−2α |regularised = iJ−2α [η−2α ], (63)

µν µν
I−2α |regularised = iJ−2α [θ−2α ], (64)
µνρσ µνρσ
I−2α |regularised = iJ−2α [κ−2α ], (65)
 19

We now impose the conditions (61) and (62) making use of the decomposition formulae (49)
and (50) giving
α h i
J−2α θ̃−2α ∼ M2 J−2(α+1) [θ−2α ], (66)
α+2
α(α + 5)
J−2α [κ̃−2α ] ∼ 2M2 J−2(α+1) [κ−2α ] − M4 J−2(α+2) [κ−2α ], (67)
(α + 2)(α + 3)

as Λ → ∞ and where
(α + 2)θ−2α − 2η−2α (α + 2)(α + 3)κ−2α − 6η−2α
θ̃−2α = , κ̃−2α = . (68)
α α(α + 5)

Note that for α 6= 0 the combination of regulators appearing in (68) are themselves regula-
tors, being Schwartz and equating to unity at the origin, θ̃−2α (0) = κ̃−2α (0) = 1. It remains
to plug in the asymptotic expressions for the scalar ILIs given by (52), (53), (54) and (55).
For the convergent ILIs, with α ≥ 1, we can use the asymptotic formula (52) to show that
the consistency relations hold automatically in the limit as Λ → ∞, as, of course, they
should. For α = 0, the conditions (66) and (67) can be understood via the α → 0 limit,
giving

J0 [θ0 ] − J0 [η0 ] ∼ M2 J−2 [θ0 ], (69)

J0 [κ0 ] − J0 [η0 ] ∼ 2M2 J−2 [κ0 ] − M4 J−4 [κ0 ]. (70)

Plugging in the asymptotic formulae, we see that there are logarithmic divergences, although
they cancel and the leftover finite parts yield the following constraints
1 5
γ[θ0 ] − γ[η0 ] = , γ[κ0 ] − γ[η0 ] = . (71)
4 12
For α = −1, we get both quadratic and logarithmic divergences. The latter cancel when we
impose the two consistency conditions (66) and (67), just as they did for α = 0, and we are
left with
 
2 2 1
Λ C1 [θ̃2 ] + 2M γ[θ2 ] − γ[η2 ] − ∼ 0, (72)
4
 
2 3 2 5
Λ C1 [κ̃2 ] + M γ[κ2 ] − γ[η2 ] − ∼ 0, (73)
2 12

where we recall that θ̃2 = 2η2 − θ2 and κ̃2 = 32 η2 − 12 κ2 . Cancellation of the quadratic
divergences requires that θ̃2 and κ̃2 are enhanced regulators of order one

C1 [θ̃2 ] = C1 [κ̃2 ] = 0. (74)

Here we see for the first time the connection between gauge invariance and the elimination
of quadratic divergences in both divergent series and ILIs. The remaining finite parts in
 20

(72) and (73) yield constraints that are very similar to (71)
1 5
γ[θ2 ] − γ[η2 ] = , γ[κ2 ] − γ[η2 ] = . (75)
4 12
At this time, we have not been able to develop any further insight into the consistency
conditions arising for the finite parts, given by (71) and (75). However, using the fact that
γ[ηλ ] = γ[η] − ln λ and Cz [ηλ ] = λ−1−z Cz [η] for any regulator ηλ (x) = η(λx) with a rescaled
cut-off, we can arrive at a rather elegant solution to the full set of consistency conditions.
For α ≥ −1, this is given by

η−2α (x) = η[1] (x), θ−2α (x) = η[1] (λx), κ−2α (x) = η[1] (µx), (76)

where λ = e−1/4 , µ = e−5/12 , and η[1] (x) is any enhanced regulator of order one.

1. Connecting to Schwinger proper time

It was noted in [58] that regularised Schwinger proper time integrals provide an elegant
framework for satisfying Wu’s consistency relations (61) and (62). In particular, we write
the regularised ILIs as follows
Z ∞
i dτ d4 k −τ (k2 +M2 )
Z
2 2 2+α
I−2α (M )|regularised = ρ(Λ τ )τ e , (77)
(α + 1)! 0 τ (2π)4
Z ∞
i dτ d4 k µ ν −τ (k2 +M2 )
Z
µν 2 2 3+α
I−2α (M )|regularised = ρ(Λ τ )τ k k e , (78)
(α + 2)! 0 τ (2π)4
Z ∞
i dτ d4 k µ ν ρ σ −τ (k2 +M2 )
Z
µνρσ 2 2 4+α
I−2α (M )|regularised = ρ(Λ τ )τ k k k k e , (79)
(α + 3)! 0 τ (2π)4

where ρ(u) → 0 faster than any power as u → 0 and ρ(u) → 1 as u → ∞. To make contact
with the current work, we note that we can write these formulae in the form of equations
(63) to (65), provided we generalise the regulators as follows:
R ∞ du 2 2
0 u
ρ(u)[u(x2 + M2Λ )]α+2 e−u(x +MΛ )
η−2α (x) → η−2α (x, MΛ ) = R ∞ du 2 α+2 −uMΛ 2 , (80)
ρ(u)[uM Λ ] e
R ∞ du 0 u 2 2 2
0 u
ρ(u)[u(x + M2Λ )]α+3 e−u(x +MΛ )
θ−2α (x) → θ−2α (x, MΛ ) = R ∞ du 2 , (81)
0 u
ρ(u)[uM2Λ ]α+3 e−uMΛ
R ∞ du 2 2
0 u
ρ(u)[u(x2 + M2Λ )]α+4 e−u(x +MΛ )
κ−2α (x) → κ−2α (x, MΛ ) = R ∞ du 2 , (82)
0 u
ρ(u)[uM2Λ ]α+4 e−uMΛ

where u = Λ2 τ , x = |k|/Λ, and MΛ = |M|/Λ. These do not have the precise form of the
η regulators described in this paper as the Λ dependence is spread between x = |k|/Λ and
MΛ = |M|/Λ. We will consider much more general regulators in our forthcoming work [59],
but for now let us examine the form of these regulators in the limit where MΛ → 0. In
 21

particular, note that in this limit, the denominators can generically be written as
Z ∞ Z ∞  
du 2
2 n −uMΛ dy y
ρ(u)[uMΛ ] e = ρ 2
y n e−y (83)
0 u y M
Z0 ∞ Λ
dy n −y
∼ y e = (n − 1)! (84)
0+ y

It follows that as MΛ → 0, the regulators fall into the class described by equation (28), with

η−2α (x, 0) = λ2+α (x), θ−2α (x, 0) = λ3+α (x), κ−2α (x, 0) = λ4+α (x). (85)

Using the expression (30), it is easy to verify that θ̃2 = 2η2 (x, 0) − θ2 (x, 0) and κ̃2 =
3
η (x, 0) − 21 κ2 (x, 0) are enhanced regulators of order one, consistent with the gauge consis-
2 2
tency condition (74). Of course, this had to be true given the claim in [58] that regularised
Schwinger proper time integrals automatically satisfy Wu’s consistency relations but it serves
as a nice check of our formalism.
We can also compute the finite integral (56) for each of the regulators in (85). Indeed,
from (28), we change the integration variable to y = ux2 and write
Z ∞
1 dy
λn (x) = ρ(y/x2 )y n e−y , (86)
(n − 1)! 0 y

to then find that

Z ∞ Z ∞
1 yn
γ[λn ] = dx dyρ′ (y/x2) 3 e−y (2 ln x) (87)
(n − 1)! 0 x
Z ∞ Z0 ∞
1
= dy duρ′(u)y n−1e−y (ln y − ln u) (88)
2(n − 1)! 0
 Z ∞ 0 
1 ′
= Ψ(n) − duρ (u) ln u , (89)
2 0

where Ψ(z) = Γ′ (z)/Γ(z) is the digamma function and we have used the fact that ρ(u) → 0
as u → 0 and ρ(u) → 1 as u → ∞. It immediately follows that
1
γ[θ−2α (x, 0)] − γ[η−2α (x, 0)] = , (90)
2α + 4
2α + 5
γ[κ−2α (x, 0)] − γ[η−2α (x, 0)] = . (91)
2 (2 + α) (α + 3)

For the logarithmically divergent ILIs corresponding to α = 0, the expressions (90) and (91)
satisfy the finite parts of the consistency condition, given by (71). However, for the quadrat-
ically divergent ILIs corresponding to α = −1, they do not agree with the corresponding
consistency conditions (75). The difference is easily understood. Although [58] has shown
that ILIs regularised using Schwinger proper time automatically satisfy Wu’s consistency
relations, recall that this is not exactly equivalent to η regularisation with regulators of the
form η(|k|/Λ) for which we derived the consistency conditions (74), (71) and (75). The
equivalence only emerges in the limit MΛ → 0, suggesting that in the limit of large Λ,
only the leading order constraints should align. This is indeed the case. For quadratically
 22

divergent ILIs, the leading order constraint is (74) and this holds for the derived regulators,
η2 (x, 0), θ2 (x, 0) and κ2 (x, 0), whereas the subleading constraint (75) does not, exactly as
expected. For logarithmically divergent ILIs, the leading order constraint is (71), which we
have just seen to hold for the derived regulators, η0 (x, 0), θ0 (x, 0) and κ0 (x, 0), just as it
should.

IV. CONCLUSIONS

We have explored possible connections between analytic number theory and the study
of divergent series and the ultra-violet regularisation of loop integrals in perturbative quan-
tum field theory. On the number theory side, we have extended Tao’s work on smoothed
asymptotics [22] which offers a tantalising taste of QFT. Both exhibit regulator dependence
of power law divergences, while the universal features of finite terms in Tao’s study of diver-
gent series mirror the universal features of logarithms in QFT. However, our analysis runs
much deeper than this elegant analogy. Inspired by Tao’s work, we have developed a new
and general method for regularising divergent integrals in QFT which we dub η regularisa-
tion. As a result, we have demonstrated a connection between the regularisation of divergent
series and the elimination of divergences in analytic number theory and the preservation of
gauge invariance at one loop in a regularised quantum field theory. Is this just a coinci-
dence, or does it signal something deeper and a possible window into how divergences are
eliminated at high energies in a complete microscopic theory?
There is certainly much more to be learnt before we can answer this question. Indeed, it is
notable that the consistency relations for preserving gauge invariance, at least in the context
of η regularisation, also include a constraint on the finite terms, as well as the divergences.
We have not been able to glean a deeper understanding of the meaning of these additional
constraints, beyond noting that they are satisfied in the appropriate limit by η regulators
inspired by Schwinger proper time integrals.
The link to Schwinger proper time integrals is particularly intriguing, not least because
of an intuitive connection between the worldline formulation of QFT and the field theory
limit of string theory. We also note how regularised Schwinger proper time techniques
have been used to improve the behaviour of functional RG equations for gauge theories
at the UV cut-off [60]. It would be interesting to see if a functional RG equation derived
using enhanced η regulators enjoyed similar properties and to explore the implications for
Weinberg’s asymptotic safety program.
In a forthcoming publication, we will extend our study of η regularisation to a much
broader class of regulators, going beyond the structure inspired by Tao [59]. This will allow
us to capture all known methods of regularisation and more beyond. Implementing Wu’s
consistency relations [25, 26, 29] in this general setting will also be shown to yield a master
equation for the regularisation scheme in the form of a differential equation. Finding solu-
tions will be the same as finding gauge invariant regulators, which will include dimensional
regularisation, as well as the enhanced η regulators discussed here. We will also describe
how η regularisation can be understood in Lorentzian signature, as well as its extension to
arbitrary loop order, consistent with unitarity, locality, and causality.
Let us say a little more about higher loops and, in particular, difficulties associated with
overlapping divergences that begin at two-loop order. Key insights in this regard were
 23

developed in the symmetry preserving regularisation scheme developed by Wu [25], which

has much in common with our generalised approach. In a future study of how to apply
η regularisation to higher loop graphs, careful treatment of overlapping divergences and
subdivergences will exploit generalisations of the ILI formalism. Moreover, at one-loop we
have seen that the ILIs encode the overall UV contribution and tensor structure of Feynman
integrals. Regularisation then follows from a simple redefinition of the measure in the master
integrals. At two-loops and higher, this shall be seen to be similarly true albeit in a much
more general way. In the repeated use of Feynman parametrisation an important step will be
to ensure no Feynman parameter integration contains UV divergences and that appropriate
subtractions can be made to show all overlapping divergences are harmless. Using the α, β, γ
technology first introduced by ‘t Hooft and Veltman [61], and following the procedures set
out in [25], including key theorems based on the ILI formalism for factorisation, subtraction,
and reduction of overlapping divergences, we will show how η regularisation can be extended
to meet such demands. It will be seen how the treatment of overlapping Feynman integrals
in generalised η regularisation requires certain technical requirements different from those
found in the loop regularisation by Wu [25].

There are several ways in which we may continue to build the bridge towards string
theory and the softening of amplitudes at high energies. Given that critical string theory is
a fundamentally higher dimensional theory, it would be interesting to explore the dimensional
dependence of Wu’s consistency relations and the implications on η regularisation. It is also
important to better understand the role played by gravity. Is there a gravitational analogue
of Wu’s consistency relations that guarantees diffeomorphism invariance is preserved by the
regulator? Further, how do we implement η regularisation in curved space?

The work of [17, 18] could open up a path connecting η regularisation (in the Schwinger
representation) to string theory via non-local particle theories that preserve the higher di-
mensional properties of strings. In this context, we are particularly interested in those
regulators that exhibit some remnant of modular invariance, such as those presented in [62].
We would also like to deepen this analysis and say something about the nature of QFT and
the effect of smoothing in position space.

Although our main focus is on developing QFT and building a bridge towards string
theory, there is much to be explored at the level of analytic number theory. For example, it
would be interesting to see if any other gauge invariant regulators identified in [59] have a role
to play in the suppression of divergences in divergent series. Further, how do these enhanced
regulators connect to analytic continuation? Tao has already provided some insight in this
regard [22]. To develop more understanding it would be useful to extend Tao’s results to
series of non-polynomial functions such as logarithms and polylogarithms, which are also
expected to appear in applications to QFT. What relation, if any, can then be made to
resurgence theory and trans-series? In particular, it would be interesting to see if any
connection can be made between η regularisation and some general resummation procedure
that might restore some, or all, non-perturbative information.
 24

Appendix A VACUUM POLARISATION AND GENERALISED WARD

IDENTITIES

Following [25], the contribution of Nf equal mass fermion loops to the gauge boson self-
energy is given by

d4 k Tr[γµ (k/ + m)γν (k/ + p/ + m)]

Z
(f )ab a b 2
Πµν (p) = 4Nf Tr(T T )(ig) . (92)
(2π)4 (k 2 + m2 − iǫ)((k + p)2 + m2 − iǫ)

After expressing the colour charge factor as Tr(T a T b ) = C2 δ ab for some constant C2 and com-
puting the Dirac algebra, we use Feynman parametrisation to write the remaining integrals
in terms of one-fold ILIs as clearly laid out in [25]. Doing so gives
Z 1  
(f )ab 2 ab 2
Πµν (p) = −4Nf g C2 δ dx 2I2µν (M) − gµν I2 (M) + 2x(1 − x)(p gµν − pµ pν )I2 (M) ,
0
(93)
where the mass factor M = m2 − x(1 − x)p2 in the ILIs includes a contribution from the
fermion mass m, the external momentum, pµ and the Feynman parameter, x. The derivation
of this formula requires several comments. Here and in [25], the formula is presented in terms
of unregulated integrals. However, these integrals are divergent. Strictly speaking, the
chain of manipulations can only be done for convergent integrals, which in our case would
correspond to a particular form of η regularisation [63]. In future work we will explore
whether or not this choice can generalise some of our conclusions. Further, one might also
worry that the regulator breaks the invariance under translations of loop momenta. However,
as we will explain in more detail in [59], η regulators can be rendered translation invariant
by redefining the origin of momentum space.

Taking the regularised form of (93), we have that

Z 1 
(f )ab
Πµν (p)|regularised ∝ dx 2I2µν (M)|regularised − gµν I2 (M)|regularised
0

2
+ 2x(1 − x)(p gµν − pµ pν )I0 (M)|regularised , (94)

As stated in the main text, we now require that this expression satisfies the generalised
Ward identity (60) asymptotically,

pµ Π(f
µν
)ab
= Π(f )ab ν
µν p = 0. (95)

We immediately see that the logarithmically divergent terms on the second line of (94)
vanish automatically when contracted with the external momenta and do not affect the
generalised Ward identity (60). However, this is not the case with the two quadratically
divergent terms on the first line of (94). These must cancel asymptotically in order to
preserve gauge invraiance, giving the condition (61) for α = −1,
1
I2µν (M)|regularised ∼ g µν I2 (M)|regularised (96)
2
 25

Note that the full set of conditions given by (61) and (62) for α = −1, 0, 1 are obtained
from the remaining generalised Ward identities coming from gauge field contributions to the
vacuum polarisation tensor and from higher point correlation functions.

ACKNOWLEDGEMENTS

We are grateful to John Barrett, Pete Millington, Killian Moehling, Lubos Motl, Ben-
jamin Muntz, Terence Tao and Paul Saffin for useful comments and discussions. RGCS
was supported by a Bell Burnell Studentship and AP by STFC consolidated grant number
ST/T000732/1. For the purpose of open access, the authors have applied a CC BY public
copyright licence to any Author Accepted Manuscript version arising. No new data were
created during this study.

[1] A. W. Moore, The infinite (Routledge, 2018).

[2] W. C. Waterhouse, Historia Mathematica 6, 430 (1979).
[3] R. L. Workman et al. (Particle Data Group), PTEP 2022, 083C01 (2022).
[4] F. J. Dyson, Phys. Rev. 85, 631 (1952).
[5] D. Dorigoni, Annals Phys. 409, 167914 (2019), arXiv:1411.3585 [hep-th].
[6] G. V. Dunne and M. Ünsal, PoS LATTICE2015, 010 (2016), arXiv:1511.05977 [hep-lat].
[7] M. Mariño, Fortsch. Phys. 62, 455 (2014), arXiv:1206.6272 [hep-th].
[8] J. R. Oppenheimer, Phys. Rev. 35, 461 (1930).
[9] S. Tomonaga, Prog. Theor. Phys. 1, 27 (1946).
[10] J. S. Schwinger, Phys. Rev. 73, 416 (1948).
[11] R. P. Feynman, Phys. Rev. 74, 1430 (1948).
[12] G. ’t Hooft and M. J. G. Veltman, Nucl. Phys. B 44, 189 (1972).
[13] E. D’Hoker and J. Kaidi, (2022), arXiv:2208.07242 [hep-th].
[14] J. Polchinski, String theory. Vol. 1: An introduction to the bosonic string , Cambridge Mono-
graphs on Mathematical Physics (Cambridge University Press, 2007).
[15] D. J. Gross and P. F. Mende, Phys. Lett. B 197, 129 (1987).
[16] D. J. Gross and P. F. Mende, Nucl. Phys. B 303, 407 (1988).
[17] S. Abel and N. A. Dondi, JHEP 07, 090, arXiv:1905.04258 [hep-th].
[18] S. Abel and D. Lewis, JHEP 12, 069, arXiv:2007.07242 [hep-th].
[19] H. Erbin, String Field Theory: A Modern Introduction , Lecture Notes in Physics, Vol. 980
(2021) arXiv:2301.01686 [hep-th].
[20] C. Chiaffrino and I. Sachs, JHEP 06, 120, arXiv:2108.04312 [hep-th].
[21] J. I. Latorre, C. Manuel, and X. Vilasis-Cardona, Annals Phys. 231, 149 (1994),
arXiv:hep-th/9303044.
[22] T. Tao, Compactness and contradiction (American Mathematical Soc., 2013).
[23] B. C. Berndt, Results in Mathematics 6, 7 (1983).
[24] G. H. Hardy, Divergent series, Vol. 334 (American Mathematical Soc., 2000).
[25] Y.-L. Wu, Int. J. Mod. Phys. A 18, 5363 (2003), arXiv:hep-th/0209021.
[26] Y.-L. Wu, Mod. Phys. Lett. A 19, 2191 (2004), arXiv:hep-th/0311082.
[27] D. Huang and Y.-L. Wu, Eur. Phys. J. C 72, 2066 (2012), arXiv:1108.3603 [hep-ph].
 26

[28] D. Huang, L.-F. Li, and Y.-L. Wu, Eur. Phys. J. C 73, 2353 (2013), arXiv:1210.2794 [hep-ph].
[29] Y.-L. Wu, Int. J. Mod. Phys. A 29, 1430007 (2014), arXiv:1312.1403 [hep-th].
[30] D. Bai and Y.-L. Wu, Eur. Phys. J. C 77, 617 (2017), arXiv:1706.06798 [hep-ph].
[31] D. J. Pengelley, arXiv:1912.03527 [math.HO].
[32] M. Santander, Unpublished Notes (2017).
[33] D. E. Knuth, Mathematics of Computation 61, 277 (1993).
[34] R. L. Graham, D. E. Knuth, O. Patashnik, and S. Liu, Computers in Physics 3, 106 (1989).
[35] B. Candelpergher et al., Ramanujan summation of divergent series, Vol. 2185 (Springer, 2017).
[36] M. D. Schwartz, Quantum Field Theory and the Standard Model (Cambridge University Press,
2014).
[37] L. Euler, On the sums of series of reciprocals (2008), arXiv:math/0506415 [math.HO].
[38] A. Knoebel, R. Laubenbacher, J. Lodder, and D. Pengelley, Mathematical masterpieces: Fur-
ther chronicles by the explorers (Springer Science & Business Media, 2007).
[39] G. Ferraro, Historia mathematica 25, 290 (1998).
[40] If F (x) is Schwartz then for each n ≥ 0, |xn F (x)| < Qn for finite constants Qn . Consider
R∞ R1 R∞
the corresponding integral 0 dxF (x) = 0 dxF (x) + 1 dxF (x). F (x) is smooth so the first
R1
integral running over a finite range, 0 dxF (x), is bounded. As for the second integral, note
R∞ R∞ R∞
that 1 dxF (x) ≤ 1 dx|F (x)| < 1 dxQn /xn = Qn /(n−1). Taking n ≥ 2 we immediately
R∞ R∞
see that 1 dxF (x) is bounded. It follows that 0 dxF (x) is bounded.
[41] To our best knowledge this remarkable regulator does not formally appear in the literature.
However, in the process of this research, we found this formula appears on Wiki without
citation or context. It is also the topic of discussion on this mathematics.stackexchange post,
in which the author notices the nature of the function being Schwartz and also relates it to
Tao’s method of smooth summation.
[42] J. S. Schwinger, Phys. Rev. 82, 664 (1951).
[43] L. Euler and A. Aycock, On divergent series (2018), arXiv:1808.02841 [math.HO].
[44] T. Appelquist and J. Carazzone, Phys. Rev. D 11, 2856 (1975).
[45] G. F. Giudice, , 155 (2008), arXiv:0801.2562 [hep-ph].
[46] C. de Rham and R. H. Ribeiro, JCAP 11, 016, arXiv:1405.5213 [hep-th].
[47] C. P. Burgess, Introduction to Effective Field Theory (Cambridge University Press, 2020).
[48] C. Branchina, V. Branchina, F. Contino, and N. Darvishi, Phys. Rev. D 106, 065007 (2022),
arXiv:2204.10582 [hep-th].
[49] J. A. Shapiro, Phys. Rev. D 5, 1945 (1972).
[50] L. P. Kadanoff, Physics Physique Fizika 2, 263 (1966).
[51] K. G. Wilson, Phys. Rev. B 4, 3174 (1971).
[52] K. G. Wilson, Phys. Rev. B 4, 3184 (1971).
[53] K. G. Wilson, Rev. Mod. Phys. 47, 773 (1975).
[54] J. Polchinski, Nuclear Physics 231, 269 (1984).
[55] R. D. Ball, Phys. Rept. 182, 1 (1989).
[56] M. Oleszczuk, Z. Phys. C 64, 533 (1994).
[57] S.-B. Liao, Phys. Rev. D 53, 2020 (1996), arXiv:hep-th/9501124.
[58] Z. Xing and L. Chang, Phys. Rev. D 107, 014019 (2023), arXiv:2210.12452 [hep-ph].
[59] A. Padilla and R. G. C. Smith, To appear, arXiv:24mm.nnnnn [hep-th].
[60] S. P. de Alwis, JHEP 03, 118, arXiv:1707.09298 [hep-th].
[61] G. ’t Hooft and M. J. G. Veltman, Nuclear Physics 44, 189 (1972).
[62] S. Abel and K. R. Dienes, Phys. Rev. D 104, 126032 (2021), arXiv:2106.04622 [hep-th].
 27

[63] We thank Killian Moehling for pointing this out in detail.