Mythic Power Series:

Notes on renormalized perturbation theory

Paul Stevenson
T.W. Bonner Laboratory, Department of Physics and Astronomy,
Rice University, Houston, TX 77251, USA

Perturbation theory in a renormalizable quantum eld theory generates expansions in powers

of a parameter, the renormalized couplant, that is a myth it doesnt exist, at least not uniquely.
In fact there are innitely many possible renormalized couplants that are all a priori on an equal
footing. Physical predictions of the theory are in principle independent of the renormalization
scheme (RS) (the renormalized couplants precise denition) but nite-order approximants
are not RS independent. These notes discuss the nature of the problem and describe how to
optimize the RS choice.

Introduction

An ordinary power series is an expansion in terms of some given, unique parameter. A mythic
power series
R = a(1 + r1 a + r2 a2 + . . .)

(1.1)

is an expansion in a parameter a that is mythic in the sense that it does not exist, at least not
uniquely; instead there are innitely many as, all fundamentally on an equal footing, related
by transformations of the form
a = a(1 + v1 a + v2 a2 + . . .)

(1.2)

with arbitrary nite coecients v1 , v2 , . . .. The mythic-series coecients simultaneously transform:

r1 = r1 v1 ,
r2 = r2 2r1 v1 + 2v12 v2 ,

(1.3)

etc. ,
so that R
R = a (1 + r1 a + r2 a2 + . . .)

(1.4)

is formally invariant.
Mythic power series arise naturally in renormalized quantum eld theory (QFT) and the
invariance of R is called renormalization-group (RG) invariance [1]. The essential point is that
the renormalization procedure for handling ultraviolet divergences in QFTs involves dening a
renormalized couplant a and the precise way in which a is dened the renormalization
scheme (RS) involves a great deal of arbitrariness. Changing from one RS to another induces
transformations of the form (1.2). However, any physically measurable quantity R must be
formally invariant under changes of RS.
Any RS involves choosing a renormalization scale , which is an arbitrary parameter with
dimensions of mass (energy and mass eectively have the same dimensions if we adopt units
with the speed of light equal to unity). The dependence of a on is given by the famous
function:

da
(a) = ba2 (1 + ca + c2 a2 + c3 a3 + . . .).
d

(1.5)

The function plays a key role in mythic power series. The coecients b, c can be unambiguously
calculated in any given QFT and are RS invariant numbers. The higher coecients c2 , c3 , . . . can
also be calculated, but their values depend on the RS choice. The coecients r1 , r2 , . . . in the
2

perturbative expansion of any specic physical quantity R can also be calculated (in the same
RS). Although the ri s and the cj s are separately RS dependent, there are certain combinations
of them that are RS invariants [2]. Sec. 2 discusses these points in detail.
Sec. 3 takes the discussion beyond the level of formal power series and considers the crucial,
practical problem how to dene, at a given order, the analog of the partial sum for a mythic
power series. In a given RS that is straightforward; one simply makes a truncation of both the R
and series (at the same order). However, the approximant thus dened is RS dependent. The
view taken here is that the right approach is to make the approximate result as RS invariant
as possible. That is, one should evaluate the approximant in an optimal RS such that it is
stationary with respect to innitesimal variations of the RS. [This is a particular instance of
the principle of minimal sensitivity [2] discussed in the appendix.] The optimal result, for a
given physical quantity at a given order, is unique and is determined by a set of optimization
equations [2].

RG invariance and its consequences

2.1

Physical quantities

QFT involves many theoretical objects various Greens functions, etc. that are not physically
measurable and are not RS invariant. For our purposes these are merely intermediate steps in
calculating physical quantities and need not be discussed. In general a perturbatively calculable
physical quantity has the form A1 R + A0 with a leading-order coecient A1 and, sometimes, a
zeroth-order term A0 . The coecients A0 and A1 (which carry the dimensions of mass to the
appropriate power) are RS invariant, so we may focus on dimensionless, normalized physical
quantities R of the form:
R = aP (1 + r1 a + r2 a2 + . . .).
The power

(2.1)

is usually 1 or 2 or 3, but need not be an integer. At the level of formal power

series there is no loss of generality in taking P = 1, since one may always replace R with another
physical quantity R1/P . However, that trick would not be natural at nite order. Later on we
will sometimes specialize to

= 1 for simplicity, before generalizing to any P.

The key property of a physical quantity is that of Renormalization Group (RG) invariance
[1], which means that a physical quantity is independent of the renormalization scheme (RS); i.e.,
it is invariant under the transformations (1.2). Heuristically, this property might be expressed
symbolically as


da R
R 
dR
+
=
,
0=

d(RS)
(RS) a d(RS) a

(2.2)

where the total derivative is separated into two pieces: (i) RS dependence at constant a (i.e.,
from the series coecients, ri ) and (ii) RS dependence from the couplant, a.
A particular case of Eq. (2.2) is the familiar RG equation expressing the renormalizationscale independence of R:

 + (a)
R = 0,
a
a

(2.3)

which involves the famous function introduced in Eq. (1.5).

What exactly is the renormalization scale ? In dening a particular RS many arbitrary
choices are made that, explicitly or implicitly, introduce many arbitrary parameters. One of these
parameters necessarily has the dimensions of mass and is called ; all the other RS parameters
can be taken as dimensionless since we can always scale them with the appropriate power of .
For our purposes, then, it is sucient to say that is the parameter with dimensions of mass
introduced in the renormalization process.
Generally a physical quantity R is not a single quantity but rather a function of some
experimentally dened parameters (e.g., R might correspond to some scattering cross section
at a centre-of-mass energy Q and scattering angle ). We may always single out one parameter
with the dimensions of mass (or energy) that we will call the physical energy scale and denote
by Q. All other experimental parameters can be taken to be dimensionless (if they were not
originally we may scale them with an appropriate power of Q to make them so). Q is needed
only to explain which quantities are, or are not, Q dependent; the precise denition of Q in any
specic case is left to the reader.
The dependence of R on Q is of great interest. The original Lagrangian of the theory if
particle masses are zero or can be neglected contains no parameter with dimensions of mass,
only a dimensionless bare coupling constant gBare . The normalized physical quantity R is then
seemingly a function only of Q and dimensionless parameters. Nave dimensional analysis implies
that the dimensionless R must be independent of Q. That conclusion, however, is false because
of the need for renormalization (for a detailed discussion see [3]). Renormalization introduces
the renormalization scale and now we nd that the coecient r1 depends logarithmically on
the ratio of Q to (and r2 involves a log-squared, etc.). Although the dependence ultimately
cancels in R, the Q dependence survives. [As we shall see in Sec. (2.3), specifying a boundary
and R is ultimately a
conditions for the -function equation introduces a mass parameter
The fact that a QFT, seemingly characterized by a dimensionless
function of the ratio Q/.
is known
bare coupling constant gbare , ends up specied by a characteristic scale parameter
as Dimensional Transmutation [4]. One also says that the apparent scale invariance of the
theory is aicted with a quantum anomaly.]

2.2

RS invariance of b and c

Theorem (t Hooft [5])

The rst two coecients, b and c, of the function are RS invariant.
Proof: In another RS there will be a dierent function:
(a )

da
= b a2 (1 + c a + c2 a2 + . . .).
d

(2.4)

The functions of two schemes are not equal ( is not a physical quantity) but are related by
(a )

da
da da
da
=

=
(a).
d
da d
da

(2.5)

Thus, since a = a(1 + v1 a + v2 a2 + . . .),

(a ) = (1 + 2v1 a + 3v2 a2 + . . .)(a).

(2.6)

Eliminating a in favour of a and then equating powers of a shows that

b = b

and

c = c,

(2.7)

completing the proof. Note that the higher coecients are not invariant; for instance c2 =
c2 + v2 v12 v1 c.
In any given renormalizable QFT the invariants b, c can be calculated and we regard them
here as givens. For instance, in Quantum Chromodynamics (QCD) with nf avours of massless
quarks one has
b=

33 2nf
,
6

c=

153 19nf
.
2(33 2nf )

(2.8)

Note that an explicit minus sign is included in front of the expansion of in Eq. (1.5) so that
b is a positive number in the most interesting (asymptotically free) QFTs. We will assume
b = 0. Incidentally, the advantage of factoring out b in (a)s expansion is that it is then absent,
or scales out of, almost all subsequent equations.

2.3

Integrated -function equation; definition of

The -function equation

da
= (a),
d

(2.9)

da
+ const. = ln .
(a)

(2.10)

integrates immediately to

 + C, where
has
We must now specify the constant of integration. By writing it as ln
dimensions of mass, we can convert the ln to a logarithm with a dimensionless argument.
Next, since (a) is known only as a small-a expansion, we naturally want to use 0 to a as the

range of integration, giving us

dx

+ C = ln(/),
(x)

(2.11)

but the integral is then divergent. This means that the constant C must be suitably innite.
is to write [2]:
One convenient way of specifying C, and thereby dening ,
(
lim

)
dx

+ C() = ln(/).
(x)

where
C() =

dx
+ cx)

bx2 (1

(2.12)

(2.13)

(In the case that c is negative use the Cauchy principal-value prescription to deal with the pole
at x = 1/c.) Note that C() involves only b and c and so is RS invariant.
is included to distinguish it from another widely-adopted denition of the
[The tilde in
parameter in the literature [6]. The two denitions are related by an RS-invariant factor:
= (c/b) ln(2c/b).]
ln(/)
It is convenient to introduce the variable

b ln(/)

(2.14)

and to dene
(x)
= 1 + cx + c2 x2 + . . .
bx2
We may regard Eq. (2.12) as equating to a function K(a):
)
(

a
dx
dx
1
1
= K(a)

.
2
x2 (1 + cx)
B(x) 1 + cx
a
0 x
B(x)

Note that the rst term gives the second-order approximation to K(a):



 ca 
1
dx
(2)

.
= + c ln 
K (a) =
x2 (1 + cx)
a
1 + ca 
a

(2.15)

(2.16)

(2.17)

Eq. (2.12), or equivalently Eq. (2.16), will be referred to as the integrated -function equation
or int- equation.

2.4

RS dependence of

we should now ask; is it RS dependent? The answer is yes, but in a simple

Having dened
and denite way.
6

Theorem (Celmaster and Gonsalves [7])

If two schemes are related, for the same value of , by
a = a(1 + v1 a + . . .)

(2.18)

/)
= v1 /b.
ln(

(2.19)

then

This relationship is exact and does not involve the v2 , v3 , . . . coecients.

Proof: We start from Eq. (2.12) and the corresponding equation in a dierent, primed, RS.
Subtracting those two equations, with = , yields
)
(
a
a

dx
dx
/)
= lim
ln(

.

0
(x)
(x )

(2.20)

Evaluating for small a leads to

/)
=
ln(

1
b

1
1

a a

)
+ O(a) = v1 /b + O(a).

(2.21)

Now, since the l.h.s. is independent of , we can choose to evaluate the r.h.s. at any value of .
Choosing (if b is positive), or 0 (if b is negative), makes a 0 and so completes
the proof.
This proof, though sound, can seem too slick. An alternative proof, due to Osborn [8] is
instructive. First, one splits each of the integrals in Eq. (2.20) into two pieces:
)
(
a

a

dx
dx
dx
dx
/)
= lim
+

.
ln(


0
(x)
(x )
(x )
(x)

(2.22)

Then one chooses to be related to by

= (1 + v1 + . . .),

(2.23)

with exactly the same coecents as in Eq. (2.18). The second and fourth terms in (2.22) can
then be shown to cancel exactly by making the change of variables
x = x(1 + v1 x + . . .),

(2.24)

and using the relation of to of Eq. (2.5). One can now take the limit 0 (choosing, say,

= 2 or = ) and use (x) bx2 as x 0 to obtain the result (2.19). This proof makes
plain that the result is non-zero only because of the 1/x2 singularity of the 1/ integrands and
that the behaviour of (x) away from x = 0 plays no role.
7

The importance of this result of Celmaster and Gonsalves [7] is its exactness. It means
parameter is scheme dependent, the s
of dierent schemes can be related
that, although the
exactly by a nite calculation of the single coecient v1 in the relation between the two schemes.
parameter of any convenient reference RS can be adopted, without prejudice,
Hence, the
as the one free parameter of the theory, taking over the role of the bare coupling constant in
the original Lagrangian specifying the QFT.

2.5

Parametrization of RS dependence

From the int- equation, (2.16), it is clear that a can depend on RS only through b ln(/)
and the scheme-dependent coecients c2 , c3 , . . . of the function. The coecients of R can
depend on RS only through these same variables, because of RG invariance: the cancellation
in Eq. (2.2) could not occur if the R coecients depended on some RS variable that a did
not depend on [2]. Thus, as far as physical quantities are concerned, these variables provide a
complete RS parametrization and we may write:
a = a(RS) = a(, c2 , c3 , . . .),

(2.25)

enters here.
Note the important point that only the ratio of to
The parametrization of RS dependence by the set of variables , c2 , c3 , . . . is not unique (see
[2] for a general discussion) but it does provide a convenient Cartesian coordinate system.
The partial derivative with respect to any one of these variables is to be taken holding all the
other variables in the set constant; that is the partial derivative is taken with all the ci s held
constant, and the cj partial derivative is taken with and the other ci (i = j) held constant.
The symbolic RG-invariance equation (2.2) can now be written out explicitly as the following
set of equations:
R
=

R
=
cj


)

(a)
+
R = 0,
a
b a

(a)
R = 0,
j
cj a
a

j = 1
(2.26)
(j = 2, 3, . . .),

involving some new functions j (a) dened by

j (a)

2.6

a
.
cj

(2.27)

The j (a) and Bj (a) functions

Requiring commutation of the second derivatives, 2 a/ cj = 2 a/cj , leads to [2]

j (a)(a) (a)j (a) = baj+2 ,
8

(2.28)

where here the prime indicates dierentiation with respect to a, regarding the coecients cj as
xed. Integrating this dierential equation leads to
a
xj+2
dx
,
j (a) = b(a)
(x)2
0

(2.29)

for j = 2, 3, . . .. (Note that b cancels out on the right-hand side.)

Alternatively, one may consider the int- equation, (2.12): Its right-hand side is /b and
depends only on the parameter, not on the cj s. Therefore, taking the partial derivative
leads to
j (a)
+
(a)

dx
0

1
(x)2

)
bxj+2 = 0,

cj

(2.30)

where the rst term arises from the cj dependence of the limit of integration a, and the second
term arises from the explicit cj dependence of the integrand. Rearranging this equation leads
directly to the result (2.29) above.
The j functions begin at order aj+1 so it is convenient to dene Bj (a) functions whose
series expansions begin 1 + O(a):
Bj (a)

(j 1)
j (a).
aj+1

(2.31)

For j = 1 it is natural to dene

B1 (a) B(a)

(a)
= 1 + ca + c2 a2 + . . . =
ci ai ,
2
ba

(2.32)

i=0

with the convention that c0 1 and c1 c. Equation (2.29) can then be re-written as
Bj (a) =

(j 1)
B(a)Ij (a),
aj1

where
Ij (a)

dx
0

xj2
.
B(x)2

(2.33)

(2.34)

(Note that this formula for Bj (a) even holds for j = 1 if the r.h.s. is interpreted as the limit
j 1 from above.)
The Bj (a) functions have power-series expansions whose coecients we write as Wij :
Bj (a)

Wij ai ,

(2.35)

i=0

with W0j 1. The other Wij coecients are xed in terms of the ci s. From the dierential
equation (2.28) it is straightforward to show that they satisfy the relation
i

j
(i + j 1 2m)cm Wim
= (j 1)i0

m=0

(2.36)

for i = 0, 1, 2, . . . and j = 1, 2, . . .. In the special case j = 1 one has Wi1 ci and the above
equation reduces to
i

(i 2m)cm cim = 0,

(2.37)

m=0

which is true identically, since the left-hand side is

i

(i m)cm cim

m=0

mcm cim

(2.38)

m=0

and the rst sum, by changing the summation variable from m to n = i m, is seen to cancel
the second.

2.7

The n invariants (P = 1 case)

The RG equations (2.26) determine how the coecients ri of R must depend on the RS variables
{, cj }. To show explicitly how this works we specialize to the

= 1 case, where

R = a(1 + r1 a + r2 a2 + . . .)
and write out the lowest-order terms to obtain
)
(
3 r2
2 r1
+a
+ . . . a2 (1 + ca + . . .)(1 + 2r1 a + . . .) = 0,
a

(
)
2 r1
3 r2
a
+a
+ . . . + a3 (1 + W12 a + . . .)(1 + 2r1 a + . . .) = 0,
c2
c2

(2.39)

(2.40)
(2.41)

and so on. (In fact the coecient W12 is zero.) Equating powers of a, one sees that r1 depends
on only, while r2 depends on and c2 only, etc., with
r1
= 1,

(2.42)

r2
r2
= 2r1 + c,
= 1,
(2.43)

c2
etc.. Upon integration one will obtain ri as a function of , c2 , . . . , ci plus a constant of integration
that is RS invariant. Thus, certain combinations of series coecients and RS parameters are
RS invariant [2]. The rst two are:
1 (Q) r1 ,

(2.44)

2 c2 + r2 cr1 r12 .

(2.45)

The rst invariant, 1 (Q), is unique in being dependent on the physical energy scale, Q (see
the discussion in subsection 2.1). A QFT calculation of the coecient r1 , in some arbitrary RS,
yields a result of the form
r1 = b ln(/Q) + r1,0 ,
10

(2.46)

whose dependence indeed conforms with Eq. (2.42). For dimensional reasons, the and Q
dependences are tied together. (The QFT calculation does not know what boundary condition
cannot explicitly appear.)
will later be applied to the -function equation, so the parameter
Similary, the higher coecients r2 , . . . depend on ln(/Q), but not on or Q separately. Hence,
for the invariants 2 , 3 , . . . the cancellation of dependence also implies the cancellation of Q
dependence. However, 1 (Q) is dierent because its denition explicitly involves , and we nd
(b ln(/Q) + r1,0 )
1 (Q) = b ln(/)
r1,0
= b ln(Q/)
R ),
b ln(Q/

(2.47)

is the -parameter

R is a
where
in the RS in which the QFT calculation was done, and
characteristic scale specic to the particular physical quantity R. Indeed, we can regard the
last step as a Celmaster-Gonsalves relation that converts between the R-specic scale and the
of some reference RS that serves as the single free parameter of the theory.
universal scale
Some convention must be adopted to uniquely dene the higher invariants j (for j 2)
because, of course, any sum of invariants is also an invariant. For example, one might quite
naturally add some multiple of c2 to Eq. (2.45). Indeed, the tildes over the j s are included
to distinguish them from an earlier denition [2]. One convenient denition is as follows [9].
For any given physical quantity R one can always dene a RS (called the fastest apparent
convergence (FAC) or eective charge scheme) such that all the series coecients ri vanish
in that scheme, so that R = aFAC (1 + 0 + 0 + . . .). Since the functions of any two RSs (see
subsection 2.2) are related by
(a ) =

da
(a),
da

(2.48)

we must have
FAC (R) =

dR
(a).
da

(2.49)

The n invariants can be dened to coincide with the coecients of the FAC-scheme function:
FAC (R) = bR

n Rn = ba2

n=0

dR
B(a)
da

(2.50)

where 0 1 and 1 c (not to be confused with the independent invariant 1 (Q) r1 ).

Re-arranging this equation as
B(a) =

(
n

n a

n=0

11

R
a

)n+2

1
dR
da

(2.51)

and equating powers of a we obtain

cj =

[(
i Cji

i=0

R
a

)i+2

R
a

(2.52)

where Cn [F (a)] means the coecient of an in the series expansion of F (a).

The rst few invariants are listed below (for
1 = c,

and

= 1):

1 (Q) = r1 ,

2 = c2 + r2 cr1 r12 ,

(2.53)

3 = c3 + 2r3 2c2 r1 6r2 r1 + cr12 + 4r13 .

2.8

The n invariants (general P)

For general

the rst few invariants are

r1
1 (Q) = ,
(
) P
P+1
r2 c r1

2 = c2 +
r12 ,
P
P
2P2
(
)
(
)
2r3 2c2 r1
P+2
c r12
2(P + 1)(P + 2) 3
3 = c3 +

2
r2 r1 + 2 +
r1 .
P
P
P2
P
3P3
1 = c,

and

The generalization of Eq. (2.52) to general P is given by

(
)i+1
j
1/P

1
R
,
cj =
i Cji P
a R
a
R a

(2.54)

(2.55)

i=0

An explicit inverse formula giving the s in terms of the cj and rj coecients can be found [9]:

(
)(n+1) (
)2
1/P
1
R
R
a
.
n = Cn B(a) 2
(2.56)
P
a
R a

12

Finite orders and optimization

3.1

Finite-order approximants

So far our discussion has been at the level of formal power series. We now need, for mythic
power series, something corresponding to the notion of the partial sums of an ordinary power
series. For mythic power series two truncations are involved, for R and for . (We need in
parameter of some reference
order to nd a in terms of the theorys one free parameter, the
scheme.)
The (k + 1)-th order, or (next-to)k -leading order (Nk LO) approximant is naturally dened
with both R and truncated after k + 1 terms:
R(k+1) aP (1 + r1 a + . . . + rk ak ),

(3.1)

where a here is shorthand for a(k+1) , the solution to the int- equation with replaced by
(k+1) :
(k+1) ba2 (1 + ca + . . . + ck ak ).

(3.2)

It is straightforward to check that the order of the error term R R(k+1) is determined by
whichever truncation, of R or , is the more severe, so it is natural to use the same number of
terms in each.

3.2

Optimization in low orders (P = 1 case)

While the exact R is RG-invariant, the nite-order approximants are not, since the truncations
spoil the cancellations in the RG equations (2.26). If R in those equations is replaced by R(k+1)
then the r.h.s. is not zero but is some remainder term O(aP+k+1 ). The idea of optimized
perturbation theory [2] is to choose an optimal RS in which the approximant R(k+1) is
stationary with respect to RS variations; i.e., the RS in which R(k+1) satises the RG equations,
(2.26), with no remainder:
(


)
(a) R(k+1)
R(k+1) 
 +

b
a
a


)
R(k+1)

 + j (a)

a

= 0.

(3.3)

R(k+1) 
cj

j = 1

opt. RS

= 0.

j = 2, . . . , k.

opt. RS

We assume here that the QFT calculations of the R and -function coecients up to and
including rk and ck have been done in some (calculationally convenient) RS. From those results
13

we can compute the values of the invariants 1 (Q) and 1 c and 2 , . . . , k . Our optimized
result will be expressed solely in terms of those invariants.
To see how this works let us consider the second-order (NLO) approximant. (For simplicity
we set

= 1 for the remainder of this subsection.)

R(2) = a(1 + r1 a),

(3.4)

(2) = ba2 (1 + ca),

(3.5)

where a here is short for a(2) , the solution to the int- equation (2.16) with replaced by (2) :


 ca 
1
(2)
.

= K (a) = + c ln 
(3.6)
a
1 + ca 
Since R(2) depends on RS only through the variable , only the j = 1 equation in Eq. (3.3)
above is non-trivial. Thus, the optimized R(2) is determined by a single optimization equation:
r1 2
a
a
2 (1 + c
a)(1 + 2
r1 a
) = 0.

(3.7)

(Overbars are used to indicate the value in the optimum RS.) As discussed in subsection 2.7,
the a2 terms must cancel in any RS, which xes

r1

= 1, leaving

1 (1 + c
a)(1 + 2
r1 a
) = 0.

(3.8)

This determines the optimized coecient r1 in terms of the invariant c and the optimized
couplant a
:
1 c
.
2 1 + c
a
But r1 is related to by the denition of the 1 (Q) invariant, Eq. (2.44):
r1 =

1 (Q) r1 .

(3.9)

(3.10)

Eliminating r1 between these last two equations and substituting into the second-order int-
equation, (3.6), gives
[
(
)
(
)]
1
c
a
1
c
a
1 + c
a ln
+
= 1 (Q).
a

1 + c
a
2 1 + c
a

(3.11)

If the values of the invariants c and 1 (Q) are known, as we assume, then we may numerically
solve this last equation to obtain a
. Substituting back in (3.9) we can nd r1 and hence obtain
(2)
=a
the optimized approximant R
(1 + r1 a
).
Note that the only approximations made here are the truncations of the R and series in
Eqs. (3.4, 3.5), which dene the second-order approximant is some general RS. We do not, for
instance, approximate Eq. (3.9) as r1 21 c. Also, we will need to solve Eq. (3.11) numerically
14

We now turn to third order. The third order approximant is dened by

R(3) = a(1 + r1 a + r2 a2 ),

(3.12)

(3) = ba2 (1 + ca + c2 a2 ),

(3.13)

where now a is short for a(3) , the solution to the int- equation with replaced by (3) . Since
R(3) depends on RS through two parameters, and c2 , there are two optimization equations
coming from Eq. (3.3). These correspond to suitably truncated versions of Eqs. (2.40, 2.41).
Using Eqs. (2.42, 2.43) they reduce to:
(
)
(3) (
1 + (2
r1 + c) a
B
a) 1 + 2
r1 a
+ 3
r2 a
2 = 0,

(3.14)

(
)
(3) (
1B
r1 a
+ 3
r2 a
2 = 0.
(3.15)
2 a) 1 + 2
(
)
(3) (
Here B
a) 1 + c
a + c2 a
2 is the B(a) function truncated at third order in the optimum
(3) (
scheme. The other function B
a) is not a polynomial; it is obtained from Eqs. (2.33, 2.34)
2

with B(x) replaced by (1 + cx + c2 x2 ).

These optimization equations, together with the denitions of the invariants 2 and 1 (Q)
(3) .
and the int- equation, fully determine the optimized result, R
Note that the optimum RS evolves from one order to the next; for instance r1 at third
(2)

order is not the same as r1 at second order (so, strictly we should have distinguished r1
(3)
r1

and

in the above).

3.3

The optimization equations

For completeness, we now formulate the optimization equations at some general, (k+1)th, order.
We also consider a general power P.
It is convenient to dene
S=

1
PaP1

R
,
a

(3.16)

whose series expansion

S = 1 + s1 a + s2 a2 + . . .
(

has coecients
sm
As we saw earlier (in the

m+P
P

(3.17)

)
rm .

(3.18)

= 1 case) all terms in the RG equations up to and including

O(aP+k ) must
of Eq. (3.3)
 cancel automatically in any RS. In the rst optimization equation
(a) R(k+1)
R(k+1) 
the
. A similar
 term is a polynomial which must cancel the rst k terms of b

a
a

15

observation applies to the other optimization equations of Eq. (3.3). Hence, we may reduce the
optimization conditions to
[
]
(k+1) (
(k+1) (
(k+1) (
(k+1) (
B
a
)
S
a
)

T
B
a
)
S
a
)
= 0,
kj
j
j

(3.19)

for j = 1, 2, . . . , k, where the notation Tn [F (a)] means truncate the series for F (a) = F0 +
F1 a + . . . immediately after the an term (i.e., Tn [F (a)] F0 + F1 a + . . . + Fn an .)
(k+1)

Note that the Bj

polynomial

B (k+1) (a)

(a) functions are given by Eqs. (2.33, 2.34) with B(a) replaced by the

1 + ca + . . . + ck ak .

For further details on solving the optimization equations and results for Re+ e at third and
fourth order see [10, 11].

References
[1] E. C. G. Stueckelberg and A. Peterman, Helv. Phys. Acta 26, 449 (1953);
M. Gell Mann and F. Low, Phys. Rev. 95, 1300 (1954).
[2] P. M. Stevenson, Phys. Rev. D23, 2916 (1981).
[3] P. M. Stevenson, Ann. Phys. 132, 383 (1981).
[4] S. Coleman and E. Weinberg, Phys. Rev. D 7, 1888 (1973).
[5] G. t Hooft, Lecture at 1977 Coral Gables Conf., Orbis Scientiae.
[6] Quantum Chromodynamics review in the Particle Physics Booklet 2010
(G. Dissertori and G. P. Salam, in K. Nakamura et al., J. Phys. G 37, 075021 (2010)).
[7] W. Celmaster and R. J. Gonsalves, Phys. Rev. D 20, 1420 (1979).
[8] H. Osborn (private communication).
[9] P. M. Stevenson, Phys. Rev. D 33, 3130 (1986).
[10] A. C. Mattingly and P. M. Stevenson, Phys. Rev. D 49, 437 (1994).
[11] P. M. Stevenson, Nucl. Phys. B 868, 38 (2013).

16

Appendix: The principle of minimal sensitivity

The principle of minimal sensitivity applies to any situation where it is known that the exact
result is independent of certain variables, but where the corresponding approximate result depends upon those variables, and hence is ambiguous. The philosophy is that such a non-invariant
approximant is most believable where it is least sensitive to small variations in the extraneous
variables, because this is where it best approximates the exact results vital property of being
completely insensitive to the extraneous variables.
A simple example is perhaps the best way to convey this idea. Consider the quantummechanical problem of computing the eigenvalues, En , of the quartic-oscillator Hamiltonian:
H = 12 p2 + x4 ,

(A.1)

where [x, p] = i. Suppose we do standard QM perturbation theory, but with:

Hint = x4 21 2 x2 .

H0 = 12 (p2 + 2 x2 ),

(A.2)

This method introduces an extraneous variable , and the approximate eigenvalues so calculated will be -dependent. For example, rst-order for the nth eigenvalue gives:
1
3
(2n2 + 2n + 1).
En(1) = (n + 12 ) +
2
42

(A.3)

However, we know that the exact eigenvalues are -independent. Therefore, it is sensible to
(1)

choose so that the approximant, En , is minimally sensitive to ; i.e.,

[
]1
(2n2 + 2n + 1) 3

= 3
.
(n + 12 )

(A.4)

This gives the optimized result:

En(1) (opt.)

]1
[
2 + 2n + 1) 3
3
(2n
= (n + 12 ) 3
.
4
(n + 21 )

(A.5)

This simple formula ts the ground-state energy to 2%, and all other energy levels to within
1%. The secret of this success is the optimal choice of , which is dierent for dierent levels.
One may proceed to the calculation of higher-order corrections for some specic eigenvalue
(e.g., the ground state). For any xed the perturbation series would be a divergent asymptotic
series, but if is chosen in each order according to the minimal sensitivity criterion (which
that gradually increases with order), one obtains a convergent sequence of results.
gives an
One may also use the same method to obtain accurate approximate wavefunctions, n (x),
from rst-order perturbation theory. Here the optimal will be a function of x; in particular, it
will be proportional to |x| at large |x|, thereby converting the Gaussian dependence exp( 21 x2 )
into the correct large-|x| behaviour.
17