Stability of QED
M. P. FrySchool of Mathematics, University of Dublin, Trinity College, Dublin 2, Ireland (Received 5 July 2011; published 19 September 2011)
It is shown for a class of random, time-independent, square-integrable, three-dimensional magnetic fields that the one-loop effective fermion action of four-dimensional QED increases faster than a quadratic inBin the strong coupling limit. The limit is universal. The result relies on the paramagnetism of charged spin-1=2fermions and the diamagnetism of charged scalar bosons.
DOI:10.1103/PhysRevD.84.065021 PACS numbers: 12.20.Ds, 11.10.Kk, 11.15.Tk
I. INTRODUCTION
Integrating out the fermion fields in four-dimensional QED continued to the Euclidean metric results in the measure for the gauge field integration
dðAÞ ¼Z1e
R
d4xð1=4F
Fþgauge fixingÞ
detrenð1eS6AÞ
Y
x;
dAðxÞ; (1.1)
where detren is the renormalized fermion determinant de-fined in Sec.II;Sis the free fermion propagator, andZis chosen so that RdðAÞ ¼1. In the limit e¼0, the Gaussian measure for the potentialA is chosen to have mean zero and covariance
Z
dðAÞAðxÞAðyÞ ¼DðxyÞ; (1.2)
where D is the free photon propagator in some fixed gauge. Naively, integration over the fermion fields pro-duces the ratio of determinants detð6PeA6 þmÞ=
detð6PþmÞ, which is not well defined;detren makes sense of this ratio. It is gauge invariant and depends only on the field strengthFand invariants formed from it.
We have chosen to introduce this paper with an abrupt intrusion of definitions in order to emphasize the central role ofdetrenin QED: it is everywhere. It is the origin of all fermion loops in QED. If there are multiple charged fer-mions, thendetrenis replaced by a product of renormalized determinants, one for each species. For our purpose here, it is sufficient to consider one fermion.
The nonperturbative calculation of detren reduces to finding the eigenvalues ofS6A,
Z
d4ySðxyÞ6AðyÞcnðyÞ ¼ 1
en
cnðxÞ: (1.3)
There are at least two complications. First,S6Ais not a self-adjoint operator, and so many powerful theorems from analysis do not apply. Second, sinceA is part of a func-tional measure, it is a random field, making the task of calculating the en for all admissible fields impossible. What can be done is to expand ln detren, the one-loop effective action, in a power series ine. Then the functional
integration can be done term-by-term to obtain textbook QED.
The first nonperturbative calculation ofdetrenwas done by Heisenberg and Euler [1] 75 years ago for the special case of constant electric and magnetic fields. Their paper gave rise to a vast subfield known as quantum field theory under the influence of external conditions. A comprehen-sive review of this body of work relevant todetrenis given by Dunne [2].
An outstanding problem is the strong field behavior of detren that goes beyond constant fields or slowly varying fields or special fields rapidly varying in one variable [2,3].1 That is, what is the strong field behavior ofdetren for a class of random fields F onR4? What ifln detren increases faster than a quadratic inFfor such fields? Is detren integrable for any Gaussian measure in this case? This is a question with profound implications for the stability of QED in isolation. Of course, QED is part of the standard model, thereby making the overall stability question a much more intricate one. Nevertheless, the stability of QED in isolation remains unknown and de-serves an answer.
In this paper, we consider the case of square-integrable, time-independent magnetic fields BðxÞ defined on R3. There are additional technical conditions onBintroduced later. The magnetic field lines are typically twisted, tangled loops. We find that
lim e!1
ln detren e2lne ¼
kBk2T
242 ; (1.4)
where kBk2¼Rd3xBBðxÞ, and T is the size of the time box. Since e always multiplies B, this means that ln detren is growing faster than a quadratic in B. In the constant field case, this result is formally equivalent to the
1We note here progress in scalarQED
4since the review [2] in
Heisenberg-Euler result [1] and to calculations relating the effective Lagrangian to the short-distance behavior of QED via its perturbative-function [2]. What is notable here is that the strong coupling limit ofln detren is universal.
To achieve universality, the derivation of (1.4) must rely on general principles. One of these is the conjectured ‘‘diamagnetic’’ inequality for Euclidean three-dimensional QED, namely
jdetQED3ð1eS6AÞj 1: (1.5)
The fermion determinant in (1.5) is defined in Sec.II. The diamagnetic inequality is known to be true for lattice formulations of QED3 obeying reflection positivity and using Wilson fermions [5–7]. Since Wilson fermions are CP invariant, there is no Chern-Simons term to interfere with the uniqueness of detQED3 [8]. And sincedetQED3 is
gauge invariant, there are no divergences when the lattice spacing for the fermions is sent to zero. As stated by Seiler [7], (1.5) is more an obvious truth than a conjecture.
SincedetQED3je¼0 ¼1anddetQED3has no zeros inefor
real values ofewhenm0[9], (1.5) can be rewritten as
0<detQED3 1: (1.6)
An inspection of Eq. (2.4) below indicates that (1.6) is a reflection of the tendency of an external magnetic field to lower the energy of a charged fermion. Therefore,
the historic heading of (1.5) and (1.6) as diamagnetic inequalities is a misnomer; paramagnetic inequalities would be a more accurate designation. The detailed justi-fication for going from (1.5) to (1.6) is given in Sec.II.
The second general principle underlying (1.4) is the diamagnetism of charged spin-0 bosons in an external magnetic field. This is encapsulated in one of the versions of Kato’s inequality discussed in Sec.III.
The final essential input to (1.4) is a restriction on the class of fields needed to obtain the limit. These restrictions are summarized in Sec. IV. As the foregoing remarks indicate,QED3 is central to the derivation of (1.4), and it is to the connection betweenQED3andQED4that we now turn.
II.QED3 ANDQED4
A. The connection
The connection has been dealt with previously [10]. In order to make this paper reasonably self-contained, we will review the relevant definitions and results. The upper bound ondetrenobtained in [10] is not optimal; it will be optimized here.
The renormalized and regularized fermion determinant in Wick-rotated Euclidean QED4 with on-shell renormal-ization,detren, may be defined by Schwinger’s proper time representation [11]
detrenð1eS6AÞ ¼1
2
Z1
0 dt
t
Tr
eP2t exp
ðD2þe
2FÞt
þe2jjFk2 242
etm2
; (2.1)
whereD¼PeA,¼ ð1=2iÞ½,y¼ ,kFk2¼
R
d4xF2ðxÞ, andeis assumed to be real. We choose the chiral representation of the-matrices so that
ij¼ k 0
0 k
;
i; j; k¼1;2;3in cyclic order. Since we will consider time-independent magnetic fields, we setA ¼ ð0;AðxÞÞwithxin R3. Then (2.1) reduces to
ln detren¼T 2
Z1
0 dt
t
2
ð4tÞ1=2 Trðe P2t
expf½ðPeAÞ2e
BtgÞ þe 2k
Bk2
122
etm2
; (2.2)
where T is the dimension of the time box, and the fac-tor 2 is from the partial spin trace. Clearly, we must have B2L2ðR3Þ. IfAis assumed to be in the Coulomb gauge rA¼0, then by the Sobolev-Talenti-Aubin in-equality [12]
Z
d3x
BðxÞ BðxÞ
274
16
1=3X3
i¼1
Z
d3xjA iðxÞj6
1=3
:
(2.3)
So we must also haveA2L6ðR3Þ.
In analogy withdetrenin (2.1), without the charge renor-malization subtraction,detQED3 may be defined by
lndetQED3ðm
2Þ
¼1 2
Z1
0 dt
t Trðe P2t
expf½ðPeAÞ2eBtgÞetm2
:
(2.4)
This definition and regularization ofdetQED3is parity con-serving and gives no Chern-Simons term. Substituting (2.4) in (2.2) and, noting that1R10 dEetE2
ln detren¼
2T
Z1
0 dE
lndetQED3ðE
2þm2Þ
þe2kBk2 243=2
Z1
0 dt t1=2e
ðE2þm2Þt
¼T
Z1
m2
dM2
ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
M2m2
p lndetQED3ðM2Þþ e2kBk2
24pffiffiffiffiffiffiffiM2
:
(2.5)
Result (2.5) will be referred to repeatedly in what follows.
B. Justification of (1.6)
Continuing our review of previous work, we turn to the derivation of the upper bound onln detrenin (1.4). Since the degrees of divergence of the first-, second-, and third-order contributions to ln detQED3 are 2, 1, and 0, respectively, these must be dealt with separately. Their definition is obtained from the expansion of (2.4) throughOðe3Þ, result-ing in
ln detQED3ð1eSA6 Þ
¼ e2 4
Z d3k
ð2Þ3j
^
BðkÞj2Z1 0
dz zð1zÞ
½zð1zÞk2þm21=2
þln det4ð1eSA6 Þ; (2.6)
where ln det4 defines the remainder and B^ is the Fourier transform ofB. Definition (2.4) assigns the value of zero to the terms of ordereande3. The argument of det
QED3 has
been changed to indicate its origin as the formal ratio of QED3 determinants detð6PeA6 þmÞ=detð6PþmÞ. Note the minus sign in (2.6) pointing to paramagnetism.
The following theorems are essential for what follows: Theorem 1 [6,13,14]. Let the operator S6A in det4 be transformed by a similarity transformation to K ¼
ðp2þm2Þ1=4SA6 ðp2þm2Þ1=4. This leaves the eigenval-ues of S6A invariant. Then K is a bounded operator on L2ðR3;d3x;C2Þ for A2LpðR3Þ for p >3. Moreover, K is a compact operator belonging to the trace ideal Ip, p >3.
The trace ideal Ipð1p <1Þ is defined as those compact operators A with kAkpp¼TrððAyAÞp=2Þ<1. From this it follows that the eigenvalues 1=en ofS6A ob-tained from (1.3) specialized to three dimensions are of finite multiplicity and satisfyP1n¼1jenjp<1forp >3. The eigenfunctions cn belong to the Sobolev space L2ðR3;pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffik2þm2d3k;CÞ. None of the e
n are real for m0[9].
Theorem 2 [15–17]. Define the regularized determinant
detnð1þAÞ ¼det
ð1þAÞexpX n1
k¼1
ð1ÞkAk=k
: (2.7)
Thendetncan be expressed in terms of the eigenvalues of A2Ipfornp.
Accordingly, det4 in (2.6) is defined and can be repre-sented as [17]
det4ð1eS6AÞ ¼Y 1
n¼1
1 e
en
expX
3
k¼1
e en
k
=k
:
(2.8)
The reality ofdet4for realeandC-invariance require that the eigenvaluesenappear in the complex plane as quartets
en, en, or as imaginary pairs when m0. As ex-pected, the expansion of ln det4 in powers of ebegins in fourth order.
We have established that det4je¼0¼1 and that det4 has no zeros for real values of e. Therefore, by (2.6) detQED3>0for all reale, thereby allowing one to go from
(1.5) and (1.6). It might be objected that this is obvious, but we will need the detailed information introduced about det4in the sequel.
The determinantdet4 is an entire function ofe consid-ered as a complex variable, meaning that it is holomorphic in the entire complex e plane. Since P1n¼1jenj3<1 for >0, its order is at most 3 [16,18]. This means that for any complex value of e, and positive constants A, K,
jdet4j< AðÞexpðKðÞjej3þÞ for any >0 From (1.6) and (2.6), and for real values ofe
ln det4 e2
4
Z d3k
ð2Þ3j
^
BðkÞj2
Z1
0
dz zð1zÞ
½zð1zÞk2þm21=2: (2.9)
This is a truly remarkable inequality. Referring to (2.9), det4’s growth is slower on the realeaxis than its potential growth in other directions. We also note thatdet4is largely unknown. Even the reduction of the fourth-order term in its expansion to an explicitly gauge invariant form involving onlyBfields requires a huge effort when the fields are not constant [19]. The sixth-order reduction has not been com-pleted as far as the author knows.
C. Upper bound ondetren
Insert (2.6) in (2.5) and get
ln detren¼e
2T
42
Z d3k
ð2Þ3jB^ðkÞj 2Z1
0
dzzð1zÞ
ln
zð1zÞk2þm2 m2
þT
Z1
m2
dM2
ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
M2m2
p ln det4ðM2Þ: (2.10)
The objective here is to obtain the behavior of lndetren when the coupling eis large, real, and positive. Since e always multiplies B, we introduce the scale parameter
is finite will be explained in Sec.III B. Then the integral in (2.10) is broken up intoReBm2 and
R1
eB.
Substitution of (2.9) into the lower range integral gives
ln detren e2T
42
Z d3k
ð2Þ3jB^ðkÞj 2Z1
0
dzzð1zÞ
ln
4eBþ2zð1zÞk22m2
m2
þT
Z1
eB
dM2
ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
M2m2
p ln det4ðeB; M2Þ: (2.11)
We have simplified the argument of the logarithm using 2pffiffiffiffiffixyxþyforx,y0. Then foreB m2,
lndetrene
2TjjBk2
242 ln
4eB m2
þT
Z1
eB dM2
ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
M2m2
p lndet4ðeB;M2Þ
þO
eTRd3x Br2B
B
: (2.12)
The integral in (2.12) can be estimated by making a large mass expansion of ln det4. This is facilitated by inserting (2.6) in (2.4) and examining the smalltregion ofln det4’s resulting proper time representation. The details of this expansion are in Sec. 3B of [10], and give the result
ln det4ðeB; M2Þ ¼ M!1
1 2
Z1
0 dt
t ð4tÞ
3=2etM2Z
d3x
2 45e
4t4ðBBÞ2þOðe4t5BBB r2BÞ
¼e4
R
ðBBÞ2
480M5 þO
e4RBBB r2B M7
: (2.13)
In the first line of (2.13), it is assumed that the heat kernel expansion is an asymptotic expansion intin the strict sense of its definition, namely [20]
< xjet½ðPeAÞ2eBjx >ð4tÞ3=2 X N
n¼0 anðxÞtn
t!0þð4tÞ3=2aNþ1ðxÞtNþ1: (2.14)
This must hold for every N. A necessary condition for (2.14) is thatB be infinitely differentiable to ensure that each coefficientanis finite. As far as the author knows, it is not known yet if this is a sufficient condition. So (2.14) is an assumption that may require additional conditions onB. Only coefficients an of Oðe2nÞ, n2 are present in ln det4’s expansion.
Thetintegration in (2.13), although extending to infin-ity, is limited to smalltsinceM! 1due to the parameter eBin (2.12). Substituting (2.13) in (2.12) results in
ln detren
e2jjBk2T
242 ln
4eB
m2
þO
e2TRðBBÞ2
B2
þO
eTR
B r2B
B
; (2.15)
or
lim e!1
ln detren
e2lne
kBk2T
242 ; (2.16)
consistent with (1.4). This bound is independent of the charge renormalization subtraction point. If the subtraction
were made at photon momentum k2¼2 instead of k2 ¼0, then the lnm2 terms in (2.10) and (2.11) would be replaced withln½zð1zÞ2þm2, which has nothing to do with strong coupling.
The scaling procedure used here is designed to obtain the least upper bound on ln detren. In [10], we chose to break up the M integral as Rem42kBk4 and
R1
e4kBk4. This
resulted in a fast1=e4falloff of theln det
4terms compared to e2 here, but gave a weaker upper bound on ln det
ren, namely
lim e!1
ln detren e2lne
kBk2T
62 : (2.17)
We mention that the coefficient 1=960 in (3.16) in [10] should be1=360.
Here we might have chosen a more general scaling, such as eðlneÞB or eðln lneÞB, etc., with 1, >0. Then the right-hand side of (2.16) would have been re-placed withkBk2T=242. The case <1causes the first remainder term in (2.15) to be no longer subdominant. Therefore, our scalingeBis an optimal one.
III. LOWER BOUND ONdetren
A. Fundamentals
On referring to (2.5) the lower bound ondetrenwill come from operations onln detQED3. We begin with the operator
ln detQED3 ¼ 1 2
Z1
0 dt
t Tr
eZt 0
dseðtsÞðPeAÞ2
BesðPeAÞ2
þe2Zt 0
ds1Zts1 0
ds2eðts1s2Þ½ðPeAÞ2eB
Bes2ðPeAÞ2Bes1ðPeAÞ2
em2t þ1
2
Z1
0 dt
t Trðe P2t
eðPeAÞ2t Þetm2
: (3.1)
The spin trace in the first term is zero, and the last term, after tracing over spin, is the one-loop effective action of scalar QED3,
ln detSQED3 ¼
Z1
0 dt
t e tm2
TrðeP2teðPeAÞ2tÞ: (3.2)
Thus,
lndetQED3¼lndetSQED3
e2
2
Z1
0 dt
t e tm2
TrZt
0 ds1
Zts1
0
ds2eðts1s2Þ½ðPeAÞ
2eB
Bes2ðPeAÞ2Bes1ðPeAÞ2
;
(3.3)
remembering that the factor1=2in the last term of (3.1) is canceled by the spin trace. LetA¼ ½ðPeAÞ2þm21. In AppendixB, it is shown that1A=2B
1=2
A 2I2; that is, it is a Hilbert-Schmidt operator providedB2L2andm0. Then (2.7) gives
ln det2ð1e1A=2B1A=2Þ ¼ln det½ð1eA1=2B1A=2ÞeeA1=2B1A=2
¼Tr ln½ð1e1A=2BA1=2ÞeeA1=2B1A=2
¼TrZ
1
0 dt
t e tm2
ðeðPeAÞ2t
e½ðPeAÞ2eBt
þeAB
¼ e2Z1 0
dt t e
tm2Zt
0 ds1
Zts1
0
ds2Trðeðts1s2Þ½ðPeAÞ
2eB
Bes2ðPeAÞ2Bes1ðPeAÞ2Þ: (3.4)
In going from the penultimate to the last line in (3.4), use was again made of the identity (A2). Substituting (3.4) in (3.3) gives
ln detQED3 ¼ 1
2 ln det2ð1e
1=2
A B
1=2 A Þ
þln detSQED3: (3.5)
Asln detQED3andln det2 are well-defined by our choice of
fields, so is ln detSQED3 in (3.5). What has been
accom-plished here is to isolate the Zeeman termBinln det2. Since 1A=2B1A=2 is Hilbert-Schmidt and self-adjoint, ln det2 is susceptible to extensive analytic analysis.
Substitute (3.5) in (2.5):
lndetren¼ T
Z1
m2
dM2
ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
M2m2
p 1
2lndet2ð1e
1=2 A B
1=2 A Þ
þlndetSQED3þ
e2kBk2
24 ffiffiffiffiffiffiffiM2
p : (3.6)
We now introduce two central inequalities. The first relies on the diamagnetism of charged scalar bosons as expressed by Kato’s inequality in the form [21,22]
TrðeðPeAÞ2t
Þ TreP2t
: (3.7)
This implies that on average the energy eigenvalues of such bosons rise in a magnetic field and hence by (3.2) that [22]
ln detSQED30: (3.8)
The second inequality is introduced beginning with the penultimate line of (3.4). Noting that the spin trace of the ABterm is zero, then
lndet2ð1e1A=2B 1=2 A Þ
¼Z1
0 dt
t e tm2
TrðetðPeAÞ2
e½ðPeAÞ2eBt
Þ: (3.9)
By the Bogoliubov-Peierls inequality [23,24]and Sec. 2.1, 8 of [25]
Tre½ðPeAÞ2eBt
TretðPeAÞ2t
etehBi; (3.10)
where
hBi ¼TrðBe
tðPeAÞ2t Þ
Hence,
ln det2ð1e1A=2B 1=2
A Þ 0; (3.12) which is consistent with (3.5) when combined with (1.6) and (3.8). There is another reason why (3.12) holds. Let C¼e1A=2B1A=2 SinceCis Hilbert- Schmidt,
ln det2ð1CÞ ¼ln det½ð1CÞeC
¼Tr½lnð1CÞ þC
¼1
2Tr lnð1C
2Þ
¼1 2
X1
n¼1
lnð1 2
nÞ: (3.13)
The third line of (3.13) follows from the second since the trace over spin eliminates all odd powers ofC. In the last
line, we introduced the real eigenvalues n of e1A=2
B1A=2. Sinceln det2is real and finite, thenj nj<1for all n, giving (3.12). Because1A=2B1A=2I2, it is a com-pact operator, and so the n are countable and of finite multiplicity.
Now consider
@
@m2lndet2ðm 2Þ
¼Z1
0
dtetm2
Trðe½PeA2eBt
eðPeAÞ2t Þ 0;
(3.14)
by (3.9), (3.10), and (3.11). Therefore,det2is a monotoni-cally increasing function of m2. Next, break up the M integral in (3.6) as in Sec.II C:
ln detren¼ T
ZeB
m2
dM2
ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
M2m2
p 1
2 ln det2ð1e
1=2
A B
1=2
A Þ þln detSQED3þ
e2k Bk2
24 ffiffiffiffiffiffiffiM2
p
þT
Z1
eB
dM2
ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
M2m2
p ln detQED3þ
e2kBk2
24pffiffiffiffiffiffiffiM2
; (3.15)
where we reinserted (3.5) into the upper-rangeMintegral. By (3.14)
ZeB
m2
dM2
ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
M2m2
p ln det2ðM2Þ ln det 2jM2¼m2
ZeB
m2
dM2
ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
M2m2
p ¼2 ln det2ð1e1A=2B1A=2ÞjM2¼m2
ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
eBm2
p
:
(3.16)
Hence, (3.8) and (3.16) result in (3.15) becoming
ln detren T
ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
eBm2
p
ln det2ð1e1A=2B 1=2
A ÞjM2¼m2þ
e2TkBk2
242 ln
eB
m2
þ e2T 122kBk
2ln
1þ
ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
1m
2
eB
s
þT
Z1
eB
dM2
ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
M2m2
p ln detQED3þ e
2k Bk2
24pffiffiffiffiffiffiffiM2
: (3.17)
We now turn to the strong coupling behavior ofln det2.
B. Strong coupling behavior ofln det2
The eigenvalues n in (3.13) are obtained from
e1A=2BA1=2’n¼ n’n; (3.18)
for ’n2L2 following the remark under (B3) in AppendixB. Letting1A=2’n¼ cngives
ðPeAÞ2eB
n
cn¼ m2cn; (3.19)
where cn 2L2 providedm0. This follows from (B5) and Young’s inequality (B7). The requirement thatm0 follows from the role of the eigenvaluesf ng1n¼1 as adjust-able coupling constants whose discrete values result in bound states with energy m2 for a fixed value of e.
Since the operator ðPeAÞ2eB0, such bound states are impossible unlessj nj<1for alln, which is the physical reason why (3.12) is true. Inspection of (3.19) suggests that as e increases j nj likewise increases for fixed nto maintain the bound state energy at m2. This is illustrated by the constant field case that is excluded from our analysis:
j nj ¼
jeBj
ð2nþ1ÞjeBj þm2; n¼0;1;. . .: (3.20)
c;
ðPeAÞ2eB
c
c;
ðPeAÞ2
e
jBj
c: (3.21)
Thus, the Hamiltonian on the left,Hþ, dominates that on the right,H. Let Nm2ðHÞ denote the dimension of the
spectral projection onto the eigenstates of HamiltonianH with eigenvalues less than or equal to m2. Because HþH, then Nm2ðHþÞ Nm2ðHÞ. Nm2ðHþÞ is
an overestimate of the number of the bound states ofHþ
at m2 for a fixed value of but satisfactory for our purpose here.
By the Cwinkel-Lieb-Rozenblum bound in the form [26]
Nm2ðHÞ C
Z
d3x
e
jBðxÞj m2 3=2
þ ; (3.22)
where½aþ¼maxða;0ÞandC¼20:1156. The factor 2 accounts for the additional spin degrees of freedom in the present estimate. Since j nj ¼Oð1Þ, we are confident that the degeneracy/multiplicity associated with each n in (3.13) does not exceed cjej3=2Rd3xjBj3=2, where c
0:2312 is another finite constant. This estimate has to be modified for values ofn > Nbeyond which nassumes its asymptotic form as discussed below. Therefore, fornN we will estimate the sum in (3.13) by factoring out the common maximal degeneracycjej3=2Rd3xjBj3=2and treat each nin the factored sum as having multiplicity equal to one. Those n, if any, that vanish as e! 1 give a sub-dominant contribution toln det2 in (3.13) since by inspec-tion their contribuinspec-tion grows at most as 2
nje= nj3=2. We now turn to the largeedependence of n. From here on we assume that cn is normalized to one. By C-invariance we may assume e >0. Now consider the expectation value of (3.19):
hnjðPeAÞ2jni e n
hnjBjni ¼ m2: (3.23)
From (3.23) ifhnjBjni>0then n>0andvice versa. Therefore, we need only consider n>0and write
n¼
hnjð
PeAÞ2jni ehnjBjni þ
m2 ehnjBjni
1
; (3.24)
wherehnjBjni0as (3.23) must be satisfied. The case n ¼0 for some n corresponding to hnjBjni ¼0 can be ignored as n¼0contributes nothing toln det2 in (3.13). An easy estimate gives
jðcn;BcnÞj ðcn;jBjcnÞ max
x jBðxÞj: (3.25) Because B2L2 and is assumed infinitely differentiable, then maxxjBj is finite. Hence, hnjBjni is a bounded function ofeandn.
Now consider the ratio Rn¼ hnjðPeAÞ2jni= ehnjBjni in (3.24). The case Rn!e 10 is ruled out
since this implies n! 1. The case Rn!e 11 implies n!0, which gives a subdominant contribution to (3.13) as discussed above. The final possibility is 1Rn<1 for e! 1. The case Rn!1 for e! 1 happens if hnjðPeAÞ2jni ehnjBjni. Since c
n2L2,
hnjðPeAÞ2eBjni ¼0 implies ðPeAÞc n¼ 0. Now this may happen for the B fields considered so far. But if we exclude zero-mode supportingBfields [27] from our analysis it cannot. By so doing, we can exclude the case n¼1nðeÞ,nð1Þ ¼0. We will see below why this is necessary.
We proceed to estimate the strong coupling limit of ln det2 in (3.13). First, consider the sum for nN. We need only consider0<j nj<1for alle, includinge¼ 1 as concluded above. Hence, on factoring out the common maximal multiplicity of the nwe get
lim e 1
XN
n¼1
lnð1 2 nÞ
c1e3=2
Z
d3xjBj3=2; (3.26)
wherec1is a constant and noting again that the eigenvalues n!0are subdominant.
Since n!0forn! 1and1=2 jlnð1 2nÞ= 2nj 3=2for 2n<1=2, the absolute convergence of the series in (3.13) requiresP1n¼1 2
n<1. Consider this sum forn > N and indicate the degeneracy factorsn explicitly:
S X 1
n>N
nðeÞ 2nðeÞ: (3.27)
We estimated from (3.22) that ncje= nj3=2
R
d3xjBj3=2. So
Sce3=2Z d3xjBj3=2 X 1
n>N
no degeneracy
j nj1=2<1: (3.28)
This implies that forn > N
j nðeÞj ¼CnðeÞ
n2þ ; (3.29)
where >0andCnis a bounded function ofnandewith lime!1CnðeÞ<1. Otherwise, j nj<1for any ncannot be satisfied. Accordingly, the series in (3.28) is uniformly convergent ineby the WeierstrassMtest and so
lim e!1
X1
n>N
lnð1 2 nÞ
=e3=2 c2
Z
d3xjBj3=2; (3.30)
wherec2is a constant. From (3.13), (3.26), and (3.30), we conclude
lim
e!1jln det2ð1e 1=2
A B
1=2 A Þj=e3=2
c3Z d3xj
Bj3=2; (3.31)
As a check on (3.31), refer to (3.5). ForB2L3=2ðR2Þ, we found [10]
ln detQED3 Ze
3=2
6
Z
d2xjBðxÞj3=2; (3.32)
forBðxÞ 0orBðxÞ 0,x2R2.Z is the dimension of the remaining space box. We know that ln det20 and
ln detSQED30 in (3.5). Specializing (3.31) to these B
fields, it is seen that the strong coupling growth ofln det2 is consistent with (3.32).
Finally, if zero-mode supportingBfields were allowed, we would have obtained ln det2¼e 1Oðe3=2lnðeÞÞ, ðeÞ!e 10, since when j nje 11nðeÞ, nð1Þ ¼0, the logarithm in (3.13) gives an additional factor lnn. As will be seen below the limit (1.4) requires lime!1ln det2=e3=2 ¼finite (or zero).
C. Strong coupling limit of (3.17)
It remains to estimate the large coupling limit of the last term in (3.17),
IT
Z1
eB
dM2
ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
M2m2
p ln detQED3þ
e2kBk2
24 ffiffiffiffiffiffiffiM2
p
¼T
Z1
eB
dM2
ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
M2m2
p e2 4
Z d3k
ð2Þ2jB^ðkÞj 2
Z1
0
dz zð1zÞ
½zð1zÞk2þM21=2þ
e2kBk2
24pffiffiffiffiffiffiffiM2
þln det4ð1eS6AÞ
; (3.33)
where we substituted (2.6) forln detQED3. Calculation of the
first two terms in (3.33) is straightforward. The last term has already been estimated in Sec.II and is given by the second term in (2.15). Hence,
I¼O
e2TRðBBÞ2
B2
þO
eTR
B r2B
B
: (3.34)
Taking into account (3.31) and (3.34), we obtain from (3.17)
lim e!1
ln detren
e2lne
kBk2T
242 : (3.35)
Equations. (2.16) and (3.35) therefore establish (1.4).
IV. SUMMARY
The two assumptions underlying (1.4) are first that the continuum limit of the lattice diamagnetic inequality co-incides with (1.5), and second that the heat kernel expan-sion of the Pauli operator in (2.14) is an asymptotic series. These assumptions can and should be proven or falsified.
In addition, the result (1.4) assumes that the vector potential and magnetic field satisfy the following conditions:
B2L2ðR3Þto defineln det
renin (2.5) and to ensure that
A1=2BA1=22I2 following Appendix B. In addition B2L3=2ðR3Þ in order that the degeneracy estimate in (3.22) is defined. To ensure that the bound in (3.31) holds, zero-mode supportingBfields are excluded. Also,Bmust be infinitely differentiableðC1Þ to ensure that the expan-sion coefficients in (2.14) are finite.
If A is assumed to be in the Coulomb gauge, then by (2.3) A2L6ðR3Þ. If
B2L3=2ðR3Þ, then
A2L3ðR3Þby the Sobolev-Talenti-Aubin inequality [12]. In order to define detQED3, it is necessary to assume A2LrðR3Þ, r >3, following the discussion under (2.6). IfA2L3ðR3Þ andL6ðR3Þ, thenA2LrðR3Þ,3< r <6also. This follows from Ho¨lder’s inequality [28]
kfgkr jjfkpkgjjq; (4.1)
withp1þq1 ¼r1,p,q,r1. SinceB¼rAand B2C1, thenA2C1.
We note that the sample functions AðxÞ supporting the Gaussian measure in (1.2) with probability one are not C1. It is generally accepted that they belong to
S0ðR4Þ, the space of tempered distributions. Therefore, we point out here that the C1 functions we introduced can be related to A 2S0ðR4Þ by the convoluted field A
ðxÞ¼
R
d4yf
ðxyÞAðyÞ2C1, providedf2SðR4Þ, the functions of rapid decrease. Then the Fourier transform of the covarianceRdðAÞA
ðxÞAðyÞderived from (1.2) is ^
DðkÞjf^ðkÞj2, where f^ 2C1. Since QED4 must be ultraviolet regulated before renormalizing, f^ can serve as the regulator by choosing, for example, f^ ¼1, k2 2, andf^ ¼0,k222. So the need to regulate can serve as a natural way to introduce C1 background fieldsA
intodetren—but not the rest ofdin (1.1)—and into whatever else one is calculating. This procedure is a generalization of that used in the two-dimensional Yukawa model [29].
Finally, the obvious generalization of (1.4) for an ad-missible class of fields onR4 is
lim e!1
ln detren
e2lne ¼
1 482
Z
d4xF2
ðxÞ: (4.2)
There is no chiral anomaly term since F falls off faster than 1=jxj2 and Rd4xF~
F¼
R
d4x@
ðAFÞ ¼0., where F~¼12F. Equation (4.2) remains to be verified.
If (1.4) and (4.2) do indeed indicate instability, then they are yet another reason why QED should not be considered in isolation.
APPENDIX A
Ft¼etðXþYÞetX:
Then
dFt dt ¼ e
tðXþYÞYetX:
Integrating gives
etðXþYÞetX¼ Zt 0
dseðtsÞðXþYÞYesX; (A1)
known as Duhamel’s formula. Iterating once gives the required identity:
etðXþYÞetX
¼ Zt
0
dseðtsÞXYesXþZt 0
ds1
Zts1
0
ds2eðts1s2ÞðXþYÞYes2XYes1X: (A2)
APPENDIX B
Here we show that the operatorK¼1A=2B1A=22I2 and hence that K is Hilbert-Schmidt. This follows [14,15,28] if and only if K is a bounded operator on L2ðR3; d3x;C2Þhaving a representation of the form
ðKfÞðxÞ ¼Z Kðx; yÞfðyÞd3y; f2L2; (B1)
where
Kðx; yÞ ¼ hxj1A=2B1A=2jyi; (B2)
and whereK2L2ðR3R3;d3xd3yÞ. Moreover,
kKk2 2 ¼
Z
jKðx; yÞj2d3xd3y: (B3)
If it can be shown that K2L2, then it trivially follows thatKmapsL2 into itself. So consider
kKkL2 ¼2
X
i
Z
d3xd3yB
iðxÞAðx; yÞBiðyÞAðy; xÞ
2X i
Z
d3xd3yjB
iðxÞkAðx; yÞkBiðyÞkAðy; xÞj:
(B4)
A form of Kato’s inequality [5,21,31] asserts that the interacting scalar propagator is bounded by the free propa-gator
jAðx; yÞj ðxyÞ; (B5)
whereðxÞ ¼ ð4jxjÞ1emjxj in three dimensions. Then
kKkL2 1 82
X
i
Z
d3xd3yjBiðxÞ
1 jxyj2e
2mjxyj
BiðyÞj: (B6)
By Young’s inequality in the form [23]
Z
d3xd3yfðxÞgðxyÞhðyÞ
kfkpkgkqjjhkr; (B7)
where p1þq1þr1¼2, p, q, r1, and kfkp¼
ðRd3xjfðxÞjpÞ1=p etc., obtain from (B6) with p¼r¼2, q¼1
kKkL2 1 82
X
i
jjBik2Z d3xe2mjxj=x2
¼ kBk2ð4mÞ1: (B8)
Therefore, by the theorem that began this appendix, 1A=2B1A=22I2 whenm0andB2L2. We men-tion that this can be proved even when m¼0 provided B2L3=2.
[1] W. Heisenberg and H. Euler, Z. Phys. 98, 714 (1936) [arXiv:physics/0605038].
[2] G. V. Dunne, in Ian Kogan Memorial Collection from Fields to Strings: Circumnavigating Theoretical Physics, edited by M. Shifman, A. Vainshtein, and J. Wheater (World Scientific, Danvers, MA, 2005), Vol. 1, p. 445. [3] Z. Haba,Phys. Rev. D29, 1718 (1984).
[4] G. V. Dunne and Q.-h. Wang, Phys. Rev. D 74, 065015 (2006).
[5] D. Brydges, J. Fro¨hlich, and E. Seiler,Ann. Phys. (N.Y.) 121, 227 (1979).
[6] E. Seiler, Lecture Notes in Physics (Springer, Berlin/ Heidelberg/New York, 1982), Vol. 159.
[7] E. Seiler, in Proceedings of the International Summer School of Theoretical Physics, Poiana Brasov, Romania, 1981, edited by P. Dita, V. Georgescu, and R. Purice, Progress in Physics Vol. 5 (Birkha¨user, Boston, 1982), p. 263.
[12] E. H. Lieb and W. E. Thirring, inStudies in Mathematical Physics, Essays in Honor of Valentine Bargmann, edited by E. H. Lieb, B. Simon, and A. S. Wightman (Princeton University Press, Princeton, NJ, 1976), p. 269.
[13] E. Seiler and B. Simon, Commun. Math. Phys. 45, 99 (1975).
[14] B. Simon, Trace Ideals and Their Applications, London Mathematical Society Lecture Note Series Vol. 35 (Cambridge University Press, Cambridge, England, 1979). [15] N. Dunford and J. Schwartz, Linear Operators Part II:
Spectral Theory(Interscience, New York, 1963).
[16] I. C. Gohberg and M. G. Kreıˇn,Introduction to the Theory of Linear Nonselfadjoint Operators, Translations of Mathematical Monographs Vol. 18 (American Mathematical Society, Providence, RI, 1969).
[17] B. Simon,Adv. Math.24, 244 (1977).
[18] B. Ja. Levin, Distribution of Zeros of Entire Functions, Translations of Mathematical Monographs Vol. 5 (American Mathematical Society, Providence, RI, 1964).
[19] R. Karplus and M. Neuman, Phys. Rev. 80, 380 (1950).
[20] C. M. Bender and S. A. Orszag, Advanced Mathematical Methods for Scientists and Engineers (Springer, New York, 1999).
[21] T. Kato,Isr. J. Math.13, 135 (1972); B. Simon,Math. Z. 131, 361 (1973); M. Schechter, J. Funct. Anal. 20, 93 (1975); H. Hess, R. Schrader, and D. A. Uhlenbrock,Duke
Math. J.44, 893 (1977); B. Simon,Indiana Univ. Math. J. 26, 1067 (1977);J. Funct. Anal.32, 97 (1979).
[22] R. Schrader and R. Seiler,Commun. Math. Phys.61, 169 (1978).
[23] E. H. Lieb,Rev. Mod. Phys.48, 553 (1976).
[24] H. Hogreve, R. Schrader, and R. Seiler,Nucl. Phys.B142, 525 (1978).
[25] W. Thirring, Lehrbuch der Mathematischen Physik 4: Quantenmechanik Grosser Systeme (Springer, Vienna, New York, 1980).
[26] E. H. Lieb, M. Loss, and J. P. Solovej,Phys. Rev. Lett.75, 985 (1995).
[27] M. Loss and H.-T. Yau,Commun. Math. Phys.104, 283 (1986); C. Adam, B. Muratori, and C. Nash,Phys. Rev. D 60, 125001 (1999);62, 085026 (2000);Phys. Lett. B485, 314 (2000); D. Elton,J. Phys. A33, 7297 (2000); L. Erdo¨s and J. P. Solovej, Rev. Math. Phys. 13, 1247 (2001); A. Balinsky and W. D. Evans, J. Funct. Anal. 179, 120 (2001); Bull. Lond. Math. Soc. 34, 236 (2002); D. Elton, Commun. Math. Phys. 229, 121 (2002); G. V. Dunne and H. Min,Phys. Rev. D78, 067701 (2008). [28] M. Reed and B. Simon,Functional Analysis(Academic,
New York, 1972).
[29] E. Seiler,Commun. Math. Phys.42, 163 (1975). [30] H. L. Cycon, R. G Froese, W. Kirsch, and B. Simon,
Schro¨dinger Operators (Springer, Berlin/Heidelberg, 1987).