These are comments on dimensional regularization of Feynman integrals, which was discussed in the lectures of K.Gawedzki and D.Gross. This material can be found in physical literature, but in a slightly di erent form. We will rst de ne a certain remarkable operator (the D-dimensional integral), which acts on functions of nonnegative symmetric bilinear forms, and then show how it applies to dimensional regularization. Acknowledgement: I am grateful to Pierre Deligne, who helped me improve this text. 1. The D-dimensional integral. Let E be a real vector space of dimension N with a volume element dvE . Denote 2 by S 2E the space of symmetric bilinear forms on E , and by S+E S 2E the set of nonnegative symmetric bilinear forms. We think of elements of S 2E as of selfadjoint operators E ! E . For any such operator A, we can canonically de ne its determinant det(A) 2 R. Let dA be a volume element on S 2E de ned by dvE . That is: identify E with E by any positive operator X : E ! E with deteminant 1, regard A 2 S 2E as a selfadjoint operator on E using this identi cation, and de ne dA to be the volume element on S 2E corresponding to the norm jjAjj2 = Tr(A2 ). (It is clear that this de nition does not depend on X ). Fix a complex number D with ReD > N 1. Lemma 1. For any B 2 S 2E such that B > 0


2 +

(det A)(D N

1)=2 e tr(AB ) dA = 1)=4

N (D=2)(det B ) D=2 ;


N (D=2) = (2 )N (N

N 1 Y l=0

D+l : 2

Proof. Denote this integral by f (B). It is easy to see that for any S 2 GL(E ), f (S BS ) = f (B)j det S j D . From this it follows that f (B) = C (D)(det B) D=2 . It remains to compute C (D). Let us pick B such that det B = 1, and identify E with E using the isomorphism B : E ! E . Then E becomes a Euclidean space with the inner product B 1. Thus the constant C (D) has the form


A:E !E;A =A;A 0

(det A)(D N

1)=2 e trA dA
Typeset by


Since this integral is invariant under conjugation by SO(E ), we can reduce it to the 2 quotient space S+E =SO(E ) = 0; 1)N =SN . It is known from the theory of sym2 2 metric spaces that the pushforward of dA from S+E to S+E =SO(E ) has the form Q 2 bN 1 i<j N jsi sj jds1:::dsN , where si are the eigenvalues of A 2 S+E =SO(E ), and N (N 1)=4 N (N +1)=4 ; bN = 2 QN l=1 (l=2) Thus, C (D) has the form (2)

bN N !




0;1)N i


N 1)=2 Y e ti Y jt i i 1 i<j N

tj jdt1:::dtN :

To compute this integral, we use (a special case of) Selberg's integral formula:


0 ti 1 i

Y Y aY jti tj jdt1:::dtN = ti (1 ti)b
1 i<j N N l Y (a + l+1 ) (b + l+1 ) ( 2 + 1) 2 2 : (a + b + 2N 2l+3 ) ( 3 ) 2 l=1



Let us set si = bti in Selberg's integral, and send b to +1. In the limit, using Stirling's formula, and setting a = (D N 1)=2, we get

C (D) = bN N !


N N=2 Y l=1

l l ( D+2 N ) ( 2 + 1) = (3) 2 D+l N l = (D=2):: N 2 2

1Y l=1


The Lemma is proved. 2 Now consider the function E on S 2E which is supported at S+E and is given D by the formula E (A) = ND=2 N (D=2) 1 (det A)(D N 1)=2 : (5) D By Lemma 1, we have (6)


2 +

E (A)e tr(AB ) dA = ND=2(det B ) D=2 ; D

ReD > N 1. If ReD > N 1, the function E (A) is a locally L1-function, so it de nes a D distribution on S 2E . Let us compute the Fourier transform of E (A), i.e. D ^E (B) = D


B 2 S 2E (the integral is understood in the sense of distributions).

S2 E

E (A)e itr(AB )dA; D

Lemma 2.
(7) ^E (B) = Jlim ND=2 det(iB + J ) D=2: D !+0
where J 2 S 2E ; J > 0. Remark. Here we use the branch of the function z D=2 which is equal to e D ln z=2 for real positive z. Proof. By analytic continuation, formula (6) holds for arbitrary B 2 (S 2E )C with ReB > 0. In particular, we can replace B with iB + J , where B 2 S 2E . In the limit when J ! +0, we obtain formula (7). Let : S 2E ! Hom(E ; E ) be the tautological embedding. Let be the polynomial of degree N on S 2E , de ned by the equation (B) = det( (B)). We regard as a di erential operator of order N on S 2E with constant coe cients. Corollary 3. If ReD > N + 1, then


E ( A) = N E ( A) : D D 2

Proof. The proof is obvious after passing to Fourier transforms. Let S (S 2E ) be the space of complex-valued Schwarz functions on S 2E . For any f 2 S (S 2E ) and ReD > N 1, de ne


E ID (f ) =


E Proof. For ReD > N 1 2K , K 2 Z+, we can de ne ID (f ) by the formula

E Corollary 4. ID (f ) extends to an entire function of D. E ID ( f ) = (

S2 E

E (A)f (A)dA: D


) KN


S2 E

K E D+2K (A) f (A)dA:

According to Corollary 3, if ReD >> 0, formula (10) de nes the same function as E (9). Thus, (10) de nes an analytic continuation of ID . Proposition 5. Let D 0 be a nonnegative integer, and V be a vector space of dimension D, with a positive de nite symmetric bilinear form . Then for any f 2 S (S 2E ) one has (11)
E ID ( f ) =


Hom(E;V )

f (x ( ))dx;

where x ( ) denotes the inverse image of under x (a nonnegative symmetric bilinear form on E ). Proof. It is enough to check that (11) holds for functions e tr(AB)1S E (A) where 1Y is the characteristic function of the set Y , and Re(B) > 0. This follows from the fact that the function f (A) admits a Fourier representation.
2 +


In the case (12)

f ( A) = e

tr(AB )1

2 S+ E (A);

evaluating the Gaussian integral on the right hand side of (11), we get ND=2(det B) D=2 . On the left hand side of (11), by Lemma 1, we get the same. The proposition is proved. Remark. From Proposition 5 it is clear why formula (9) works only for ReD > N 1. Indeed, for integer D between 0 and N 1 the distribution D (A)dA is concentrated on forms of rank D, so it is a purely singular distribution. Thus, the measure D (A)dA is not absolutely continuous with respect to dA and hence cannot be represented as a function times dA. In fact, it turns out that this distribution is singular (not locally L1 ) for all D with ReD N 1. Proposition 5 motivates the following de nition. E De nition. ID (f ) is called the D-dimenional integral of f . Another way of computing the D-dimensional integral is using Fourier transforms. By Plancherel formula, it follows from (9) that (13)
E ID (f ) = (2 ) N (N +1)=2



f^(B)^E (B)dB; D

where f^ is the Fourier transform of f . (It is clear that ^E (B) = limJ !+0 ND=2 det( iB+ D J ) D=2). For ReD < 2, this formula can be written as (14)
E ID (f ) = 2 N (N +1)=2 N (D N 1)=2





(B ) j det(B )j D=2 dB;

where (B) is the signature of B (the number of pluses minus the number of minuses). If ReD 2, the distribution ^E (B) is singular, so in order to de ne D E ID (f ) by a direct integration formula, we need to use the operator . Namely, from Lemma 1 it follows that (15) ^E (B) = ( 1)N pN (D)^E +2(B); D D

pN (D) =

N N YD+l 2 l=1


E So if ReD < 2 + 2K , we can de ne ID by the formula K E (f ) = 2 N (N +1)=2 N (D N 1)=2 Y pN (D 2J ) 1 ID J =1 Z K f (B )e i(D4 2K ) (B ) j det(B )j( D+2K )=2 dB; ^ 2E S


just the Mellin transform of f :

Example. If N = 1, the D-dimensional integral of f is (up to normalization)
E ID (f ) = D=2 (D=2) 1


x D 1 f (x)dx:

2 2 Let S (S+E ) be the space of smooth functions f on S+E , such that all the derivatives of f are rapidly decaying. It is easy to see that the restriction map 2 E 2 S (S 2E ) ! S (S+E ) is surjective. So, ID de nes a linear functional S (S+E ) ! C.

Now we will extend our construction to a more general situation. Let F E be a subspace of dimension M with a volume element dvF . Any symmetric bilinear form A on E de nes a form AF on F (the restriction of E ). Also, for any symmetric bilinear form B on E such that BF ? is nondegenerate we can de ne a form BF on F as follows. Let (F ? )?B be the B-orthogonal complement of F ? . As BF ? is nondegenerate, this space is naturally identi ed with F = E =F ?. By the de nition, the form BF is the restriction of B to (F ? )?B = F . Let us nd an explicit expression of BF . Choose a positive de nite inner product on E and thus identify E with E . Then we have a decomposition E = F F ?, and B becomes a selfadjoint operator E ! E with a decomposition B = B11 B12 ; B21 B22 where Bij = Bji . Now the operator corresponding to BF ? is B22, and the operator corresponding to BF is B11 B12B221 B21. From this formula it is obvious that for this operator to be de ned, it is su cient for B22 to be invertible. The following Lemma is a generalization of Lemma 1. Lemma 6. For ReD > N 1 and any positive B 2 S 2E , C 2 S 2F

2. D-dimensional integral with parameters.



function (AF C ) is well de ned. Proof. Denote the left hand side of (17) by I (B; C ). Let S 2 GL(E ) be an automorphism which preserves F . Make a change of variable A ! S AS . Then AF ! SF AF SF , where SF is the restriction of S to F . Thus we get (18) I (B; C ) = j det S jN M I (SBS ; (SF ) 1 CSF 1): Now choose B with det BF ? = 1, and identify E with E using the map B : E ! E . Then E becomes a Euclidean space with the inner product B 1 . Now we can identify F with F using the map BF : F ! F , thus F is also a Euclidean space.

Remark. The space S 2F inherits a volume element from F , so the delta-

N (D=2) e tr(CB F ) (det C )(D M 1)=2(det B ? ) D=2 ; F M (D=2)

2 +

(AF C )(det A)(D N

1)=2 e tr(AB ) dA =

After these identi cations, integral (17) takes the form (19)


Using the decomposition E = F F ? , we can write A in the form A A (20) A = A11 A12 ; 21 22 where Aij = Aji . This allows us to rewrite (19) as (21)

A2 S E
2 +


C )(det A)(D

N 1)=2 e trA dA:


To compute this integral, let us use the identity det A = det A11 det(A22 A21A111 A12). Since we are integrating over such A that A11 = C , we get det A = det C det(A22 A21 C 1A12 ). After the change of variable X = A22 A21 C 1A12, integral (21) takes the form (22) Z Z (D N 1)=2e trC (D N 1)=2e trX dX (det C ) (det X ) e trA C A dA12 The second integral in (22) is Gaussian and equals (23) (2 )M (N M )=2(det C )(M N )=2: The rst integral in (22) was computed in Lemma 1 and equals N M ((D M )=2). Using the identity N (D=2) = (2 )M (N M )=2 (24) N M ((D M )=2): M (D=2) we obtain the statement of the Lemma. Now consider the distribution on S 2E (depending on C as a parameter) E EF (A; C ) = (AF C ) D (A) = D F (C ) D (D=2) (25) (N M )D=2 M (D=2) (AF C )(det C ) (D M 1)=2(det A)(D N 1)=2; for ReD > N 1. 2 EF For f 2 S (S+E ), de ne a function ID (f ) on C > 0, by setting (26)
EF ID (f )(C ) = N X 0
Hom(F ? ;F )
12 1 12

A2S E :A11 =C
2 +

(det A)(D N

1)=2 e trA dA:


Since (det C ) 1 to C 0, i.e. de nes a continuous, rapidly decaying function on S+F . Remark. Although the power of (det C ) 1 in (25) is more than that of det A, for small det C the intersection of the region of integration in (26) with any xed ball has area (det C )(N M )=2 (because of presence of the delta-function), which helps compensate the negative exponent of det C . Using Lemma 6, it is easy to compute the Fourier transform ^EF (B; C ) of D EF (A; C ) with respect to A. D

S E EF (det A) 1 (trA)N M , the function ID (f ) extends continuously 2
2 +

EF (A; C )f (A)dA: D

Lemma 7.
F ^EF (B; C ) = Jlim (N M )D=2e itr(CB ) det(iBF ? + J ) D=2: D !+0

Proof. Analogous to the proof of Lemma 2. Let 2 be the operator acting on S (S 2F ?). Corollary 8. If ReD > N + 1, then


2 EF (A; C ) = N EF 2 (A; C ): D D

(here 2 acts on A). Proof. The proof is immediate after passing to Fourier transforms. 2 EF Corollary 9. For any f 2 S (S+E ), ID (f ) is a Schwarz function of C , and EF for any C 0 the function ID (f )(C ), regarded as a function of D, analytically continues to an entire function. Proof. Similar to the Proof of Corollary 4. Proposition 10. Let D 0 be a nonnegative integer, and V be a vector space of dimension D with a positive de nite symmetric bilinear form . Then for any 2 f 2 S (S+E ) and y : F ! V one has


EF ID (f )(y ( )) =


x2Hom(E;V ):xjF =y

f (x ( ))dx:

Proof. As in Proposition 5, it is enough to consider f (A) = e tr(AB). In this case, on the r.h.s. of (19) is a Gaussian integral, whose value is
(N M )D=2e tr(y ( )B F ) (det B ? ) D=2 : F

Thus, the Proposition follows from Lemma 6. EF De nition. ID is called the D-dimensional integral with external parameters. As before, the D-dimensional integral can be computed using Fourier transform. Let f 2 S (S 2E ). By Plancherel formula, (30)
EF ID (f )(C ) = (2 ) N (N +1)=2


S2 E

f^(B)^EF (B; C )dB; D

where f^ is the Fourier transform of f . Here ^EF (B; C ) = Jlim (N M )D=2eitr(CBF ) det( iBF ? + J ) D=2 : D !+0

For ReD < 2, this formula can be written as EF ID (f )(C ) = (31) Z N (N +1)=2 ((N M )D N (N +1))=2 f^(B)e iD (BF ? )eitr(CBF )j det(BF ? )j D=2 dB: 2
EF If ReD 2, the distribution ^EF (B; C ) is singular, so in order to de ne ID (f ) by D a direct integration formula, we need to use the operator . Namely, from Lemma 6 it follows that (det C ) 1 ^EF (B; C ) = ( 1)N M pNM (D) EF (B); D D+2 N Y D+l N (32) pNM (D) = M N 2 l=M +1 EF So if ReD < 2 + 2K , we can de ne ID by the formula EF ID (f )(C ) = 2 N (N +1)=2 ((N M )D N (N +1))=2(det C ) K K Y S2 E


EF 2 2 Thus we have de ned an operator ID : S (S+E ) ! S (S+F ). It is clear that FG EF EG E E ID ID = ID if E F G, and ID0 = ID . Now we can de ne the D-dimensional integral of functions with polynomial growth. Let f be a smooth function on S 2E such that there exists a 2 R for which f and all its derivatives satisfy the condition (A) = O(jjAjja ), jjAjj ! 1 EF (for any norm on S 2E ). In this case, the integral ID (f ) is de ned for ReD << 0. Indeed, the condition on f implies that the Fourier transform f^ is a distribution dened on functions with su ciently many continuous derivatives having polynomial EF growth at in nity. Thus, we can use formula (30) to de ne ID (f ) for ReD << 0. EF Remark. The operator ID has rather strange properties. For example, if E (P (A)) = 0. Indeed, the Fourier transform of P is P is any polynomial, then ID a distribution supported at 0, so for su ciently negative D formula (14) gives 0. E Thus, ID is not continuous in any topology in which polynomials are dense. Example. Let N = 1, M = 0, and f be a rational function, regular at x 0. P Write f as a sum of simple fractions: f (x) = f0 (x) + i i (x + ai ) ni , where f0 is a polynomial, ai 2 C n R , and ni are positive integers. Then X (ni D=2) ni D E ID (f ) = D=2 i (ni ) ai : i Thus we see that if f is satis es the above condition of polynomial growth then EF (f )(C ) is holomorphic in D for ReD << 0. In general, we cannot expect that ID this function will analytically continue to the whole complex plane. However, for functions of a certain algebraic nature, it extends meromorphically to the whole complex plane.


K f (B )e ^

itr(CB F






J =1 (B ) j det(B )j( D+2K )=2 dB:

pN (D 2J )


3. D-dimensional integral of functions arising from Feynman diagrams.

2 De nition. A function f on S+E is called a function of Feynman type if it has

the form

f ( A) = Ql

where P is a polynomial, Bj and mj 6= 0 are real numbers. EF Proposition 11. If f is of Feynman type then ID (f ) extends meromorphically to the whole complex plane. Proof. We will prove the statement for P = 1. For general P , the proof is similar to the case P = 1. R EF Using the formula b 1 = 01 e tb dt, we can rewrite ID (f ) in the form

P ( A) 2 ; j =1 (tr(ABj ) + mj ) P 0, such that ( tiBi )F ? is nondegenerate for ti > 0,


EF ID ( f ) =


Set B(t) = tj Bj . Using Lemma 6, we can rewrite this in the form (34) Z P EF (f )(C ) = (N M )=2 e tj mj tr(CBF (t))(det BF ? (t)) D=2 dt: ID


t1 ;:::;tl2 0;1)


P tj mj


EF ID (e

tr(A(P tj Bj )) )dt:

t1 ;:::;tl2 0;1)

Now we will use the following theorem of Bernstein, which is proved with the help of the theory of D-modules. Bernstein's theorem. Let Q(t) be any polynomial in l variables. Then there exists a polynomial di erential operator L(D) in l variables, with coe cients depending polynomially on D, and a polynomial q in D, such that L(D)Q D=2 = q(D)Q 1 D=2 . This theorem implies the following statement, which is a special case of a result of Bernstein. Corollary. Assume that Q takes positive values for tj > 0, and let g be a rapidly decaying smooth function de ned for ti 0, such that its derivatives are also rapidly decaying. Then the integral (35)

I (D; g) =


ti 2 0;1)


D=2 dt;

convergent for ReD << 0, extends meromorphically to the whole complex plane. The corollary is proved by induction in l (the number of tj ). The base of induction is trivial. Now, using Bernstein's theorem and integration by parts, we can write

I (D + 2; g) = q(D)




D=2 )dt =

(36) Z 1 (L(D) g (t))Q(t) D=2 dt + C (D) = q (D) 1 I (D; L(D) g ) + C (D); q (D )

where C (D) is the integral over the boundary, which is meromorphic by the assumption of induction. Now we apply the corollary to our situation, and obtain the proof of the Proposition for P = 1. Remark 1. Using Hironaka's theorem about the existence of a smooth resolution of singularities for a compact projective variety, it is possible to prove that EF ID is meromorphic in D in a much more general situation than in Proposition 11. For example, as was explained to me by M.Kontsevich, one can prove that if f is a 2 EF rational function without singularities in S+E , then ID (f ) is meromorphic. EF It is also possible to show that for rational f as above the poles of ID (f ) are all rational, and belong to a nite number of arithmetic progressions. Proposition 11 allows to de ne the procedure of dimensional regularization of Feynman integrals in quantum eld theory. Suppose that we have a quantum eld theory given by some Lagrangian, and we want to compute its correlation functions. According to Feynman rules, such a correlation function equals to the sum of amplitudes of all Feynman diagrams (graphs), which are given by some nite-dimensional integrals. We will work in the Euclidean setting and in momentum space. For simplicity we will assume that our theory contains only scalar massive elds. Then the amplitude of any Feynman graph is given by the formula (37)

4. Dimensional regularization of Feynman integrals.


x2Hom(E;V ):xjF =y

g(x)dx; y 2 Hom(F; V );

where V is the momentum space (a Euclidean space of dimension d with inner 1 product ), E = Hc ( n @ ; R) (the cohomology with compact supports), F is 1 the image of the map H 0 (@ ; R) ! Hc ( n @ ; R) in the long exact sequence of cohomology, and g is a function on Hom(E; V ) of the form g(z) = f (z ( )), where f is a certain function of Feynman type attached to . Remark. It is clear from the long exact sequence that E=F = H 1 ( ; R). Since cohomology groups are de ned over Z, they have natural volume elements. This de nes the volume element dx in (37). The problem is that integral (37) is often divergent, since the function g does not decay rapidly enough. Dimensional regularization is a method to give meaning to the integral (37). This EF is done as follows. Recall that since f is of Feynman type, ID (f ) is meromorphic in D. EF De nition. Dimensional regularization of integral (37) is the regular part of ID (f ) H EF at D = d, i.e. 21 i jD dj=" ID (f )d ln(D d). Remark. In general, we have to consider functions f which (even before the regularization) depend meromorphically on D. This is caused by the fact that our graph may contain divergent proper subgraphs, which have to have been regularized before we start working with . Since regularization is performed by adding D-dependent counterterms, we will have to work with functions which depend meromorphically on D.

Let F E be vector spaces with volume elements, dim E = N , dim F = M , and D be any complex number. De ne the space top of (Schwarz) top degree EF top 2 forms to be the space S (S+E ), and the space EF 1 of forms of degree top 1 2 to be the space of Schwarz functions ! : S+E ! Hom(E=F; E ). That is, top = EF 2 Hom(E=F; E ) S (S+E ). If D is an integer, and V is a Euclidean space of dimension D with inner product , then we have a linear map : Hom(E; V ) ! Hom(F; V ) given by restriction, and any f 2 top indeed de nes a top di erential form f (x ( ))dx on each ber of EF (dx is the natural volume form on the ber). top Also, any ! 2 EF 1 de nes a top 1-form on the ber. Indeed, let z : (E=F ) ! V be a linear map. We regard z as a constant 1-form on the bers of . Then the top 1-form corresponding to is de ned by the formula (! ^ z)(x) = tr(! (x; z))dx, where (x; z) 2 Hom(E (E=F ) ; R) is regarded as an operator E ! E=F . 2 Recall that the group GL(E ) acts on S+E by A ! S AS , S 2 GL(E ). Thus, 2 any element X of the Lie algebra End(E ) of GL(E ) de nes a vector eld on S+E : d LX f (A) = dt jt=0f (eX t AeXt). P 2 For any ! = Xi fi 2 top 1 , Xi 2 Hom(E=F; E ), fi 2 S (S+E ), de ne the EF form d! 2 top by the formula EF (38)

5. D-dimensional Stokes formula.

d! =


(LXi fi + Dtr(Xi )fi );

where Xi are regarded as elements of End(E ). Thus, we have de ned a linear map top d : EF 1 ! top . It is easy to see that if D is an integer, d coincides with the EF usual di erential. top EF Proposition 12. (Stokes formula.) ID (d!) = 0 for any ! 2 EF 1. Proof. It is enough to prove the formula for ReD >> 0; for other values of D the formula follows by analytic continuation. For ReD >> 0, the formula follows by integration by parts from the de nition EF (26) of ID , as LX ((det A)(D N 1)=2dA) = Dtr(X )(det A)(D N 1)=2dA.