This is a very good question which, however, shows the extent of the reigning confusion about renormalization even four decades after Wilson's Nobel Prize winning theory on the matter. I essentially answered the OP's question, and much more, about constructing continuum QFTs in Wilson's framework in my expository article "QFT, RG, and all that, for mathematicians, in eleven pages" but in a very condensed fashion (one needs to do computations on the side to follow what is being said). So let me give more details pertaining to the OP's specific question. I should preface this by saying that what follows is a "cartoon" for renormalization. I will oversimplify things by ignoring anomalous dimensions, marginal operators, and nonlocal terms generated by the RG. You will not find technical details but hopefully the conceptual picture and logical structure of renormalization will become clearer.
The OP is right to point out that in the setting of ODEs and dynamical systems a first order equation can be run backwards in time.
So let me start by recalling some important terminology from that area.
Consider a first order nonautonomous ODE of the form
dXdt=f(t,X) .
It generates a flow (groupoid morphism from time pairs to diffeomorphisms of phase space) I will denote by
U[t2,t1] which sends the initial value
X(t1) to the value of the solution at time
t2. It trivially satisfies
∀t,U[t,t]=Id and the semigroup property
∀t1,t2,t3, U[t3,t2]∘U[t2,t1]=U[t3,t1] .
This time-dependent situation is to be distinguished from the
autonomous ODE case
dXdt=f(X)
where
U[t2,t1]=U[t2−t1,0]=:U[t2−t1].
In Wilson's RG, time is scale or more precisely, t=−logΛ where the UV cutoff is implemented in momentum space by a condition like |p|≤Λ or in position space by Δx≥Λ−1=et. The high energy physics literature usually works in a nonautonomous setting while it is essential to translate the equation to autonomous form for a proper understanding of Wilson's RG. The latter imported tools and concepts from dynamical system theory like fixed points, stable and unstable manifolds etc. It is possible to do some contortions to try to make sense of these concepts in the nonautonomous setting, but these truly are notions that are congenial to autonomous dynamical systems.
Let μ=:μ−∞,∞ denote the probability measure corresponding to the free Euclidean theory.
Its propagator is
∫ϕ(x)ϕ(y) dμ−∞,∞(ϕ)=⟨ϕ(x)ϕ(y)⟩−∞,∞=∫dp(2π)Deip(x−y)pD−2Δ
where
Δ is the scaling dimension of the field
ϕ. Normally,
Δ=D−22 but I will allow more general
Δ's
in this discussion.
Now let me introduce a mollifier, i.e., a smooth function of fast decay
ρ(x) such that
∫ρ(x) dx=ρˆ(0)=1.
For any
t, let me set
ρt(x)=e−Dtρ(e−tx), so in particular
ρ0=ρ.
Let
μt,∞ be the law of the field
ρt∗ϕ where
ϕ is sampled according to
μ−∞,∞ and we used a convolution with the rescaled mollifier. In other words,
μt,∞ is the free cutoff measure at
ΛH=e−t and propagator
∫ϕ(x)ϕ(y) dμt,∞(ϕ)=⟨ϕ(x)ϕ(y)⟩t,∞=∫dp(2π)D|ρˆt(p)|2 eip(x−y)pD−2Δ .
Note that
ρˆt(p)=ρˆ(etp) which we assume to have decreasing modulus with respect to
t.
We have
ρˆ−∞=1 and
ρˆ∞=0 and
|ρˆt1(p)|2−|ρˆt2(p)|2≥0 whenever
t1≤t2. One can thus define a more general family of modified free/Gaussian theories
μt1,t2 with
t1≤t2 by the propagator
∫ϕ(x)ϕ(y) dμt1,t2(ϕ)=⟨ϕ(x)ϕ(y)⟩t1,t2=∫dp(2π)D(|ρˆt1(p)|2−|ρˆt2(p)|2) eip(x−y)pD−2Δ .
One has the semigroup property for convolution of (probability) measures
μt1,t2∗μt2,t3=μt1,t3
when
−∞≤t1≤t2≤t3≤∞. This means that for any functional
F(ϕ),
∫F(ϕ) dμt1,t3=∫∫dμt1,t2(ζ) dμt2,t3(ψ) F(ζ+ψ) .
The other key players are scale transformations
St. Their action on fields is given by
(Stϕ)(x)=e−Δtϕ(e−tx)
and obviously satisfies
St1∘St2=St1+t2.
Using the notion of push-forward/direct image of measures, one has
(St)∗μt1,t2=μt1+t,t2+t, i.e.,
∫dμt1,t2(ϕ) F(Stϕ)=∫dμt1+t,t2+t(ϕ) F(ϕ) .
Since these are centered Gaussian measures, it is enough to check the last property on propagators, i.e.,
F(ϕ)=ϕ(x)ϕ(y)
where this follows from a simple change of momentum integration variable from
p to
q=e−tp in the above formula for the propagator.
This also covers the infinite endpoint case with the conventions
t+∞=∞,
t−∞=−∞ for finite
t.
The high energy physics Wilsonian RG is the transformation of functionals RG[t2,t1] for pairs t1≤t2 obtained as follows.
Using the convolution semigroup property
∫e−V(ϕ)dμt1,∞(ϕ)=∫e−V(ζ+ψ)dμt1,t2(ζ) dμt2,∞(ψ)
=∫e−(RG[t2,t1](V))(ϕ)dμt2,∞(ϕ)
after renaming the dummy integration variable
ψ→ϕ and introducing the definition
(RG[t2,t1](V))(ϕ):=−log∫e−V(ζ+ϕ)dμt1,t2(ζ) .
If
V is the functional of
ϕ corresponding to the bare action/potential with UV cutoff
ΛH=e−t1, then
RG[t2,t1](V) is the effective potential at momentum/mass scale
ΛL=e−t2.
Trivially (Fubini plus associativity of convolution of probability measures) one has, for
t1≤t2≤t3,
RG[t3,t2]∘RG[t2,t1]=RG[t3,t1]
which is indicative of a
nonautonomous dynamical system structure, to be remedied shortly.
At this point one can already state the main goal of renormalization/taking continuum limits of QFTs: finding a correct choice of
cutoff-dependent potentials/actions/integrated Lagrangians,
(Vbaret)t∈R
such that
∀t2, limt1→−∞RG[t2,t1](Vbaret1) =: Vefft2 exists.
The OP's intuition is correct in seeing this as a backwards shooting problem: choosing the right initial condition at
ΛH to arrive where we want at
ΛL.
A difficulty here (related to scattering in classical dynamical systems) is this involves an IVP at
t=−∞ instead of finite time.
Note that the continuum QFT, its correlations, etc. should be completely determined by the collection of its effective theories indexed by scales
(Vefft)t∈R. This is most easily seen when considering correlations smeared with test functions with compact support in Fourier space and with a sharp cutoff
ρˆ(p) given by the indicator function of the condition
|p|≤1 (or at least one which satisfies
ρˆ(p)=1 in a neighborhood of zero momentum).
Switching to an autonomous setting involves some twisting by the scaling maps St. Given a potential V (bare or effective) which "lives at" scale t1, one has
∫e−V(ϕ) dμt1,∞(ϕ)=∫e−V(St1ϕ) dμ0,∞(ϕ)=∫e−(S−t1V)(ϕ) dμ0,∞(ϕ)
where we now define the action of rescaling maps on
functionals by
(StV)(ϕ):=V(S−tϕ) .
As maps on functionals, one has the identity
RG[t2,t1]=St1∘RG[t2−t1,0]∘S−t1 .
Wilson's Wilsonian RG is WRG[t]:=S−t∘RG[t,0], for t≥0. It acts on the space of "unit lattice theories" (I put quotes because I am using Fourier rather than lattice cutoffs). Thus the previous identity becomes
RG[t2,t1]=St2∘WRG[t2−t1]∘S−t1 .
The identity can be derived as follows (note the orgy of parentheses due to the increase of abstraction from functions to functionals, then to maps on functionals):
[(RG[t2−t1,0]∘S−t1)(V)](ϕ)=−log∫dμ0,t2−t1(ζ)exp[−(S−t1V)(ϕ+ζ)]
=−log∫dμ0,t2−t1(ζ)exp[−V(St1ϕ+St1ζ)]
=−log∫dμt1,t2(ξ)exp[−V(St1ϕ+ξ)]
where we changed variables to
ξ=St1ζ.
From this one gets
[(St1∘RG[t2−t1,0]∘S−t1)(V)](ϕ)=[(RG[t2,t1,]∘S−t1)(V)](St1ϕ)
and the identity follows from the trivial fact
St1(S−t1ϕ)=ϕ.
Note that (Vt)t∈R is trajectory of RG, i.e.,
∀t1≤t2, Vt2=RG[t2,t1](Vt1)
if and only if
Wt:=S−tVt is a trajectory of
WRG, i.e.,
∀t1≤t2, Wt2=WRG[t2−t1](Wt1) .
The semigroup property for
RG readily implies that for
WRG, namely,
∀t1,t2≥0, WRG[t1+t2]=WRG[t1]∘WRG[t2] .
Now define
Wstartt:=S−t∘Vbaret. Then assuming continuity of all these RG maps one has
Vefft2=limt1→−∞RG[t2,t1](Vbaret1)=St2(Wefft2)
where
Wefft2:=limt1→−∞WRG[t2−t1](Wstartt1) .
The definiteness of the continuum QFT can also be rephrased as the existence of the potentials
Wefft.
A common source of confusion is the failure to see that while
(Wefft)t∈R is (by definition, the semigroup property and continuity) a trajectory of
WRG, the family of bare potentials
(Wbaret)t∈R is not.
The same statement is true, by undoing the "moving frame change of coordinates", when replacing
W's with
V's and
WRG with
RG.
For concreteness, we need coordinates on the space where the RG acts. Assume the bare potential Vbaret
is determined by a collection of coordinates or couplings (gi)i∈I via
Vbaret(ϕ)=∑i∈Igbarei(t) ∫Oi(x) dx
for local operators of the form
Oi(x)=:∂α1ϕ(x)⋯∂αkϕ(x):t .
The Wick/normal ordering is
with respect to the free cutoff measure
μt,∞. More precisely, for every functional
F,
:F(ϕ):t :=exp[−12∫dxdy δδϕ(x) Ct,∞(x,y) δδϕ(y)] F(ϕ)
where we denoted the propagator by
Ct,∞(x,y):=⟨ϕ(x)ϕ(y)⟩t,∞.
Note that changing
−12 to
+12 followed by setting
ϕ=0 is integration with respect to
μt,∞.
For instance
:ϕ(x)2:t=ϕ(x)2−Ct,∞(x,x) and
:ϕ(x)4:t=ϕ(x)4−6Ct,∞(x,x)ϕ(x)2+3Ct,∞(x,x)2.
An easy change of variables
y=e−tx shows that
(S−tVbaret)(ϕ)=∑i∈Igstarti(t)∫:∂α1ϕ(y)⋯∂αkϕ(y):0 dy
where
gstarti(t):=e(D−Δi)t gbarei(t)
and I used the notation
Δi=kΔ+|α1|+⋯+|αk| for the scaling dimension of the local operator
Oi. The switch
gbarei→gstarti corresponds to that from
dimensionfull to
dimensionless coupling constants.
The indexing set splits as
I=Irel∪Imar∪Iirr, respectively corresponding to the three
possibilities for operators:
D−Δi>0 or relevant,
D−Δi=0 or marginal,
D−Δi<0 or irrelevant.
W=0 is a fixed point of the autonomous dynamical system WRG. The behavior near this (trivial/Gaussian/free) fixed point is governed by the linearization or differential at W=0, i.e., the maps DWRG[t] given by
[DWRG[t](W)](ϕ):=∫W(Stϕ+ζ) dμ0,t(ζ)
as follows from the definition
[WRG[t](W)](ϕ)=−log∫e−W(Stϕ+ζ) dμ0,t(ζ)
and the crude approximations
ez≃1+z and
log(1+z)≃z.
If
W has coordinates
(gi)i∈I (with
:∙:0 Wick ordering), then one can show (good not so trivial exercise) that
DWRG[t](W) has coordinates given exactly by
(e(D−Δi)tgi)i∈I,
in the same frame, i.e., with the same
t=0 Wick ordering.
If instead of flows one prefers talking in terms of the vector field
V generating the dynamics, then a trajectory
(Wt)t∈R of
WRG satisfies
dWtdt=V(Wt) with
V:=ddtWRG[t]∣∣t=0
admitting a linear plus nonlinear splitting
V=D+N. The linear part, in coordinates,
is
D(gi)i∈I=((D−Δi)gi)i∈I .
Assume the existence of
WUV:=limt→−∞Wefft, the UV fixed point, and
WIR:=limt→∞Wefft, the infrared fixed point (they have to be fixed points by continuity). The discussion of perturbative renormalizability
always refers to the situation where
WUV=0 corresponding to continuum QFTs
obtained as perturbations of the free CFT
μ−∞,∞.
By definition, the QFT or the trajectory
(Wt)t∈R of its "unit lattice"-rescaled effective theories
lies on the
unstable manifold Wu of the
W=0 fixed point. In what follows I will assume for simplicity there are no
marginal operators so the fixed point is hyperbolic and there are no subtleties due to center manifolds.
The tangent space
TWu is then spanned by functionals
ϕ⟼∫Oi, for
i in
Irel
which is typically
finite.
Note that, in principle, knowing a QFT is the same as knowing a trajectory (Wefft)t∈R and thus the same as knowing just one point of that trajectory say Weff0 (if the t=0 IVP is well-posed forwards and backwards in time, which is another delicate issue as explained in Arnold's answer). The point Weff0 can be made to sweep the unstable manifold which can be identified with the space of continuum QFTs obtained by perturbing the W=0 fixed point. On the other hand our control parameter is the choice of cut-off dependent starting points (Wstartt)t∈R. These belong to the bare surface TWu. This is why when considering say the ϕ4 model only a small finite number of terms are put in the bare Lagrangian, otherwise we would be talking about some other (family of) model(s) like ϕ6, ϕ8, etc.
So after all these explanations, it shoud be clear that renormaliztion in Wilson's framework can be seen as a parametrization of the nonlinear variety Wu by the linear subspace TWu.
If we denote the stable manifold by Ws and its tangent space by TWs then, assuming hyperbolicity of the trivial fixed point, the full space where the RG acts should be TWu⊕TWs. The stable manifold theorem gives a representation of Wu as the graph of a map from TWu into TWs.
The main problem is to find (Wstartt)t∈R so that the limit
Weff0=limt→−∞WRG[−t](Wstartt) exists. The stable manifold theorem is the t=−∞ case
of a mixed boundary problem where on a trajectory one imposes conditions (on coordinates) of the form gstarti(t)=0, i∈Iirr, and geffi(0)=λRi, i∈Irel. Irwin's proof is a nice way to slove this and it works even if the RG is not reversible. This method can be applied for finite negative t, and this should produce a collection (Wt)t<0 (all that is needed in fact) dependent on the renormalized couplings λRi. Let us assume for instance that Irel={1,2} and
Iirr={3,4,…}. Consider the map Pt given by
(λB1,λB2)⟼(gi{WRG[−t](λB1,λB2,0,0,…)})i=1,2
where
gi{W} denotes the
i-th coordinate of
W.
A possible choice of starting points is thus
Wstartt:=(P−1t(λR1,λR2),0,0,…)i∈I .
The above is more like a road map for what needs to be done but it does not quite provide a recipe for doing it. In the perturbative setting, one trades numbers in R for formal power series in R[[ℏ]]. The propagators of the μ measures get multiplied by ℏ and there is now 1ℏ in front of the V's or W's in the exponential. All the couplings gi now also become elements of R[[ℏ]]. The invertibility of Pt in this setting is easy and follows by analogues of the implicit/inverse function theorem for formal power series (e.g. in Bourbaki, Algebra II, Chapters 4-7, Berlin, Springer-Verlag, 1990).
All the work is in showing that for i≥3, the quantities
fi(λR1,λR2):=limt→−∞gi{WRG[−t](P−1t(λR1,λR2),0,0,…)}
converge to finite values. This gives the wanted parametrization
(λR1,λR2)↦(λR1,λR2,f3(λR1,λR2),f4(λR1,λR2),…) of
Wu by
AccidentalFourierTransform
knzhou
kakaz
AccidentalFourierTransform
Abdelmalek Abdesselam