Renormalization Group by ConditionalExpectations (Harmonic Extensions)
Hao Shen (Princeton University)
Harmonic analysis and Renormalization group Workshop
April 21, 2014
Gaussian measure
Measures on space of functions (distributions) {� : Rd ! R}
exp✓�1
2
ˆRd
(r�(x))2◆
D�.N
Gaussian measure + interactions
Measures on space of functions (distributions) {� : Rd ! R}
exp
�1
2
ˆRd
(r�(x))2 + �"
small
ˆRd
V (r�(x))!
D�.N
e.g.V (r�(x)) =
X
µ
cos(rµ�(x)) V (r�(x)) = �X
µ
(rµ�(x))4
Finite dimensional approximation: let ⇤ ⇢ Zd be a finitelattice,
exp
�1
2
X
x2⇤
(r�(x))2 + �X
x2⇤
V (r�(x))!Y
x2⇤
d�(x).N
The idea of renormalization group (RG)
“Coarse-graining”: let B be a block,
�(B) =1|B |X
x2B
�(x)
Integrating out the small scale details⇣(x) = �(x)� �(B)
F (�)D�! F 0(�)D�Iterate the above operations
F ! F 0 ! F 00 ! · · ·This is a transformation on the space of measures (models).
The study of measures =) Dynamical system
Heuristic statement of the result
At large scale,
exp✓�1
2
X(r�)2 + �
XV (r�)
◆
⇡ exp✓�1 + O(�)
2
X(r�)2
◆
Precise statement of the result
Recall the model on ⇤ = [�LN/2, LN/2]d \ Zd
exp
(�1
2
X
x2⇤
(r�(x))2 + �X
x2⇤
V (r�(x)))Y
x2⇤
d�(x)�
N
Squeeze ⇤ into a fixed (continuum) torus
⇤ ,! Td
Fix a mean zero smooth function f on Td .
Precise statement of the result
Moment generating functional (Laplace transform)
Z (f ) =⌦eP
x
f (x)�(x)↵
=
´e� 1
2P
⇤(r�)2+�P
⇤ V (r�)+P
x2⇤ f (x)�(x)Qx2⇤ d�(x)´
e� 12P
⇤(r�)2+�P
⇤ V (r�)Q
x2⇤ d�(x)
Here,f ( x
"point in ⇤
) := L�(d+2)N/2 f ( L�Nx"
point in Td
)
Precise statement of the result: scaling limit
Theorem
Assume that V (r�) = cos(r�), and f is sufficiently smooth,and |�| is sufficiently small. Then, there exists a constant �depending on � and
limN%1
Z (f ) = exp✓
12
ˆTd
f (x)(���)�1f (x)ddx◆
RG: If F (�) = e� 12 (r�)2+�V (r�) then
F ! F 0 ! F 00 ! · · · ! F ⇤
where F ⇤ is Gaussian.
Rigorous RG methods
The above idea of coarse-graining by “block averaging”was rigorously carried out Balaban, Gawedzki, Kupiainen (1980’s).Decomposition of Gaussian Abdessalem, Bauerschmidt, Brydges,
Dimock, Falco, Guadagni, Hurd, Slade, Yau etc. (1990’s): Let µC be theGaussian measure with covariance C = (��)�1,ˆ
F (�)dµC (�)
Decompose C into positive definite pieces C =P
j Cj s.t.
� =X
j
�j
Do integrations step by stepˆ· · ·ˆ
F (�) dµC1(�1)dµC2(�2) · · · dµCj
(�j) · · ·
The new method (conditional expectations)Write
´F (�)dµC (�) = E [F (�)]. Let ⌦ ⇢ ⇤,
EhF (�)
i= E
hE⇥F (�)
���(⌦c)⇤ i
where the conditional expectation is conditioned on all{�(x)|x /2 ⌦} being fixed.(Integrate out the field in ⌦, keeping the field out of ⌦ fixed.)Thinking of the Gaussian measure as e
12 (�,��),
�X
⇤
��� = Q(�(⌦)) + Q(�(⌦c)) + crossing term
= ” x2 + y 2 + xy ”
With ”y” fixed, shift ”x” to the minimizer of the quadraticform to cancel the crossing term.
The new method
The minimizer is P⌦� =P
y2@⌦ P⌦(·, y)�(y) i.e. theharmonic extension of � from ⌦c into ⌦,
� = P⌦�+ ⇣
where P⌦ is the Poisson kernel for domain ⌦. Then,
�X
x2⇤
�(x)��(x) = �X
x2X
⇣(x)�D⌦⇣(x)�
X
x2⇤
P⌦�(x)�P⌦�(x)
where �D⌦ is the Laplacian on ⌦ with Dirichlet b.c.
The conditional expectation = integrating out ⇣:
E⇥F (�)
���(⌦c)⇤= E⇣ [F (P⌦�+ ⇣)]
Covariance of ⇣ is the Dirichlet Green’s function of �� on ⌦.
The new method
Note that the functional
E⇣ [F (P⌦�+ ⇣)]
depends on � via P⌦�. Namely, the functional, after theintegration, depends on the field in a smoother way(smoothing effect of the Poisson kernel). Intuitively,
” P⌦ ⇡ the block averaging operator ”
The new method
The new method has the following advantages:the fluctuation field ⇣ automatically has finite rangesupport (In the block-average method, the fluctuation field
only has exponentially decaying correlation; in Brydges’
method, constructing the decomposition with finite range
properties is a nontrivial task.)
RG =) local elliptic PDE problems. (In the
non-translation-invariant models, the difficulty would be
merely estimating solutions to elliptic PDEs with non-constant
coefficients).
As we will see, the norm on the functionals will besimplified.
The a priori tuning of the GaussianBased on the above heuristics, we split the quadratic part:
�12(r�)2 =� �
2
X
x2⇤
(r�(x))2 + � � 12
X
x2⇤
(r�(x))2
�!�/p�
�����!� 12
X
x2⇤
(r�(x))2 + �
2
X
x2⇤
(r�(x))2
where � ⇡ 1 (will be determined in the end); � := 1 � ��1.
exp⇢�
2
X
x2⇤
(r�(x))2 + �X
x2⇤
W (r�(x))�
dµ(�)
dµ(�) = e� 12P
x2⇤(r�(x))2Y
x2⇤
d�(x)�N
An expansion over polymersTake f = 0 for simplicity, to focus on the main ideas of ourRG method. Let E be the Gaussian expectation.
Z = EY
x2⇤
e�2 (r�(x))2
✓�e�W (r�(x)) � 1| {z }
”O(�)”
�+ 1◆�
= EX
X✓⇤
I (⇤\X )K (X )
�
where for any sets X ,Y ,
I (Y ) =Y
x2Y
e�2 (r�(x))2 K (X ) =
Y
x2X
e�2 (r�(x))2
✓e�W (r�(x))�1
◆
Overview of the rest of the talk
EX
X✓⇤
I (⇤\X ,�)K (X ,�)
�I (Y ,�) =
Y
x2Y
e�2 (r�(x))2
|K (X ,�)| A�|X |eP
X
(r�)2 A � 1
Two questions:How to keep this algebraic structure at all the scales(sites ! blocks, � ! smoother fields)?How to maintain similar analytical bounds for all thescales?
Propagate algebraic structure to next scale
At scale j , one has Z =P
X2Pj
Ij(⇤\X ,�)Kj(X ,�)
Z = EX
X2Pj
Ij lives on the grey blocks, and Kj lives on the black regions.
Propagate algebraic structure to next scale
We would like to find a way to rewrite the expression onscale j + 1:
Z = E X
U2Pj+1
Ij+1(⇤\U)Kj+1(U)
�
while Ij+1,Kj+1 still have the above forms andfactorization properties.The idea is to postulate that Ij+1 will be exponential of anew quadratic form and then find Kj+1.
Propagate algebraic structure to next scale
E X
X2Pj
Ij(⇤\X )Kj(X )
�= E
X
X2Pj
Y
B2⇤\X
⇣Ij+1(B)+�Ij(B)
⌘Kj(X )
�
= E X
X2Pj
X
Y✓⇤\X
Ij+1(⇤\(X [ Y )) �Ij(Y ) Kj(X )
�
Consider the smallest j+1 scale polymer that contains X [ Y ,
written as U := X [ Y 2 Pj+1, and then resum:
E X
U2Pj+1
Ij+1(⇤\U)
K#j
(U)z }| {X
XqY=U
�Ij(Y )Kj(X )
�
Conditional expectation
E X
U2Pj+1
Ij+1(⇤\U)Y
V2c.c(U)
K#j (V )
�
= E X
U2Pj+1
Ij+1(⇤\U)
⇥Y
V2c.c(U)
E⇥K#
j (V )���(V+)c
⇤�
= E X
U2Pj+1
Ij+1(⇤\U)Kj+1(U)
�
Conditional expectation
Recall that “the conditional expectation = integration out ⇣”:
E⇥K (�)
���V c
⇤= E⇣ [K (PV�+ ⇣)]
where covariance of ⇣ is the Dirichlet Green’s function for V .After the integration, the resulting functional depends on � viaPV� which is expected to be “smoother” than �; however,
The Poisson kernel PU(x , y) for x 2 U, y 2 @U only hassmoothing effect for x sufficiently far from @U!If one directly took conditional expectation fixing the fieldoutside the “red line”, Ij+1 would also get involved in theintegration...
Conditional expectation (making a corridor)The idea to settle the above dilema is still “expand and resum”.
X
U2Pj+1
Y
B2Bj+1(⇤\U)
✓(Ij+1(B)� 1) + 1
◆K 0
j+1(U)
=X
U2Pj+1
X
D✓⇤\U
Y
B2Bj+1(D)
(Ij+1(B)� 1)K 0j+1(U)
Then we “glue” the Ij+1(B)� 1 which touch U onto K 0j+1
until the set U [ ([B) is surrounded by a “corridor” filledby only “1” ’s.After that, we resum up all the rest of Ij+1(B)� 1 out ofthe corridor, and we obtain
Z = E X
U2Pj+1
Ij+1(⇤\U)Kj+1(U)
�
where U is a set slightly larger than U.
Propagate analytical bound to next scale
We want:Kj(X ) . A�|X |
j 8X 2 Pj
) Kj+1(U) . A�|U|j+1 8U 2 Pj+1
However,
Kj+1(U) ⇡ EhX
B=U
�Ij(B) +X
X=U
Kj(X )��� · · ·
i
So many finer scale polymer X ,Y ’s are contained in afixed coarse scale polymer U!In fact, only O(Ld) terms...
The important scalingHow to deal with this O(Ld) growth of K? Let’s look at thecovariance of rPX�:Lemma: Let x 2 X , and C be the covariance of the Gaussianfield �. If d(x , @X ) � 1
3Lj , then
X
y1,y22@X
(rxPX )(x , y1)C (y1, y2)(rxPX )(x , y2) O(1)L�dj
This means that rPX� “scales down by O(L� d
2 )”.
” K = c0 + c2(rPX�)2 + O((rPX�)
4) ”
As long as we choose
Ij+1 = eEj+1+�
j+1(rPU
�)2
in the way that the c0 + c2(rPX�)2 part is absorbed into Ij+1,then O((rPX�)4) will scale down by O(L�2d).
The important scaling
Lemma: Let x 2 X , and C be the covariance of the Gaussianfield �. If d(x , @X ) � 1
3Lj , then
X
y1,y22@X
(rxPX )(x , y1)C (y1, y2)(rxPX )(x , y2) O(1)L�dj
Idea of proof:
LHS =X
y22@X
rxC (x , y2)| {z }.|x�y2|1�d=O(L�(d�1)j )
L�j
z}|{rx PX (x , y2)| {z }
O(L�(d�1)j )
Propagate analytical bound to next scale
Again, we want:
Kj(X ,�) . eP
X
(r�)2 8�
) Kj+1(U,�0) . eP
X
(r�0)2 8�0
In the field decomposition method,
E⇣eP
X
(r�)2 = E⇣eP
X
(r�0+r⇣)2 E⇣e2
PX
⇣(r�0)2+(r⇣)2
⌘
Bounds deteriorate (lose control) ... One has to put inmany extra terms.
The norm for K
The weight in “the method of decomposition of Gaussian field”was roughly
G (X ) ⇠ eP
X
(r�)2
but complete form is complicated
G (X ) = exp✓
c1X
X
(r�j+1)2 + c2
X
B2Bj
(X )
supB?
��L2jr2�j+1��2
+ c3X
@X
(r�j+1)2 + c4 max
0p2supB?
��Lpjrp⇣��2◆
where �j+1 = �j + ⇣.
Propagate analytical bound to next scale
We bound Kj(X ,�) . G (X ,X+) where
G (X ,Y ) := Ehe
2P
X
(r�)2���Y c
i �N(X ,Y )
for X ⇢ Y where
N(X ,Y ) := Ehe
2P
X
(r�)2���Y c = 0
i
An interesting property
exp
2
X
X
(@ 2)2
! G (X ,Y ) exp
2
X
X
(@ 1)2
!
where1 1 minimizes
PY (@�)2 �
PX (@�)
2with �Y c
fixed,
2and 2 minimizes
PY (@�)2 with �Y c
fixed.
Propagate analytical bound to next scale
The conditional expectation automatically takes care of theintegration of G :
E⇥G (X ,Y )
���Uc
⇤
=EhEhe
2P
X
(r�)2���Y c
i ���Uc
i �N(X ,Y )
=Ehe
2P
X
(r�)2���Uc
i �N(X ,Y )
=G (X ,U)N(X ,U)
N(X ,Y )| {z }cL
�dj |X |
Linearization of RG map
The map (�j , �j+1,Kj) ! Kj+1 is smooth (w.r.t. the norms wedefined).The linearized part of Kj+1(U) at (0, 0, 0) contains three parts:
The contributions from Kj(X ) where X ✓ U is “large” -which can be shown negaligible.For small X , the Taylor remainder after the second orderTaylor expansion of E
⇥Kj(X )
���(U+)c⇤
is negaligible.One can choose �j+1 so that the leading Taylor terms ofE⇥Kj(X )
���(U+)c⇤
are absorbed into Ij+1.
Determine �
The dynamical system
�j+1 = �j + ↵(Kj)
Kj+1 = LKj + f (�j ,Kj)(1)
satisfies kLk < 1. By standard stable manifold theorem, thereexists an initial tuning � so that |�j | ! 0, kKjkj ! 0.In other words,
Ehe
�2P
x2⇤(r�(x))2+zW (r�)i! ”econstE [1] ” (j ! 1)
Thank you!