1
Summary of lessons on MC techniques
Sara Bolognesi 31 May 2006
Matrix Element Monte Carlo
• amplitude calculation
• phase space integration
• approximations
Parton shower
Hadronization (or Fragmentation)
1
Parton level simulation
Increasing in collider energy
Increasing number of final state particles
Increasing the complexity of the parton level simulation:
More Feynman diagrams and with a complex topology :
• you need to correctly manage the phase space
• more efficient integration techniques have to be developed
• increasing of the difficulty in the amplitude calculation
Sara Bolognesi 31 May 20062
Amplitude calculation
1
2diagr diagr
terms
N NN
2 2 2 †
1 2 1 2 1 22ReA A A A A A
terms diagrN N
Standard method: TRACE EVALUATION
the number of terms grows with a power 2 of the number of diagrams
EX.: with 2 diagrams -> 3 terms
New technique: HELICITY AMPLITUDE METHODS
you loose spin info by summing on polarization
working with amplitude (not squared!)
you keep the spin info:
smaller cancellations between different terms, the cancellations become more easy to manage
so in reality terms diagrN N spin configurationsyou need to use approximation (ex. m~0)
Sara Bolognesi 31 May 2006
3
Spinor formalism
A u p Xu p Re-write the amplitude as a product of terms
each of which represents a “piece” of diagram
you can construct the diagrams for
Very simple to computerize the process in an automatic program
Sara Bolognesi 31 May 2006
EXAMPLE: with these pieces
e e ude v
4
Phase space integration For each intermediate particle you have a propagator so you have a resonance in the amplitude
The dimension of the phase space (i.e. the number of variables on which you have to integrate) grows with 3N-4(N = number of final particles)
You want to have a stable and reliable result and you need it quickly!!
•adaptive iterative algorithm•multi-channel technique
The main strategy is
Sara Bolognesi 31 May 2006
choose as integration variables those which have the peaks (i.e. align the peaks with the axis)
= PHASE SPACE SPLITTING
make a change of variable to flat the bumps in the amplitude
= IMPORTANCE / STRATIFIED SAMPLING
5
Sara Bolognesi 31 May 2006
Phase space splitting
21, 1
1
, 12
nj j
j
dMdps n dps j j
Re-write the phase space of N outgoing particles as the product of phase space of N/2 couples of particles
where M(j,j+1) are the variables where you expect the bumps in the amplitude
EXAMPLE:
so you have to choose carefully the splitting accordingly to physics process(eventually you need to use different splitting for the same amplitude -> multi-channel techniques)
ud bbe v
6
Numerical integrationWe can’t computerize an analytical integration but we can apply a numerical integration(better we know our function -> more efficient and precise could be the integration)
MC integration is the best technique:
1E
N E = estimate of the result
N = number of points used to integrate
the error on the estimate converges to the real result not very quickly but it’s independent from the integral dimensions (i.e. the number of variables)
Techniques to improve the convergence are more than welcome!!
importance sampling stratified sampling
Sara Bolognesi 31 May 20067
Importance sampling
f x f x
I f x dx P x dx dPP x P x
1f x
P x
put more points where the function peaks more
change of variable to pass from x (distributed according with f) to P (rather uniformly distributed)
1P x dx
• choose a probability density function P(x) with those properties
• generate uniform random number z
PRATICALLY:
MATHEMATICS:
THE MAIN IDEA:
• change the variable:
• sample f with x
1x P z
0P x with
Sara Bolognesi 31 May 20068
Sara Bolognesi 31 May 2006
Stratified sampling
Put more points where the function varies more quickly
Split your integration interval in N different sub-intervals (thanks to the linearity):
• put the same amount of points in each intervals• use narrower intervals where the function varies more quickly
The error on the total integral is lower if the error on different sub-intervals is of the same order
THE MAIN IDEA:
PRATICALLY:
MATHEMATICS:
9
Adaptive-iterative algorithmBecause you don’t know the shape of the function, you need to learn about that during the integration itself:
split the integration domain in sub-intervals where you put the same amount of points
at the first iteration points equally distributed on all the domain
see where the first integration is less precise
at the second iteration put more points here
and so on…
… but we need to have already the peaks aligned with the axis (thanks to the phase splitting)
Sara Bolognesi 31 May 200610
Multichannel techniques…but typically you have many peaks in different direction (one peak on the mass of each intermediate resonance)
For each channel use a different mapping i.e. a different set of variables to map your phase space
M(bev) ~ Mtop
M(bb) ~ MZ
define a probability (i) of pick up that channel(the best i are the ones which give roughly the same error for the different channels)
Sara Bolognesi 31 May 2006
M(ev) ~ MW
M(ev) ~ MW
EXAMPLE:
ud bbe v
11
Approximations in M.E. MC
2
1M
q
22 21 3 1 3 1 3 132 2 2 coseq p p m E E p p
0em
MASSLESS APPROXIMATION
+ to reduce the number of helicity configuration that you have to consider
- could raise some fictitious singularity -> instability in the MC
Anyone MC can’t cover all the phase space
you have to apply clever cuts on the phase space in order to avoid singularities and instability
NOTE: also without massless approximation the mass of an electron can be a trouble!
Sara Bolognesi 31 May 2006
2 0q 1,3 0E 13 0
se e
22 4~ 10em GeV
12
In general you can’t consider only a subset of diagrams of your process because • you can loose big interference terms or cancellation effects
• you break the Gauge Invariancebut you can save the G.I. by forcing your intermediate particles to be on shell = Narrow Width Approximation
qq->e+e-e+e- ~ qq->ZZ->e+e-e+e-
PRODUCTION TIMES DECAY
Mathematically this corresponds to • neglect spin correlation between production and decay of the intermediate particles• approximate the resonance with a Dirac delta
- you can’t esteem the acceptance of cuts on the invariant mass- you loose spin info (that affects the angular distributions) and all the irreducible backgrounds+ you can use this approx. to introduce NLO corrections
NARROW WIDTH APPROXIMATION AND PRODUCTION TIMES DECAY
Sara Bolognesi 31 May 2006
~
13
Sara Bolognesi 31 May 2006
The extension of the NWA from s to t channel (i.e. virtual intermediate particle)
A sort of PDF = probability to find a photon in an electron
scattering of a quasi real photon with a positron
q2 ~ 0
EQUIVALENT VECTOR BOSON APPROXIMATION
WEIZSECKER-WILLIAM APPROXIMATION
The extension of the WWA from the photon to massive vector boson
A sort of PDF =
scattering of real vector boson
q2 ~ MV
A little more tricky because MV doesn’t belong to the allowed values of q2 (q2<0)
probability to find a V in a fermion
14
Parton shower
Sara Bolognesi 31 May 2006
HARD SCATTERING PARTONIC PROCESS
HADRONIZATION / FRAGMENTATION
high energy (>~100 GeV) “low” energy (~100 MeV)
PARTON SHOWER = radiation from initial and final state (ISR/FSR)
• bulk of the corrections to the LO process
• scale evolution2ˆ hadronss Q M
Approximated with only 1->2 split 1->3 processes
virtual correctionsi.e. only O(s) considered at each split:
15
Sara Bolognesi 31 May 2006
Bulk of corrections From a perturbative point of view, we have divergences due to
• virtual corrections• soft emission (IR divergences)• collinear emission
equal and opposite, they cancel one with the other
2lnn n s
m
ordinate emission from big angles at larger scale to small angles at lower scale:
Multiple emission:
Physical interpretation:
emitted and intermediate particles as partons (constituent) of the initial particleDGLAP equation = probability of splitting, a sort of PDF
MC algorithm
generate a random number 0<r<1 if r < Prob. of non splitting stop here
else r=Pnosplit(Q2,K2) extract K and simulate splitting (with 4-mom. conservation)
16
Sara Bolognesi 31 May 2006
HadronizationReconstruct the final stable hadrons from partons radiated during the Parton Shower
Models to describe the reality (i.e. to fit the data) not a coherent theory (no Lagrangian): PS + hadronization
String Fragmentation (Lund model)
color exchange (i.e strong force) between quarks represented with a string
potential proportional to the quarks distance (string length)
when the energy is high, the string crack and give rise to a new couple of quark
MC contains an iterative algorithm to reproduce this process with a probabilistic approach
17
hadron
THE ENDTHE END