Universality in Multiparameter Fitting: Sloppy Modelsweb.mit.edu/sea06/agenda/talks/Sethna.pdf ·...
Transcript of Universality in Multiparameter Fitting: Sloppy Modelsweb.mit.edu/sea06/agenda/talks/Sethna.pdf ·...
-
Universality in Universality in MultiparameterMultiparameterFitting: Sloppy ModelsFitting: Sloppy Models
James P. Sethna, Josh Waterfall, Ryan Gutenkunst, Fergal Casey, Kevin S. Brown, Chris Myers, Veit Elser, Piet Brouwer
Cell Dynamics
Fitting Exponentials, Polynomials
Fits good: measured bad
-
FitEnsemble:InterpolationEnsemble: Extrapolation
Fitting Decaying ExponentialsFitting Decaying Exponentials
ttt eAeAeAty 321 321),,(γγγ −−− ++=γA
Classic Ill-Posed Inverse Problem
Given Geiger counter measurements from a
radioactive pile, can we recover the identity of the elements and/or
predict future radioactivity? Good fits with bad decay rates!
∑=
−=
DN
i i
iyyC1
2
2))(()(σθθ
P, S, I3532 125
6 Parameter Fit
-
Fitting ExponentialsFitting Exponentials
Sloppy direction
Stiff
dire
ctio
n
χ2 contours
eigen
param
s
Best Fit
Hessian ∂2C/∂θ2 at Best FitSloppy Directions ⇔Small Eigenvalues
Horizontal scale shrunk by 1000 times!
Aspect ratio = Human hairFits Good
Measuring Parameters Bad
bare params
-
Is Sloppiness Universal?Is Sloppiness Universal?
Syst
ems
Bio
QM
C
Rad
ioac
tivity
Man
y ex
ps
GO
E
Prod
uct
Fitti
ng p
lane
Mon
omia
ls
Eig
enva
lue
NOT SLOPPY
Random Matrix Theory Connection:Explain Eigenvalue
Distributions
Sloppy Systems:
• Enormous range of eigenvalues
• Roughly equal density in log
• Observed in broad range of systems
-
Sloppy Systems BiologySloppy Systems Biology
Fits good: measured
bad
-
Where is Sloppiness From?Where is Sloppiness From?Fitting Polynomials to Data
Fitting Monomials to Datay = ∑an xn
Functional Forms SameHessian Hij = 1/(i+j+1)Hilbert matrix: famous
Orthogonal Polynomialsy = ∑bn Ln(x)
Functional Forms DistinctEigen ParametersHessian Hij = δij
Sloppiness arises when bare parameters skew in eigenbasis Small Determinant!|H| = ∏ λn
-
Why are they Sloppy?Why are they Sloppy?The Vandermonde Ensemble
Assumptions:i. Parameters are nearly degenerate: θι = θ0 + εiii. Residuals symmetric in parameters: mi = ∑j εjiiii.Cost is sum of squares of residuals: C(θ) = ∑rk2({mi})
⎥⎥⎥⎥
⎦
⎤
⎢⎢⎢⎢
⎣
⎡
=
dN
dd
NV
εεε
εεε
21
21
111
2/)1()()det( −<
∝−=∏ NNji
jiV εεε
kjikkkj
k
i
j
iij VAkm
rrJ =∂∂
=∂∂
= ∑ −1εθ
AVAVJJH TTT ==
Vandermonde Matrixd=N-1 Random instance εi / V ?
Random model A?Also, equal spacings in logs, level repulsion
-
ReferencesReferences“The sloppy model universality class and the Vandermonde matrix”, Joshua J. Waterfall, Fergal P. Casey, Ryan N. Gutenkunst, Kevin S. Brown, Christopher R. Myers, Piet W. Brouwer, Veit Elser, and James P. Sethna, http://arxiv.org/abs/cond-mat/0605387.“Sloppy systems biology: tight predictions with loose parameters”, Ryan N. Gutenkunst, Joshua J. Waterfall, Fergal P. Casey, Kevin S. Brown, Christopher R. Myers & James P. Sethna (submitted).“The Statistical Mechanics of Complex Signaling Networks: Nerve Growth Factor Signaling”, Kevin S. Brown, Colin C. Hill, Guillermo A. Calero, Christopher R. Myers, Kelvin H. Lee, James P. Sethna, and Richard A. Cerione, Physical Biology 1, 184-195 (2004) . “Statistical Mechanics Approaches to Models with Many Poorly Known Parameters”, Kevin S. Brown and James P. Sethna, Phys. Rev. E 68, 021904 (2003). “Bayesian Ensemble Approach to Error Estimation of InteratomicPotentials”, Søren L. Frederiksen, Karsten W. Jacobsen, Kevin S. Brown, and James P. Sethna, Phys. Rev. Letters 93, 165501 (2004).“Bayesian Error Estimation in Density Functional Theory”, J. J. Mortensen, K. Kaasbjerg, S. L. Frederiksen, J. K. Norskov, James P. Sethna, K. W. Jacobsen, Phys. Rev. Letters 95, 216401 (2005).
-
Universality in Universality in MultiparameterMultiparameterFitting: Sloppy ModelsFitting: Sloppy Models
James P. Sethna, Josh Waterfall, Ryan Gutenkunst, Fergal Casey, Kevin S. Brown, Chris Myers, Veit Elser, Piet Brouwer
Cell Dynamics
Fitting Exponentials, Polynomials
Fits good: measured bad
Universality in Multiparameter Fitting: Sloppy ModelsFitting Decaying ExponentialsFitting ExponentialsIs Sloppiness Universal?Sloppy Systems BiologyWhere is Sloppiness From?Why are they Sloppy?ReferencesUniversality in Multiparameter Fitting: Sloppy Models