grain size distributions application nal comments · grain size distributions application nal...
Transcript of grain size distributions application nal comments · grain size distributions application nal...
grain size distributions appli ation �nal ommentsDoing statisti s of (grain size) distributions(or why doing it easy if it an be in�nitely ompli ated?)Raimon Tolosana-DelgadoSedi-Seminar6 November 2007
Raimon Tolosana-Delgado Doing statisti s of (grain size) distributions
grain size distributions appli ation �nal ommentsoutline1 grain-size distribution hara terization2 geometry and statisti s of distributions hara teristi sCoDa analysisHilbert spa e3 appli ationgeneralitiesa gla ial data set from the Aar-Gotthard massif (Alps)the Darss sill data set4 �nal omments Raimon Tolosana-Delgado Doing statisti s of (grain size) distributions
grain size distributions appli ation �nal ommentsa grain-size distribution example
sample RT1− 19
Den
sity
0 5 10 15
0.00
0.05
0.10
0.15
0.20
Raimon Tolosana-Delgado Doing statisti s of (grain size) distributions
grain size distributions appli ation �nal ommentsa grain-size distribution example
sample RT1− 19
Den
sity
0 5 10 15
0.00
0.05
0.10
0.15
0.20
Raimon Tolosana-Delgado Doing statisti s of (grain size) distributions
two measurement typesquasi ontinuous LPS (x 5)binned sieve data
grain size distributions appli ation �nal ommentsa grain-size distribution example
sample RT1− 19
Den
sity
0 5 10 15
0.00
0.05
0.10
0.15
0.20
Raimon Tolosana-Delgado Doing statisti s of (grain size) distributions
distribution!in reasingbounded in [0,1℄
grain size distributions appli ation �nal ommentstraditional pro edure (Folk, 1979; Petrology of Sedimentary Ro ks)1 ompute the ummulative urve (distribution)2 read some quantiles form the line, and derive distribution measures: entral tenden ymode: most frequent φmedian: φ50
1
2(φ16 + φ84) ombined 1
3(φ16 + φ84 + φ50)sortinginterquartile range: 1
2(φ75 − φ25) (bad)
1
2(φ84 − φ16) ombined 1
4(φ84 − φ16) + 1
6.6(φ95 − φ5)Argument: statisti s require many al ulations ! omputers are here to do some dirty workRaimon Tolosana-Delgado Doing statisti s of (grain size) distributions
skewnesskurtosis
grain size distributions appli ation �nal omments hara teristi s CoDa analysis Hilbert spa ewhat must have a urve to be a density?it must be positiveits integral (area under the urve) must be one
Raimon Tolosana-Delgado Doing statisti s of (grain size) distributions
grain size distributions appli ation �nal omments hara teristi s CoDa analysis Hilbert spa ewhat must have a urve to be a density?it must be positiveits integral (area under the urve) must be onewhat an have/be a distribution?several modes vs. one single maximum ontinuous vs. dis ontinuoussmooth urve vs. fra tal hara ter (extreme irregularity)symmetry vs. skewnessvirtually any shape (not just a normal one)Raimon Tolosana-Delgado Doing statisti s of (grain size) distributions
grain size distributions appli ation �nal omments hara teristi s CoDa analysis Hilbert spa ewhat must have a urve to be a density?it must be positiveits integral (area under the urve) must be onewhat an have/be a distribution?several modes vs. one single maximum ontinuous vs. dis ontinuoussmooth urve vs. fra tal hara ter (extreme irregularity)symmetry vs. skewnessvirtually any shape (not just a normal one)Raimon Tolosana-Delgado Doing statisti s of (grain size) distributions
grain size distributions appli ation �nal omments hara teristi s CoDa analysis Hilbert spa e ompositional data: a brief summary (I) omposition: de�nitionsa omposition is . . . a ve tor x = [x1, . . . xD] withD positive omponents (xi ≥ 0)of onstant sum (P xi = 100)
Raimon Tolosana-Delgado Doing statisti s of (grain size) distributions
grain size distributions appli ation �nal omments hara teristi s CoDa analysis Hilbert spa e ompositional data: a brief summary (I) omposition: de�nitionsa omposition is . . . a ve tor x = [x1, . . . xD] withD positive omponents (xi ≥ 0)of onstant sum (P xi = 100)its sample spa e is:the set of possible measurementsthe simplex (SD), e.g. if D = 3 the ternary diagram
Raimon Tolosana-Delgado Doing statisti s of (grain size) distributions
grain size distributions appli ation �nal omments hara teristi s CoDa analysis Hilbert spa e ompositional data: a brief summary (I) omposition: de�nitionsa omposition is . . . a ve tor x = [x1, . . . xD] withD positive omponents (xi ≥ 0)of onstant sum (P xi = 100)its sample spa e is:the set of possible measurementsthe simplex (SD), e.g. if D = 3 the ternary diagramits s ale is:a subje tive assessment of the di�eren e between two valuesa relative one, an in rement from 100 ppm to 200 ppm is moresigni� ant that from 1000 ppm to 1100 ppm
Raimon Tolosana-Delgado Doing statisti s of (grain size) distributions
grain size distributions appli ation �nal omments hara teristi s CoDa analysis Hilbert spa e ompositional data: a brief summary (I) omposition: de�nitionsa omposition is . . . a ve tor x = [x1, . . . xD] withD positive omponents (xi ≥ 0)of onstant sum (P xi = 100)its sample spa e is:the set of possible measurementsthe simplex (SD), e.g. if D = 3 the ternary diagramits s ale is:a subje tive assessment of the di�eren e between two valuesa relative one, an in rement from 100 ppm to 200 ppm is moresigni� ant that from 1000 ppm to 1100 ppm
Raimon Tolosana-Delgado Doing statisti s of (grain size) distributions
grain size distributions appli ation �nal omments hara teristi s CoDa analysis Hilbert spa e ompositional data: a brief summary (II)the Simplex: a meaningful Eu lidean spa e stru turethe losure: C [·] : RD+ −→ S
D : x′ = xP
xi
Raimon Tolosana-Delgado Doing statisti s of (grain size) distributions
grain size distributions appli ation �nal omments hara teristi s CoDa analysis Hilbert spa e ompositional data: a brief summary (II)the Simplex: a meaningful Eu lidean spa e stru turethe losure: C [·] : RD+ −→ S
D : x′ = xP
xi
Raimon Tolosana-Delgado Doing statisti s of (grain size) distributions
grain size distributions appli ation �nal omments hara teristi s CoDa analysis Hilbert spa e ompositional data: a brief summary (II)the Simplex: a meaningful Eu lidean spa e stru turethe losure: C [·] : RD+ −→ S
D : x′ = xP
xi hange operation: perturbation x ⊕ y = C [x1 · y1, . . . , xD · yD]
Raimon Tolosana-Delgado Doing statisti s of (grain size) distributions
grain size distributions appli ation �nal omments hara teristi s CoDa analysis Hilbert spa e ompositional data: a brief summary (II)the Simplex: a meaningful Eu lidean spa e stru turethe losure: C [·] : RD+ −→ S
D : x′ = xP
xi hange operation: perturbation x ⊕ y = C [x1 · y1, . . . , xD · yD]�zero�, neutral element n = [1/D, . . . 1/D] omplement operation: power transformation λ⊙x = Cˆ
xλ1 , . . . , xλ
D
˜
Raimon Tolosana-Delgado Doing statisti s of (grain size) distributions
grain size distributions appli ation �nal omments hara teristi s CoDa analysis Hilbert spa e ompositional data: a brief summary (II)the Simplex: a meaningful Eu lidean spa e stru turethe losure: C [·] : RD+ −→ S
D : x′ = xP
xi hange operation: perturbation x ⊕ y = C [x1 · y1, . . . , xD · yD]�zero�, neutral element n = [1/D, . . . 1/D] omplement operation: power transformation λ⊙x = Cˆ
xλ1 , . . . , xλ
D
˜
Raimon Tolosana-Delgado Doing statisti s of (grain size) distributionsx2
x1
x3
0
-1
-2
1
2
-3
3
grain size distributions appli ation �nal omments hara teristi s CoDa analysis Hilbert spa e ompositional data: a brief summary (II)the Simplex: a meaningful Eu lidean spa e stru turethe losure: C [·] : RD+ −→ S
D : x′ = xP
xi hange operation: perturbation x ⊕ y = C [x1 · y1, . . . , xD · yD]�zero�, neutral element n = [1/D, . . . 1/D] omplement operation: power transformation λ⊙x = Cˆ
xλ1 , . . . , xλ
D
˜Ait hison distan e: d2A(x,y) =
∑
i>j
(
ln xi
yi− ln
xj
yj
)2
Raimon Tolosana-Delgado Doing statisti s of (grain size) distributions
grain size distributions appli ation �nal omments hara teristi s CoDa analysis Hilbert spa e ompositional data: a brief summary (II)the Simplex: a meaningful Eu lidean spa e stru turethe losure: C [·] : RD+ −→ S
D : x′ = xP
xi hange operation: perturbation x ⊕ y = C [x1 · y1, . . . , xD · yD]�zero�, neutral element n = [1/D, . . . 1/D] omplement operation: power transformation λ⊙x = Cˆ
xλ1 , . . . , xλ
D
˜Ait hison distan e: d2A(x,y) =
∑
i>j
(
ln xi
yi− ln
xj
yj
)2Ait hison s alar produ t 〈x,y〉A =P
i>j ln xi
yi· ln xj
yjAit hison norm ||x||A = d(n,x)
Raimon Tolosana-Delgado Doing statisti s of (grain size) distributions
grain size distributions appli ation �nal omments hara teristi s CoDa analysis Hilbert spa e ompositional data: a brief summary (II)the Simplex: a meaningful Eu lidean spa e stru turethe losure: C [·] : RD+ −→ S
D : x′ = xP
xi hange operation: perturbation x ⊕ y = C [x1 · y1, . . . , xD · yD]�zero�, neutral element n = [1/D, . . . 1/D] omplement operation: power transformation λ⊙x = Cˆ
xλ1 , . . . , xλ
D
˜Ait hison distan e: d2A(x,y) =
∑
i>j
(
ln xi
yi− ln
xj
yj
)2Ait hison s alar produ t 〈x,y〉A =P
i>j ln xi
yi· ln xj
yjAit hison norm ||x||A = d(n,x)
Raimon Tolosana-Delgado Doing statisti s of (grain size) distributions
parallel and orthogonal linesx2
x1
x3
n
grain size distributions appli ation �nal omments hara teristi s CoDa analysis Hilbert spa e ompositional data: a brief summary (II)the Simplex: a meaningful Eu lidean spa e stru turethe losure: C [·] : RD+ −→ S
D : x′ = xP
xi hange operation: perturbation x ⊕ y = C [x1 · y1, . . . , xD · yD]�zero�, neutral element n = [1/D, . . . 1/D] omplement operation: power transformation λ⊙x = Cˆ
xλ1 , . . . , xλ
D
˜Ait hison distan e: d2A(x,y) =
∑
i>j
(
ln xi
yi− ln
xj
yj
)2Ait hison s alar produ t 〈x,y〉A =P
i>j ln xi
yi· ln xj
yjAit hison norm ||x||A = d(n,x)
Raimon Tolosana-Delgado Doing statisti s of (grain size) distributions
ellipses and ir les
grain size distributions appli ation �nal omments hara teristi s CoDa analysis Hilbert spa e ompositional data: a brief summary (II)the Simplex: a meaningful Eu lidean spa e stru turethe losure: C [·] : RD+ −→ S
D : x′ = xP
xi hange operation: perturbation x ⊕ y = C [x1 · y1, . . . , xD · yD]�zero�, neutral element n = [1/D, . . . 1/D] omplement operation: power transformation λ⊙x = Cˆ
xλ1 , . . . , xλ
D
˜Ait hison distan e: d2A(x,y) =
∑
i>j
(
ln xi
yi− ln
xj
yj
)2Ait hison s alar produ t 〈x,y〉A =P
i>j ln xi
yi· ln xj
yjAit hison norm ||x||A = d(n,x)basis: a set of D− 1 ompositions (orthonormal) e1, . . . , eD−1 ∈ SD
Raimon Tolosana-Delgado Doing statisti s of (grain size) distributions
grain size distributions appli ation �nal omments hara teristi s CoDa analysis Hilbert spa e ompositional data: a brief summary (II)the Simplex: a meaningful Eu lidean spa e stru turethe losure: C [·] : RD+ −→ S
D : x′ = xP
xi hange operation: perturbation x ⊕ y = C [x1 · y1, . . . , xD · yD]�zero�, neutral element n = [1/D, . . . 1/D] omplement operation: power transformation λ⊙x = Cˆ
xλ1 , . . . , xλ
D
˜Ait hison distan e: d2A(x,y) =
∑
i>j
(
ln xi
yi− ln
xj
yj
)2Ait hison s alar produ t 〈x,y〉A =P
i>j ln xi
yi· ln xj
yjAit hison norm ||x||A = d(n,x)basis: a set of D− 1 ompositions (orthonormal) e1, . . . , eD−1 ∈ SD
Raimon Tolosana-Delgado Doing statisti s of (grain size) distributions
2 systems of axisx y
z
x y
z
grain size distributions appli ation �nal omments hara teristi s CoDa analysis Hilbert spa e ompositional data: a brief summary (II)the Simplex: a meaningful Eu lidean spa e stru turethe losure: C [·] : RD+ −→ S
D : x′ = xP
xi hange operation: perturbation x ⊕ y = C [x1 · y1, . . . , xD · yD]�zero�, neutral element n = [1/D, . . . 1/D] omplement operation: power transformation λ⊙x = Cˆ
xλ1 , . . . , xλ
D
˜Ait hison distan e: d2A(x,y) =
∑
i>j
(
ln xi
yi− ln
xj
yj
)2Ait hison s alar produ t 〈x,y〉A =P
i>j ln xi
yi· ln xj
yjAit hison norm ||x||A = d(n,x)basis: a set of D− 1 ompositions (orthonormal) e1, . . . , eD−1 ∈ SD oordinates: log-ratios (relative s ale) ξ1, . . . ξD−1 ∈ R su h that
x = ξ1 ⊙ e1 ⊕ ξ2 ⊙ e2 ⊕ · · · ⊕ ξD−1 ⊙ eD−1
Raimon Tolosana-Delgado Doing statisti s of (grain size) distributions
grain size distributions appli ation �nal omments hara teristi s CoDa analysis Hilbert spa e ompositional data: a brief summary (II)the Simplex: a meaningful Eu lidean spa e stru turethe losure: C [·] : RD+ −→ S
D : x′ = xP
xi hange operation: perturbation x ⊕ y = C [x1 · y1, . . . , xD · yD]�zero�, neutral element n = [1/D, . . . 1/D] omplement operation: power transformation λ⊙x = Cˆ
xλ1 , . . . , xλ
D
˜Ait hison distan e: d2A(x,y) =
∑
i>j
(
ln xi
yi− ln
xj
yj
)2Ait hison s alar produ t 〈x,y〉A = 〈clr(x), clr(y)〉eucAit hison norm ||x||A = d(n,x)basis: a set of D− 1 ompositions (orthonormal) e1, . . . , eD−1 ∈ SD oordinates: log-ratios (relative s ale) ξ1, . . . ξD−1 ∈ R su h that
x = ξ1 ⊙ e1 ⊕ ξ2 ⊙ e2 ⊕ · · · ⊕ ξD−1 ⊙ eD−1 entered log-ratio transformation:clr(x) = ln x
D√
Q
i xi
= ln(x) − 1
D
P
i lnxiRaimon Tolosana-Delgado Doing statisti s of (grain size) distributions
grain size distributions appli ation �nal omments hara teristi s CoDa analysis Hilbert spa efrom ompositions to distributionsdistributions: ompositions of in�nite partsa distribution is . . . a urve f(x) withpositive value f(x) > 0 for all x
Raimon Tolosana-Delgado Doing statisti s of (grain size) distributions
grain size distributions appli ation �nal omments hara teristi s CoDa analysis Hilbert spa efrom ompositions to distributionsdistributions: ompositions of in�nite partsa distribution is . . . a urve f(x) withpositive value f(x) > 0 for all xof sum equals 100% =⇒ integral equal to one R
f(x)dx = 1
Raimon Tolosana-Delgado Doing statisti s of (grain size) distributions
grain size distributions appli ation �nal omments hara teristi s CoDa analysis Hilbert spa efrom ompositions to distributionsdistributions: ompositions of in�nite partsa distribution is . . . a urve f(x) withpositive value f(x) > 0 for all xof integral equal to one R
f(x)dx = 1its sample spa e is an in�nite-dimensional �simplex�, one of theAit hison spa es of measures (Egoz ue and Diaz-Barrero, 2003; vanden Boogaart, 2005)Raimon Tolosana-Delgado Doing statisti s of (grain size) distributions
grain size distributions appli ation �nal omments hara teristi s CoDa analysis Hilbert spa efrom ompositions to distributionsdistributions: ompositions of in�nite partsa distribution is . . . a urve f(x) withpositive value f(x) > 0 for all xof integral equal to one R
f(x)dx = 1its sample spa e is an in�nite-dimensional �simplex�, one of theAit hison spa es of measures (Egoz ue and Diaz-Barrero, 2003; vanden Boogaart, 2005)its s ale is?Raimon Tolosana-Delgado Doing statisti s of (grain size) distributions
grain size distributions appli ation �nal omments hara teristi s CoDa analysis Hilbert spa efrom ompositions to distributionshow do we ompare grain-size urves?sample RT1− 10
Den
sity
0 5 10 15
0.00
0.05
0.10
0.15
0.20
0.25
0.30
sample RT1− 16
Den
sity
0 5 10 15
0.0
0.1
0.2
0.3
0.4
sample RT1− 17
Den
sity
0 5 10 15
0.00
0.05
0.10
0.15
0.20
Raimon Tolosana-Delgado Doing statisti s of (grain size) distributions
grain size distributions appli ation �nal omments hara teristi s CoDa analysis Hilbert spa efrom ompositions to distributionsthe Hilbert spa e of distributions: operationsHilbert spa e: �Eu lidean spa e of in�nite dimension� (but . . .)not unique!
Raimon Tolosana-Delgado Doing statisti s of (grain size) distributions
grain size distributions appli ation �nal omments hara teristi s CoDa analysis Hilbert spa efrom ompositions to distributionsthe Hilbert spa e of distributions: operationsHilbert spa e: �Eu lidean spa e of in�nite dimension� (but . . .)not unique!perturbation (translation, sum), powering (multipli ation)the neutral element must be hosen: n(x) (normal, exponential,uniform)Raimon Tolosana-Delgado Doing statisti s of (grain size) distributions
grain size distributions appli ation �nal omments hara teristi s CoDa analysis Hilbert spa efrom ompositions to distributionsthe Hilbert spa e of distributions: operationsHilbert spa e: �Eu lidean spa e of in�nite dimension� (but . . .)not unique!perturbation (translation, sum), powering (multipli ation)the neutral element must be hosen: n(x) (normal, exponential,uniform)s alar produ t ⇒ angles, distan es, lengthsRaimon Tolosana-Delgado Doing statisti s of (grain size) distributions
grain size distributions appli ation �nal omments hara teristi s CoDa analysis Hilbert spa efrom ompositions to distributionsthe Hilbert spa e of distributions: operationsHilbert spa e: �Eu lidean spa e of in�nite dimension� (but . . .)not unique!perturbation (translation, sum), powering (multipli ation)the neutral element must be hosen: n(x) (normal, exponential,uniform)s alar produ t ⇒ angles, distan es, lengths〈f(x), g(x)〉 =
∫
dom(n)
clr (f(x)) · clr (g(x))n(x)dx
Raimon Tolosana-Delgado Doing statisti s of (grain size) distributions
grain size distributions appli ation �nal omments hara teristi s CoDa analysis Hilbert spa efrom ompositions to distributionsthe Hilbert spa e of distributions: operationsHilbert spa e: �Eu lidean spa e of in�nite dimension� (but . . .)not unique!perturbation (translation, sum), powering (multipli ation)the neutral element must be hosen: n(x) (normal, exponential,uniform)s alar produ t ⇒ angles, distan es, lengths〈f(x), g(x)〉 =
∫
dom(n)
clr (f(x)) · clr (g(x))n(x)dx lr transformationclr (f(x)) = ln
f(x)
n(x)−
∫
dom(n)
lnf(x)
n(x)n(x)dxRaimon Tolosana-Delgado Doing statisti s of (grain size) distributions
grain size distributions appli ation �nal omments hara teristi s CoDa analysis Hilbert spa efrom ompositions to distributionsthe Hilbert spa e of distributions: basis and oordinatesin�nite dimension ⇒ in�nite oordinates, in�nite elements in a basisorthonormal:∫
d(n)
f∗(x) · g∗(x)n(x)dx = 0 and
∫
d(n)
(f∗(x))2· n(x)dx = 1
Raimon Tolosana-Delgado Doing statisti s of (grain size) distributions
grain size distributions appli ation �nal omments hara teristi s CoDa analysis Hilbert spa efrom ompositions to distributionsthe Hilbert spa e of distributions: basis and oordinatesin�nite dimension ⇒ in�nite oordinates, in�nite elements in a basisorthonormal:∫
d(n)
f∗(x) · g∗(x)n(x)dx = 0 and
∫
d(n)
(f∗(x))2· n(x)dx = 1series of polynomials (standard physi al-mathemati al tool)normal density: Hermite polynomialsexponential density: Laguerre polynomialsuniform density: Legendre polynomials (beta density: Ja obipolynomials)
Raimon Tolosana-Delgado Doing statisti s of (grain size) distributions
grain size distributions appli ation �nal omments hara teristi s CoDa analysis Hilbert spa efrom ompositions to distributionsthe Hilbert spa e of distributions: basis and oordinatesin�nite dimension ⇒ in�nite oordinates, in�nite elements in a basisorthonormal:∫
d(n)
f∗(x) · g∗(x)n(x)dx = 0 and
∫
d(n)
(f∗(x))2· n(x)dx = 1series of polynomials (standard physi al-mathemati al tool)normal density: Hermite polynomialsexponential density: Laguerre polynomialsuniform density: Legendre polynomials (beta density: Ja obipolynomials) oordinatesHermite polynomials: µ
σ2 , σ2−1
√
2σ2, et . (deviation from normality)Laguerre polynomials: λ − 1Raimon Tolosana-Delgado Doing statisti s of (grain size) distributions
grain size distributions appli ation �nal omments hara teristi s CoDa analysis Hilbert spa eexample of Hermite polynomialspolynomials: H0 = 1, H1 = x, H2 = 1√2(1 − x2)de omposition: φ(x; µ, σ) = φ(x; 0, 1) ⊕ µσ2 ⊙ eH1 ⊕ σ2−1√
2σ2⊙ eH2
−15 −10 −5 0 5
0.00
0.04
0.08
0.12
mean: −5
−15 −10 −5 0 5
050
100
var: 10
−8 −6 −4 −2
0.0
0.1
0.2
0.3
0.4
mean: −5
−8 −6 −4 −2
010
2030
var: 1
−4 −2 0 2 4 6
0.00
0.10
0.20
0.30
mean: 0
−4 −2 0 2 4 6
02
46
810
var: 2
−1.0 −0.5 0.0 0.5 1.0
0.0
0.4
0.8
1.2
mean: 0
−1.0 −0.5 0.0 0.5 1.0
−4
−3
−2
−1
01
var: 0.1
2 3 4 5 6 7
0.0
0.2
0.4
0.6
mean: 5
2 3 4 5 6 7
05
1020
var: 0.5
0 5 10
0.00
0.10
0.20 mean: 5
0 5 10
020
4060
80
var: 5
Raimon Tolosana-Delgado Doing statisti s of (grain size) distributions
grain size distributions appli ation �nal omments hara teristi s CoDa analysis Hilbert spa eexample of Hermite polynomials
−10 −5 0 5 10
0.00
0.02
0.04
0.06
0.08
X
Y
0, −0.25, 0, 0, −0.050, −0.25, 0.25, 0, −0.050, −0.25, 0.5, 0, −0.050, −0.25, 0.75, 0, −0.050, −0.25, 1, 0, −0.05
Raimon Tolosana-Delgado Doing statisti s of (grain size) distributions
grain size distributions appli ation �nal omments generalities Aar-Gotthard Darssalgorithm1 look at the data2 hoose a reasonable �zero� distribution, referen e distribution
Raimon Tolosana-Delgado Doing statisti s of (grain size) distributions
grain size distributions appli ation �nal omments generalities Aar-Gotthard Darssalgorithm1 look at the data2 hoose a reasonable �zero� distribution, referen e distributiondi�usion ⇒ normal distribution omminution ⇒ fra tal, exponential distributionbounded phenomena ⇒ uniform distribution
Raimon Tolosana-Delgado Doing statisti s of (grain size) distributions
grain size distributions appli ation �nal omments generalities Aar-Gotthard Darssalgorithm1 look at the data2 hoose a reasonable �zero� distribution, referen e distributiondi�usion ⇒ normal distribution omminution ⇒ fra tal, exponential distributionbounded phenomena ⇒ uniform distribution3 ompute the �standardized� density:ln( your histogram/zero density )=Y4 �x (whi h, how many) polynomials to use, orthonormal with respe tto the zero density (X)Raimon Tolosana-Delgado Doing statisti s of (grain size) distributions
grain size distributions appli ation �nal omments generalities Aar-Gotthard Darssalgorithm1 look at the data2 hoose a reasonable �zero� distribution, referen e distributiondi�usion ⇒ normal distribution omminution ⇒ fra tal, exponential distributionbounded phenomena ⇒ uniform distribution3 ompute the �standardized� density:ln( your histogram/zero density )=Y4 �x (whi h, how many) polynomials to use, orthonormal with respe tto the zero density (X)5 Y = X · β =⇒ regression =⇒ β̂ new �data� (meaningless β0!)Raimon Tolosana-Delgado Doing statisti s of (grain size) distributions
grain size distributions appli ation �nal omments generalities Aar-Gotthard Darssalgorithm1 look at the data2 hoose a reasonable �zero� distribution, referen e distributiondi�usion ⇒ normal distribution omminution ⇒ fra tal, exponential distributionbounded phenomena ⇒ uniform distribution3 ompute the �standardized� density:ln( your histogram/zero density )=Y4 �x (whi h, how many) polynomials to use, orthonormal with respe tto the zero density (X)5 Y = X · β =⇒ regression =⇒ β̂ new �data� (meaningless β0!)6 apply the method you want to β̂prin ipal omponent analysis, non-supervised lassi� ationmapping Raimon Tolosana-Delgado Doing statisti s of (grain size) distributions
grain size distributions appli ation �nal omments generalities Aar-Gotthard Darssa gla ial data set from the Aar-Gotthard massif (Alps)goal and lo ation hara terize the e�e t of (weathering and) omminution ongrain-size/geo hemistry/mineralogy ompositionsfo us: omminution ∼ gla ial omminution (�uvial sorting)Aar-Gotthard graniti massif, entral Alps (Switzerland); RhoneGlets her, Damma Glets her, Tiefer Glets her�uvial sediments, re ent entral morraines, older lateral morrainesgrain size analyses:laser parti le sizer (LPS), 116 lasses of φ ∈ [−0.865, 14.61]sieves, 11 lasses from φ < −1 to φ > 8analyses of major oxides and tra e element geo hemistry (not now)Raimon Tolosana-Delgado Doing statisti s of (grain size) distributions
grain size distributions appli ation �nal omments generalities Aar-Gotthard Darssmeasured densitiessample RT1− 1b
0 5 10 15
0.0
0.2
0.4
0.6
0.8
1.0
sample RT1− 2
0 5 10 15
0.0
0.2
0.4
0.6
0.8
1.0
sample RT1− 3
0 5 10 15
0.0
0.2
0.4
0.6
0.8
1.0
sample RT1− 7a
0 5 10 15
0.0
0.2
0.4
0.6
0.8
1.0
sample RT1− 7b
0 5 10 15
0.0
0.2
0.4
0.6
0.8
1.0
sample RT1− 7c
0 5 10 15
0.0
0.2
0.4
0.6
0.8
1.0
sample RT1− 8
0 5 10 15
0.0
0.2
0.4
0.6
0.8
1.0
sample RT1− 9
0 5 10 15
0.0
0.2
0.4
0.6
0.8
1.0
sample RT1− 10
0 5 10 15
0.0
0.2
0.4
0.6
0.8
1.0
sample RT1− 11
0 5 10 15
0.0
0.2
0.4
0.6
0.8
1.0
sample RT1− 16
0 5 10 15
0.0
0.2
0.4
0.6
0.8
1.0
sample RT1− 17
0 5 10 15
0.0
0.2
0.4
0.6
0.8
1.0
sample RT1− 19
0 5 10 15
0.0
0.2
0.4
0.6
0.8
1.0
sample RT1− 20
0 5 10 15
0.0
0.2
0.4
0.6
0.8
1.0
Raimon Tolosana-Delgado Doing statisti s of (grain size) distributions
grain size distributions appli ation �nal omments generalities Aar-Gotthard Darssmeasured densitiessummary of hara teristi ssample gla ier pla ement oarsest modes �nest1b Rhone side yes 2 yes2 Rhone side yes 1 no3 Rhone side yes 1 no7a Damma side (older) yes 3 yes7b Damma side (older) yes 3 no7 Damma side (older) yes many yes8 Damma side (older) yes 2 yes9 Damma front no 1 yes10 Damma front yes 1 no11 Damma front yes 1 no16 Rhone washed no 1 yes17 Rhone washed yes 2 yes19 Tiefer front yes 2 yes20 Tiefer front yes 2 noRaimon Tolosana-Delgado Doing statisti s of (grain size) distributions
grain size distributions appli ation �nal omments generalities Aar-Gotthard Darss lr-transformed densities�tted model: 11 Laguerre polynomials (exponential referen e)0 5 10 15
−6
−4
−2
02
RT1−1b
phi
rel.d
ens.
0 5 10 15
−6
−4
−2
02
4
RT1−2
phi
rel.d
ens.
0 5 10 15
−8
−6
−4
−2
02
4
RT1−3
phi
rel.d
ens.
0 5 10 15
−6
−4
−2
02
RT1−7a
phi
rel.d
ens.
0 5 10 15
−6
−4
−2
02
RT1−7b
phi
rel.d
ens.
0 5 10 15
−6
−4
−2
02
RT1−7c
phi
rel.d
ens.
0 5 10 15
−6
−4
−2
02
RT1−8
phi
rel.d
ens.
0 5 10 15
−6
−4
−2
02
RT1−9
phi
rel.d
ens.
0 5 10 15
−6
−4
−2
02
RT1−10
phi
rel.d
ens.
0 5 10 15
−6
−4
−2
02
RT1−11
phi
rel.d
ens.
0 5 10 15
−8
−6
−4
−2
02
RT1−16
phi
rel.d
ens.
0 5 10 15
−6
−4
−2
02
RT1−17
phi
rel.d
ens.
0 5 10 15
−6
−4
−2
02
RT1−19
phi
rel.d
ens.
0 5 10 15
−6
−4
−2
02
4
RT1−20
phi
rel.d
ens.
RT
1−17
RT
1−10
RT
1−7c
RT
1−8
RT
1−20
RT
1−1b
RT
1−19
RT
1−3
RT
1−7a
RT
1−7b
RT
1−2
RT
1−11
RT
1−9
RT
1−16
020
060
010
00
Cluster Dendrogram
hclust (*, "ward")dist(t(beta))
Hei
ght
Raimon Tolosana-Delgado Doing statisti s of (grain size) distributions
grain size distributions appli ation �nal omments generalities Aar-Gotthard Darssoriginal densities�tted model: 11 Laguerre polynomials (exponential referen e)0 5 10 15
0.00
00.
010
0.02
0
RT1−1b
phi
dens
ity
0 5 10 15
0.00
0.01
0.02
0.03
0.04
0.05 RT1−2
phi
dens
ity
0 5 10 15
0.00
0.01
0.02
0.03
0.04
0.05
0.06
RT1−3
phi
dens
ity
0 5 10 15
0.00
0.01
0.02
0.03
0.04
RT1−7a
phi
dens
ity
0 5 10 15
0.00
00.
010
0.02
00.
030
RT1−7b
phi
dens
ity
0 5 10 15
0.00
00.
005
0.01
00.
015
0.02
0
RT1−7c
phi
dens
ity
0 5 10 15
0.00
00.
010
0.02
0
RT1−8
phi
dens
ity
0 5 10 15
0.00
0.01
0.02
0.03
0.04
RT1−9
phi
dens
ity
0 5 10 15
0.00
0.01
0.02
0.03
0.04
RT1−10
phi
dens
ity
0 5 10 15
0.00
0.01
0.02
0.03
0.04
0.05
RT1−11
phi
dens
ity
0 5 10 15
0.00
0.01
0.02
0.03
0.04
0.05
0.06
RT1−16
phi
dens
ity
0 5 10 15
0.00
00.
010
0.02
00.
030
RT1−17
phi
dens
ity
0 5 10 15
0.00
00.
005
0.01
00.
015
0.02
00.
025
RT1−19
phi
dens
ity
0 5 10 15
0.00
0.01
0.02
0.03
0.04
0.05
0.06
RT1−20
phi
dens
ity
−0.6 −0.4 −0.2 0.0 0.2 0.4 0.6
−0.
6−
0.4
−0.
20.
00.
20.
40.
6
explained var: 0.97
Comp.1
Com
p.2 RT1−1b
RT1−2
RT1−3
RT1−7aRT1−7b
RT1−7c
RT1−8
RT1−10
RT1−11
RT1−17
RT1−19
RT1−20
−3 −2 −1 0 1 2 3
−3
−2
−1
01
23
C2
C3 C4C5
C6
C7
C8
C9
C10
C11
Raimon Tolosana-Delgado Doing statisti s of (grain size) distributions
grain size distributions appli ation �nal omments generalities Aar-Gotthard Darss omparing densities: luster analysis of β̂'snon-supervised lassi� ation: Ward luster analysisR
T1−
17
RT
1−10
RT
1−7c
RT
1−8
RT
1−20
RT
1−1b
RT
1−19
RT
1−3
RT
1−7a
RT
1−7b
RT
1−2
RT
1−11
RT
1−9
RT
1−16
040
080
012
00
cluster analysis, Ward method
hclust (*, "ward")dist(t(beta))
Hei
ght
Raimon Tolosana-Delgado Doing statisti s of (grain size) distributions
grain size distributions appli ation �nal omments generalities Aar-Gotthard Darss omparing densities: 9 and 16 against the other?sample RT1− 9
Den
sity
0 5 10 15
0.0
0.1
0.2
0.3
sample RT1− 16
Den
sity
0 5 10 15
0.0
0.1
0.2
0.3
0.4
sample RT1− 17
Den
sity
0 5 10 15
0.00
0.05
0.10
0.15
0.20
sample RT1− 20
Den
sity
0 5 10 15
0.0
0.1
0.2
0.3
0.4
Raimon Tolosana-Delgado Doing statisti s of (grain size) distributions
grain size distributions appli ation �nal omments generalities Aar-Gotthard Darss omparing densities: a �map� of di�eren es
Raimon Tolosana-Delgado Doing statisti s of (grain size) distributions−0.6 −0.4 −0.2 0.0 0.2 0.4 0.6
−0.
6−
0.4
−0.
20.
00.
20.
40.
6
explained var: 0.97
Comp.1
Com
p.2
RT1−1b
RT1−2
RT1−3
RT1−7aRT1−7b
RT1−7c
RT1−8
RT1−10
RT1−11
RT1−17
RT1−19
RT1−20
−3 −2 −1 0 1 2 3
−3
−2
−1
01
23
C2
C3 C4
C5
C6
C7
C8
C9
C10
C11
hara teristi smorraines:old side: 7(a,b, ), 8re ent side: 1b, 2, 3front: (9) 10, 11, 19,20washed sediment:(16) 17unimodal:2, 3, 10, 11tail to �nest:heavy: 7 , (16), 17,19yes: 1b, 7a, 8, (9)no: 2, 3, 7b, 10, 11,20C2: pure exponential
grain size distributions appli ation �nal omments generalities Aar-Gotthard Darss omparing densities using the normal referen e
Raimon Tolosana-Delgado Doing statisti s of (grain size) distributions
RT
1−7c
RT
1−8
RT
1−3
RT
1−10
RT
1−11
RT
1−2
RT
1−17
RT
1−7a
RT
1−7b
RT
1−20
RT
1−1b
RT
1−19
RT
1−9
RT
1−16
050
000
1000
0015
0000
Cluster Dendrogram
hclust (*, "ward")dist(t(beta))
Hei
ght
−0.6 −0.4 −0.2 0.0 0.2 0.4 0.6
−0.
6−
0.4
−0.
20.
00.
20.
40.
6
explained var: 1
Comp.1
Com
p.2
RT1−1b
RT1−2
RT1−3
RT1−7aRT1−7b
RT1−7cRT1−8
RT1−10RT1−11
RT1−17
RT1−19
RT1−20
−10000 −5000 0 5000 10000
−10
000
−50
000
5000
1000
0
C2C3
C4
C5
C6
C7C8
C9 C10C11
hara teristi sold side morraines: 7(a,b, ), 8; re ent side: 1b, 2, 3; front: (9) 10, 11, 19, 20;washed sediment: (16) 17heavy tail to �nest: 7 , (16), 17, 19; some �ne: 1b, 7a, 8, (9); no �ne: 2, 3,7b, 10, 11, 20unimodal: 2, 3, 10, 11; C2: mean; C3: inverse varian e
grain size distributions appli ation �nal omments generalities Aar-Gotthard Darssthe Darss sill data setDrogden Sill
Darss Sill
Darss
peninsula
Falster
Moen
PlantagenetGround
Arkona Basin
Meck
lenburg
Bay
Fehmarn Belt
Oere
so
un
d
Gre
at
Belt
Little
Belt
Kattegat
Kadet C
hannel
Lolland
Germany
Denmark Sweden
Investigation areaRaimon Tolosana-Delgado Doing statisti s of (grain size) distributions10000 20000 30000 40000 50000 60000 70000
4180
000
4200
000
4220
000
4240
000
Easting
Nor
thin
g
grain size distributions appli ation �nal omments generalities Aar-Gotthard Darssthe Darss sill data setthe Darss sill data setBalti Sea (412,560 km2): world large bra kish water bodieso eanographi onditions ontrolled by: (i) large river freshwaterinput, (ii) restri ted o ean onne tions (Danish Straits)∼ 73% of water ex hange is via the Darss Sill (minimum waterdepth: 18 m bsl)prevailing out�ow of bra kish Balti waters in the upper part of thewater olumn usually along the Danish oastin�ow of more saline water in the southern part of the area and inthe deeper part of the Kadet Channeltidal urrents are negligible; sediment-transporting bottom urrentsare intermittentout ropping till geneti ally related to an i e marginal zone, shortre-advan e of the retreating Late Wei hselian i e sheet (∼ 13,500years BP); thin over of lag sediments (lo ally, stones and blo ks >1 m) Raimon Tolosana-Delgado Doing statisti s of (grain size) distributions
grain size distributions appli ation �nal omments generalities Aar-Gotthard Darss onventional analysis: average (µ) and sorting (σ2)10000 20000 30000 40000 50000 60000 70000
4180
000
4200
000
4220
000
4240
000
Easting
Nor
thin
g
mean phi:
0.18−0.810.85−1.51.51−2.17
2.17−2.832.83−3.453.49−4.15
10000 20000 30000 40000 50000 60000 70000
4180
000
4200
000
4220
000
4240
000
Easting
Nor
thin
g
phi variance:
0.03−1.121.13−2.192.23−3.31
3.37−4.314.46−4.836.63−6.63Kadet Channel, Danish oast (↑ φ̄; ↑ σ2
φ;)Kadet Channel, German oast (∼ φ̄; ↓ σ2φ;)Ground, Plantagenet basin? (↑ φ̄; ↓ σ2
φ;)Bar, Darss Sill itself? (↓ φ̄; ↓ σ2φ;)Raimon Tolosana-Delgado Doing statisti s of (grain size) distributions
grain size distributions appli ation �nal omments generalities Aar-Gotthard Darssproposed analysis (I): average (β1 = µσ2 ) and sorting (β2 = σ2−1√
2σ2)
10000 20000 30000 40000 50000 60000 70000
4180
000
4200
000
4220
000
4240
000
Easting
Nor
thin
g
first coordinate:
0.04−18.8218.91−37.2238.04−54.22
57.1−75.1276.05−83.44113.55−113.55
10000 20000 30000 40000 50000 60000 70000
4180
000
4200
000
4220
000
4240
000
Easting
Nor
thin
g
second coordinate:
−26.81 :: −26.81−19.58 :: −17.83−17.63 :: −13.13
−12.58 :: −8.72−8.51 :: −3.96−3.94 :: 0.6two areas with extremely low varian e (↓↓ σ2
φ;)undis riminated ground (∼ σ2φ;)Raimon Tolosana-Delgado Doing statisti s of (grain size) distributions
grain size distributions appli ation �nal omments generalities Aar-Gotthard Darssproposed analysis (II): a non-redundant hara terization−0.2 −0.1 0.0 0.1 0.2
−0.
2−
0.1
0.0
0.1
0.2
Comp.1
Com
p.2
−400 −200 0 200 400
−40
0−
200
020
040
0
b1 b2
β1 and β2 almost proportional (inverse)Raimon Tolosana-Delgado Doing statisti s of (grain size) distributions
grain size distributions appli ation �nal omments generalities Aar-Gotthard Darssproposed analysis (II): a non-redundant hara terization−0.2 −0.1 0.0 0.1 0.2
−0.
2−
0.1
0.0
0.1
0.2
Comp.1
Com
p.2
−400 −200 0 200 400
−40
0−
200
020
040
0
b1 b2
10000 20000 30000 40000 50000 60000 70000
4180
000
4200
000
4220
000
4240
000
Easting
Nor
thin
g
second PC:
−0.86 :: −0.26−0.26 :: 0.350.35 :: 0.93
0.96 :: 1.53Inf :: −Inf2.33 :: 2.76
β1 and β2 almost proportional (inverse) omplementary dire tion to β1||β2 (non-normality?)Raimon Tolosana-Delgado Doing statisti s of (grain size) distributions
grain size distributions appli ation �nal omments generalities Aar-Gotthard Darssproposed analysis (II): a non-redundant hara terization10000 20000 30000 40000 50000 60000 70000
4180
000
4200
000
4220
000
4240
000
Easting
Nor
thin
g
first coordinate:
0.04−18.8218.91−37.2238.04−54.22
57.1−75.1276.05−83.44113.55−113.55
10000 20000 30000 40000 50000 60000 70000
4180
000
4200
000
4220
000
4240
000
EastingN
orth
ing
second PC:
−0.86 :: −0.26−0.26 :: 0.350.35 :: 0.93
0.96 :: 1.53Inf :: −Inf2.33 :: 2.76Kadet Channel, Danish oast (low departure)Kadet Channel, German oast (average departure)two singular areas (extreme sorting, average-high departure)undis riminated ground (low departure) ∼ Danish oastRaimon Tolosana-Delgado Doing statisti s of (grain size) distributions
grain size distributions appli ation �nal omments�nal omments1 grain size urves an be seen as (in�nite-dimensional) ompositionshigh-dimensionalfurther stru ture (ordered bins, smoothness)relative s ale
Raimon Tolosana-Delgado Doing statisti s of (grain size) distributions
grain size distributions appli ation �nal omments�nal omments1 grain size urves an be seen as (in�nite-dimensional) ompositionshigh-dimensionalfurther stru ture (ordered bins, smoothness)relative s ale2 grain size information an be summarized in a few parametersdepend on a referen e urveorthogonal =⇒ independentin reasingly omplexRaimon Tolosana-Delgado Doing statisti s of (grain size) distributions
grain size distributions appli ation �nal omments�nal omments1 grain size urves an be seen as (in�nite-dimensional) ompositionshigh-dimensionalfurther stru ture (ordered bins, smoothness)relative s ale2 grain size information an be summarized in a few parametersdepend on a referen e urveorthogonal =⇒ independentin reasingly omplex3 these parameters may be statisti ally treatedto un over groups ( luster)to display patterns of variation (biplot)Raimon Tolosana-Delgado Doing statisti s of (grain size) distributions
grain size distributions appli ation �nal omments�nal omments1 grain size urves an be seen as (in�nite-dimensional) ompositionshigh-dimensionalfurther stru ture (ordered bins, smoothness)relative s ale2 grain size information an be summarized in a few parametersdepend on a referen e urveorthogonal =⇒ independentin reasingly omplex3 these parameters may be statisti ally treatedto un over groups ( luster)to display patterns of variation (biplot)to explain them (regression, anova)to make several measures ompatible (weighted average)Raimon Tolosana-Delgado Doing statisti s of (grain size) distributions