Introduction

This document provides solutions for semivariogram models #1 through #9/For each model, we identify:

  • Nugget
  • Partial Sill
  • Total Sill
  • Range (or effective range)
  • Covariance function \(C(d)\) (if it exists)

The effective range is defined as the distance \(d\) at which the semivariogram reaches 95% of the total sill \((\tau^2 + \sigma^2)\):

\[ \gamma(d_{eff}) = 0.95(\tau^2 + \sigma^2) \]


Model #1: Linear

Definition

\[ \gamma(d) = \begin{cases} \tau^2 + \sigma^2 d & d > 0, \tau^2 > 0, \sigma^2 > 0 \\ 0 & d = 0 \end{cases} \]

(a) Nugget, Sill, and Range

Parameter Value Explanation
Nugget \(\tau^2\) \(\lim_{d \to 0^+} \gamma(d) = \tau^2 + \sigma^2(0) = \tau^2\)
Partial Sill Does not exist No finite sill - grows linearly without bound
Total Sill Does not exist \(\lim_{d \to \infty} \gamma(d) = \infty\)
Range Does not exist Never reaches a sill
Effective Range Does not exist No sill to define percentage of

(b) Covariance Function

\[ \boxed{\text{The covariance function } C(d) \text{ does NOT exist for the linear model.}} \]

The model is intrinsic (not stationary), so it has no valid covariance function.


Model #2: Spherical

Definition

\[ \gamma(d) = \begin{cases} \tau^2 + \sigma^2 \left[ \frac{3}{2}\phi d - \frac{1}{2}(\phi d)^3 \right] & 0 < d \le 1/\phi \\ \tau^2 + \sigma^2 & d \ge 1/\phi \\ 0 & d = 0 \end{cases} \]

(a) Nugget, Sill, and Range

Parameter Value
Nugget \(\tau^2\)
Partial Sill \(\sigma^2\)
Total Sill \(\tau^2 + \sigma^2\)
Range \(1/\phi\)
Effective Range \(1/\phi\) (same as range)

(b) Covariance Function

For \(d = 0\): \[ C(0) = \tau^2 + \sigma^2 \]

For \(0 < d \le 1/\phi\): \[ C(d) = \sigma^2 \left[ 1 - \frac{3}{2}\phi d + \frac{1}{2}(\phi d)^3 \right] \]

For \(d \ge 1/\phi\): \[ C(d) = 0 \]

Thus: \[ \boxed{ C(d) = \begin{cases} \tau^2 + \sigma^2 & d = 0 \\ \sigma^2 \left[ 1 - \frac{3}{2}\phi d + \frac{1}{2}(\phi d)^3 \right] & 0 < d \le 1/\phi \\ 0 & d \ge 1/\phi \end{cases} } \]

Effective Range Derivation

For the spherical model, the range is exact (not asymptotic). The model reaches the sill exactly at \(d = 1/\phi\):

At \(d = 1/\phi\): \[ \gamma(1/\phi) = \tau^2 + \sigma^2 \left[ \frac{3}{2}(1) - \frac{1}{2}(1)^3 \right] = \tau^2 + \sigma^2 \left[ \frac{3}{2} - \frac{1}{2} \right] = \tau^2 + \sigma^2 \]

So the range is \(1/\phi\), and since it reaches the sill exactly, the effective range equals the range.


Model #3: Exponential

Definition

\[ \gamma(d) = \begin{cases} \tau^2 + \sigma^2 \left(1 - \exp(-\phi d)\right) & d > 0 \\ 0 & d = 0 \end{cases} \]

(a) Nugget, Sill, and Range

Parameter Value
Nugget \(\tau^2\)
Partial Sill \(\sigma^2\)
Total Sill \(\tau^2 + \sigma^2\)
Range Does not exist (asymptotic)
Effective Range \(3/\phi\)

(b) Covariance Function

For \(d = 0\): \[ C(0) = \tau^2 + \sigma^2 \]

For \(d > 0\): \[ C(d) = \sigma^2 e^{-\phi d} \]

Thus: \[ \boxed{ C(d) = \begin{cases} \tau^2 + \sigma^2 & d = 0 \\ \sigma^2 e^{-\phi d} & d > 0 \end{cases} } \]

Effective Range Derivation

We want \(\gamma(d) = 0.95(\tau^2 + \sigma^2)\):

\[ \tau^2 + \sigma^2(1 - e^{-\phi d}) = 0.95(\tau^2 + \sigma^2) \]

Subtract \(\tau^2\) from both sides:

\[ \sigma^2(1 - e^{-\phi d}) = 0.95\sigma^2 \]

Divide by \(\sigma^2\):

\[ 1 - e^{-\phi d} = 0.95 \]

Solve for \(e^{-\phi d}\):

\[ e^{-\phi d} = 0.05 \]

Take the natural logarithm:

\[ -\phi d = \ln(0.05) \]

Solve for \(d\):

\[ d = -\frac{\ln(0.05)}{\phi} \]

Since \(-\ln(0.05) \approx 2.9957 \approx 3\):

\[ \boxed{d_{eff} \approx \frac{3}{\phi}} \]


Model #4: Gaussian

Definition

\[ \gamma(d) = \begin{cases} \tau^2 + \sigma^2 \left(1 - \exp(-\phi^2 d^2)\right) & d > 0 \\ 0 & d = 0 \end{cases} \]

(a) Nugget, Sill, and Range

Parameter Value
Nugget \(\tau^2\)
Partial Sill \(\sigma^2\)
Total Sill \(\tau^2 + \sigma^2\)
Range Does not exist (asymptotic)
Effective Range \(\sqrt{3}/\phi\)

(b) Covariance Function

For \(d = 0\): \[ C(0) = \tau^2 + \sigma^2 \]

For \(d > 0\): \[ C(d) = \sigma^2 e^{-\phi^2 d^2} \]

Thus: \[ \boxed{ C(d) = \begin{cases} \tau^2 + \sigma^2 & d = 0 \\ \sigma^2 e^{-\phi^2 d^2} & d > 0 \end{cases} } \]

Effective Range Derivation

We want \(\gamma(d) = 0.95(\tau^2 + \sigma^2)\):

\[ \tau^2 + \sigma^2(1 - e^{-\phi^2 d^2}) = 0.95(\tau^2 + \sigma^2) \]

Subtract \(\tau^2\) from both sides:

\[ \sigma^2(1 - e^{-\phi^2 d^2}) = 0.95\sigma^2 \]

Divide by \(\sigma^2\):

\[ 1 - e^{-\phi^2 d^2} = 0.95 \]

Solve for \(e^{-\phi^2 d^2}\):

\[ e^{-\phi^2 d^2} = 0.05 \]

Take the natural logarithm:

\[ -\phi^2 d^2 = \ln(0.05) \]

Solve for \(d\):

\[ d^2 = -\frac{\ln(0.05)}{\phi^2} \]

\[ d = \frac{\sqrt{-\ln(0.05)}}{\phi} \]

Since \(-\ln(0.05) \approx 2.9957 \approx 3\):

\[ \boxed{d_{eff} \approx \frac{\sqrt{3}}{\phi}} \]


Model #5: Power Exponential

Definition

\[ \gamma(d) = \begin{cases} \tau^2 + \sigma^2 \left(1 - \exp(-|\phi d|^p)\right) & d > 0 \\ 0 & d = 0 \end{cases} \]

where \(0 < p \le 2\).

(a) Nugget, Sill, and Range

Parameter Value
Nugget \(\tau^2\)
Partial Sill \(\sigma^2\)
Total Sill \(\tau^2 + \sigma^2\)
Range Does not exist (asymptotic)
Effective Range \([-\ln(0.05)]^{1/p}/\phi \approx 3^{1/p}/\phi\)

(b) Covariance Function

For \(d = 0\): \[ C(0) = \tau^2 + \sigma^2 \]

For \(d > 0\): \[ C(d) = \sigma^2 e^{-|\phi d|^p} \]

Thus: \[ \boxed{ C(d) = \begin{cases} \tau^2 + \sigma^2 & d = 0 \\ \sigma^2 e^{-|\phi d|^p} & d > 0 \end{cases} } \]

Effective Range Derivation

We want \(\gamma(d) = 0.95(\tau^2 + \sigma^2)\):

\[ \tau^2 + \sigma^2(1 - e^{-|\phi d|^p}) = 0.95(\tau^2 + \sigma^2) \]

Subtract \(\tau^2\) from both sides:

\[ \sigma^2(1 - e^{-|\phi d|^p}) = 0.95\sigma^2 \]

Divide by \(\sigma^2\):

\[ 1 - e^{-|\phi d|^p} = 0.95 \]

Solve for \(e^{-|\phi d|^p}\):

\[ e^{-|\phi d|^p} = 0.05 \]

Take the natural logarithm:

\[ -|\phi d|^p = \ln(0.05) \]

Multiply by -1:

\[ |\phi d|^p = -\ln(0.05) \]

Take the \(p\)-th root:

\[ |\phi d| = [-\ln(0.05)]^{1/p} \]

Since \(d > 0\) and \(\phi > 0\):

\[ \boxed{d_{eff} = \frac{[-\ln(0.05)]^{1/p}}{\phi}} \]

Since \(-\ln(0.05) \approx 2.9957 \approx 3\):

\[ \boxed{d_{eff} \approx \frac{3^{1/p}}{\phi}} \]

Special Cases

\(p\) Value Model Effective Range
\(p = 1\) Exponential \(3/\phi\)
\(p = 2\) Gaussian \(\sqrt{3}/\phi\)

Model #6: Rational Quadratic

Definition

\[ \gamma(d) = \begin{cases} \tau^2 + \sigma^2 \left(\frac{d^2}{\phi + d^2}\right) & d > 0 \\ 0 & d = 0 \end{cases} \]

(a) Nugget, Sill, and Range

Parameter Value
Nugget \(\tau^2\)
Partial Sill \(\sigma^2\)
Total Sill \(\tau^2 + \sigma^2\)
Range Does not exist (asymptotic)
Effective Range \(\sqrt{19\phi}\)

(b) Covariance Function

For \(d = 0\): \[ C(0) = \tau^2 + \sigma^2 \]

For \(d > 0\): \[ C(d) = \sigma^2 \left(\frac{\phi}{\phi + d^2}\right) \]

Thus: \[ \boxed{ C(d) = \begin{cases} \tau^2 + \sigma^2 & d = 0 \\ \sigma^2 \left(\frac{\phi}{\phi + d^2}\right) & d > 0 \end{cases} } \]

Effective Range Derivation

We want \(\gamma(d) = 0.95(\tau^2 + \sigma^2)\):

\[ \tau^2 + \sigma^2\left(\frac{d^2}{\phi + d^2}\right) = 0.95(\tau^2 + \sigma^2) \]

Subtract \(\tau^2\) from both sides:

\[ \sigma^2\left(\frac{d^2}{\phi + d^2}\right) = 0.95\sigma^2 \]

Divide by \(\sigma^2\):

\[ \frac{d^2}{\phi + d^2} = 0.95 \]

Multiply both sides by \((\phi + d^2)\):

\[ d^2 = 0.95\phi + 0.95d^2 \]

Subtract \(0.95d^2\) from both sides:

\[ d^2 - 0.95d^2 = 0.95\phi \]

\[ 0.05d^2 = 0.95\phi \]

Solve for \(d^2\):

\[ d^2 = \frac{0.95}{0.05}\phi = 19\phi \]

Take the square root:

\[ \boxed{d_{eff} = \sqrt{19\phi}} \]


Model #7: Wave

Definition

\[ \gamma(d) = \begin{cases} \tau^2 + \sigma^2 \left(1 - \frac{\sin(\phi d)}{\phi d}\right) & d > 0 \\ 0 & d = 0 \end{cases} \]

(a) Nugget, Sill, and Range

Parameter Value
Nugget \(\tau^2\)
Partial Sill \(\sigma^2\)
Total Sill \(\tau^2 + \sigma^2\)
Range Does not exist (asymptotic with oscillations)
Effective Range Not well-defined due to oscillations

(b) Covariance Function

For \(d = 0\): \[ C(0) = \tau^2 + \sigma^2 \]

For \(d > 0\): \[ C(d) = \sigma^2 \frac{\sin(\phi d)}{\phi d} \]

Thus: \[ \boxed{ C(d) = \begin{cases} \tau^2 + \sigma^2 & d = 0 \\ \sigma^2 \frac{\sin(\phi d)}{\phi d} & d > 0 \end{cases} } \]

Effective Range Derivation

For the wave model, the effective range is not well-defined because:

  1. The model oscillates around the sill
  2. It reaches the sill at \(d = \pi/\phi\), then drops below, then rises again, etc.
  3. The function \(\frac{\sin(\phi d)}{\phi d}\) crosses zero at \(d = \pi/\phi, 2\pi/\phi, \dots\)

If we attempt to solve for 95% of the sill:

\[ \tau^2 + \sigma^2\left(1 - \frac{\sin(\phi d)}{\phi d}\right) = 0.95(\tau^2 + \sigma^2) \]

This simplifies to:

\[ \frac{\sin(\phi d)}{\phi d} = 0.05 \]

This transcendental equation has multiple solutions due to the oscillatory nature of the sine function. Therefore, there is no single effective range.

Key Distances

Distance \(C(d)\) \(\gamma(d)\)
\(d = \pi/\phi\) \(0\) \(\tau^2 + \sigma^2\)
\(d = 3\pi/(2\phi)\) Negative \(> \tau^2 + \sigma^2\)
\(d = 2\pi/\phi\) \(0\) \(\tau^2 + \sigma^2\)

Model #8: Power Law

Definition

\[ \gamma(d) = \begin{cases} \tau^2 + \sigma^2 d^\lambda & d > 0 \\ 0 & d = 0 \end{cases} \]

where \(0 < \lambda < 2\).

(a) Nugget, Sill, and Range

Parameter Value
Nugget \(\tau^2\)
Partial Sill Does not exist
Total Sill Does not exist (unbounded)
Range Does not exist (unbounded)
Effective Range Does not exist

(b) Covariance Function

\[ \boxed{\text{The covariance function } C(d) \text{ does NOT exist for the power law model.}} \]

The model is intrinsic (not stationary), so it has no valid covariance function.

Why Effective Range Doesn’t Exist

The power law model has no sill because:

\[ \lim_{d \to \infty} \gamma(d) = \lim_{d \to \infty} (\tau^2 + \sigma^2 d^\lambda) = \infty \]

Since there is no finite sill, we cannot define a distance at which the model reaches 95% of the sill.

Special Cases

\(\lambda\) Value Model Name
\(\lambda = 1\) Linear (Model #1)
\(0 < \lambda < 1\) Sub-linear
\(1 < \lambda < 2\) Super-linear

Model #9: Matérn

Definition

\[ \gamma(d) = \begin{cases} \tau^2 + \sigma^2 \left[1 - \frac{(2\sqrt{\nu}d\phi)^\nu}{2^{\nu-1}\Gamma(\nu)} K_\nu(2\sqrt{\nu}d\phi)\right] & d > 0 \\ 0 & d = 0 \end{cases} \]

where: - \(\nu > 0\) is the smoothness parameter - \(K_\nu(\cdot)\) is the modified Bessel function of the second kind - \(\Gamma(\nu)\) is the Gamma function

(a) Nugget, Sill, and Range

Parameter Value
Nugget \(\tau^2\)
Partial Sill \(\sigma^2\)
Total Sill \(\tau^2 + \sigma^2\)
Range Does not exist (asymptotic)
Effective Range Depends on \(\nu\); approximately \(\sqrt{8\nu}/\phi\)

(b) Covariance Function

For \(d = 0\): \[ C(0) = \tau^2 + \sigma^2 \]

For \(d > 0\): \[ C(d) = \sigma^2 \frac{(2\sqrt{\nu}d\phi)^\nu}{2^{\nu-1}\Gamma(\nu)} K_\nu(2\sqrt{\nu}d\phi) \]

Thus: \[ \boxed{ C(d) = \begin{cases} \tau^2 + \sigma^2 & d = 0 \\ \sigma^2 \frac{(2\sqrt{\nu}d\phi)^\nu}{2^{\nu-1}\Gamma(\nu)} K_\nu(2\sqrt{\nu}d\phi) & d > 0 \end{cases} } \]

Effective Range Derivation

For the Matérn model, the effective range is the distance where:

\[ \gamma(d) = 0.95(\tau^2 + \sigma^2) \]

This means:

\[ \frac{(2\sqrt{\nu}d\phi)^\nu}{2^{\nu-1}\Gamma(\nu)} K_\nu(2\sqrt{\nu}d\phi) = 0.05 \]

This equation cannot be solved analytically due to the Bessel function. Instead, we use numerical approximations.

Common Approximation

A widely used approximation for the Matérn effective range is:

\[ \boxed{d_{eff} \approx \frac{\sqrt{8\nu}}{\phi}} \]

This approximation works well for moderate to large \(\nu\).

More Accurate Values

For different values of \(\nu\), the effective range is:

\(\nu\) Effective Range (approx)
0.5 \(3/\phi\)
1.0 \(3.37/\phi\)
1.5 \(3.67/\phi\)
2.0 \(3.92/\phi\)
3.0 \(4.32/\phi\)
5.0 \(4.90/\phi\)
10.0 \(5.62/\phi\)
\(\infty\) \(\sqrt{3}/\phi\)

Special Cases

\(\nu\) Value Model Name Effective Range
\(\nu = 0.5\) Exponential \(3/\phi\)
\(\nu = 1\) Whittle \(3.37/\phi\)
\(\nu \to \infty\) Gaussian \(\sqrt{3}/\phi\)

Summary Table: Complete Models #1-9

Model Nugget Partial Sill Total Sill Range Effective Range Covariance \(C(d)\)
#1 Linear \(\tau^2\) None None None None Does not exist
#2 Spherical \(\tau^2\) \(\sigma^2\) \(\tau^2 + \sigma^2\) \(1/\phi\) \(1/\phi\) \(\sigma^2[1 - \frac{3}{2}\phi d + \frac{1}{2}(\phi d)^3]\)
#3 Exponential \(\tau^2\) \(\sigma^2\) \(\tau^2 + \sigma^2\) None \(3/\phi\) \(\sigma^2 e^{-\phi d}\)
#4 Gaussian \(\tau^2\) \(\sigma^2\) \(\tau^2 + \sigma^2\) None \(\sqrt{3}/\phi\) \(\sigma^2 e^{-\phi^2 d^2}\)
#5 Power Exponential \(\tau^2\) \(\sigma^2\) \(\tau^2 + \sigma^2\) None \([-\ln(0.05)]^{1/p}/\phi\) \(\sigma^2 e^{-|\phi d|^p}\)
#6 Rational Quadratic \(\tau^2\) \(\sigma^2\) \(\tau^2 + \sigma^2\) None \(\sqrt{19\phi}\) \(\sigma^2 \phi/(\phi + d^2)\)
#7 Wave \(\tau^2\) \(\sigma^2\) \(\tau^2 + \sigma^2\) None Not defined \(\sigma^2 \sin(\phi d)/(\phi d)\)
#8 Power Law \(\tau^2\) None None None None Does not exist
#9 Matérn \(\tau^2\) \(\sigma^2\) \(\tau^2 + \sigma^2\) None Depends on \(\nu\) \(\sigma^2 \frac{(2\sqrt{\nu}d\phi)^\nu}{2^{\nu-1}\Gamma(\nu)} K_\nu(2\sqrt{\nu}d\phi)\)

Classification Summary

Type Models Covariance Exists?
Stationary (bounded) #2 Spherical, #3 Exponential, #4 Gaussian, #5 Power Exponential, #6 Rational Quadratic, #7 Wave, #9 Matérn ✅ Yes
Intrinsic (unbounded) #1 Linear, #8 Power Law ❌ No

Summary of Effective Range Derivations

Model Equation to Solve Effective Range
Spherical Exact at sill \(1/\phi\)
Exponential \(1 - e^{-\phi d} = 0.95\) \(3/\phi\)
Gaussian \(1 - e^{-\phi^2 d^2} = 0.95\) \(\sqrt{3}/\phi\)
Power Exponential \(1 - e^{-|\phi d|^p} = 0.95\) \([-\ln(0.05)]^{1/p}/\phi\)
Rational Quadratic \(\frac{d^2}{\phi + d^2} = 0.95\) \(\sqrt{19\phi}\)
Wave \(\frac{\sin(\phi d)}{\phi d} = 0.05\) Not well-defined
Matérn \(\frac{(2\sqrt{\nu}d\phi)^\nu}{2^{\nu-1}\Gamma(\nu)} K_\nu(2\sqrt{\nu}d\phi) = 0.05\) Numerically determined
Linear No sill Does not exist
Power Law No sill Does not exist