Â
Oberseminar of Christoph Thäle's working group
at the
Ruhr University Bochum, Germany
Â
gilles.bonnet@rub.deÂ
\(\widehat{X}\) Â ... hyperplane process.
The complement \(\mathbb{R}^d \setminus ( \cup_{H\in \widehat{X}} H )\)Â is a collection of (open) polytopes.
Their closures are the cells of the corresponding hyperplane tessellation \(X\)
Text
Note in [Schneider-Weil]:
Common assumptions:
Theorem: Intensity of stationary flat process [SW.4.4.3]
Let \(X\) be a stationary \(k\)-flat process in \(\mathbb{R}^d\) with intensity \(\gamma\). Then
Â
where \(\lambda_E\) is the \(k\)-dimensional Lebesgue measure in \(E\).
      \(\lambda\) is the Lebesgue measure in \(\mathbb{R}^d\).
\(\widehat{X}\) is a point process in the affine Grassmannian \(A(d,k)\)
\(\widehat{X}\) is a point process in the affine Grassmannian \(A(d,d-1)\)
o
Common assumptions:
\( \widehat{X} \) Â ... point process in the affine Grassmannian \(A(d,d-1)\)
\(\widehat{X}_m\) ... intersection process of order \(m\) of \(X\), \(m\in\{1,\ldots,d \}\).
\[ \widehat{X}_m := \{ H_1 \cap\cdots\cap H_m : H_i \in X \ , \ H_i \neq H_j  \ ,\ 1\leq i<j\leq m\} . \]
Theorem: Intensity of intersections [SW.4.4.8]
Let \(\widehat{X}\) be a stationary hyperplane process in \(\mathbb{R}^d\) with intensity \(\gamma\) and directional distribution \(\varphi\).
Then \(\widehat{X}_m\) is stationary and isotropic of intensity
\[ \frac{\gamma^k}{m!} \int_{S^{d-1}} \cdots \int_{S^{d-1}} \nabla_k(u_1,\ldots, u_m) \varphi (\mathrm{d} u_1) \ldots \varphi (\mathrm{d} u_m)\]
\(\widehat{X}\) Â ... point process in the affine Grassmannian \(A(2,1)\),
\(n_R := F_0^1(B_R) \) ... number of line intersecting the ball \(B_R\),
\(I_R := F_0^2(B_R) \) ... number of intersection points in \(B_R\),
\(L_R := F_1^1(B_R) \) ... sum of the lengths of segments in \(B_R\).
Theorem: CLT (planar, binomial) [Paroux '06]
Let \(\widehat{X}\) be a stationary and isotropic Poisson line process in \(\mathbb{R}^2\) and \(R>0\). As \(R\to\infty\),
Â
Â
Â
\(\widehat{X}\) Â ... Poisson hyperplane process
\(\widehat{X}_m\)Â ... intersection process of order \(m\) of \(\widehat{X}\),
\(F_i^m(W) :=\!\frac{1}{m!} \sum V_i (H_1 \cap \cdots \cap H_m \cap W) \) ...Intrinsic volume of \(\widehat{X}_m\) in \(W\!.\)
The proof relies on the so called Hoeffding’s decomposition of \(U\)-statistics.
In the isotropic setting, precise asymptotics for the variance are obtained when \(K\) is a ball, and ellipse or a rectangle.
In general \(\mathbb{V}F_k(rK) \) is of order \(r^{2d-1}\).
as \(r\to\infty\).
Theorem: CLT (count&volume)[Heinrich+2 Schmidt'06][Heinrich'09]
Let \(\widehat{X}\) be a stationary Poisson hyperplane process and \(W\) a fixed convex body. Assume that \(i\in\{0,d-k\}\), then
A different parametrization of the hyperplanes:
Intersections:
Result generalized to other expending window \(\rho K\) instead of \(B_r\).
Schulte and Reitzner show first a much more general CLT for \(U\)-statistics of Poisson processes. The proof relies on Malliavin calculus and the Wiener–Itô chaos expansion of \(U\)-statistics.
Theorem: CLT (all \(V_i\)'s + Wasserstein) [Schulte, Reitzner '13]
Let \(\widehat{X}\) be a Poisson hyperplane process of intensity measure \(\lambda \theta\), \(\lambda >0\) and \(W\) a fixed convex body. Assume that \(i\in\{0,\ldots,d-m\}\), then
Â
Recall:Â \(d_W(Y,Z) := \sup_{h\in\mathrm{Lip}(1)} |\mathbb{E}h(Y) - \mathbb{E}h(Z)|\).
\(\widehat{X}\) Â ... Poisson hyperplane process
\(\widehat{X}_m\)Â ... intersection process of order \(m\) of \(\widehat{X}\),
\(F_i^m(W) :=\!\frac{1}{m!} \sum V_i (H_1 \cap \cdots \cap H_m \cap W) \) ...Intrinsic volume of \(\widehat{X}_m\) in \(W\!.\)
\(\widehat{X}\) Â ... Poisson flat process in \(A(d,k)\)
\(\widehat{X}_m\)Â ... intersection process of order \(m\) of \(\widehat{X}\),
\(\psi \colon \mathcal{C}_o^d \subset \mathcal{C}^d \to \mathbb{R} \) satisfying mild assumptions,
\(F_\psi^m(W) :=\!\frac{1}{m!} \sum \psi (H_1 \cap \cdots \cap H_m \cap W) \) ... \(\psi\)-content of \(\widehat{X}_m\) in \(W\!,\)
\(\hat{F}_\psi^m (W) := t^{-m-\frac{1}{2}}(F_\psi^m(W)- \mathbb{E}F_\psi^m(W))\)
Theorem: Multivariate CLT [Last, Penrose, Schulte, Thäle '14]
1. Under reasonable assumption, one has
Â
where \(\xi\) is a Gaussian field whose covariance function can be written as a geometric integral.
2. If \(\psi(r \, \cdot) = r^\alpha \psi(\cdot)\) then for \(t=1\),
Â
\(\widehat{X}\) Â ... Poisson process of hyperplanes (totally geodesic subspaces)
\(\widehat{X}_m\)Â ... intersection process of order \(m\) of \(\widehat{X}\),
\(F^m(W) :=\!\frac{1}{m!} \sum \mathcal{H}^{d-m} (H_1 \cap \cdots \cap H_m \cap W) \)...Hausdorff measure of \(X_m\)
Methods:
Increase the intensity
Fix the window
Â
CLT for any \(d\geq 2\) and any \(m\)
Â
Increase the window
Fix the intensity
CLT for \(d\in\{2,3\}\) and any \(m\)
No CLT for \(d\geq 4\) and \(m=1\)
No CLT for \(d\geq 7\) and any \(m\)
\(\nLeftrightarrow\)
\(\widehat{X}_m\)... intersection process of Poisson/Binomial great subspheres
\(W\) ... window = spherical cap
\(F^m(W)\) ... analogous as before
Question:
Literature: I'm not sure about what is already know...
\(\widehat{X}\) Â ... Poisson hyperplane process.
The complement \(\mathbb{R}^d \setminus ( \cup_{H\in \widehat{X}} H )\)Â is a collection of (open) polytopes.
Their closures are the cells of the corresponding hyperplane tessellation \(X\)
For us, Tessellation = Marked Point Process where
If one chooses the most left point, the collection of centers is \(\widehat{X}_d\).
\(\Rightarrow\) the number of cells with center in \(W\) is \(F_0^d(W) \).
For a general notion of "tipicality" one needs to use Palm measure which definition can be found in:
Kallenberg (2007) "Random measure"
\(X\) Â ... stationary Poisson hyperplane tessellation
\(\gamma^{(d)}\) ... cell intensity := the intensity of the point process formed by the centers of all cells.
Your favorite book:
\(X\) Â ... Poisson hyperplane tessellation
\(x\in\mathbb{R}^d\!\!\) ... typical signal from a data set (\(\mathbb{R}^d\) or a Poisson process in \(\mathbb{R}^d\))
\((H(u_i,t_i))_{i\in I}\) ... stationary and isotropic hyperplane process
\(y_i = \mathrm{sign}(\langle u_i,x \rangle - t_i) \) ... measurements in \(\{-1,1\}\)
Quality of the compression:
\(x = o \in\mathbb{R}^d\) ... typical signal from a data set (\(\mathbb{R}^d\) or a Poisson process in \(\mathbb{R}^d\))
\((H(u_i,t_i))_{i\in I}\) ... stationary and isotropic hyperplane process
\(Z_o\)... the zero cell is the set of all points in \(\mathbb{R}^d\) with the same enconding as \(o\).
Quality of the compression:
\(x=o\in\mathbb{R}^d\!\!\) ... typical signal from a data set (in this slide: \(\mathbb{R}^d\)),
\((H(u_i,t_i))_{i\in I}\) ... Poisson hyperplane process: stationary, isotropic, intensity \(\gamma_d\).
The norm of the closest point to \(o\) in \(\mathbb{R}^d \setminus Z_o\) is \(r_\textrm{in}(Z_o)\),
where \(r_\textrm{in}(P) := \sup \{ r : B(o,r) \subset P \}\) for any polytope \(P\) containing the origin.
Proposition: Inradii of \(Z_o\) and \(Z\) [Baccelli, O'Reilly '19]
Assume that \(\gamma_d\to\infty\), then for any fixed \(r>0\)
  \(\lim_{d\to\infty} \mathbf{P}(r_{\mathrm{in}}(Z_o)>r) = 0\),
  \(\lim_{d\to\infty} \mathbf{P}(r_{\mathrm{in}}(Z)>r) = 0\).
Questions:
1) What if \(\gamma_d \to 0\) ? (super easy)
2) What if the data is a Poisson Process ?
\(x = o \in\mathbb{R}^d\!\!\) ... typical signal from a data set (in this slide: \(\mathbb{R}^d\)),
\((H(u_i,t_i))_{i\in I}\) ... Poisson hyperplane process: stationary, isotropic, intensity \(\gamma_d\).
The norm of the farthest point to \(x=0\) in \(\mathbb{R}^d \cap Z_o\) is \(r_\textrm{out}(Z_o)\),
where \(r_\textrm{out}(P) := \inf \{ r : P \subset B(o,r) \}\) for any polytope \(P\) containing the origin.
Theorem: Outradius of \(Z_o\) [Baccelli, O'Reilly '19]
Assume that \(\gamma_d \sim \rho d^\alpha\) and let \(R>0\).
Then there exist \(0<\rho_\ell < \frac{\sqrt{pi}}{R \sqrt{2}} < \rho_u\) such that
  \(\lim_{d\to\infty} \mathbf{P}(r_{\mathrm{out}}(Z_o)\leq d^{\frac{3}{2}-\alpha} R) = 0\), if \(\rho<\rho_\ell\),
  \(\lim_{d\to\infty} \mathbf{P}(r_{\mathrm{out}}(Z_o)\geq d^{\frac{3}{2}-\alpha} R) = 0\), if \(\rho>\rho_u\).
Method: Dualisation \(\to\) Poisson convex hull + Fine estimates
Questions:
1) Other regimes for \(\gamma_d\),
2) Typical cell ?
\(x = o \in\mathbb{R}^d\!\!\) ... typical signal from a data set,
\((H(u_i,t_i))_{i\in I}\) ... Poisson hyperplane process: stationary, isotropic, intensity \(\gamma_d\).
What if the data set is a stationary Poisson process \(\eta\) ?
Â
This question makes sense only if \(Z_o\) contains points of the process.
Â
We will come back to it in a few slides...
\(x = o \in\mathbb{R}^d\!\!\) ... typical signal from a data set (here: a stationary Poisson process \(\eta\) of intensity \(\lambda_d\)),
\((H(u_i,t_i))_{i\in I}\) ... Poisson hyperplane process: stationary, isotropic, intensity \(\gamma_d\),
Theorem: Separation of displaced point [Baccelli, O'Reilly '19]
Assume that \(\gamma_d \sim \rho d^\alpha\).
\(x = o \in\mathbb{R}^d\!\!\) ... typical signal from a data set (here: \(\mathbb{R}^d\)),
\((H(u_i,t_i))_{i\in I}\) ... Poisson hyperplane process: stationary, isotropic, intensity \(\gamma_d\).
Theorem: Expected volumes of \(Z_o\) and \(Z\)
\[ \mathbf{E} V_d(Z_o) = d! \kappa_d \left( \frac{d \kappa_d}{2 \gamma \kappa_{d-1}} \right)^d, \quad \mathbf{E} V_d(Z) = \frac{1}{\kappa_d} \left( \frac{d \kappa_d}{\gamma \kappa_{d-1}} \right)^d. \]
Corollary: Limits of \(\mathbf{E} Z_o\) and \(\mathbf{E} Z\) [Baccelli, O'Reilly '19]
Assume that \(\gamma_d = \rho d\), then
\[ \lim_{d\to\infty} \mathbf{E} V_d(Z_o) = \begin{cases} 0,& \rho<\frac{\pi}{\sqrt{e}}\\1,&\rho>\frac{\pi}{\sqrt{e}}\end{cases}, \quad \lim_{d\to\infty}\mathbf{E} V_d(Z) = \begin{cases} 0,& \rho<\frac{1}{\sqrt{e}}\\1,&\rho>\frac{1}{\sqrt{e}}.\end{cases} \]
\(x = o \in\mathbb{R}^d\!\!\) ... typical signal from a data set (here: a stationary Poisson process \(\eta\) of intensity \(\lambda_d\))
\((H(u_i,t_i))_{i\in I}\) ... Poisson hyperplane process: stationary, isotropic, intensity \(\gamma_d\).
Theorem: Number of Poisson points in \(Z_o\) [Baccelli, O'Reilly '19]
Assume \( \lambda_d = d^{d(\alpha-1)} e^{d \lambda}\) and \(\gamma_d \sim \rho d^\alpha\), with \(\lambda \in\mathbb{R}\), \(\alpha\in \mathbb{R}\) and \(\rho>0\).
Set \(\rho^* := \frac{e^{\lambda}\pi}{\sqrt{e}}\). We have
\[ \lim_{d\to\infty} \mathbf{E} |(\eta\cap Z_o)\setminus\{o\}| = \begin{cases} 0,& \rho>\rho^*\\\infty,&\rho<\rho^* .\end{cases} \]
Thus, for \(\rho>\rho^*\),
\[ \lim_{d\to\infty} (\eta\cap Z_o = o) = 1 .\]
\(x = o \in\mathbb{R}^d\!\!\) ... typical signal from a data set (here: a stationary Poisson process \(\eta\) of intensity \(\lambda_d\))
Theorem: Farthest Poisson point in \(Z_o\) [Baccelli, O'Reilly '19]
Assume \( \lambda_d = d^{d(\alpha-1)} e^{d \lambda}\) with \(\lambda \in\mathbb{R}\), and \(\gamma_d \sim \rho d^\alpha\) with \(\rho < \rho^* := \frac{e^{\lambda}\pi}{\sqrt{e}}\).
\(M_d := \max \{\|x\| : x \in \eta \cap Z_o\}\).
Theorem: Concentration in \(Z_o\) [O'Reilly'20]
Assumptions:
Then, there exist \(0<R_\ell < R_u := e^{-\rho} \sqrt{\pi e/2}\) such that
Â
Â
Â
Â
Method:
Similar result for Poisson Voronoi.
H1. For any of the model show concentration results
H2. Compute the variance in the spherical Binomial/Poisson setting
H... Suggestions ?
T1. Concentration for the number of cells in a window \(\Rightarrow\) H1.
T... Suggestions ?