\documentclass{article}
\usepackage{fullpage, amsmath, amssymb}

\title{Extensions}

\begin{document}

\section{Gram matrix}

In the ICML paper, we showed that the graph could be recovered as long as the
expected gram matrix $\mathcal{G}$ of observations had its restricted
eigenvalues lower bounded away from $0$.  Recall that the gram matrix $G$ is
given by $X X^T$, where $X$ is an $n\times m$ matrix, where $n$ is the number of
measurements and $m$ is the number of nodes and if $X_i$ is a column vector
indicating the nodes infected at step $i$, then the rows of $X$ are given by
$X_i^T$. It follows that $$\mathcal{G} = \mathbb{E}[\sum_i X_i X_i^T]$$

Note that the indices of the sum  are themselves random variables!  Furthermore,
recall the definition of a restricted eigenvalue as $\|X_{S^c}\|_1 \leq 3
\|X_S\|_1$.

\subsection{Voter model}

In the voter model, $\mathbb{P}[X_j^{t+1} = 1 | X^t] = \sum_{i=1}^m A_{i,j}
X^t_i$, where $A$ is the weighted adjacency matrix of the graph. Furthermore,
the model runs indefinitely until time $T$, a hyperparameter of the model.
Therefore, $\mathbb{E}[X_{n+1} | X_n] = A^T X_n$ and $\mathcal{G} = \sum_i A^i
\mathbb{E}[X_0 X_0^T] (A^T)^i$.  In the case of the single-source model,
$\mathbb{E}[X_0X_0^T] = \frac{1}{m} I_m$ and it follows that $$\mathcal{G} =
\frac{1}{m} \sum_{i=1}^T A^i (A^T)^i$$

\section{Submodularity of Generalized Linear Cascades}

For which link functions is the resulting influence function submodular? We know
it to be true for the IC and LT model, what about the Logistic cascade model?

The answer is no. If we take the example of three nodes $(A, B, C)$ with the
possiblity to influence a remaining node D with respective edge weights: $a$,
$b$, and $c$, then we see that the following equality must be verified:
$$2\sigma(a+b) \geq \sigma(a+b+c) + \sigma(a)$$

We see that for $a=.5$, $,b=.5$, and $c=1$, the equality is violated.
Interestingly however is that if we let the scale parameter of the sigmoid go to
infinity, it is harder to violate the inequality. (TH) note that the LT model is
NOT submodular for every fixed value of thresholds, but only in expectation over
the random draw of these thresholds.

\section{Logistic regression}

Let's try to fit a logit model to the cascades. The variance of the parameters
is approximated by $\hat V (\hat \beta) = (\sum p_i(1-p_i)X_i X_i^T)^{-1}$

Let's have a look at the matrix we are taking the inverse of. If none of the
parents of node $i$ are active at step t, then the $t^{th}$ term is $0$.
Similarly, if the probability that node $i$ becomes active is $1$, then the term
cancels out. We can therefore write:$$A = \sum_{g_i \notin \{0,1\}} g_i (1-g_i)
X_i X_i^T$$

Furthermore, we notice that $x^TAx = \sum_{g_i \notin\{0,1\}} g_i(1-g_i)x^TX_i
X_i^Tx = \sum_{g_i \notin \{0,1\}} g_i(1-g_i)\|X_i^Tx\|_2^2$. This matrix-vector
product is zero as soon as I can find one node $a$ which can never be active
when the parents of my node are active (think line graph or circle with single
source law).

Suppose now that with high probability, each parent has been active at least
once. If I consider the union of all nodes which were active just before $X_i$
was, then by considering only this sub-space of parameters, I can avoid the
previous pitfall. But is it invertible nonetheless?

Let's consider a line and node $i \notin \{2, n-2\}$ along that line. The
parents cannot be infected at the same time, depending on whether the source $s
\geq i$ or $s \leq i$. The matrix $A$ is diagonal-by-block.

\end{document}