Search

New and existing methods for generating, and especially detecting, deepfakes are investigated and compared on the simple problem of authenticating coin flip data. Importantly, an alternative approach to deepfake generation and detection, which uses a Markov Observation Model (MOM) is introduced and compared on detection ability to the traditional Generative Adversarial Network (GAN) approach as well as Support Vector Machine (SVM), Branching Particle Filtering (BPF) and human alternatives. MOM was also compared on generative and discrimination ability to GAN, filtering and humans (as SVM does not have generative ability). Humans are shown to perform the worst, followed in order by GAN, SVM, BPF and MOM, which was the best at the detection of deepfakes. Unsurprisingly, the order was maintained on the generation problem with removal of SVM as it does not have generation ability.

A continuous-time Markov chain rate change formula for simulation, model selection, filtering and theory is proven. It is used to develop Markov chain importance sampling, rejection sampling, branching particle filtering algorithms and filtering equations akin to the Duncan-Mortensen-Zakai equation and the Fujisaki-Kallianpur-Kunita equation but for Markov signals with general continuous-time Markov chain observations. A direct method of solving these filtering equations is given that, for example, applies to trend, volatility and/or parameter estimation in financial models given tick-by-tick market data. All the results also apply to continuous-time Hidden Markov Models (CTHMM), which have become important in applications like disease progression tracking, as special cases and the corresponding CTHMM results are stated as corollaries.

A topological neural network (TNN), which takes data from a Tychonoff topological space instead of the usual finite dimensional space, is introduced. As a consequence, a distributional neural network (DNN) that takes Borel measures as data is also introduced. Combined these new neural networks facilitate things like recognizing long range dependence, heavy tails and other properties in stochastic process paths or like acting on belief states produced by particle filtering or hidden Markov model algorithms. The veracity of the TNN and DNN are then established herein by a strong universal approximation theorem for Tychonoff spaces and its corollary for spaces of measures. These theorems show that neural networks can arbitrarily approximate uniformly continuous functions (with respect to the sup metric) associated with a unique uniformity. We also provide some discussion showing that neural networks on positive-finite measures are a generalization of the recent deep learning notion of deep sets.

Training a neural network (NN) typically relies on some type of curve-following method, such as gradient descent (GD) (and stochastic gradient descent (SGD)), ADADELTA, ADAM or limited memory algorithms. Convergence for these algorithms usually relies on having access to a large quantity of observations in order to achieve a high level of accuracy and, with certain classes of functions, these algorithms could take multiple epochs of data points to catch on. Herein, a different technique with the potential of achieving dramatically better speeds of convergence, especially for shallow networks, is explored: it does not curve-follow but rather relies on 'decoupling' hidden layers and on updating their weighted connections through bootstrapping, resampling and linear regression. By utilizing resampled observations, the convergence of this process is empirically shown to be remarkably fast and to require a lower amount of data points: in particular, our experiments show that one needs a fraction of the observations that are required with traditional neural network training methods to approximate various classes of functions., Comment: 16 pages

Herein, the Hidden Markov Model is expanded to allow for Markov chain observations. In particular, the observations are assumed to be a Markov chain whose one step transition probabilities depend upon the hidden Markov chain. An Expectation-Maximization analog to the Baum-Welch algorithm is developed for this more general model to estimate the transition probabilities for both the hidden state and for the observations as well as to estimate the probabilities for the initial joint hidden-state-observation distribution. A believe state or filter recursion to track the hidden state then arises from the calculations of this Expectation-Maximization algorithm. A dynamic programming analog to the Viterbi algorithm is also developed to estimate the most likely sequence of hidden states given the sequence of observations., Comment: Added Reference. Corrected Typos

The mean-field stochastic partial differential equation (SPDE) corresponding to a mean-field super-Brownian motion (sBm) is obtained and studied. In this mean-field sBm, the branching-particle lifetime is allowed to depend upon the probability distribution of the sBm itself, producing an SPDE whose space-time white noise coefficient has, in addition to the typical sBm square root, an extra factor that is a function of the probability law of the density of the mean-field sBm. This novel mean-field SPDE is thus motivated by population models where things like overcrowding and isolation can affect growth. A two step approximation method is employed to show existence for this SPDE under general conditions. Then, mild moment conditions are imposed to get uniqueness. Finally, smoothness of the SPDE solution is established under a further simplifying condition., Comment: 55 pages

In this paper, the strong existence and uniqueness for a degenerate finite system of quantile-dependent McKean-Vlasov stochastic differential equations are obtained under a weak H\"{o}rmander condition. The approach relies on the apriori bounds for the density of the solution to time inhomogeneous diffusions. The time inhomogeneous Feynman-Fac formula is used to construct a contraction map for this degenerate system.

Herein, a methodology is developed to replicate functions, measures and stochastic processes onto a compact metric space. Many results are easily established for the replica objects and then transferred back to the original ones. Two problems are solved within to demonstrate the method: (1) Finite-dimensional convergence for processes living on general topological spaces. (2) New tightness and relative compactness criteria are given for the Skorokhod space $D(\mathbf{R}^{+};E)$ with $E$ being a general Tychonoff space. The methods herein are also used in companion papers to establish the: (3) existence of, uniqueness of and convergence to martingale problem solutions, (4) classical Fujisaki-Kallianpur-Kunita and Duncan-Mortensen-Zakai filtering equations and stationary filters, (5) finite-dimensional convergence to stationary signal-filter pairs, (6) invariant measures of Markov processes, and (7) Ray-Knight theory all in general settings., Comment: 203 pages, 5 figures

An explicit weak solution for the 3/2 stochastic volatility model is obtained and used to develop a simulation algorithm for option pricing purposes. The 3/2 model is a non-affine stochastic volatility model whose variance process is the inverse of a CIR process. This property is exploited here to obtain an explicit weak solution, similarly to Kouritzin (2018). A simulation algorithm based on this solution is proposed and tested via numerical examples. The performance of the resulting pricing algorithm is comparable to that of other popular simulation algorithms., Comment: 21 pages, 2 figures

Marcinkiewicz strong law of large numbers, ${n^{-\frac1p}}\sum_{k=1}^{n} (d_{k}- d)\rightarrow 0\ $ almost surely with $p\in(1,2)$, are developed for products $d_k=\prod_{r=1}^s x_k^{(r)}$, where the $x_k^{(r)} = \sum_{l=-\infty}^{\infty}c_{k-l}^{(r)}\xi_l^{(r)}$ are two-sided linear processes with coefficients $\{c_l^{(r)}\}_{l\in \mathbb{Z}}$ and i.i.d. zero-mean innovations $\{\xi_l^{(r)}\}_{l\in \mathbb{Z}}$. The decay of the coefficients $c_l^{(r)}$ as $|l|\to\infty$, can be slow enough for $\{x_k^{(r)}\}$ to have long memory while $\{d_k\}$ can have heavy tails. The long-range dependence and heavy tails for $\{d_k\}$ are handled simultaneously and a decoupling property shows the convergence rate is dictated by the worst of long-range dependence and heavy tails, but not their combination. The Marcinkiewicz strong law of large numbers is also extended to the multivariate linear process case., Comment: 38 pages

The use of sequential Monte Carlo within simulation for path-dependent option pricing is proposed and evaluated. Recently, it was shown that explicit solutions and importance sampling are valuable for efficient simulation of spot price and volatility, especially for purposes of path-dependent option pricing. The resulting simulation algorithm is an analog to the weighted particle filtering algorithm that might be improved by resampling or branching. Indeed, some branching algorithms are shown herein to improve pricing performance substantially while some resampling algorithms are shown to be less suitable in certain cases. A historical property is given and explained as the distinguishing feature between the sequential Monte Carlo algorithms that work on path-dependent option pricing and those that do not. In particular, it is recommended to use the so-called effective particle branching algorithm within importance-sampling Monte Carlo methods for path-dependent option pricing. All recommendations are based upon numeric comparison of option pricing problems in the Heston model.

In a market with stochastic volatility and jumps, we consider a VIX-linked fee structure for variable annuity contracts with guaranteed minimum withdrawal benefits (GMWB). Our goal is to assess the effectiveness of the VIX-linked fee structure in decreasing the sensitivity of the insurer's liability to volatility risk. Since the GMWB payoff is highly path-dependent, it is particularly sensitive to volatility risk, and can also be challenging to price, especially in the presence of the VIX-linked fee. In this paper, we present an explicit weak solution for the value of the VA account and use it in Monte Carlo simulations to value the GMWB guarantee. Numerical examples are provided to analyze the impact of the VIX-linked fee on the sensitivity of the liability to changes in market volatility.

A class of Fleming-Viot processes with decaying sampling rates and $\alpha$-stable motions that correspond to distributions with growing populations are introduced and analyzed. Almost sure long-time scaling limits for these processes are developed, addressing the question of long-time population distribution for growing populations. Asymptotics in higher orders are investigated. Convergence of particle location occupation and inhabitation time processes are also addressed and related by way of the historical process. The basic results and techniques allow general Feller motion/mutation and may apply to other measure-valued Markov processes.

A general class of non-Markov, supercritical Gaussian branching particle systems is introduced and its long-time asymptotics is studied. Both weak and strong laws of large numbers are developed with the limit object being characterized in terms of particle motion/mutation. Long memory processes, like branching fractional Brownian motion and fractional Ornstein-Uhlenbeck processes with large Hurst parameters, as well as rough processes, like fractional processes with with smaller Hurst parameter, are included as important examples. General branching with second moments is allowed and moment measure techniques are utilized.

Strong solutions of p-dimensional stochastic differential equations that can be represented locally in explicit simulation form are considered. The following three-way equivalence is established: 1) There exists such a representation from all starting points, 2) the representation pair satisfies a set differential equations, and 3) the stochastic differential equation coefficients satisfy commutation relations. Next, construction theorems, based on a diffeomorphism between the original equation solutions and the strong solutions to a simpler Ito integral equation, with a possible deterministic component, are given. Finally, motivating examples are provided and reference to its importance in filtering and option pricing is given., Comment: 27 pages

New simulation approaches to evaluating path-dependent options without matrix inversion issues nor Euler bias are evaluated. They employ three main contributions: Stochastic approximation replaces regression in the LSM algorithm; Explicit weak solutions to stochastic differential equations are developed and applied to Heston model simulation; and Importance sampling expands these explicit solutions. The approach complements Heston (1993) and Broadie and Kaya (2006) by handling the case of path-dependence in the option's execution strategy. Numeric comparison against standard Monte Carlo methods demonstrate up to two orders of magnitude speed improvement. The general ideas will extend beyond the important Heston setting., Comment: 42 pages

The Marcinkiewicz Strong Law, $\displaystyle\lim_{n\to\infty}\frac{1}{n^{\frac1p}}\sum_{k=1}^n (D_{k}- D)=0$ a.s. with $p\in(1,2)$, is studied for outer products $D_k=X_k\overline{X}_k^T$, where $\{X_k\},\{\overline{X}_k\}$ are both two-sided (multivariate) linear processes ( with coefficient matrices $(C_l), (\overline{C}_l)$ and i.i.d.\ zero-mean innovations $\{\Xi\}$, $\{\overline{\Xi}\}$). Matrix sequences $C_l$ and $\overline{C}_l$ can decay slowly enough (as $|l|\to\infty$) that $\{X_k,\overline{X}_k\}$ have long-range dependence while $\{D_k\}$ can have heavy tails. In particular, the heavy-tail and long-range-dependence phenomena for $\{D_k\}$ are handled simultaneously and a new decoupling property is proved that shows the convergence rate is determined by the worst of the heavy-tails or the long-range dependence, but not the combination. The main result is applied to obtain Marcinkiewicz Strong Law of Large Numbers for stochastic approximation, non-linear functions forms and autocovariances., Comment: 23 pages

Almost sure convergence rates for linear algorithms $h_{k+1} = h_k +\frac{1}{k^\chi} (b_k-A_kh_k)$ are studied, where $\chi\in(0,1)$, $\{A_{k}\}_{k=1}^\infty$ are symmetric, positive semidefinite random matrices and $\{b_{k}\}_{k=1}^\infty$ are random vectors. It is shown that $|h_n- A^{-1}b|=o(n^{-\gamma})$ a.s. for the $\gamma\in[0,\chi)$, positive definite $A$ and vector $b$ such that $\frac{1}{n^{\chi-\gamma}}\sum\limits_{k=1}^n (A_{k}- A)\to 0$ and $\frac{1}{n^{\chi-\gamma}}\sum\limits_{k=1}^n (b_k-b)\to 0$ a.s. When $\chi-\gamma\in\left(\frac12,1\right)$, these assumptions are implied by the Marcinkiewicz strong law of large numbers, which allows the $\{A_k\}$ and $\{b_k\}$ to have heavy-tails, long-range dependence or both. Finally, corroborating experimental outcomes and decreasing-gain design considerations are provided., Comment: 27 pages

Herein, we introduce and study a new class of discrete random fields designed for quick simulation and covariance inference under inhomogeneous condition. Simulation of these correlated fields can be done in a single pass instead of relying on multi-pass convergent methods like the Gibbs Sampler or other Markov Chain Monte Carlo methods. The fields are constructed directly from specified marginal probability mass functions and covariances between nearby sites. The proposition on which the construction is based establishes when and how it is possible to simplify the conditional probabilities of each site given the other sites in a manner that makes simulation quite feasible yet maintains desired marginal probabilities and covariances between sites. Special cases of these correlated fields have been deployed successfully in data authentication, object detection and image generation. The limitations that must be imposed on the covariances and marginal probabilities in order for the algorithm to work are studied. What's more, a necessary and sufficient condition that guarantees the permutation property of correlated random fields are investigated. In particular, Markov random fields as a subclass of correlated random fields are derived by a general and natural condition. Consequently, a direct and flexible single pass algorithm for simulating Markov random fields follows.

Herein, we analyze an efficient branching particle method for asymptotic solutions to a class of continuous-discrete filtering problems. Suppose that $t\to X_t$ is a Markov process and we wish to calculate the measure-valued process $t\to\mu_t(\cdot)\doteq P\{X_t\in \cdot|\sigma\{Y_{t_k}, t_k\leq t\}\}$, where $t_k=k\epsilon$ and $Y_{t_k}$ is a distorted, corrupted, partial observation of $X_{t_k}$. Then, one constructs a particle system with observation-dependent branching and $n$ initial particles whose empirical measure at time $t$, $\mu_t^n$, closely approximates $\mu_t$. Each particle evolves independently of the other particles according to the law of the signal between observation times $t_k$, and branches with small probability at an observation time. For filtering problems where $\epsilon$ is very small, using the algorithm considered in this paper requires far fewer computations than other algorithms that branch or interact all particles regardless of the value of $\epsilon$. We analyze the algorithm on L\'{e}vy-stable signals and give rates of convergence for $E^{1/2}\{\|\mu^n_t-\mu_t\|_{\gamma}^2\}$, where $\Vert\cdot\Vert_{\gamma}$ is a Sobolev norm, as well as related convergence results., Comment: Published at http://dx.doi.org/10.1214/105051605000000539 in the Annals of Applied Probability (http://www.imstat.org/aap/) by the Institute of Mathematical Statistics (http://www.imstat.org)