Markov Modulated Poisson Process (MMPP)

An MMPP is defined by two interacting layers: an unobservable continuous-time Markov chain that modulates the arrival rate, and an observable Poisson process whose instantaneous rate is a function of the current Markov state. This architecture captures the bursty, self-correlated nature of primary user (PU) traffic in cognitive radio networks, where idle periods cluster together and busy periods exhibit temporal persistence.

The underlying Markov chain is fully characterized by its infinitesimal generator matrix Q, where off-diagonal entries q_ij represent the transition rate from state i to state j, and diagonal entries q_ii = -Σ_{j≠i} q_ij ensure row sums equal zero. Each state i is associated with a Poisson arrival rate λ_i, forming the rate vector that maps Markov states to instantaneous traffic intensity.

MMPP excels at modeling overdispersed arrival processes where the variance-to-mean ratio exceeds unity—a hallmark of real spectrum traffic. By alternating between high-rate and low-rate states, MMPP generates clustered arrivals that match empirical PU ON/OFF patterns more accurately than a simple Poisson process. This makes it essential for spectrum occupancy prediction and proactive handoff algorithm design.

Fitting an MMPP to observed spectrum data typically employs the Expectation-Maximization (EM) algorithm or its forward-backward variant. The E-step computes the posterior probabilities of the hidden Markov state sequence given the observed inter-arrival times, while the M-step updates the Q matrix and rate vector λ to maximize the likelihood. This iterative procedure converges to a local maximum of the incomplete-data likelihood function.

The inter-arrival times in an MMPP follow a phase-type (PH) distribution, which generalizes the exponential distribution. The cumulative distribution function is given by F(t) = 1 - π exp((Q - Λ)t) 1, where π is the initial state probability vector, Λ = diag(λ_1, ..., λ_n), and 1 is a column vector of ones. This matrix-exponential form enables tractable analysis of channel holding time and spectrum availability windows.

When modeling a cognitive radio's channel access, the MMPP/M/1 queue captures both bursty PU arrivals and exponential service times. The queue's stationary distribution is computed via matrix-geometric methods, yielding key performance metrics such as forced termination probability and expected handoff delay. This framework directly informs the design of proactive spectrum handoff policies under realistic traffic assumptions.

A stochastic framework used to mathematically represent the temporal behavior of licensed spectrum users. The MMPP is a specific type of PU activity model that captures bursty traffic patterns by modulating the Poisson arrival rate with an underlying Markov chain. Common alternatives include simple ON/OFF models with exponentially distributed holding times, which fail to capture the self-similarity observed in real spectrum measurements.

A statistical model that infers unobservable channel occupancy states from observable signal measurements. While an MMPP generates arrivals based on a hidden state, an HMM is used for Bayesian inference to estimate that hidden state from noisy observations. In spectrum prediction, an HMM can decode the underlying Markov chain of an MMPP to forecast future channel states.

A probability distribution constructed from a Markov chain that models complex channel holding time and inter-arrival patterns. It generalizes the exponential distribution and is intimately related to the MMPP: the inter-arrival times in an MMPP follow a phase-type distribution. This allows the MMPP to capture non-exponential, heavy-tailed idle and busy periods observed in empirical spectrum data.

A matrix defining the probabilities of a frequency channel transitioning between idle and busy states. In an MMPP, this matrix governs the underlying Markov chain that modulates the Poisson arrival rate. Each row specifies the probability of moving from one hidden state to another, directly controlling the burstiness and correlation structure of primary user arrivals.

A measure of long-range dependence in a spectrum occupancy time series. A Hurst exponent greater than 0.5 indicates self-similar or bursty traffic, which simple Poisson models fail to capture. The MMPP addresses this by introducing a Markov-modulated rate, producing arrivals with tunable correlation structures that better match empirical Hurst parameter estimates from real spectrum measurements.

The likelihood that an ongoing secondary user transmission is prematurely dropped due to a collision with a returning primary user. This is a key performance metric for evaluating spectrum handoff strategies. An MMPP-based PU activity model provides more accurate estimates of forced termination probability than memoryless models, as it captures the temporal correlation in primary user arrivals that leads to clustered channel busy periods.