Monte Carlo Methods

Monte Carlo is a collection of methods that simulate a system (or experiment) by selecting random numbers from one or more probability distributions and approximately calculating the desired quantity through these simulations. The earliest applications were based on problems of nuclear transport and neutron diffusion; the name is attributed to von Neumann and Ulam. Unlike gambling concepts, the method computes values using random samples drawn from probability density functions that define the physical system.

Simulation is a methodology that transfers the cause-effect relationships of a real system to a computer and allows examination of system behavior under different conditions on a model; it does not guarantee an optimal solution but is suitable for comparing alternatives and testing worst-case scenarios. In early applications, the term “analog simulation” was also used; in current usage, Monte Carlo provides a distinct technical framework separate from model sampling and games of chance.

• Static/Dynamic: Models that represent single-time snapshots without a time dimension are static; those that model processes evolving over time are dynamic. Static simulation models are frequently referred to in the literature as Monte Carlo simulations.
• Deterministic/Stochastic: Models without randomness are deterministic; models in which events occur with associated probabilities are stochastic. Stochastic models represent inputs and processes using distributions and are often preferred for simulation because analytical solutions are frequently difficult or impossible to obtain.

For an event producing outcomes a ≤ x ≤ b, the probability density is p(x), and the cumulative distribution is defined as:

$Q(x)={\int }_a^{x}p(x^{\prime })d{x}^{\prime }$

The “fundamental principle” of Monte Carlo is the transformation $T=Q(x)$, which ensures that $T$ is uniformly distributed over $[0,1]$, and the inverse solution $X=Q^{-}1(T)$ yields samples from the target distribution (inverse transform method). This structure is the standard approach for generating samples from a desired distribution.

Numbers generated on computers are deterministic; however, when they exhibit statistical “random” properties, they are called pseudo-random numbers. A widely used generator is the mixed congruential method:

$xi=(a_{x}i-1+c)modm,ui=xi/m.$

If the sequence parameters $m,a,c$ are chosen appropriately, the period can be extended; each number is scaled to the range 0–1 for use in simulation.

For discrete distributions, the cumulative probability function is first constructed; the value corresponding to the interval into which the number $U$ falls is selected. For continuous distributions, the same principle applies: the value corresponding to the interval into which $U$ falls is selected.

For continuous distributions, $F_X(x)=P(X\le x)$ is computed; a value $U\sim U(0,1)$ is generated, and the equation $F_X(x)=U$ is solved for $x$ (inverse transform). Ready-made RNG functions (e.g., RANDU, RN(0), RND(($x$)) are frequently used in practice.

Even under identical parameter conditions, each simulation run produces a different set of estimates; this is the fundamental error inherent in Monte Carlo and is managed through sample size and methodological choices. Taking the average of independent repetitions reduces the standard error to the order of $1/\sqrt{n}$ under stationary conditions.

Input samples drawn by simple random sampling (SRS) represent the known input distribution with random fluctuations at the histogram/moment level; this increases the variance of simulation estimates. Techniques such as antithetic variables, stratification, common random numbers, and Latin hypercube sampling have therefore been developed to address this.

In this context, descriptive sampling is based on the principle of “deterministic data × random sequence”; input values are selected deterministically, and only their assignment is random. Its implementation and programming are simple; it consists of two main steps: (i) generation of a set of descriptive variables (sample value set), and (ii) matching these values with random assignment sets. A typical formulation for the negative exponential distribution is:

$x_i=F^{-}1((i-0.5)/n),i=1,\dots ,n$

In empirical comparisons, the variance of estimates obtained by descriptive sampling is significantly lower than that of SRS; in some applications, no additional variance reduction techniques are needed.

The algorithmic framework of descriptive sampling relies on using the set of sample values $X_d$ generated by inverse transform, without replacement in each run, and matching them with a random index at each step.

A simple thought experiment: When random numbers generated on [0,1] simulate the steps of a neutron through matter—for example, values 0.23, 0.71, and 0.62 indicate that after the third step the total distance exceeds 1, meaning the particle has left the medium—repeated simulations approximate quantities such as “number of steps to exit.”

MCNP is a general-purpose Monte Carlo code that solves three-dimensional, time-dependent, continuous-energy transport of neutrons, photons, and electrons; it covers fixed-source and subcritical problems. Its origins lie in work at Los Alamos; its development continues with strong emphasis on quality assurance, documentation, and improved physical models.

Monte Carlo and general simulation techniques are widely used in fields such as production/manufacturing, inventory and distribution, repair and maintenance systems, communication and transportation infrastructure, risk minimization in space missions, military defense, nuclear technology, measurement and evaluation, and socioeconomic systems; the method’s ability to represent risk is the reason for its preference in these areas.

Advantages: Flexibility; interpretability; incremental implementation capability; effectiveness in solving large and complex problems where classical methods are inapplicable; ability to model constraints and conditions difficult to analyze with other methods; rapid analysis of long-term outcomes; opportunity to test new policies and ideas without altering the system; high control over experimental conditions.

Disadvantages: Developing a good model requires time and cost; no guarantee of an optimal solution; each model is context-specific; ease of application may lead to neglect of analytical solutions; errors in modeling and analysis of results can produce incorrect conclusions.