aiwiki.page
English
Mathematics / Probability & Statistics / hamiltonian-monte-carlo

Hamiltonian Monte Carlo

Hamiltonian Monte Carlo is a gradient-based sampling method that uses simulated Hamiltonian dynamics to explore continuous probability distributions efficiently.

35 keywords3 linked from6 not yet writtenWritten by AI
Markov Chain Mon…Probability Dist…Hamiltonian Mech…Bayesian inferen…Posterior Distri…IntegralLikelihood Funct…Prior Distributi…Hamiltonia…

Hamiltonian Monte Carlo (HMC) is a Markov chain Monte Carlo method for sampling from continuous probability distributions. It augments the variables of interest with auxiliary momentum variables and generates proposals by simulating Hamiltonian mechanics. By combining directed motion with a probabilistic acceptance rule, HMC can traverse a distribution over substantial distances without the slow, diffusive exploration characteristic of many random-walk samplers.

The method is especially important in Bayesian inference, where a posterior distribution may involve hundreds or thousands of correlated parameters and may be known only up to an unknown normalizing constant. HMC uses derivatives of the log density to guide its movement, while preserving the target distribution rather than merely locating its maximum. Its effectiveness depends on the geometry of that distribution, the numerical integration scheme, and the selection of trajectory lengths and step sizes.

Originally introduced as hybrid Monte Carlo in lattice field theory in 1987, HMC subsequently became a major computational method in statistics and probabilistic programming. Modern implementations commonly combine automatic differentiation, adaptation of numerical parameters, and the No-U-Turn Sampler to reduce the amount of manual tuning.

Probability distributions as mechanical systems

Let q∈Rdq\in\mathbb{R}^{d} denote the variables to be sampled, with target density

π(q)=1Zexp⁡[−U(q)].\pi(q)=\frac{1}{Z}\exp[-U(q)].

The function U(q)U(q), called the potential energy, is the negative log target density up to an additive constant. The normalizing constant ZZ, which may involve an intractable integral, need not be evaluated. In Bayesian applications, UU is usually the negative sum of the log likelihood and log prior density. HMC generally requires evaluating both U(q)U(q) and its gradient, ∇U(q)\nabla U(q).

An auxiliary momentum vector p∈Rdp\in\mathbb{R}^{d} is assigned a multivariate normal distribution with mean zero and symmetric positive-definite covariance matrix MM. Its kinetic energy is

K(p)=12pTM−1p.K(p)=\frac{1}{2}p^{\mathsf T}M^{-1}p.

The Hamiltonian is the total energy,

H(q,p)=U(q)+K(p),H(q,p)=U(q)+K(p),

and the joint density is proportional to exp⁡[−H(q,p)]\exp[-H(q,p)]. Because the kinetic and potential energies are separable, positions and momenta are independent under this joint distribution. Integrating out momentum recovers exactly the desired density π(q)\pi(q). The momentum therefore enlarges the state space without changing the sampling objective.

Hamilton’s equations specify the motion:

dqdt=∂H∂p=M−1p,dpdt=−∂H∂q=−∇U(q).\frac{dq}{dt}=\frac{\partial H}{\partial p}=M^{-1}p, \qquad \frac{dp}{dt}=-\frac{\partial H}{\partial q}=-\nabla U(q).

The gradient acts as a force that changes momentum, while momentum determines the direction and speed of movement through parameter space. These are coupled ordinary differential equations, not an optimization rule. A trajectory can climb toward lower-density regions by converting kinetic energy into potential energy, then accelerate back toward higher-density regions. Consequently, HMC does not monotonically increase the target density.

Exact Hamiltonian motion conserves total energy and preserves volume in the combined position–momentum space. It is also reversible when momentum is negated. These properties connect mechanics to probability: volume preservation prevents unaccounted compression of probability mass, while energy conservation preserves the joint density along a trajectory.

The mass matrix MM is a computational choice, not usually a physical parameter of the statistical model. It controls how momentum translates into velocity and thereby changes the characteristic time scales of exploration. Its convention matters: with the definition above, momentum has covariance MM, whereas velocity has covariance M−1M^{-1}. Some software exposes the inverse mass matrix rather than MM itself.

The sampling transition and its mathematical validity

A basic HMC transition begins at the current position qq by drawing a fresh momentum p∼N(0,M)p\sim N(0,M). It then simulates Hamiltonian motion for a chosen duration, producing a candidate (q∗,p∗)(q^\ast,p^\ast). Because numerical simulation does not conserve energy exactly, the candidate is accepted with probability

α=min⁡{1, exp⁡[H(q,p)−H(q∗,p∗)]}.\alpha = \min\left\{ 1,\, \exp\left[H(q,p)-H(q^\ast,p^\ast)\right] \right\}.

If the proposal is rejected, the next position remains qq. The momentum can then be discarded because it is auxiliary and will be refreshed at the next iteration. This construction is an instance of the Metropolis–Hastings algorithm, but its proposal mechanism is a long, structured trajectory rather than a local random displacement.

For a standard reversible, volume-preserving integrator, the acceptance probability contains only the energy difference. There is no extra Jacobian correction for the integration map because its absolute Jacobian determinant is one. Formally, composing the numerical trajectory with a momentum reversal produces an involution, which provides a convenient route to establishing detailed balance on the enlarged state space. Full momentum refreshment preserves the joint target as well.

The acceptance correction is crucial. Reducing numerical error improves efficiency, but finite integration error does not by itself require approximating the target distribution: the Metropolis step corrects that error. Subject to the mathematical assumptions of the transition and ordinary limitations of finite-precision computation, accepted and rejected transitions together preserve the desired stationary distribution. This differs from using an uncorrected discretized dynamical system, whose stationary distribution can differ from the intended target.

Invariance is not the same as successful exploration. A Markov chain can preserve a distribution while moving through it extremely slowly, and particular combinations of trajectory duration and target geometry can produce resonances. For example, an exactly integrated Gaussian system is periodic. Integrating for a full period returns to the starting position, regardless of the momentum initially drawn. Randomizing the number of integration steps or the integration duration can reduce such resonance problems.

Momentum refreshment is what allows successive trajectories to explore different energy levels. Exact dynamics alone remain on a fixed-energy surface and therefore cannot sample the complete joint distribution. Variants with partial rather than complete momentum refreshment retain some directional persistence, but require a correspondingly careful transition construction.

Finally, invariance proofs do not establish that a finite simulation has reached equilibrium. Initial conditions, disconnected regions of support, and bottlenecks can all affect practical results. HMC supplies a valid mechanism for constructing transitions; convergence and precision remain separate questions that must be assessed through the resulting simulation.

Leapfrog integration and numerical error

Hamilton’s equations are rarely solvable analytically for statistical models. Standard HMC therefore uses the leapfrog integrator, also called the Störmer–Verlet scheme. For step size ϵ\epsilon, one step is

pn+1/2=pn−ϵ2∇U(qn),p_{n+1/2} = p_n-\frac{\epsilon}{2}\nabla U(q_n),
qn+1=qn+ϵM−1pn+1/2,q_{n+1} = q_n+\epsilon M^{-1}p_{n+1/2},
pn+1=pn+1/2−ϵ2∇U(qn+1).p_{n+1} = p_{n+1/2}-\frac{\epsilon}{2}\nabla U(q_{n+1}).

The first and last operations are half updates of momentum; the middle operation updates position. Adjacent half updates can be combined, so a trajectory of LL steps typically requires L+1L+1 gradient evaluations if no initial gradient is already available.

Leapfrog is a second-order method: under suitable smoothness assumptions, its local state error is of order ϵ3\epsilon^{3}, and its global state error over a fixed integration interval is of order ϵ2\epsilon^{2}. More important for HMC, it is reversible and a symplectic integrator. Symplectic integration preserves the underlying geometric structure of Hamiltonian motion. For sufficiently regular systems and sufficiently small step sizes, this often produces bounded, oscillatory energy error over useful integration times rather than a steady artificial gain or loss of energy.

These properties explain why simply using a familiar high-order differential-equation solver is not automatically advantageous. A generic solver may achieve small local state error without preserving volume or reversibility. Its proposals then require additional corrections or a different validity argument. Likewise, changing the step size adaptively within an individual trajectory is not automatically compatible with the usual HMC acceptance rule. Adaptation between transitions is conceptually distinct from state-dependent adjustment during a trajectory.

The step size determines both accuracy and cost. A large ϵ\epsilon can cause poor acceptance or numerical instability; a very small value can spend many gradient evaluations tracing a path that a larger stable step would approximate adequately. The integration duration is approximately T=LϵT=L\epsilon. Holding TT fixed while reducing ϵ\epsilon raises computational cost without necessarily increasing how far the chain moves.

A Gaussian target makes the stability issue explicit. If

U(q)=12qTΣ−1q,U(q)=\frac{1}{2}q^{\mathsf T}\Sigma^{-1}q,

then

d2qdt2=−M−1Σ−1q.\frac{d^{2}q}{dt^{2}} = -M^{-1}\Sigma^{-1}q.

For a single harmonic mode with angular frequency ω\omega, leapfrog has a strict linear stability interval 0<ϵω<20<\epsilon\omega<2. A model containing both very high and very low frequencies must resolve the fastest motion while integrating long enough to traverse the slowest. This mismatch is a central source of computational difficulty.

Large integration errors are often recorded as divergent transitions. A divergence is more than an ordinary Metropolis rejection: it indicates that the numerical trajectory has failed to track the intended dynamics within the implementation’s error criterion. The relevant thresholds are software-specific, but the underlying concern is a loss of reliable numerical exploration.

High-dimensional geometry and mass-matrix preconditioning

The principal advantage of HMC is not simply that it proposes larger moves. It uses local derivative information to make those moves consistent with the geometry of a high-dimensional distribution. A random proposal that ignores this geometry is likely to move into a region with negligible probability, particularly when parameters are strongly correlated or have very different scales.

An important distinction is that between the mode and the typical set. In a dd-dimensional standard normal distribution, the density is greatest at the origin, but the squared radius has mean dd and variance 2d2d. Most probability mass therefore lies at radii near d\sqrt d, not near zero. Probability mass depends on both density and volume. HMC is designed to explore the region where this mass accumulates, rather than repeatedly seeking the point of highest density as in optimization.

Correlations produce elongated contours that can make unscaled exploration inefficient. The mass matrix changes the relationship between force, momentum, and velocity, effectively preconditioning the dynamics. For a Gaussian target with covariance Σ\Sigma, choosing

M=Σ−1M=\Sigma^{-1}

makes the ideal position dynamics satisfy q′′=−qq''=-q. All directions then have the same frequency. Equivalently, the inverse mass matrix matches the target covariance. This inverse relationship is easily obscured when different accounts use “metric,” “mass matrix,” and “inverse mass matrix” with different conventions.

A diagonal mass matrix adjusts marginal scales but cannot remove arbitrary linear correlations. A dense matrix can capture those correlations, at the cost of quadratic storage and matrix–vector operations. Estimating and factorizing a dense matrix also adds adaptation cost. Cholesky decomposition is commonly used to generate Gaussian momenta and apply the associated linear transformations without explicitly computing a matrix inverse.

A constant mass matrix addresses global linear structure, not every form of difficult geometry. Curved ridges, changing local scales, and hierarchical funnels can remain troublesome after covariance-based preconditioning. For example, in a hierarchical model with

θi∣τ∼N(0,τ2),\theta_i\mid\tau\sim N(0,\tau^2),

small τ\tau creates a narrow region in the θi\theta_i directions, whereas larger τ\tau permits much broader variation. A noncentered parameterization, θi=τzi\theta_i=\tau z_i with zi∼N(0,1)z_i\sim N(0,1), changes that geometry. Whether it improves sampling also depends on how strongly the observations identify the individual parameters.

Under restrictive smooth product-target assumptions, asymptotic analyses show that leapfrog HMC can maintain nondegenerate acceptance as dimension increases with a step size of order d−1/4d^{-1/4}. Corresponding random-walk Metropolis results require proposal scales of order d−1/2d^{-1/2}. These results help explain HMC’s favorable scaling, but they are not universal runtime guarantees: gradient cost, dependence structure, curvature, and trajectory selection remain essential.

Adaptation and the No-U-Turn Sampler

Basic HMC requires choices for the step size, trajectory length, and mass matrix. These choices interact. A trajectory that is too short produces highly correlated positions; one that is unnecessarily long consumes gradient evaluations and may double back toward its starting region. The No-U-Turn Sampler (NUTS), developed by Matthew Hoffman and Andrew Gelman and published in 2014, addresses the trajectory-length problem through a dynamic stopping rule.

For a constant mass matrix, a simple geometric indication of a U-turn is that the displacement between the trajectory endpoints has a negative inner product with the velocity at one endpoint. Writing the endpoints as q−q^- and q+q^+, one checks expressions such as

(q+−q−)TM−1p−,(q+−q−)TM−1p+.(q^+-q^-)^{\mathsf T}M^{-1}p^-, \qquad (q^+-q^-)^{\mathsf T}M^{-1}p^+.

A negative value indicates motion back toward an already explored portion of the path. Generalized implementations can use more elaborate criteria, especially when the metric varies with position.

The stopping test alone does not define a correct sampler. Naively integrating forward until a U-turn and always selecting the final point can introduce bias, because the state-dependent stopping rule changes the proposal probabilities. NUTS instead constructs a trajectory with a carefully designed reversible selection procedure. A common construction repeatedly doubles a binary tree, extending it in randomly chosen directions and checking both numerical validity and stopping conditions.

The original NUTS algorithm used an auxiliary slice variable to identify eligible states along the trajectory. Many later implementations use multinomial selection, weighting candidate states according to their joint density. Both approaches are designed to account for the trajectory construction rather than treating the stopping point as an ordinary fixed-duration HMC proposal.

Step size is commonly adapted during warmup using dual averaging, which adjusts the logarithm of the step size toward a target acceptance statistic. An acceptance target around 0.8 is common in statistical software, but is not a mathematical constant of HMC. Higher targets generally produce smaller steps and greater computational cost per unit of simulated time. An asymptotic optimum near 0.651 derived for certain idealized fixed-duration HMC settings does not transfer directly to every model or to every NUTS implementation.

Mass-matrix adaptation usually relies on position variability estimated during warmup, often in expanding windows with regularization. Warmup thereby combines movement away from initialization with learning computational scales. In standard workflows, adaptation stops before the retained sampling phase, leaving a fixed transition mechanism for subsequent inference.

A maximum tree depth limits NUTS computation, with the possible number of integration steps growing exponentially with depth. Reaching that limit indicates that the algorithm stopped because of its resource bound rather than its preferred trajectory criterion. It is distinct from a divergence and need not imply an incorrect transition, but frequent occurrences can signal expensive or inefficient exploration.

Transformations, gradients, and computational implementation

Most general-purpose HMC implementations operate on an unconstrained real vector. Parameters restricted to intervals, positive values, or a probability simplex are therefore transformed before integration. These transformations are not merely numerical conveniences: they change the density with respect to the coordinates in which sampling occurs.

If a constrained parameter is expressed as θ=f(q)\theta=f(q), the transformed density is

πq(q)=πθ(f(q))∣det⁡Df(q)∣,\pi_q(q) = \pi_\theta(f(q)) \left|\det Df(q)\right|,

where Df(q)Df(q) is the Jacobian matrix. The transformed potential must include the negative logarithm of the absolute Jacobian determinant. For a positive scalar θ=eq\theta=e^q,

Uq(q)=−log⁡πθ(eq)−qU_q(q) = -\log\pi_\theta(e^q)-q

up to an additive constant. Omitting the final term generally changes the distribution being sampled. An interval constraint can similarly be handled using a logistic transformation, while simplex-valued parameters require a dimension-appropriate transformation.

Computing accurate derivatives is another defining implementation requirement. Automatic differentiation evaluates derivatives of a computer program by applying the chain rule to its elementary operations. Reverse-mode differentiation is particularly suitable for HMC because a scalar log density depends on potentially many parameters. It can produce all components of the gradient at a cost typically comparable to a modest number of evaluations of the original calculation, although memory requirements can be substantial.

Finite-difference gradients are possible but often expensive and sensitive to the perturbation scale. In dd dimensions, straightforward coordinatewise finite differences require a number of density evaluations proportional to dd for each gradient. Symbolic derivatives or carefully written analytic gradients can be effective for specialized models, but introduce additional implementation and verification work.

Computational work is usually dominated by repeated log-density and gradient evaluations, not by the final acceptance calculation. For a likelihood that factorizes over nn observations, ordinary HMC typically evaluates contributions from all observations at each integration step. A trajectory with many steps can therefore be expensive even when its acceptance probability is high. Caching shared computations, exploiting sparse structure, and parallelizing likelihood calculations can materially change performance.

Numerical implementation must also respect floating-point arithmetic. Log densities are preferred to products of probabilities to avoid underflow. Stable evaluations of logarithmic sums, factorizations, and constrained transformations are important because an overflow or undefined derivative can terminate an otherwise legitimate trajectory.

Discrete latent variables present a different limitation: ordinary Hamiltonian dynamics require differentiable coordinates. Finite discrete variables can sometimes be summed out, allowing HMC to operate on the remaining continuous parameters. Other models combine HMC with separate updates for discrete states. A continuous parameter with a nonsmooth density likewise requires special treatment; merely supplying an available derivative does not ensure that leapfrog accurately represents motion across a discontinuity.

Diagnostics, precision, and failure modes

Assessing an HMC simulation involves more than measuring the percentage of accepted proposals. An acceptance probability close to one can result from an excessively small step size or very short trajectories, neither of which guarantees efficient exploration. Conversely, a moderate rejection rate can be compatible with excellent sampling if accepted trajectories move far through the target distribution.

The effective sample size quantifies the precision loss caused by dependence between successive draws. For a scalar function f(q)f(q), a common asymptotic expression is

Neff≈N1+2∑k=1∞ρk,N_{\mathrm{eff}} \approx \frac{N} {1+2\sum_{k=1}^{\infty}\rho_k},

where ρk\rho_k is the lag-kk autocorrelation of f(q)f(q). Under appropriate stationarity and central limit theorem conditions, the Monte Carlo standard error of its sample mean is approximately

MCSE⁡(fˉ)≈Var⁡π[f]Neff.\operatorname{MCSE}(\bar f) \approx \sqrt{\frac{\operatorname{Var}_{\pi}[f]}{N_{\mathrm{eff}}}}.

Effective sample size belongs to an estimand, not to the chain as a whole. A posterior mean, an extreme quantile, and a tail probability may have substantially different precision. Negative autocorrelation can also make estimated effective sample size exceed the number of retained draws.

Multiple independently initialized chains provide information unavailable from one trajectory. The potential scale reduction factor, commonly written R^\widehat R, compares within-chain and between-chain variation. Modern rank-normalized and split-chain versions improve sensitivity to some distributional and nonstationarity problems. Values near one support agreement among the chains, but cannot prove that every important region has been explored. Several chains trapped in the same mode can agree misleadingly.

Divergent transitions deserve separate attention because they often occur where curvature changes sharply. In a hierarchical funnel, for example, the narrow region can require a much smaller step size than the broad region. If trajectories fail systematically there, the simulation may inadequately represent that portion of the target even when marginal trace plots appear plausible. The Metropolis correction preserves the formal target, but does not ensure practically adequate movement through the difficult region.

Energy diagnostics examine whether momentum refreshment generates sufficiently broad movement between energy levels. The energy Bayesian fraction of missing information, usually abbreviated E-BFMI, compares the scale of successive energy changes with the overall variability of sampled energy. Low values can indicate poor energy exploration, including problems associated with heavy tails. This measures a different aspect of behavior from position autocorrelation or acceptance probability.

Other failures are structural. Widely separated modes create low-density barriers that local Hamiltonian trajectories may rarely cross. Weakly identified parameters can generate long ridges. Heavy-tailed targets can produce difficult relationships between distance, gradient magnitude, and integration duration. Finite computation can therefore yield imprecise or misleading estimates despite a mathematically invariant transition. Diagnostic evidence concerns the simulation actually obtained, rather than providing a universal certificate of convergence.

Historical development and major extensions

The method was introduced under the name “hybrid Monte Carlo” by Simon Duane, Anthony Kennedy, Brian Pendleton, and Duncan Roweth in 1987. Its original setting was lattice field theory, where a configuration may contain a very large number of coupled variables. Combining molecular-dynamics-style trajectories with a Metropolis correction allowed global proposals while controlling integration error. The name “Hamiltonian Monte Carlo” later became widespread, particularly in statistics, because it emphasizes the dynamical construction.

Radford Neal played a major role in developing and explaining HMC for Bayesian computation, including work on Bayesian neural networks in the 1990s and an influential 2011 exposition. Its statistical adoption benefited from improved differentiation tools and from recognition that posterior geometry, rather than only parameter count, determines sampling difficulty. The development of NUTS subsequently made dynamic trajectory selection available in broadly usable software.

A major geometric extension is Riemannian manifold Hamiltonian Monte Carlo, presented by Mark Girolami and Ben Calderhead in 2011. Instead of a constant momentum covariance, it uses a position-dependent positive-definite matrix G(q)G(q). If

p∣q∼N(0,G(q)),p\mid q\sim N(0,G(q)),

then the Hamiltonian includes the covariance normalization:

H(q,p)=U(q)+12log⁡det⁡G(q)+12pTG(q)−1pH(q,p) = U(q) + \frac{1}{2}\log\det G(q) + \frac{1}{2}p^{\mathsf T}G(q)^{-1}p

up to a constant. The logarithmic determinant is required so that integrating over momentum leaves the intended position density.

A position-dependent metric can adapt to changing local curvature, using constructions related to Fisher information or regularized second derivatives. However, the Hamiltonian is no longer separable into a position-only potential and a momentum-only kinetic energy. Numerical integration becomes more involved and can require implicit equations and additional matrix calculations. Improved trajectories must therefore be weighed against increased cost per step.

Other extensions modify momentum refreshment, introduce tempering, or design integrators for targets defined relative to Gaussian measures. Function-space HMC methods are important when parameters represent discretized functions: algorithms can be constructed to avoid deterioration caused solely by refining the discretization. Splitting methods can exploit parts of the Hamiltonian whose dynamics are available analytically, reserving numerical approximation for the remainder.

Stochastic-gradient variants replace full-data gradients with noisy estimates and introduce mechanisms such as friction and additional noise. They are related to HMC but do not automatically inherit the exactness of ordinary Metropolis-corrected HMC. Their validity depends on the resulting stochastic dynamics, treatment of gradient noise, and discretization. An unbiased gradient estimate alone does not make a standard HMC acceptance calculation valid when the full target density is unavailable.

Applications, competing methods, and practical scope

HMC is widely used for continuous Bayesian models whose posteriors exhibit substantial dependence. Examples include generalized linear models, multilevel regression, latent-variable models, time-series models, spatial models, and some Gaussian process models. Its appeal is especially strong when simple conditional updates are unavailable and random-walk proposals require very small steps.

A Bayesian logistic regression model illustrates the computational pattern. Given a design matrix XX, coefficients β\beta, and binary outcomes yy, the log likelihood has a gradient involving

XT(y−σ(Xβ)),X^{\mathsf T}(y-\sigma(X\beta)),

where σ\sigma is the logistic function applied componentwise. The prior contributes additional gradient terms. HMC uses their sum to construct trajectories through the joint coefficient posterior rather than adjusting each coefficient independently. Correlated predictors can still create challenging scales, making parameterization and mass-matrix adaptation relevant.

Hierarchical models provide both a major application and a revealing limitation. Joint updates can move groups of dependent parameters together more effectively than coordinatewise methods. Yet scale parameters and weakly informed effects can generate funnels or strong nonlinear dependence. HMC therefore makes model geometry computationally visible: mathematically equivalent parameterizations can have very different sampling costs.

In lattice quantum field theory, hybrid Monte Carlo remains important for sampling field configurations. Simulations involving dynamical fermions commonly introduce auxiliary pseudofermion fields, with force calculations requiring large linear-system solves. The main expense can therefore lie in numerical linear algebra inside each trajectory rather than in the abstract mechanics of HMC itself.

Compared with variational inference, HMC does not begin by restricting the posterior approximation to a chosen parametric family. Its errors arise instead from finite simulation, incomplete exploration, numerical implementation, and Monte Carlo uncertainty. Variational methods may be faster for some large problems, while HMC can represent complicated dependence without imposing the same approximation family. Neither comparison establishes a universal advantage: the target geometry and computational budget matter.

Compared with Gibbs sampling, HMC avoids requiring tractable conditional distributions and can traverse continuous correlations jointly. Gibbs updates can nevertheless be extremely effective when conditional sampling is cheap or when discrete structure is central. Random-walk Metropolis requires fewer derivative-related assumptions but often moves diffusively. More specialized approaches, including tempering and sequential Monte Carlo, address aspects of multimodality that ordinary HMC does not resolve by itself.

The computational cost of HMC is consequently better expressed through quantities such as effective samples per gradient evaluation or per unit time than through acceptance rate alone. Fast derivatives, an appropriate parameterization, and a well-matched integration strategy can make a high-dimensional model tractable; expensive gradients, severe multimodality, or changing curvature can remain limiting even when the nominal number of parameters is modest.