Monte Carlo fails in a wonderfully democratic way.

It does not care much about dimension. Give it a 2-dimensional integral or a 200-dimensional integral and the basic error rate is still roughly (N^{-1/2}). That robustness is why it survives everywhere from particle transport to derivatives desks.

It is also why Monte Carlo can feel stubbornly slow.

Quasi-Monte Carlo asks a less democratic question: what if the problem is not really as high-dimensional as its notation says?

An Asian option with 32 monitoring dates can be written as a 32-dimensional Gaussian integral. A mortgage-backed security model can involve hundreds of random draws. But the payoff may be mostly sensitive to a few broad features: the terminal level, the average level, a handful of coarse path movements. If so, we should not spend our best coordinates on late, local wiggles.

This is the quiet job of the Brownian bridge:

same Brownian path law
different coordinates
different numerical integration problem

The stochastic process does not change. The order in which the simulator spends randomness does.

Even Points Need the Right Coordinates

Monte Carlo estimates an integral by sampling independent points:

\[\hat I_N = \frac{1}{N}\sum_{i=1}^N f(U_i), \qquad U_i \sim \mathrm{Uniform}([0,1]^d).\]

The error is random. Its root-mean-square scale is typically (O(N^{-1/2})). Caflisch’s Acta Numerica review emphasizes the virtue and the cost: the rate is dimension-independent, but slow.1

Low-discrepancy sequences replace independent random points with deterministic or randomized point sets that cover the unit cube more evenly. Halton’s 1960 paper is one of the classical constructions.2 Sobol sequences later became especially important in finance and engineering.3 The theoretical story often invokes discrepancy and variation: if the points are uniformly spread and the integrand is tame, the quadrature error can improve.

But this is not magic. A low-discrepancy point set covers early coordinates especially well. If the payoff depends equally and violently on every coordinate, there may be little advantage. If the payoff is mostly a low-dimensional smooth function in disguise, quasi-Monte Carlo can look almost unfair.

That word “disguise” is doing work.

The same Brownian path can be generated from uniforms in multiple equivalent ways. QMC sees the coordinate map, not only the distribution.

Build the Same Path in Another Order

For a geometric Brownian asset,

\[S_t = S_0 \exp\{(r-\sigma^2/2)t+\sigma W_t\},\]

an arithmetic Asian call has discounted payoff

\[e^{-rT}\max\left(\frac{1}{m}\sum_{j=1}^m S_{t_j}-K,0\right).\]

The usual sequential construction draws independent normals (Z_1,\ldots,Z_m) and sets

\[W_{t_j}=W_{t_{j-1}}+\sqrt{\Delta t}\,Z_j.\]

That is natural. It is also a very specific coordinate system. Coordinate 1 controls the first tiny increment. Coordinate 2 controls the second tiny increment. The terminal level is spread across all coordinates.

A Brownian bridge starts differently. Draw (W_T) first. Then draw the midpoint conditional on the endpoints. Then draw quarter points, then eighth points, and so on. Every filled point has the correct Gaussian conditional distribution:

\[W_t \mid W_a,W_b \sim \mathcal{N}\left( \frac{(b-t)W_a+(t-a)W_b}{b-a}, \frac{(t-a)(b-t)}{b-a} \right).\]

The final path distribution is identical. But the first coordinate now controls the terminal Brownian level. The next few coordinates control coarse path shape. Fine-scale noise waits its turn.

Caflisch, Morokoff, and Owen used this idea to explain why quasi-Monte Carlo could work surprisingly well in high-dimensional mortgage-backed-security valuation: the Brownian bridge can reduce effective dimension, making the important variance arrive in coordinates that low-discrepancy sequences cover well.4

A Tiny Asian-Option Experiment

The lab below prices an arithmetic Asian call under geometric Brownian motion. It compares three estimators:

  1. MC: pseudorandom normals and sequential Brownian increments.
  2. QMC sequential: shifted Halton points mapped to normals, then sequential increments.
  3. QMC bridge: the same shifted Halton construction, but coordinates are consumed by a Brownian bridge.

The reference price is a larger shifted-Halton Brownian-bridge estimate. The plotted errors are RMSE across deterministic replicates. This is a toy lab, not a production pricer: it uses Halton rather than a modern scrambled Sobol net, does not estimate Greeks, and ignores dividends, smile, rates curves, and discretization bias. The experiment is narrow on purpose so the coordinate effect has nowhere to hide.

MC Sequential QMC Brownian bridge QMC Reference

Deterministic toy experiment. QMC uses shifted Halton points; MC uses pseudorandom normals. The right-lower panel compares the mean estimate at max paths against a larger bridge-QMC reference. The variance-coordinate panel is computed analytically for the average Brownian level, not fitted from the option payoff.

Try raising Monitoring dates to 32. The formal dimension doubles from 16 to 32, but the bridge curve in the variance panel still jumps high immediately. For the average Brownian level, the terminal coordinate explains a large chunk of variance because most monitoring dates share it.

Now raise Volatility. The payoff becomes more nonlinear and more sensitive to tail paths. QMC still often helps, but the gap can narrow or wobble across seeds. This is a useful humility test: low discrepancy is not a universal accelerator. It likes smoothness and effective low dimension.

Lower Max paths. At very small budgets, all estimates can be lumpy. QMC’s advantage is not that every prefix is perfect. It is that carefully spread points tend to waste fewer samples on empty corners and accidental crowds.

Dimension Depends on the Interface

The dangerous simplification is:

QMC works in low dimension
MC works in high dimension

The better sentence is:

QMC works when the important variation is visible to early coordinates

That is why coordinate construction matters. A Brownian bridge does not reduce the number of monitoring dates. It changes which independent normal coordinate controls which path feature. For path averages, barriers, and other payoffs with large dependence on coarse path shape, this can dramatically change effective dimension. For payoffs driven by local discontinuities at many dates, it may do less.

L’Ecuyer’s finance review frames QMC as a variance-reduction family for integrals that are well approximated by sums of low-dimensional smooth functions, and notes that such integrals appear often in finance and risk.5 That phrasing is exactly right. The method is not a blessing sprinkled over every high-dimensional integral. It is a bet on structure.

Randomness Has Plumbing

There is an engineering lesson hiding in the numerical one.

When we say “simulate the model,” we often imply that randomness is a primitive object. Draw uniforms. Transform them. Run the model. Average the payoff.

But in a computer, randomness always enters through coordinates. Those coordinates are part of the estimator. Sequential Brownian increments, Brownian bridges, principal components, antithetic variables, control variates, common random numbers, and low-discrepancy sequences are not afterthoughts. They are plumbing choices that decide where variance is allowed to flow.

Monte Carlo is robust because it does not need the plumbing to be beautiful.

Quasi-Monte Carlo is fast when you can give it one.

Further Reading

  1. Russel E. Caflisch, “Monte Carlo and quasi-Monte Carlo methods,” Acta Numerica, 1998. Cambridge Core, PDF. ↩

  2. J. H. Halton, “On the efficiency of certain quasi-random sequences of points in evaluating multi-dimensional integrals,” Numerische Mathematik, 1960. Springer, EuDML. ↩

  3. I. M. Sobol, “On the distribution of points in a cube and the approximate evaluation of integrals,” USSR Computational Mathematics and Mathematical Physics, 1967. MathNet, DOI metadata. ↩

  4. Russel E. Caflisch, William Morokoff, and Art Owen, “Valuation of mortgage-backed securities using Brownian bridges to reduce effective dimension,” Journal of Computational Finance, 1997. Risk.net, PDF. ↩

  5. Pierre L’Ecuyer, “Quasi-Monte Carlo methods with applications in finance,” Finance and Stochastics, 2009. Springer. ↩