Monte Carlo Better than Quadrature
Table of Contents
Core Intuition
Q: Why Monte Carlo method is often the default method for high-dimensional integral approximation, instead of Quadrature methods? I got this question while reading Fearnhead et al. (2025, p. 6).
This comes directly from the following:
- Suppose we apply a cubature rule based on a grid of \(m+1\) equally spaced points in each dimension.
- The spacing of these points will be \(\delta = 1/m\)
- Then there will be \(n=(m+1)^d\) points in total
- If we have a cubature whose error decays in \(\delta^r\), for some power \(r\), then the error decays at a rate of \(m^{-r} \simeq n^{-r/d}\)
- Note for Monte Carlo method: its error decays at a rate of \(n^{1/2}\), by the central limit theorem
Mathematical Foundation
Q: Why cubature whose error decays in \(\delta^r\)? Taylor’s theorem with integral remainder.
Taylor’s Theorem with Remainder
For a function \(f\) with \(r\) continuous derivatives on an interval containing \(x_0\) and \(x\):
\[f(x) = \sum_{k=0}^{r-1} \frac{f^{(k)}(x_0)}{k!}(x-x_0)^k + R_r(x)\]where the remainder can be written as:
\[R_r(x) = \frac{1}{(r-1)!} \int_{x_0}^{x} (x-t)^{r-1} f^{(r)}(t) \, dt\]Proof of the Error Rate Bound in δʳ
Step 1: Take absolute value of the remainder:
\[|R_r(x)| = \left| \frac{1}{(r-1)!} \int_{x_0}^{x} (x-t)^{r-1} f^{(r)}(t) \, dt \right|\]Step 2: Apply triangle inequality:
\[|R_r(x)| \leq \frac{1}{(r-1)!} \int_{x_0}^{x} |x-t|^{r-1} |f^{(r)}(t)| \, dt\]| Step 3: Use $$ | f^{(r)}(t) | \leq | f^{(r)} | _{\infty}$$ (definition of sup norm): |
| Step 4: Evaluate the integral. Let $$h = | x - x_0 | $$: |
Step 5: Combine:
\[|R_r(x)| \leq \frac{|f^{(r)}|_{\infty}}{(r-1)!} \cdot \frac{h^r}{r} = \frac{|f^{(r)}|_{\infty}}{r!} h^r\]| Step 6: Set $$\delta = h = | x - x_0 | \(and\)C = \frac{1}{r!}$$: |
Derivation of Taylor’s Theorem with Remainder Formula
The key idea is to repeatedly apply integration by parts to express the error in terms of the highest derivative.
Starting Point: Fundamental Theorem of Calculus
\[f(x) = f(x_0) + \int_{x_0}^{x} f'(t) \, dt\]This is just: \(f(x) - f(x_0) = \int_{x_0}^x f'(t) \, dt\)
Step 1: Integration by Parts (once)
Apply integration by parts to \(\int_{x_0}^{x} f'(t) \, dt\) with:
- \(u = f'(t)\), so \(du = f''(t) \, dt\)
- \(dv = dt\), so \(v = t - x\)
So:
\[f(x) = f(x_0) + f'(x_0)(x-x_0) + \int_{x_0}^{x} (x-t) f''(t) \, dt\]Step 2: Integration by Parts (again)
Apply integration by parts to \(\int_{x_0}^{x} (x-t) f''(t) \, dt\) with:
- \(u = f''(t)\), so \(du = f'''(t) \, dt\)
- \(dv = (x-t) \, dt\), so \(v = -\frac{(x-t)^2}{2}\)
So:
\[f(x) = f(x_0) + f'(x_0)(x-x_0) + \frac{f''(x_0)}{2!}(x-x_0)^2 + \int_{x_0}^{x} \frac{(x-t)^2}{2!} f'''(t) \, dt\]Pattern Recognition
After k integration by parts, we get:
\[f(x) = \sum_{j=0}^{k-1} \frac{f^{(j)}(x_0)}{j!}(x-x_0)^j + \int_{x_0}^{x} \frac{(x-t)^{k-1}}{(k-1)!} f^{(k)}(t) \, dt\]Setting k = r:
\[\boxed{R_r(x) = \frac{1}{(r-1)!} \int_{x_0}^{x} (x-t)^{r-1} f^{(r)}(t) \, dt}\]Intuition
The integral form shows that the remainder is a weighted average of \(f^{(r)}(t)\) over the interval \([x_0, x]\), with weights \((x-t)^{r-1}\) that give more importance to points closer to \(x_0\).
Key Equation
For curvature: If we have a cubature whose error decays in \(\delta^r\). for some power \(r\), then the error decays at a rate of \(m^{-r} \simeq n^{-r/d}\).
\[f(x) = \sum_{k=0}^{r-1} \frac{f^{(k)}(x_0)}{k!}(x-x_0)^k + R_r(x),\] \[R_r(x) = \frac{1}{(r-1)!} \int_{x_0}^{x} (x-t)^{r-1} f^{(r)}(t) \, dt,\] \[|R_r(x)| \leq C \cdot \delta^r \cdot |f^{(r)}|_{\infty}\]As \(\delta = 1/m\), \(n=(m+1)^d\), so \(m^{-r} \simeq n^{-r/d}\).
For Monte Carlo: The error decays at a rate of \(n^{1/2}\), by the central limit theorem.
Component of
- Numerical Integration #todo
- Bayesian Computation #todo
- Scalable Machine Learning #todo
Insights
- Curse of dimensionality breaks quadrature: cubature error scales as \(n^{-r/d}\), so for fixed \(n\) and \(r\), doubling \(d\) halves the exponent. Monte Carlo error \(n^{-1/2}\) is completely dimension-free.
- Crossover dimension: Monte Carlo beats an order-\(r\) cubature rule when \(d > 2r\). For a second-order rule (\(r=2\)), the crossover is \(d > 4\); for most practical integrands in ML/Bayesian settings, \(d\) is in the hundreds or thousands.
- Higher-order rules don’t rescue quadrature at scale: increasing \(r\) shifts the crossover dimension, but collecting enough smooth derivatives of \(f\) in high \(d\) is itself intractable.
- Monte Carlo’s slow 1D rate is a feature in high \(d\): \(n^{-1/2}\) looks poor compared to \(n^{-r}\) in 1D, but its constant does not grow exponentially with \(d\).
Pitfalls
- Monte Carlo is not always preferable: in \(d \leq 2\) or \(d \leq 4\) (for \(r \geq 2\)), a well-chosen quadrature rule is faster to converge; defaulting to Monte Carlo in low dimensions wastes samples.
- Rate assumes i.i.d. samples: the \(n^{-1/2}\) CLT rate breaks down for correlated samples (e.g., MCMC chains), which introduce an effective-sample-size penalty.
- Ignoring variance: the CLT bound hides the integrand variance \(\sigma^2\); a high-variance integrand can make Monte Carlo impractical even when the dimension favors it. Variance reduction (importance sampling, control variates) is then essential.
- Confusing \(n\) (total points) with \(m\) (points per dimension): the grid has \(n = (m+1)^d\) points, so comparing quadrature and Monte Carlo at the same \(n\) (not the same \(m\)) is critical.
Connections
- Quasi-Monte Carlo #todo: low-discrepancy sequences beat i.i.d. draws, achieving \(O((\log n)^d/n)\) error for moderate \(d\).
- MCMC #todo: Markov chains replace i.i.d. draws when the target is known up to a constant; \(n^{-1/2}\) rate survives asymptotically via ergodic theory.
- Importance Sampling #todo: reweights a surrogate distribution to reduce variance; \(n^{-1/2}\) rate holds with a smaller constant.
- Curse of Dimensionality #todo: \(n^{-r/d}\) cubature rate is the canonical example of exponential cost growth with \(d\).
- Central Limit Theorem #todo: source of Monte Carlo’s \(n^{-1/2}\) guarantee for i.i.d. estimators.
References
- Fearnhead et al. (2025). Scalable Monte Carlo for Bayesian Learning. https://arxiv.org/abs/2407.12751