Why is Gauss-Legendre quadrature so accurate with few points?

Gauss-Legendre quadrature optimally chooses both the evaluation points and weights so that an n-point rule integrates polynomials of degree up to 2n−1 exactly. Unlike equally-spaced rules (trapezoidal, Simpson), the Gauss points are non-uniform and positioned at roots of Legendre polynomials. A 5-point rule is exact for degree-9 polynomials and achieves much lower errors for smooth functions with the same number of function evaluations.

What is Romberg integration?

Romberg integration applies Richardson extrapolation repeatedly to the Trapezoidal rule results at successively halved step sizes T(h), T(h/2), T(h/4), … building a triangular table. Each extrapolation step increases the order of accuracy by 2, reaching O(h^(2k)) after k steps. For smooth integrands it achieves very high accuracy with relatively few function evaluations.

Is Simpson's rule always more accurate than the Trapezoidal rule?

Typically yes. Simpson's 1/3 rule has error O(h⁴) versus O(h²) for the Trapezoidal rule, so for the same number of subintervals it is generally more accurate. However, for non-smooth functions (discontinuities, sharp peaks) the advantage diminishes. For very few subintervals the function shape can occasionally make the Trapezoidal rule competitive.

How is numerical integration used in FEM (Finite Element Method)?

FEM uses numerical integration (Gauss quadrature) to compute element stiffness matrices and load vectors. Linear elements typically use 1-point Gauss integration; quadratic elements use higher-order Gauss points. Stresses are computed at Gauss points and then extrapolated to nodes for post-processing. The number and location of Gauss points directly affects both accuracy and computational cost.

Numerical Integration Methods Comparator (Trapezoidal, Simpson, Gauss) — Free Online Calculator

Integrand f(x)

True value: —

Bounds & Subintervals

Lower bound a

Upper bound b

Subintervals n

Results Comparison

Method	Value	Rel. Error

Results

—

Trapezoidal

—

Simpson 1/3

—

Gauss 5-point

—

Romberg

Func

Error vs n (log-log)

Error Orders

Trapezoidal: $E = O(h^2)$, Simpson: $E = O(h^4)$

Gauss $n$-point: $E = O(h^{2n})$, Romberg: $E = O(h^{2k})$ after $k$ extrapolations

$$\int_a^b f(x)\,dx \approx \frac{h}{2}[f(x_0)+2\sum_{i=1}^{n-1}f(x_i)+f(x_n)]$$

What is Numerical Integration?

🙋

What exactly is numerical integration, and why can't we just use the exact formulas from calculus?

🎓

Basically, it's a set of techniques to approximate the area under a curve when you can't find an exact antiderivative. In practice, many real-world functions from physics and engineering are too complex to integrate by hand. For instance, the stress-strain curve for a new composite material might only exist as experimental data points. Try moving the "Subintervals n" slider in the simulator above to see how more slices give a better approximation of the area.

🙋

Wait, really? So all these methods (Trapezoidal, Simpson's) are just different ways of slicing up the area? Why are some so much better than others?

🎓

Exactly! They differ in how they approximate the slice of area. The Trapezoidal rule uses straight lines, but Simpson's rule uses parabolas, which fits curves better. The real game-changer is the Gauss-Legendre method, which picks optimal sampling points. When you change the bounds `a` and `b` in the simulator, watch the error log-log chart. You'll see Simpson's error drop much faster than Trapezoidal's as you increase `n`.

🙋

Okay, and what's Romberg integration doing? It says it uses "extrapolation" – is it cheating to get a better answer?

🎓

Great question! It's not cheating; it's a clever trick. Romberg starts with the Trapezoidal rule at different `n` (like coarse and fine meshes), then mathematically cancels out the leading error terms. A common case is in high-precision CAE, where you need a very accurate result with minimal computational cost. In the simulator, Romberg often reaches machine precision with far fewer function evaluations, which is why it's so powerful.

Physical Model & Key Equations

The core idea is to approximate the definite integral of a function $ f(x) $ over $[a, b]$. The simplest method, the Composite Trapezoidal Rule, divides the interval into `n` subintervals of width $ h = (b-a)/n $.

$$ \int_a^b f(x)\,dx \approx \frac{h}{2}\left[f(x_0) + 2\sum_{i=1}^{n-1}f(x_i) + f(x_n)\right] $$

Here, $ x_i = a + i \cdot h $. The formula sums the area of `n` trapezoids. The error is proportional to $ h^2 $, meaning if you halve $ h $ (double `n`), the error reduces by roughly a factor of 4.

Gauss-Legendre Quadrature uses a fundamentally different approach. It doesn't use equally spaced points. Instead, it selects optimal points (Gauss points) and weights to exactly integrate polynomials of the highest possible degree.

$$ \int_{-1}^{1}f(x)\,dx \approx \sum_{i=1}^{n}w_i f(x_i) $$

For an `n`-point rule, it can integrate a polynomial of degree $ 2n-1 $ exactly. The points $ x_i $ are the roots of the Legendre polynomials, and $ w_i $ are carefully chosen weights. For a general interval $[a, b]$, a change of variable is applied. This is why its error order, $ O(h^{2n}) $, is so superior.

Frequently Asked Questions

In a log-log graph, when the error is proportional to h^k with respect to the number of divisions n, it is visualized as a straight line with slope k. This allows for quick comparison and confirmation of the convergence order of each method (e.g., O(h^2) for the trapezoidal rule, O(h^4) for Simpson's rule, etc.).

If the integrand has singularities (points where it becomes infinite) or rapid changes, convergence deteriorates with uniform division. Additionally, if the function is discontinuous, the accuracy of higher-order methods decreases. In such cases, consider using adaptive integration or variable transformation.

Generally, Simpson's rule is more accurate, but it cannot be applied if the integrand is not smooth or if the number of divisions is odd. Moreover, if the computational cost is similar, Gaussian quadrature is often even more accurate.

Gaussian quadrature optimizes the integration points and weights based on the roots of orthogonal polynomials, achieving a much higher algebraic precision (2n-1 order) than the trapezoidal rule or Simpson's rule with the same number of evaluation points. It is particularly effective for smooth functions suitable for polynomial approximation.

Real-world Applications

Finite Element Method (FEM) Stiffness Matrices: When solving structural problems, FEM breaks a part into small elements. Calculating the stiffness of each element requires integrating complex shape functions over its volume. Gauss quadrature is the industry standard here because it provides high accuracy with very few evaluation points, drastically speeding up simulations.

Fatigue Life (S-N Curve) Integration: Predicting how many cycles a component can endure under variable loads involves integrating under its S-N (stress-life) curve. Numerical integration allows engineers to handle real, irregular load spectra measured from sensors, not just idealized textbook functions.

CFD Pressure Force Calculation: In computational fluid dynamics, the total force on a wing or blade is found by integrating the pressure distribution over its surface. Accurate and fast numerical integration is critical for computing lift and drag forces during aerodynamic optimization.

Probability & Statistics (PDF Normalization): For a continuous random variable, the total probability must be 1, which means the area under its Probability Density Function (PDF) must integrate to 1. Numerical integration is used to normalize experimental data and calculate probabilities for complex, non-standard distributions.

Common Misconceptions and Points to Note

When you start using this tool, there are a few common pitfalls to watch out for. First is the misconception that "higher-order methods are always the correct answer". While Simpson's rule and Gaussian quadrature are indeed efficient, if the function is not smooth (e.g., has non-differentiable points, sharp variations), you might not only fail to achieve the expected accuracy but could even get worse results than with the trapezoidal rule. For example, trying it on a function with a sharp peak will clearly show you this difference.

Next is the pitfall in setting the number of segments 'n'. Simpson's rule won't work correctly unless n is even, and for Romberg integration, you won't extract its maximum performance unless you increase n in powers of two (2, 4, 8, 16…). If you try Simpson's rule with "n=5" in the tool, the calculation will run, but an approximation process is actually happening internally, so it's not a true performance evaluation.

Finally, understand that "zero error" is an illusion. Log-log graphs make the error appear to decrease linearly, but calculations always involve rounding errors. If you recklessly increase the number of segments too much (e.g., n=1,000,000), the increased number of calculation points causes rounding errors to accumulate, leading to "computational saturation" where accuracy actually degrades. If you set an extremely large n in the tool, you should observe that the graph line doesn't keep falling but rather flattens out or even starts to rise.

Numerical Integration Methods Comparator