How Convolution Erases Edges into a Bell Curve

Open with a single sharp triangular distribution and ask why convolving it with copies of itself turns its sharp tip into a smooth, rounded bell curve.

• A triangle has a sharp peak and sharp corners
• Convolution with an identical triangle changes the shape
• The puzzle: where does the sharpness go?

Let the learner commit to one guess about the shape after convolving a triangle with itself once before any explanation is revealed.

• Make one independent prediction
• Compare against the observed result later

A simulation where the learner picks a starting shape (triangle, rectangle, or exponential), chooses the number of self-convolutions from 1 to 20, and watches the density morph from angular to smooth bell-shaped.

• Pick triangle, rectangle, or exponential as the seed
• Adjust the number of self-convolutions n from 1 to 20
• Overlay the matching Gaussian for comparison
• Watch the sharp features flatten step by step

Show side-by-side bar histograms: one die produces a flat uniform block, two dice show a triangle, three dice show a smooth hump, and twenty dice look indistinguishable from a Gaussian bell curve.

• 1 die: flat block (uniform)
• 2 dice: triangle with a single peak
• 3 dice: rounded mound
• 20 dice: indistinguishable from a Gaussian

Explain that the density of X + Y, where X and Y are independent, is the convolution of their densities. Each convolution therefore adds one more independent copy, and the convolution operation is associative.

• f_{X+Y}(z) equals the convolution of f_X and f_Y
• Independence is what licenses this combination rule
• n-fold convolution equals the density of the sum of n independent copies
• Convolution is associative: order of adding does not matter

Walk through the intuition: each new independent variable averages with the running sum, so a single extreme value gets diluted by the typical values around it. Sharp corners correspond to rare extreme configurations, and many independent additions make those rare configurations vanish in the limit.

• Adding an independent variable averages the outcome
• Sharp peaks arise from a single extreme event; many variables suppress extremes
• The distribution's variance grows by sigma-squared each step
• Relative spread shrinks, so fine features smooth out

Test the rule on a deliberately nasty seed: a density with a single sharp spike and a long flat tail. Let the learner apply repeated convolution and confirm that even this pathological shape still converges to the same Gaussian family, only with different mean and variance.

• Start with a spiky-plus-flat density
• Apply many self-convolutions
• Confirm convergence to a Gaussian
• Note that mean and variance depend on the seed

Mark the limits: convolving Cauchy or other heavy-tailed distributions never settles into a Gaussian because they have no finite variance, and convolving dependent variables breaks the averaging argument.

• Cauchy distribution: self-convolution stays Cauchy, not Gaussian
• Infinite variance prevents the standard CLT from applying
• Strong dependence between summands blocks the averaging
• Finite mean and variance plus independence is the safe regime

Close the loop: convolution of independent distributions is summing independent variables; after enough such additions the Central Limit Theorem forces any finite-variance seed toward a Gaussian, erasing sharp edges.

• Convolution = sum of independent random variables
• Repeated convolution = many independent additions
• CLT guarantees convergence to a Gaussian
• Sharp edges vanish because extremes are averaged away