Sites lie along a hidden gradient t (a lakeshore, time, development) and each species is abundant only near its own optimum. Ordinations of such tables put the sites on an arch: the first axis is close to cos(πt) and the second to cos(2πt) = 2cos²(πt) − 1, a parabola in the first. When similarity between sites decays fast (a kernel with small ρ, or PCA of species with narrow niches) the ends curl in and the arch becomes a horseshoe; CA of Gaussian niches gives an almost exact parabola. The second axis is not a new factor; the order along the arch recovers the gradient.
Simulation: n sites at ti = (i − 0.5)/n; 30 species with optima evenly spread over [−0.1, 1.1] and expected count 20 exp(−(t − μ)²/(2w²)). Kernel: the similarity ρ|i−j| between sites, centred (it uses the positions, not the table).
code/ch09_heterogeneous_data_a.R (Figure 9.25).