To tackle papers like Larman et al. (1994), you need three core mathematical building blocks. Here is what they actually mean in plain English:
Test how the lower-bound sieve works. Click the button to run shapes through the four directional binary relations (Relation 1 through 4) to see how the pool filters down to a guaranteed uniform group.
Initial State: 1000 random convex shapes loaded.
The lower bound f(n) >= n^0.2 is achieved by sequentially filtering disjoint pairs through four distinct transitive binary relations based on their projections onto the X-axis.
For any two disjoint convex sets A and B, the authors define:
By Dilworth’s Lemma, applying this 4-step sieve forces either a chain of size n^(1/5) (disjoint sets) or leaves an antichain of size n^(1/5) where every pair intersects because they are incomparable in all 4 relations.
To prove the upper bound f(n) <= n^(log 2 / log 5) approx n^0.431, the authors construct a recursive geometric gadget based on a 5-cycle intersection graph.
Start with 5 segments forming a 5-cycle graph. Each segment p_i q_i intersects only p_(i-1)q_(i-1) and p_(i+1)q_(i+1). To scale up to step k, replace each segment with a miniature copy of the 5-segment gadget scaled down by a tiny epsilon > 0.
The total number of segments grows as n = 5^k, while the maximum independent or complete subgraph size grows as 2^k. Solving for the exponent yields:
When shapes are restricted to axis-parallel rectangles, the bound jumps dramatically to n / (2 log n) because rectangles cannot wrap around each other like line segments.
Let psi(m) be the minimum n such that D(F)I(F) >= m for axis-parallel rectangles. The authors prove by induction that psi(2^k) <= k * 2^k.
The Induction Step: For a family of k * 2^k rectangles, choose vertical line v at the t-th rightmost vertical side. Either >= 2^k rectangles cross v (yielding an interval-intersection reduction), or they split into left and right subfamilies, each satisfying the inductive hypothesis.