Reciprocal estimates retain the exact leading coefficients when solving the normalized determinant inequality, even when its bracket is negative.
theorem
CCDN.auxiliary_kernel_quotient_bound
{r d M X a k : ℝ}
(hr : 0 ≤ r)
(hd : 1 ≤ d)
(hM : 0 < M)
(hX : 0 < X)
(ha : 0 < a)
(hk : 0 ≤ k)
(hupper : a ≤ X * (M + r) / ((r + 1) * M))
(hlower : 1 / (d * (r + 1)) - 1 / M ≤ k)
:
If 0 < a ≤ X (M + r) / ((r + 1) M) and k ≥ 1 / (d (r + 1)) - 1 / M is nonnegative, then
k / a ≥ 1 / (d X) - (2 r + 1) / (X M).
theorem
CCDN.auxiliary_small_height_bracket_bound
{r d M X a k u b l : ℝ}
(hr : 0 ≤ r)
(hd : 1 ≤ d)
(hM : 2 * d ≤ M)
(hX : 0 < X)
(halow : X * (M - d) / ((r + 1) * M) ≤ a)
(haup : a ≤ X * (M + r) / ((r + 1) * M))
(hk : 0 ≤ k)
(hklow : 1 / (d * (r + 1)) - 1 / M ≤ k)
(hb : 0 ≤ b)
(hl : 0 ≤ l)
(hheight : l ≤ 2 * d * (r + 1) * b)
(hscale : d * b ≤ X * M)
(hcomparison : a * u ≤ b - k * l)
:
Solve the normalized inequality without imposing a sign on its bracket. The small-height assumptions absorb both reciprocal errors into C(r).