Documentation

FirstOrderMethodsOptimization_Beck_2017.Chap11.Lemma_11_7

@[reducible, inline]
noncomputable abbrev halvingCount (a : ) (m : ) :

Helper for Lemma 11.7: the number of halving steps among the first m transitions.

Instances For
    @[reducible, inline]
    noncomputable abbrev strictHalfRatioCount (a : ) (m : ) :

    Helper for Lemma 11.7: the number of strict-half-ratio steps among the first m transitions.

    Instances For
      theorem quadratic_step_recurrence_antitone {a : } {γ : PosReal} (hstep : ∀ (k : ), a k - a (k + 1) 1 / γ * a (k + 1) ^ 2) :
      Antitone a

      Helper for Lemma 11.7: the quadratic recurrence makes the sequence antitone.

      theorem reciprocal_increment_ge_one_div_two_gamma_of_strict_half_ratio {a : } {γ : PosReal} (hstep : ∀ (k : ), a k - a (k + 1) 1 / γ * a (k + 1) ^ 2) (k : ) (hak : 0 < a k) (hak_succ : 0 < a (k + 1)) (hhalf : a k / 2 < a (k + 1)) :
      1 / (2 * γ) 1 / a (k + 1) - 1 / a k

      Helper for Lemma 11.7: a strict-half-ratio step yields a uniform reciprocal increment.

      theorem geometric_prefix_bound_of_halving_count {a : } {γ : PosReal} (hstep : ∀ (k : ), a k - a (k + 1) 1 / γ * a (k + 1) ^ 2) (m : ) :
      a m (1 / 2) ^ halvingCount a m * a 0

      Helper for Lemma 11.7: every halving step contributes one factor 1 / 2 to the prefix bound.

      theorem reciprocal_prefix_bound_of_strict_half_ratio_count {a : } {γ : PosReal} (hstep : ∀ (k : ), a k - a (k + 1) 1 / γ * a (k + 1) ^ 2) (m : ) (hm_pos : 0 < a m) :
      (strictHalfRatioCount a m) / (2 * γ) 1 / a m - 1 / a 0

      Helper for Lemma 11.7: every strict-half-ratio step contributes one reciprocal increment of size at least 1 / (2γ).

      theorem geometric_term_le_epsilon_of_log_bound {a0 : } (ha0 : 0 < a0) (ε : PosReal) {m : } (hlog : 2 / Real.log 2 * (Real.log a0 + Real.log (1 / ε)) m) :
      (1 / 2) ^ (m / 2) * a0 ε

      Helper for Lemma 11.7: the logarithmic lower bound on the iteration count forces the geometric term below ε.

      theorem nonnegative_sequence_le_max_geometric_or_sublinear_of_quadratic_step_recurrence {a : } {γ : PosReal} (ha_nonneg : ∀ (k : ), 0 a k) (hstep : ∀ (k : ), a k - a (k + 1) 1 / γ * a (k + 1) ^ 2) {n : } (hn : 2 n) :
      a n max ((1 / 2) ^ ((n - 1) / 2) * a 0) (4 * γ / (n - 1))

      Lemma 11.7 (1): if a nonnegative scalar sequence satisfies a k - a (k + 1) ≥ (1 / γ) * a (k + 1)^2 for every k and some positive γ, then for every n ≥ 2 one has a n ≤ max {((1 / 2)^((n - 1) / 2)) * a 0, 4γ / (n - 1)}.

      theorem nonnegative_sequence_le_epsilon_of_quadratic_step_recurrence {a : } {γ : PosReal} (ha_nonneg : ∀ (k : ), 0 a k) (hstep : ∀ (k : ), a k - a (k + 1) 1 / γ * a (k + 1) ^ 2) (ε : PosReal) {n : } (hn : max (2 / Real.log 2 * (Real.log (a 0) + Real.log (1 / ε))) (4 * γ / ε) + 1 n) :
      a n ε

      Lemma 11.7 (2): if ε > 0 and n ≥ max {(2 / log 2) * (log (a 0) + log (1 / ε)), 4γ / ε} + 1, then the same recurrence implies a n ≤ ε.